• FOIA documents revealed disputed Pentagon plans for a less restrictive AI system
• OpenAI denied agreeing to language requiring minimal refusal rates from models
• A Pentagon lawyer later withdrew an earlier confirmation about the contract

Documents released through a Freedom of Information Act (FOIA) lawsuit reportedly revealed that the Pentagon sought out an OpenAI system with fewer restrictions.

The paperwork described models intended for national security work that would have “minimal refusal rates” during certain military applications.

OpenAI and the Pentagon have both disputed that such wording appeared in the final agreement signed between them.

Dispute emerges over AI restrictions in military contract documents

The disputed language appeared in a contract document describing work involving testing, evaluation, and improvement of OpenAI models for defense-related activities.

The document referred to specialized models created for government security uses, including systems designed to reject fewer requests during operation.

However, officials from both organizations later said the document obtained through the lawsuit was not the final contract version.

"OpenAI has never agreed to contract language requiring 'minimal refusal rates.' This language does not appear in our executed contract," said Nate Evans, an OpenAI spokesperson.

“The document you received appears to be an earlier draft proposed by the Department before we provided feedback.”

Evans added that OpenAI rejected the wording and said the final executed agreement did not contain that requirement.

A Department of Defense representative later stated that the phrase “minimal refusal rates” does not appear in any active contract.

The disagreement continued after a Pentagon lawyer initially confirmed the document before later withdrawing that statement.

Trevor Tiedeman, a Pentagon research official, said there may have been confusion between released records and executed agreements.

Tiedeman also indicated he would examine how the document passed through review procedures before becoming publicly available.

The original document was released after The Intercept requested contract records connected with military artificial intelligence agreements.

The request specifically sought completed agreements, while officials later argued the released material represented an earlier version.

OpenAI did not provide the complete contract publicly, leaving some details about the agreement unavailable for independent review.

Concerns grow over military AI safeguards and operational limits

The dispute has renewed discussion about how much control technology companies should retain over artificial intelligence used by governments.

Commercial AI systems typically include restrictions designed to prevent assistance with certain harmful activities, including weapons development.

OpenAI's public chatbot, for example, refuses requests involving assistance with planning specific attacks or selecting targets for strikes.

“Minimal refusal is likely referring to little or no safeguards on the model being used,” said Heidy Khlaaf, chief scientist at the AI Now Institute.

She argued that private technology companies face difficult decisions when their systems become connected with military operations.

The Pentagon has also faced separate disagreements with other AI firms over restrictions involving defense applications.

Despite those disputes, OpenAI expanded its military agreement, allowing use of its services across classified government networks.

The updated agreement retained references to mission-focused AI models, although significant sections of the contract were redacted.

OpenAI has said its agreement includes restrictions against autonomous weapons use and certain surveillance activities involving Americans.

The company and government agencies have not fully disclosed how those restrictions would apply across different military scenarios.

Via The Intercept

Unless you’ve done an extensive amount of research, And potentially even for those who have done an extensive amount of research, it would be shocking for most how many standing desk options there are in the wild today.

It’s one thing to look at how the vast majority of these standing desk companies are all shockingly similar. In fact, in my experience, I've noticed some even reuse parts or have quite possibly the same main distributor, with just a white label on it.

But it’s an entirely different game when you look at how Vari chooses to set itself apart. The company has a unique setup process, easy assembly, and now, with the included TechTrack, Vari has decided to step up its game once more.

the TechTrack system has actually changed what having a desk could look like. You're not just putting a desk shelf or accessories on top of your desk; you can actually integrate products into your desk.

There are only a few other desks that I've seen that have done something like this, where they have changed from the standardized slab of wood or material. Every time, I've loved it, as it is a decision to change the game without reinventing the entire wheel.

Vari Electric Standing Desk with TechTrack: Price and availability

The Vari Electric Standing Desk with TechTrack is available direct from Vari in a range of styles, with prices starting at $949. However, at the time of review, it's discounted to $760.

I chose the walnut desltop with black frame, although I am a big fan of some of the other color options too. Speaking of options, there is also a ton of TechTrack accessories including a phone holder, a shelf, a laptop stand, and a tablet stand.

The only thing I can compare it to is something almost like taking the ingenuity and brilliance of the Balolo Desk Shelf and building it into a desktop. It allows for sleek, customizable modularity in a way that is not otherwise attainable.

Vari Electric Standing Desk with TechTrack: Unboxing and First Impressions

Vari does a phenomenal job, or should I say a very good job, at packaging and creating an unboxing experience. I have unboxed a lot of desks and gear at this point, and Vari is still one of the top unboxing experiences. It's seamless, it's easy, and it feels very well thought through.

For this desk specifically, Vari recommends assembling the legs while the desktop is still in the shipping box to help reduce the risk of scratches from the TechTrack system.

I didn't think that was possible until I was putting together another desk I'm currently testing, and I scratched both the floor I was assembling on and the desktop itself when a rogue screw fell underneath the desktop. It turns out following instructions is a brilliant idea. Who would have thought it?

The assembly process may not be everything, but it does have a good indicator of how the rest of the product will operate. For most of its life, all of these tasks are very similar. It's a wooden slab on top of some legs where you can put all your desk accessories.

However, with this desk, the process is extremely easy to follow, ensuring you do it right without any mess-ups and quickly. This already sets you off for success, making you believe that this desk is the best from the very moment you have it.

Vari Electric Standing Desk with TechTrack: Design & Build Quality

Obviously, the big deal here is the TechTrack system, but beyond that, another big deal for this desk is that the desktop and legs are also great to start with. Anybody can toss a flashy new gadget on a desk and think that it's special.

Vari decided to upgrade one of its highest-end and one of their already popular desks. It was a true upgrade. That, combined with TechTrack accessories that are actually useful and helpful for professionals made this release a slam dunk.

Vari Electric Standing Desk with TechTrack: In use

When I use standing desks, I use them to their fullest extent. By this, I mean I will, multiple times an hour:

- change the height of the desk

- utilize the presets

- load down the desks with quite a bit of gear

- build out each desk to be pretty advanced

Some people use a standing desk to feel better and stand up once a day or once a week. I use it to change my thoughts, make sure I'm staying fresh, or even just to keep my toes under the fire and get all my standing time in.

By being honest, I've even had times when I'll stand just to make sure I don't have to leave a meeting early to go to the bathroom. All that said, I've put this desk through a very similar test, and so has my team.

We've used this desk day in and day out for almost 150 days now, so that time has been with me. That time has been with other people I have helping test these desks.

The point is that we've actually put it through the wringer. We've tried it in more than just a studio setting for an hour under special lighting. Through that time, we've proven and have the experience that the stability on this desk is spectacular, even at full extension.

For me, I usually have at least one preset at full extension, especially if I have more than two presets. That is because then I can frequently use them to put chairs under or feel like the desk is up and out of the way. At the absolute highest the electronics will let it go, I still don't worry about my gear wobbling or a monitor tipping over.

Another thing to note is that even with our 60-inch-wide desk, which is fairly large, I don't experience any sway while playing or using it day-to-day. While moving up and down, the motors operate smoothly and quietly.

Vari Electric Standing Desk with TechTrack: Final verdict

In short, this is a desk you should consider if you have a knick-knack or 17 on your desk and you're still trying to find a way to organize your series workflow.

This is the kind of desk you want if you are into gadgets. This is also the kind of desk you should consider if you have a few gadgets on your desk and want to project a professional image.

However, if you're the kind of person who just throws things on their desk, doesn't really care, or is going to stay seated at all times, then you can probably find another desk that fits your budget and desires a little better.

This is the desk meant for those who like to be organized, get things done, and do it in style. There's a handful of areas that always make sense to spend a little more, and in my view, I stay in this because I'm quickly becoming one of them.

For more office furniture, we've tested the best standing desks and the best office chairs.

• AI memory demand is making even broken smartphones increasingly valuable to buyers
• Old phones are gaining value as replacement components become harder to source
• Grade D devices remain abundant despite stronger prices for salvageable hardware

A global shortage of memory chips driven by artificial intelligence demand is pushing millions of consumers toward the resale market.

Owners who once let broken or outdated smartphones sit in drawers are now finding that even damaged units carry unexpected monetary value.

Buyers across the refurbishment and repair industries are competing for parts that were previously considered worthless scrap.

Salvage values climb as parts become scarce

Marketplace figures from B-Stock, a large business-to-business platform for used smartphones, show pricing for both Apple and Samsung devices rose in the second quarter of 2026 compared with the prior two quarters.

Grade D handsets, the lowest condition tier, made up the bulk of units sold last year, with millions changing hands through the platform.

In the first half of 2026, mid-tier C-condition phones actually overtook the lowest tier by volume, accounting for 37.8% of sales versus 31.6% for the bottom grade.

The iPhone 17 Pro Max fetched the strongest average price among Apple models, reaching $940 across all condition levels during the quarter.

Samsung's top-selling models were the S23 Ultra, the S22 Ultra and the S24 Ultra, with the S22 Ultra pushing the older Z Fold5 out of second place.

After the Galaxy S26 launched, prices for previous-generation Samsung flagships climbed as well, with a well-kept S23 Ultra reaching $425 in June, its strongest month of the year.

Repair industry sees new value in broken parts

B-Stock has cautioned against assuming shoppers are simply rushing to dump unwanted hardware.

The company points out that people are keeping their phones for extended periods, a pattern that itself drives up the volume of low-grade listings as devices age.

Manufacturers are not yet routinely pulling memory chips from damaged phones to install in new products, since testing, compatibility and reliability checks complicate that approach.

According to the company, the economics of recovering usable components are shifting as memory prices climb, making it worthwhile to salvage parts that once were not worth the labor involved.

Displays, camera assemblies, batteries, charging ports, housings, speakers, cables and circuit boards can all retain value even when a phone itself is beyond repair.

A single low-grade device does not necessarily mean every part inside it shares that same poor rating.

A phone traded in after four or more years often fails battery health testing alone, which is enough to place it in the bottom tier under standard platform rules.

The current trend feels less like a sudden environmental awakening and more like a market correction.

Consumers are not necessarily recycling out of ecological conviction, and retailers are sweetening trade-in offers to keep the supply of parts flowing.

Whether the shift meaningfully reduces electronic waste, or simply redirects it toward a more profitable pipeline, remains an open question worth watching.

• Anduril and Poland's PGZ are preparing local production of the new Barracuda cruise missile
• Anduril displayed a full-scale Barracuda mock-up during Poland's major defense exhibition
• Production details remain unsettled as Poland and Anduril continue negotiations

Anduril has partnered with Poland's state-owned conglomerate PGZ to jointly produce a new cruise missile system.

Dubbed Barracuda-500M, the ground-launched weapon will reportedly be manufactured locally as part of a long-term arrangement between both companies.

The missile deal comes as Warsaw continues to expand its arsenal amid rising concern over Russian military activity near NATO's eastern border.

Barracuda-500M production takes shape

A full-scale mock-up of the missile was placed on public display for visitors attending this week's defense trade event held in Kielce, Poland.

The unveiling took place during the annual MSPO defense exhibition, where company representatives outlined the scope of the missile partnership.

The partnership covers both local assembly and eventual full-scale production of the missile inside Poland.

Company representatives have not yet said whether components will initially be shipped from the United States for domestic assembly work.

They have also not yet publicly disclosed a timeline for full-scale production or for when the missile might become operational.

Autonomous aircraft added to the pitch

Separately, Anduril is also pitching two autonomous aircraft designed to operate alongside Poland's future F-35 fighters and AH-64E Apache helicopters.

Brian Moran, who oversees Anduril's business operations across Europe, outlined the pitch during a separate interview at the MSPO show.

According to Moran, the Fury combat drone could accompany Poland's incoming batch of 32 F-35 fighters as a loyal wingman.

He explained that the aircraft has already been chosen by the U.S. Air Force for a formal, funded procurement effort.

"Fury is tailor-made to partner with an F-35 as a loyal wingman and as a force multiplier to keep men and women in uniform out of harm's way," Morgan said.

The Thunder rotorcraft, a separate autonomous platform built by the company, is designed specifically to operate alongside Poland's incoming fleet of 96 Apache helicopters.

Thunder was first unveiled publicly during the Farnborough Airshow held this past July, several months before Anduril intensified its talks with Polish defense officials.

"In July, we launched Thunder, which is an autonomous collaborative platform for the Apache attack helicopter," said Moran.

"And just like Fury operating with an F-35, with Poland getting all these Apaches, having Thunder is a force multiplier, especially as Poland will have to train a lot of pilots for 96 Apaches."

Moran noted that Poland is set to become the second-largest operator of Apache helicopters anywhere on the globe, trailing only the United States.

Delivery of all 96 helicopters, along with their associated equipment and weapons systems, is scheduled to formally begin during 2028.

That gap gives the Polish military several years to train pilots and prepare support systems before the Apaches actually arrive.

Beyond the 96 helicopters, Poland is also trying to expand its arsenal with ongoing discussions with Anduril regarding a possible purchase of the Thunder drone.

Via Defense News

• Microsoft and two key US teaching federations have agreed safeguards on AI use within education across America
• The American Federation of Teachers and the United Federation of Teachers have made the deal with Microsoft after months of negotiations with all parties
• The AFT is in talks with other companies including OpenAI and Anthropic

After the failure to pass federal legislation to establish protections against AI overreach in schools, two teaching federations have taken the step of establishing legally-enforceable rules with Microsoft, covering the use of AI in schools.

Protections within the National AI Safety & Privacy Standard include the agreement not to use student or teacher data to train or improve AI models. Furthermore, any collected data will not be monetized, as educators attempt to find a middle ground that reduces screen time and balances technology use across all age groups.

This agreement follows various localized actions against AI use in schools, including a moratorium on AI use within the New York City school system, and a similar ban in the Los Angeles Unified School District.

Iron-clad privacy

“We have forged a hard-fought, iron-clad privacy agreement with real teeth that protects students and families, because no one else, including the federal government, has stepped up to do the real work,” said AFT President Randi Weingarten. “We can get angrier and angrier, or we can act decisively; anything less than legally enforceable provisions is simply a wish list.”

The concerns of the American Federation of Teachers (AFT) and the United Federation of Teachers (UFT) are multiple, and ultimately center on their teachers’ comfort, focus, and ability to conduct lessons without the specter of AI intrusions. So, issues about training data, monetization, screen time, and other AI-related challenges have been sewn up by the National AI Safety & Privacy Standard, which is legally enforceable.

“This agreement ensures three foundational goals: protect the privacy and data of children, give schools real control over AI tools in the classroom, and provide transparency and information to parents," Weingarten continued.

"And with New York City adopting our framework to ban screens and AI use in the younger years, this agreement adds to the momentum around online safety for kids and educators.”

Legally enforceable guardrails

The guardrails put in place by the agreement concern protecting privacy of educators, parents and students, enhancing safety of students and educators, and providing transparency and control about how AI is used in schools.

This also includes information about what data is collected and the protections that are in place. It places schools in control over the use of AI, and forces AI providers to communicate product changes and risks.

“This standard sets a high bar for child privacy and AI safety, and we’ll extend this agreement to every school district across the country,” said Microsoft Vice Chair and President Brad Smith.

Along with banning the use of student and teacher data for AI model training, the data will not be sold or repurposed, and again schools will control data use, along with how it is stored and deleted. The guardrails also ensure AI systems are protected by human oversight, and are designed to block harmful actions and protect sensitive information.

The AFT reports it is in talks with OpenAI and Anthropic to also adopt the National AI Safety & Privacy Standard.

• Claude turned a famous mathematical proof into millions of checkable code lines
• Anthropic says Claude completed years of expected work in 11 days
• The massive proof contains 13 million lines of Lean code

Anthropic has used its Claude artificial intelligence system to produce a fully computer-checked version of a famous, centuries-old mathematical proof.

The proof addresses Fermat's Last Theorem, a hypothesis first proposed by the mathematician Pierre de Fermat back in the year 1637.

Mathematician Andrew Wiles produced the very first full mathematical proof of the theorem back in 1995, spanning 129 pages total in length.

A proof rebuilt for machines

Formalizing a proof simply means converting its mathematical reasoning into code that computers can check automatically without any human assistance.

Anthropic says it expected the entire task to take several years, based on how mathematicians first described the project.

Instead, the company says its internal research model finished the entire proof in only 11 days of continuous, largely unsupervised work.

The finished proof runs to 13 million lines of specialized code written in a programming language called Lean, used by mathematicians.

Along the way, Claude's agents reportedly proved roughly 30,300 separate theorems, ultimately using 29,500 of them in the final version.

Human input was reportedly limited to occasional high-level guidance, rather than any direct hands-on coding throughout the entire eleven-day process.

At 13 million lines, the resulting proof is over five times larger than Mathlib, the community's own main proof library.

"This extraordinary autoformalization achievement…proves Fermat's Last Theorem with no assumptions other than the axioms of mathematics," said Kevin Buzzard, a mathematician at Imperial College London.

“Along the way we see autoformalization of algebra, harmonic analysis, geometry and number theory, and we learn that AI autoformalization artefacts are now robust enough to be built upon; the proof is multi-layered.”

Anthropic attempted the formalization several times before succeeding, with those efforts contributing roughly 7% of the final proof’s non-boilerplate lines.

Not Anthropic's first math breakthrough

The formalization arrives just one month after Anthropic detailed a separate breakthrough involving the Riemann zeta function, a well-studied mathematical object.

That function sits at the very center of the Riemann hypothesis, considered one of mathematics' hardest unsolved problems worldwide.

Rival lab OpenAI is also pursuing similar work, using its newest Astra model to solve several classic Erdős problems.

That same OpenAI effort also reportedly narrowed several long-standing open questions within the field of theoretical computer science.

Anthropic says its breakthrough came only after giving Claude access to an open-source software tool named Prove2Me, built by outside collaborators.

The software helps AI agents pick the most useful next step during a long, multi-stage research workflow, while also cutting inference costs.

Anthropic has also expanded free access and research credits for mathematicians working specifically on formalization projects, alongside dedicated larger grants.

Despite the record pace set here, an eleven-day timeline still shows how labor-intensive full formalization remains, even with today's most capable systems.

Prusa is at the heart of the open-source 3D printing community, and this machine really shows it, as its partnership with Bondtech has produced one of the finest machines on the market today.

What's even more surprising is that you can buy the simple four-tool kit for less than £2,000/$2,000 assembled, and the eight-tool set that I'm looking at in this review is a touch over that £2,000/$2,000.

That money not only gets you access to eight-tool printing, but also to one of the most reliable and solid-performing 3D printers I've ever used. The machine is based around the CORE One, a machine that I'm just upgrading to the One+, and is still the most used and reliable 3D printer in the workshop.

Whilst that's machine is only one filament, upgrading that reliable system into this eight-tool-changer machine absolutely makes sense.

I went for the fully assembled version because I just don't have time to assemble a machine, but if you do fancy the kit, you can save almost £300/$300 over the pre-assembled version that I'm looking at.

While the eight-tool-changer printing is obviously the major headline here, it's only when you start using it and put it side-by-side against the likes of the Bambu Lab H2C and, at the cheaper end, the Snapmaker U1 that you really start to see the full potential.

The speed is a significant upgrade on both of those machines, and even with the H2C's far larger print bed, in a head-to-head printing a 5-color print, whilst on the Core One+ INDX it had to be done in several batches, the new 3D printer was able to produce substantially more models than the larger machine in less time.

At present, considering the price, the multifilament printing reliability and speed, this is the best 3D printer I've ever tested (and I've tested many). It makes it hard to recommend any other machine, and with the CORE One L now also adapted to utilise the INDX system, Prusa once again leads the field for this scale of machine, from small to large-scale projects.

Prusa CORE One+ (Gen 2) INDX 8-Tool: Price and Availability

The Original Prusa CORE One+ (Gen 2) INDX 8-Tool is available directly from Prusa. Prices start at $2249 / £2039 assembled, and $1949 / £1769 for the kit.

Shipping starts October 15.

If you already own a CORE One or CORE One+, you can buy the upgrade kit, which is $999 / £909 for the 8-tool version (a 4-toolhead is also available). Other INDX conversion kits from Prusa can also be found here.

Essentially, like the very first model, it doesn't matter what Original Prusa machine you have; there is always an upgrade route to the latest version.

Prusa CORE One+ (Gen 2) INDX 8-Tool: Design

I looked at the original Prusa CORE One back in January 2025, and at the time it was a significant update to the MK4S, their previous machine. Since then, we've seen the larger version, the original Prusa CORE One L, and more recently the Plus upgrades to both. Now we also have the INDX from Bondtech and Prusa Research.

Since the machine is from Prusa, you have several ways to buy it, unlike most companies these days. The first is to buy it off the shelf as a fully assembled, ready-to-go option, which is exactly the option I went with in this review.

If, however, you really want to see the inner workings of the machine, and for many people running print farms or small businesses, I would probably recommend at least buying one machine in kit form so that you can see how it works. This gives you a good idea of how to repair it if something does go wrong.

Taking a look at the machine, the first thing you notice is that, from the outside, it looks almost identical to the CORE One. In fact, the main body and build area are the same, and it's only the 8 reels on the outside that really give it away.

I've used them without the dry boxes you can attach as well, really because I only have two and need to purchase another six to ensure my machine looks balanced. Still, aside from the four reel holders on each side, the major difference here is the INDX system bolted onto the front top of the machine.

This added dock essentially covers the view to the platform and contains the eight new toolheads, which are all heated through induction rather than traditional wires. This means that they can be attached to their magnetic dock along the front of the machine cold, heated up in just a few seconds, and printing following a filament change can start within a few seconds, far faster than any other system.

The other major difference here is that, obviously, the CORE One was a fully enclosed machine. This carries over to the INDX expansion, with a hinged top lid, secured in place by two rotating holders that are, of course, 3D printed.

The area is slightly smalled the original Prusa CORE One because because of the new additions but only marginally at 248 x 205 x 270 mm, dropping from 250 x 220 x 270 mm, so 2mm in one direction and 15mm in the other while the height remains the same.

Essentially, the Prusa CORE One+ INDX is the original Prusa CORE One+, the second generation, with the Bondtech INDX system incorporated.

While it may look very similar asthetically, in practice, this machine's potential goes far beyond what was available before and at this price.

• Design: 4.5/5

Prusa CORE One+ (Gen 2) INDX 8-Tool: Features

One of the great features of the Prusa machines is their customizability, so one of the first decision you have when looking at a Prusa machine is do you buy it assembled or as a kit. If you're a business and buying several, I would personally buy the majority as assembled and one as a kit, it will just give you that insight that can be invaluable.

Either way, if you need to keep costs down, you can opt for the assembled version without the camera, advanced filtration system, or additional print sheets, which keeps it at its base level. Or, if you want those more advanced features, they can be added on as well, so you can have the camera, advanced filtration, and different sheet finishes depending on the materials you're actually printing.

The machine is also available as a 4, or 8-tool option, with about £300 difference between the two. If you can stretch to it, I would definitely opt for the eight-tool version, which I'm looking at in this review, but if you can't, you can always buy the eight-tool upgrade kit later.

When it comes to the print volume, the machine offers an area of 248 x 205 x 270 mm, which isn't the largest for a machine of this size and a slight reduction over the Core One+, but it's still pretty decent.

Like the CORE One and One+, it's designed for engineering materials such as PC and nylon and has a heated chamber that reaches up to 55°C.

The main feature is its advanced tool changer. Unlike other multifilament machines that have one or two nozzles, or, in the case of the H2C 6, U1 4, here we see 8 individual nozzles that are preloaded with filaments during the initial calibration process.

Those nozzles sit cold in a docking station across the front of the machine, and only when the head picks up each nozzle does it use induction rather than wiring to heat it to the right temperature in about 12 seconds per material, so it's ready for use. This process dramatically cuts both filament waste and the time it takes to get up to speed.

What's impressive here is that when those tools are back in the dock, they cool right down. They're actually just passive toolheads, so whilst the filament is loaded within them, there are no electronics, making it even easier to change those nozzles when needed.

One feature Prusa highlights is that bed levelling is done through pressure, so the nozzles touch the print bed before printing starts; a load cell reads the force, and the machine knows exactly where the print bed is. This differs from many newer printers that have started using optical solutions. Whilst these are generally good, they're not always 100% accurate.

One of the big changes for Prusa, and something the company has tried to avoid until now, is a silicone wiper. Positioned just to the right of the machine, there are some silicone pads which each of the nozzles wipes across to remove any blobs before printing, and this is positioned right next to the purge collector, which, as I found through the test, even after hundreds of hours of printing, gathers very little waste material.

The INDX head features conductive electronics that heat each passive toolhead. Still, it's more than just technology: the extruder is self-adjusting, so it can alter its grip for different materials. This essentially means that, using the same system, you can print with soft materials such as TPU and rigid engineering filaments without having to stop the extruder for each.

Finally, because each one is passive, it takes literally seconds to change them. So, if you want to go from the standard 0.4 to 0.25 at the finest, up to a big, thick 0.8, then the tool-change process takes a couple of seconds. Simply unload the filament, remove the toolhead from the dock with a pull and release the feeder tube; plug in the new one, pop it back in the dock, load the filament, and you're set to go.

Each toolhead needs to be calibrated, and an automatic tool-offset calibration sensor is mounted on the outside of the print area. The speed at which this works, and the fact that it all works automatically, is impressive. To keep things running smoothly, I rerun the full calibration at the end of the week after clearing out any debris from the machine, which is probably one reason this machine has been so reliable.

Once again, to reduce vibration from movement in the machine, and so it sits in with most workshop devices, it features an all-steel build. As you pick up the machine and move it around, it is relatively heavy but easy to manoeuvre, with decent grip handles on either side, so you can move it from one side of the workshop to the other without too much issue.

And these days security is a huge issue, and it's nice to see that Prusa has taken this seriously and offers a 100% offline and, if you want, private solution to transferring files over to the machine. If you want to remove the Wi-Fi module, that's very easy to do, and if you want higher security, Prusa again offers a range of advanced features.

• Features: 5/5

Prusa CORE One+ (Gen 2) INDX 8-Tool: Performance

There's been a lot of anticipation around the Prusa CORE One+ INDX 8T in the 3D printing community. Essentially, it blends the robust printing abilities of the Prusa CORE One with Bondtech's tool changer, and the result is an absolutely stunning printer that produces exceptionally fine multicolor prints at a very reasonable price, considering the competition in this sector.

There are, of course, a few small downsides, such as the fact that the print bed on the CORE version is relatively small compared with what we've been used to seeing recently, and on my initial few runs I did have a couple of misprints. But as the printer itself bedded in and I became more familiar with the system, I was able to resolve all issues and also learn more about the printer, which gave me greater confidence that, as with their previous printers, if anything goes wrong, I would be able to fix it without the wait that I've experienced with many other multitool machines.

That getting-started process is, in essence, no more complicated than with any other machine, and from the box to sitting on the workshop bench ready to produce its first print took no more than 10 minutes. It was faster than I expected, especially since I set up the XL a couple of years ago.

Essentially, the main machine is complete, and then it's just a matter of opening the box with all the nozzles, popping them into the front of the machine, and inserting each feeder tube. Then you're essentially set and ready to go once the power cable has been inserted.

Also supplied is a USB-C key with plenty of models to get you started. As I loaded the eight reels onto the side of the machine, it did expand the space needed either side of the printer, but it still fit neatly next to some of the machines in the test area.

What I also really like is that I can easily shift this printer around the workshop without too much issue. It's heavy, but nothing compared with some of the larger machines. It is that stopgap between the Snapmaker and the Bambu Lab H2C.

As I started the test, the big question that I had was: with a standard-size print bed, would this printer actually be a viable option for people running print farms? Was it fast enough even though that platform wouldn't be able to print the same volume of models? As the first few test models that Prusa supplied came off the print platform, I was instantly impressed by the complexity of the models produced and by the time it took to print.

I tried out a series of multi-filament and, in some cases, multi-material prints, and they took far less time than I would've expected, even with the H2C set to its fastest usable print speeds.

I ended up running a head-to-head between the two with a dragon model, and this is where the speed of the new Prusa CORE One+ INDX 8T really stood out.

On the H2C, the platform was able to take 14 dragons, which was impressive, whilst the Prusa INDX was restricted to 9. When it came to the time it would take for the dragons to print, this is where the biggest difference came, and whilst there were five additional dragons on the H2C, it takes 3 days and 10 hours to make that print, while the Prusa will took one day and 35 minutes according to the software. When translated to the actual time these printers took to run through those prints, they were quite accurate, plus or minus about 30 minutes. Notably, by the time the H2C, with much the same speed setting as the Prusa INDX, had finished its 15 prints, I had 27 from the CORE One INDX. Obviously there is the inconvenience of clearing the platform after each print run.

Another point here is that there is no purge tower. Instead, a small amount of filament is scraped from the toolhead before it's used, but the tool-change speed is far faster than other tool changers. Essentially, that induction system works incredibly quickly, meaning sometimes you can't even really tell that the tool change has happened, as the printer is only not printing for a few seconds.

This test was also a great time to ensure calibration and accuracy for both printers were spot-on, and it actually took a couple of days to get 100% reliability from both the Prusa and the H2C through tuning. But since I ran the test right at the start, both machines have been running side-by-side flawlessly.

Setting down the models from each printer, what's instantly apparent is how little quality difference there is between the two. Layer by layer, the walls look smooth and the transition between colors is neat. You can obviously still see the layers. These prints are as fine as any.

Through the test, as always, there were plenty of firmware updates, and some of the early issues that I had were with tool eight, which for some reason would always print just a little bit high was resolved.

After an update, a good vacuum of the inner chamber and a full recalibration, that seemed to kick into life and resolved the issue. It was a bit of an odd one. It worked in the initial stages of the test, then after a firmware update and a particularly messy misprint, it stopped working correctly, and then it worked again after the next update. It's now been three weeks, and that particular nozzle is now working absolutely fine.

There are a few other small issues that I've come across. My biggest issue whilst using the machine is the filament mount on the side of the machine, not so much that it's on the side, but the fact that those filaments are exposed. Prusa does sell individual dry cases for each of those filaments, which is a good thing, but these are very much bolt-on, and while they work, there are neater solutions.

Then there's the filament feed, and this is the part that started to get to me over the weeks with plenty of filament changes. Essentially, unlike the Snapmaker or Bambu multi systems, you still have to feed the filament into the machine, and with four filament reels on each side, this can get a little bit messy.

If space is tight, getting your hands in and feeding the reel around while pushing the filament up into the feeder tube all the way to the hot end is one of those tedious, fiddly things. I would love to see some sort of auto-feed system.

Then, for less experienced users, the camera and filtration system on the back need to be fitted, and it's not as simple as plugging them in and attaching them. You do need to remove the panel, feed the cable through, and plug it into the board, which takes about 15 to 20 minutes.

It's not complicated; it's just one additional step rather than being ready out of the box. The idea is that you can customise the machine to your exact needs. Not everyone wants a camera, but since cameras are such a standard part of many setups, it's something you expect.

The other surprise is that the nozzles aren't designed for abrasive printing, so if you do want to use PETG-CF, PC-CF, or another abrasive material, the nozzles will wear out relatively quickly. Essentially, they're rated for 10 kg of PETG-CF or 5 kg of PC-CF. While checking all the different material types, I noted that if you're using Prusament Ultra Glow, you'll get through just 0.5 kg of filament before the nozzle likely needs replacing. These are not hardened nozzles, although those are on the way very soon.

The point is that the print area is just 248 x 205 x 270 mm. That reduces the CORE One's build plate from 250 x 220 x 270 mm, giving a slightly smaller build area again.

One of the good things is that many of the early issues, such as stringing issues with the nozzles carrying filament back into the machine and other early performance problems, have all been addressed by each firmware update.

Obviously, for this review sample, it is an early production model, and what is striking is that with every firmware update that came in, and as I became more familiar with the machine and it bedded in, the machine became more accurate and reliable.

It's now been three weeks since I've had to do anything to it, and it's been running continuously printing ready for a new term. I will obviously give it a bit of oil as it requires, but otherwise this is a multitool machine that will print thousands of models without much maintenance, and that is one of the key points of Prusa machines.

Dimensional accuracy - score of 4

Target 25 = X: 24.76mm / 0.24mm Error | Y: 24.70mm / 0.30mm Error

Target 20 = X: 19.82mm / 0.18mm Error | Y: 19.80mm / 0.20mm Error

Target 15 = X: 14.75mm / 0.25mm Error | Y: 14.72mm / 0.28mm Error

Target 10 = X: 9.88mm / 0.12mm Error | Y: 9.83mm / 0.17mm Error

Target 5 = X: 4.88mm / 0.12mm Error | Y: 4.78mm / 0.22mm Error

X Error Average = 0.182

Y Error Average = 0.234

X&Y Error Average = 0.208

Fine Flow Control - score of 4

Fine Negative Features - score of 5

Overhangs - score of 5

Bridging - score of 4

XY resonance - score of 2.5

Z-axis alignment - score of 2.5

Add the totals for a final score of 27 out of 30.

Before I leave the performance section, another point about the prints is the new color-blending mode within PrusaSlicer known as Prusa Color Mix, and this is something I'm still trying to get my head around. It's quite incredible how it works.

Essentially, like a dot matrix, it blends the colors you have together to create more colors by creating fine layers of filament rather than physically blending the two or more. With two fish models I printed very early on, I was first impressed that they printed.

They're quite complex models, and I set them aside for a few days. Then it occurred to me that there were about three or four different greens within the print, yet I only had one green loaded onto the machine. At that point, I realised that the color mix was in full flow with these models.

I went back to PrusaSlicer and started trying some color-blending modes myself and, again, was impressed by how easy it was to use. I would say that randomly loading different-colored filaments onto the spools will get you some pretty interesting results. If you really want headline-worthy features, careful filament choice is definitely needed. However you approach it, the results are incredibly impressive.

Whilst I've just touched on the beginning of how this technology works, it's probably one of the most exciting enhancements from any 3D printer that I've seen, meaning that those eight filament spools, especially if you get the color combinations right, can produce thousands of different colors.

• Performance: 5 / 5

Original PRUSA CORE One: Final verdict

The Original Prusa CORE One+ INDX 8T is a 3D printer in a different league, and while the build area is smaller than even the CORE One, the speed at which you can do a multicolor print is much faster than any off-the-shelf machine I've tried.

Most of the initial misprints and small issues I had were mainly the machine bedding in rather than any major problems, and this is an early review sample with fixes coming through the firmware updates. Still, even then as an early production sample its performance has been absolutely solid and rivals and beats any of the machines I have in the workshop.

Unusually, it has also been of major interest to several friends who run 3D print farms and engineering businesses, particularly for prototyping with multiple materials and colors. The production times of those models are vastly reduced, and whilst 3D printing already cut down their prototyping time dramatically, the Prusa CORE One+ INDX once again cuts down that time even further.

What's surprising here is that, even with a fully assembled machine pretty much ready to go out of the box, and with all those new firmware updates and obviously the knowledge that's out there as well, this becomes one of the cheapest multitool machines on the market whilst also being easily one of the best.

Some machines have a refined, fully product-designed, more plastic aesthetic that gives them a high-quality feel, and they are obviously extremely impressive.

The Prusa's more workshop-style design means it's a far better fit for many businesses; if anything ever goes wrong, you have full access to the plans, a huge knowledge base, and one of the best support networks available from any 3D printer manufacturer. You'll never be stuck without a printer.

Should you buy the Original PRUSA CORE One?

Buy it if...

Don't buy it if...

We've also tested out the best laser engravers

Camera technology has come on leaps and bounds throughout the 20th and 21st century, but one of the most important developments was the invention of the CMOS image sensor. This invention, dubbed the ‘camera-on-a-chip’, is the technology that allows everyone to carry high-resolution cameras around in their pockets.

But the history of the CMOS can be traced back to NASA’s desire to install small, low-powered cameras capable of capturing high-resolution imagery on interplanetary spacecraft.

Dr. Eric R. Fossum is the inventor of the camera-on-a-chip, and he has not only fathered an era of space-based photography, but also the many industries and technologies that revolve around the micro-cameras we know today.

The camera-on-a-chip has changed the world

Before the development of the CMOS image sensor, photography and videography often relied on photographic film, or the more advanced Charge-Coupled Devices (CCDs). The Apollo 11 mission famously used purpose-built Hasselblad cameras - which were too heavy for the return trip to Earth - requiring the astronauts to transport just the film back for development. They’re still there to this day.

CCDs on the other hand were the precursor to Dr. Fossum’s camera-on-a-chip. They detected photons using individual pixels that fed an electrical charge to a single readout amplifier in a row-by-row configuration. These cameras improved significantly in reducing weight and capturing much more vivid imagery than their film-based counterparts, but suffered from high power-usage and heat generation.

CMOS image sensors improved further on the CCD design, giving each individual pixel its own local conversion circuit and amplifier allowing each pixel to be read simultaneously while significantly reducing the energy required per image taken and the heat generated by the camera.

CMOS image sensors are the basis upon which most modern cameras work - from the camera on your phone and in your doorbell, to the Advanced Driver Assistance Systems that power self-driving cars, and even the cameras still heading into space today.

I spoke to Dr. Fossum to understand how CMOS image sensors were developed and their impact on the world we know today.

• Eric, did you ever think that your invention - the CMOS-based image sensor - would have such an impact on society worldwide?

I thought it could become ubiquitous and thought it should become so, but my confidence that it would actually happen was low, due to all kinds of unknowns. I did not foresee at all the social-media application, nor its impact on society, nor its impact on citizen-newsgathering, nor its impact on social justice, whether to protect good actors in law enforcement or to identify bad actors in law enforcement. I also knew there could be impact on privacy and liberty, and I am uncomfortable with how that has evolved.

• I note that you worked at NASA's legendary JPL (Jet Propulsion Lab) at the time. We don't usually associate JPL with cutting-edge digital imaging. How did that happen? Was JPL essentially another Xerox PARC?

NASA and JPL is associated with cutting edge digital imaging, from its earliest activities in the robotic exploration of space. JPL engineers found ways to beam back images from distant places using digital imaging and data transmission. Eugene F. Lally at JPL in 1961 was probably the first advance the idea of a ‘mosaic’ image capture, where mosaic refers to an array of pixels. JPL developed many digital imaging technologies for scientific use in the harsh environment of space with high precision and reliable image capture across a larger range of wavelengths than consumer products. My job at JPL was created by JPL management to further grow JPL’s bench-depth in CCD technology.

• I am sure that you followed avidly the visual journey of Artemis II a few weeks ago: from the iconic photos to the videos that the crew and the entire team produced during that trip. Do you remember what your thoughts were when you first saw the wonderful pictures the crew delivered?

My initial thoughts focused on the impressive images, and I was pleased that humans have returned to the moon, at least into orbit for now. But it did occur to me that that it was also cool that NASA –invented image sensor technology, employed in a consumer smartphone (and digital camera), was used on the this NASA mission, further completing its promise to make cameras faster, cheaper and better.

• Your current research at Dartmouth centers heavily on the Quanta Image Sensor (QIS). Can you tell us more about what it is, how is it different from CMOS image sensors and why it is called Quanta (which I presume is the plural of Quantum).

In the CMOS QIS, due to its ultra-low readout noise, it is possible for the first time to discern individual photons via their generated photoelectrons, and to count with high accuracy the number of photoelectrons at each pixel, at room temperature, at low voltages, and without the use of avalanche gain. The CMOS QIS we developed at Dartmouth and later at the spin-off company Gigajot is, in essence, a CMOS image sensor that uses intra-pixel charge transfer (invented at JPL) and extremely low readout-node capacitance, to generate a large enough voltage signal from a single photoelectron to overcome background thermal noise in silicon.

The QIS concept, itself though, is technology agnostic, since it was about creating great images one photon at a time. Single-photon-avalanche photodiode (SPAD) technology has significantly advanced in the past 20+ years so that it is also becoming useful for image sensor implementation. A SPAD QIS has larger pixels and consumes much more power than the CMOS QIS, but offers time-tagging of pixel arrival far more accurate than CMOS QIS and is being produced as a foundry-process that is also easily integrated with CMOS circuits.

• How has the IEEE impacted your professional journey?

I have been a member of IEEE since I was a student and have since graduated to Life Fellow. Profession societies are very important for the dissemination of leading research among experts in the field, and IEEE is no exception. I have published many papers under the IEEE, in the sub-silos of Electron Devices and Solid-State Circuits, and have benefitted from IEEE meetings such as the International Electron Devices Meeting (IEDM) and International Solid-State Circuits Conference (ISSCC). IEEE was also a partner in the early days of what is now known at the International Image Sensor Workshop, that I and others founded in 1986.

• What is the future of consumer image sensors from your point of view given that mainstream smartphones regularly boasts sensors with hundreds of millions of pixels already and we've reached a plateau in terms of returns?

Certainly in the near term, image sensors and consumers have benefitted from 3D integration and stacked wafer devices. The on-chip integration of more advanced functions possibly including smart-image-sensor functions or edge offers new avenues for improving image capture and understanding right at the focal plane. This was something I dreamed of since the 1980’s when I was exploring, researching and writing about smart image sensors and focal-plane image processing, but it was too early in those days.

• What are your thoughts on PIS (Processing-in-sensor) which fuses compute, the image sensor and memory in one holistic system at the edge?

As mentioned, I have been thinking about this topic since I was a graduate student in the early 1980’s at Yale and then at Columbia University. I am not familiar with the acronym PIS (which might really be a poor choice!) but I do think it is very application dependent. Furthermore, there is a data communication and parallel processing problem where local pixel computing is easy, but global image computing, where different regions have to be analysed, is still difficult on the focal plane. This may result in avoiding the addition of extra power dissipation to the image sensor chip and instead partitioning the imaging system into two parts, local computing on the focal plane and regional or global computing off the focal plane.

• Today there are billions of cameras that capture the world in two dimensions, generating millions of Gigabytes worth of data. Have you ever been interested in exotic storage technologies (some of which use image sensors)?

I have not really worked much in exotic storage systems, although at Photobit, our spinoff from JPL, we worked with companies, from other startups to giants like AT&T Bell Labs, on readout of holographic optical memories. Although, at that time, we were just focused on the image sensor chip and high speed readout.

• Can you tell me about the significance of the IEEE Jun-Ichi Nishizawa Medal, and what this recognition means to you?

Gosh, who doesn’t like being recognised by your peers for your accomplishments? Still, it is a very humbling event and makes me reflect on the great people that I have had the chance to work with over the years. It has been such a privilege to work with these folks and take this journey together. As I always try to mention, the amazing CMOS image sensor technology we all use every day was made possible by the hard and innovative work of thousands of image sensor engineers around the globe.

• We're planning a special series towards the end of the year, looking at tools that thought leaders use in their everyday lives. Name one (either physical or software-based) that you consider to be critical to your daily work routine and why you chose it.

I think AI tools are great time-savers in finding and curating information, and I often learn much from their apparent knowledge, and sometimes from their insights. These days I use Co-Pilot a lot. But I am always concerned with whom, directly or indirectly, I am sharing information and ideas with. I think this is something we should all ponder and ask hard questions of the AI giants.

• A £3 Wi-Fi extender carried hidden administrator access beyond ordinary user controls
• Every device running the firmware shared the same concealed administrator password
• Changing the visible administrator password could not disable the secret account

A £3 Wi-Fi extender bought through Temu has exposed security problems that challenge the idea of cheap connected devices being simple bargains.

Security researcher Keiran Smith examined the device and discovered hidden access features that ordinary users would never see during normal operation.

Smith, who holds a penetration-testing certification, picked up the six-antenna extender after seeing it promoted through a targeted ad on the shopping app.

The cheap extender contained access users could not control

The examination began with the hardware, where he identified a MediaTek MT7620 processor commonly used in low-cost networking products.

After extracting the firmware stored inside the device, Smith found a hidden administrator account with complete control over its functions.

The account used a fixed password embedded inside the software, meaning every unit using that firmware carried the same credentials.

Changing the normal administrator password through the device settings would not remove this separate hidden access.

Smith also found a remote login service that accepted the concealed credentials without requiring physical access to the extender itself.

“It’s worth being precise about what makes this as bad as it is, because ‘hardcoded password’ covers a wide range of sins,” Smith said

The researcher said this case was more serious because the password remained identical across devices rather than being generated individually.

“A default credential is something the owner can see, is told about and can change,” he said. “What we have here is the opposite on every count.”

“This one is a compile-time constant rather than something derived from the MAC address or serial number, so it is identical on every unit ever sold.”

The combination of hidden access, unchanged credentials, and remote availability creates a security concern for ordinary owners.

Even if technically skilled users discovered the account, Smith found that changes could disappear after restarting the extender.

Additional flaws raise questions about cheap connected hardware

The investigation also uncovered a command injection weakness that could allow attackers to execute unauthorized instructions through the device.

Smith found that the extender lacked strong protection around software updates, creating possible opportunities for tampered firmware installation.

He admitted that these issues did not prove that manufacturers intentionally created unsafe features for malicious purposes.

They could have originated from factory testing processes and remained active accidentally before consumer sales.

This Temu extender shows how extremely cheap smart devices can create security challenges beyond their purchase price.

Consumers may focus on immediate savings while having little visibility into the software decisions built inside connected equipment.

The findings do not mean every inexpensive networking device contains similar weaknesses, though they show why basic security checks matter.

As more homes add connected products, hidden software features could become a larger concern for users and manufacturers.

Via CyberNews

The American engineer Mark Cooper is considered one of the leading pioneers of the wireless communications industry, envisioning a world in which people would communicate wherever they were and not from fixed locations. Without his work in the mid-20th century, the mobile communications industry would arguably be non-existent.

Hello Moto

On the 38th anniversary of the first cell phone call, journalist Bob Greene wrote an article for CNN commemorating Cooper's pioneering work. During this interview, he recounted the way that Cooper explained the thinking behind the concept of the mobile phone.

The concept for the "personal telephone", as Cooper described it, came from shifting the concept of a telephone number from something that was fixed to a location or infrastructure to something that represented an individual.

This notion, which broke away from the absolute necessity of being tethered to a copper wire, gave individuals (initially working professionals) absolute mobility and freedom to communicate from wherever they were.

Call me maybe

Cooper headed Motorola's communications systems division in the 1970s and conceived of the first portable cellular phone in 1973, which began a 10-year process of bringing the device to market.

He was the lead inventor named on the 'radio telephone system' patent filed in October 1973, but several months earlier he had demonstrated to the press the first handheld cellular phone call in public. He used a prototype DynaTAC device, which he used to call a base station that had been installed on the rooftop of the then-Burlington House in New York.

It's now impossible to envisage how the business world could have been shaped without his pioneering work, with hundreds of millions of units shipping each year. Indeed, this capability to communicate from anywhere – and anytime – has led to immeasurable productivity benefits over the last few decades.

"Lyrics are where people find meaning in music."

So says Marco Paglia, Chief Product Officer at Musixmatch - the platform delivering time-synced lyrics and translations across all major music streaming services.

At a time when passive background streaming feels like it dominates the landscape, I spoke to Marco to find out how lyrics are changing the game for musicians and their listeners.

In an era of passive streaming, why are lyrics still an important tool for turning casual listeners into diehard fans?

Lyrics represent one of the few moments in streaming when the listener stops being passive. A playlist plays at you, while lyrics make you lean in, follow the words and recognise something about yourself in what an artist is saying. That moment of connection is often where a casual listener starts to become a fan.

This experience has changed significantly over the last few decades. Lyrics used to be a souvenir: you bought the CD, pulled out the booklet, and read along. Now they are a fundamental part of the listening journey itself - time synced, translated and available immediately on screen as the song plays.

The scale of that behaviour is significant, with 88% of premium subscribers actively using lyrics. People read along, search for a chorus to find a track, share a line on Instagram or TikTok and print it on shirts.

Lyrics are where people find meaning in music. We see that reflected in listening behaviour too: tracks with lyrics generate 3.5x more saves, one of the clearest signals that someone wants to come back to a track. None of that is passive.

Ultimately, that’s why lyrics are such a powerful connective tissue between a song, an artist and the people who become invested in it.

On social media, a track can blow up in 24 hours. How much momentum does an artist lose if their lyrics aren’t synced on day one?

More than most artists realise, and that loss isn’t recoverable! The first few days of a release are a unique window: attention is at its highest, fans are actively searching for the track, and the moments that shape its trajectory are starting to build. Once that window passes, you can’t recreate it.

There are three things happening all at once in that initial period. Firstly, discovery: 81% of listeners search for lyrics online, so if a track’s lyrics aren’t available, it risks missing people actively trying to find it through words they’ve heard.

Secondly, algorithmic weighting: days one to seven are when saves and completion count most, and lyrics drive both of these things.

Thirdly, social engagement: lyrics give fans something to interact with and share. If there is no lyric card to screenshot, no sticker, no karaoke moment, you’re leaving fan engagement on the table right as curiosity peaks.

Then the window closes. Add lyrics on day eight and the opportunity to capture that initial release-week momentum has already passed. At that point, the focus shifts to longer-term discovery - and there’s no way to recreate that first week.

That’s exactly why we built Pre-Release, which means lyrics, sync and translations can all be in place and live from the moment a track drops.

For example, when the Rolling Stones released their latest album it launched fully covered, with lyrics, sync and translations in eight languages. Day-one lyrics aren’t a nice-to-have. They have to be part of the release itself.

How much manual effort is actually required from an artist or manager to get their lyrics and metadata release-ready across all platforms?

Historically an enormous amount, and almost all of it is invisible, unglamorous work: transcribing, time-syncing line by line, getting the structure and credits right.

Our job is to collapse that process. Customers can get verified in one step, we then pull their catalogue automatically from distributor data and audit it against what’s live on every Digital Service Provider (DSP). That produces a coverage report showing exactly where lyrics are missing, and prioritising the tracks where filling those gaps will have the greatest impact.

From there, customers can use our AI tools to transcribe and sync lyrics, or hand it to our expert curators. Our transcription engine is built specifically for the singing voice, and is more than twice as accurate as general speech tools.

That technology is backed by a global curator network that can help with proofreading across 100+ languages. Once everything is ready, it can be distributed to 25+ platforms simultaneously, with live status tracking. What could have taken months, now takes days.

There’s another part of this that catalogue managers sometimes underestimate: the same process can power the campaign around a release. Once the lyrics are ready ahead of release day, that file can also produce the vertical video loops for TikTok, Instagram, Spotify Canvas assets, and the YouTube lyric videos ready to go live alongside the track. One deliverable, every surface, no video shoot.

You describe Music Lens as ‘the world’s first Music Agent’. What does that mean in practice, who benefits and how?

Music Lens allows users to stop querying a database and to have a conversation with their catalogue. In practice, that means four things.

It can match a catalogue to a creative brief, it turns global trends into readable insight so you can see a breakthrough moment while it’s still happening, it gives clear visibility into royalties, splits and DSP performance (market by market), and it generates campaign-ready visuals from songs in seconds.

I like to think of it as an agent that develops through four stages: first, a musicologist, reading lyrics and extracting meaning, moods and themes. Then a DJ, enabling search by feeling, rather than simply based on the artist or title. Then it becomes a music expert, reasoning through a brand brief in the way a human sync agent would. Finally, it is an analyst, mapping how an artist's themes evolve and identifying patterns across countries and genres.

The biggest beneficiaries are publishers and labels with catalogues too large to know intimately, sync teams doing the matching process from memory, and independent artists who’ve never had this class of analysis available to them at all.

And the part I care most about is that it’s non-generative: it analyses meaning, but it doesn’t write songs. It’s built on derived data, meaning new information created by processing a catalogue of over 100 million works. It’s permission-based, with no scraping, no unlicensed content and no raw lyric ever surfaces in the output. That was a hard licensing constraint and it shaped every architectural decision downstream.

Do platforms like this level the playing field for artists to build a career without needing the support of a major label?

We give independent artists access to the tools and the infrastructure that historically were much harder to access without the support of a major label. We can offer reach across every platform, control over metadata, professional-grade catalogue intelligence, a merchandising operation, and the one people often forget: translation.

For example, an independent artist can verify their profile, own their lyrics, sync them for free, distribute them to 25+ platforms, ask Music Lens which of their songs has untapped potential, turn a lyric into merchandise, and reach fans in their own language on day one.

More than 1.5 million artists already use Musixmatch Pro to do exactly this. Ten years ago, much of that capability would have sat behind a label services deal.

Translation is the most underrated part of that shift. An artist making music in a bedroom in Bologna with listeners in São Paulo used to have no way to close that gap. Now the words travel with the music.

There are of course elements of labels that we don’t replace, and which are still a real advantage, such as capital, generating radio and playlist relationships and dedicated teams focused on building an artist’s career.

But we have certainly helped level the playing field by putting more information, tools and infrastructure directly into the hands of independent artists.

Beyond backend metadata and licensing, how do you plan to turn your platform into an active community and a richer experience for fans?

Musixmatch has always been a community, a large one, made of people who care enough about a song to get its words right. That passion is our asset. What we’re building now is the layer where fans and artists actually meet.

Translation is one of the most powerful examples. Most people know lyrics in languages they don’t speak, and translation turns that from a novelty into belonging.

Last November, we powered Spotify’s first global rollout of lyric translations on Rosalía’s LUX: 14 languages on one record, 23 translations synced for release day. It became the most-streamed album in a single day by a Spanish-speaking female artist. When language stops being a barrier, fandom stops being local.

Merchandise, and specifically fan-personalised merchandise, is another opportunity. Our Lyrics Merch tool allows an artist to turn a lyric into something a fan wears: designed, printed and sold through us, with the artist keeping 100% of sales. But the part I find most interesting is what happens when an artist opens that experience up to their fans.

If they want to offer it, the fan can choose their own favourite line from the song and have that printed on a shirt. It’s a completely different relationship from buying tour merch. The fan isn’t picking from what the artist decided to print, they’re telling the artist which words mattered most to them. Every shirt becomes a small piece of feedback about which lyric actually landed, and the fan ends up wearing a line they chose themselves.

Finally, there’s participation. We have a global community made up of music lovers, who transcribe, sync and translate lyrics. Our curators help us to ensure lyrics are accurate, properly synced and accessible to fans around the world. The more credit and better tools we give them, the better the data gets, which improves the experience for everyone.

• NEC has reportedly ended quantum hardware development after nearly three decades of research
• Several NEC quantum researchers have reportedly moved to rival Fujitsu
• NEC once ranked first globally in semiconductor sales in 1990

NEC has reportedly stopped building physical quantum computers after roughly three decades of research into superconducting circuits and solid-state qubit technology.

The Japanese technology company is instead expected to concentrate on quantum applications while avoiding the costs associated with developing physical machines.

Several engineers connected to the initiative, including a senior research figure, subsequently moved across to rival firm Fujitsu to continue their work.

A legacy built on chip dominance

NEC once dominated the global chip industry, ranking first worldwide in semiconductor sales back in 1990, ahead of several rivals.

Companies such as Hitachi, Toshiba, Fujitsu, Motorola and Intel all trailed well behind NEC's chip business throughout that period.

Intel eventually surpassed NEC as personal computer processors grew into a massive and highly profitable segment of the global electronics market.

NEC's 1999 breakthrough in superconducting qubits emerged from a firm that possessed vast internal expertise in solid-state device engineering and manufacturing.

Its researchers produced the earliest known solid-state qubit, a milestone still referenced widely across the quantum computing field.

Variants of that early breakthrough now quietly underpin quantum processors currently built separately by both technology giants Google and IBM.

Unlike NEC, IBM remained committed to quantum hardware and grew into a leading global developer of quantum computing systems overall.

That contrast makes NEC's exit relatively rare among the original corporate pioneers of quantum computing hardware development.

A retreat that took nearly a decade

NEC did not exit the semiconductor business overnight; the retreat instead unfolded gradually across roughly a full decade of restructuring.

The company effectively abandoned its DRAM memory operations in 1999, folding them into a joint venture that was later renamed Elpida.

In November 2002, NEC spun off its remaining chip units, covering microcontrollers and system-on-chip devices, into an entirely new firm.

That entity was subsequently renamed NEC Electronics, eventually becoming the major chipmaker now known globally as Renesas Electronics.

NEC's current business portfolio now centers on IT services, enterprise systems, aerospace, defense, telecom infrastructure, AI and security work.

Quantum hardware development apparently did not fit naturally alongside that existing and increasingly diversified corporate business focus of the company today.

NEC has not formally described the move as a complete withdrawal from quantum computing research.

The company says it will keep assessing real-world uses and commercialization paths connected to quantum technology.

Quantum computing hardware remains a notably costly pursuit, and returns on investment can take many additional years to materialize fully.

Meanwhile, Fujitsu, working alongside research partner Riken, already launched a quantum computer built around 256 superconducting qubits last year.

The firm aims to build a system exceeding 1,000 qubits sometime within 2026, according to earlier public statements from company officials.

Longer term, Fujitsu has reportedly set a goal of building machines that carry over 10,000 superconducting qubits before fiscal 2030 arrives.

NEC's departure suggests that even firms with decades of deep expertise increasingly view quantum hardware as commercially unviable in the short term.

Via Diamond Online (originally in Japanese)

I’ve been a huge fan of Dell displays for a long time. So, when I first saw the UltraSharp 32 4K QD-OLED (U3226Q), I thought two things: What a long name, and Wow, that looks incredible.

I have got really used to my 52-inch beast from Dell, so I know they are capable of delivering spectacular quality and novel ideas. So, the idea of Dell jumping feet first into a creative monitor is really exciting… and they did it very, very well.

This monitor has DisplayHDR True Black; it’s a QD-OLED anti-glare coating; it's a 120Hz, 4K panel; and it’s built from the ground up for color-critical workflows that are vital to getting the colors just right. There’s a magnetic shade for use in areas with harsh or inconsistent lighting, a built-in colorimeter, and up to 1000 nits of peak brightness.

The specs are incredible, and rightfully so; this monitor is truly brilliant in all of the best ways. As long as you forget that unseemly price tag.

Dell UltraSharp 32 4K QD-OLED: Price and availability

The Dell UltraSharp 32-inch 4K QD-OLED can be purchased for $2,440 at Dell in the US. And in the UK, it's got an equivalent cost of £2,405 at Dell UK. It's also available from a range of other retailers, but not reliably so.

Dell UltraSharp 32 4K QD-OLED: Unboxing and First Impressions

I haven't really used it on a monitor or with shaders before; however, I've recently seen several. I didn't really understand why it would be used, but now, after seeing half a dozen monitor shaders in the last six months, I have a pretty good idea of why you would want them, why you would need them, and why you wouldn't need them.

Right away, I noticed just how sleek and polished this monitor looks. Granted, I have very good experience with Dell monitors at this point, but I still recognize the build quality on this thing, even from first on most

The packaging is pretty straightforward. After testing all some of the best business monitors over the last couple of years, I've learned to expect a certain level of standard in packaging. The monitor itself had front and back protection to ensure that it won't break during transport.

Dell UltraSharp 32 4K QD-OLED: Design & Build Quality

There is a reason why Dell monitors are insanely popular, especially in the business world. That's because, even though they don't have all the gimmicks that many other providers do, they have built their reputation on dependability.

This monitor, even though it is a high-end performance monitor, is built the very same way. It has the essentials covered, so the user can rely on it for any use case. It's able to fit in a really small project space all around.

The last thing I'll mention about design and build quality is the shader itself. It's made from a nice material. It doesn't look like it will sag at all, and it magnetically attaches, which is fantastic.

You may not always want to use this shader, but when you need it, you can easily slip it on to the display.

Dell UltraSharp 32 4K QD-OLED: In use

Using this monitor has been fantastic. It looks super sharp. It's clean. It's simple. It's not overstated. It mounts great on modern monitor arms. It has great ports. All these things result in very smooth, easy usability for me.

I was able to get this all set up very quickly, and once everything was set up, I was able to use it right away thanks to its single USB-C cable for connecting a laptop or desktop.

Now, if I want to add more to this monitor, such as a secondary machine or something, it does have great ports and several options, though. It doesn't have nearly as many options as some of the more advanced monitors I have seen.

My team and I have used this monitor for photo editing, for video editing, and for basic website design. So far, it's been spectacular to use. Its picture quality is spectacular, and the shader is super easy to disable when I don't need it.

One thing that I'm very thankful for is that this monitor is still bright. Most of the monitors and shaders I have seen have a really dim max brightness, but this one still looks great with max brightness.

Another thing worth touching on is the ports for the Dell Ultrasharp for the desktop setup in the photo. I use a single USB-C cable running directly from my laptop to the monitor.

If you've a more complicated setup where you're using one of the mini Thunderbolt docking stations or something similar, you can plug directly into it via USB-C or, of course, HDMI.

For the longest time, I would recommend Samsung or BenQ panels for creative work - in my experience, those companies produce some of the best monitors for photo editing and the like.

However, several more companies are stepping up when it comes to high-quality creative panels. Dell is clearly one of them, and honestly, I don't think its are too late to the game either. I think it's right on time, bringing its background of reliability and dependability to a growing demographic.

Dell UltraSharp 32 4K QD-OLED: Final verdict

If you know that you are looking for a creative monitor with specific specs, the best panel possible, and you even know what a shader is and what it does, then this is the kind of 4K monitor you should be looking into.

If you have to wonder whether this is worth it, it probably isn’t. It's total overkill for running standard office software and streaming the occasional movie.

If you know you're looking for an absurdly nice monitor because you are a video editor, a photo editor, or a content creator in some way, and you need a high-quality one, then absolutely make sure you check this one out.

• Meta’s Danish facility converts server heat into household heating for thousands
• Odense uses existing infrastructure to transform wasted data center energy
• Heat pumps help raise server warmth to temperatures suitable for homes

A data center operated by Meta in Odense, Denmark is channeling its cooling heat straight into nearby homes.

Server operations typically send almost all of their electricity out as warmth, with one industry estimate placing that figure near 90%.

Most facilities simply vent that energy skyward, but the Odense site routes it through the city's own heating infrastructure instead.

How the recovery system actually works

Roughly 40% of a data center's total energy consumption goes toward cooling servers, and nearly all of it is released as low-grade heat.

That heat arrives too cool for direct use, since server exhaust never reaches temperatures high enough to warm buildings on its own.

Water from the municipal district-heating network runs through the facility's cooling units, absorbing that low-level warmth from the servers, and then moves to a nearby plant packed with ammonia heat pumps that raise the temperature further before distribution.

Engineering firm Ramboll designed this recovery system for district-heating operator Fjernvarme Fyn, lifting water temperatures between 70 and 75°C.

That range makes the water hot enough for household radiators once it re-enters the wider distribution network.

"Our project is relatively unique because we have a very large distribution network," said Jan Strømvig, head of Fjernvarme Fyn, adding Meta placed its facility directly inside that network, allowing highly efficient energy transfer throughout the system.

On its part, Meta described the achievement as one of scale rather than any fundamentally new engineering breakthrough.

"Heat pumps are not a new invention, nor are coils for capturing heat. What is new and innovative is the pairing of these two in hyper scale," said Lauren Edelman, energy specialist at Meta

Numbers vary, but scale remains substantial

Ramboll's published figures describe roughly 45 megawatts flowing directly into Odense's heating network, making it Denmark's largest heat-pump installation.

Recovery figures cited by Ramboll point to about 215,000 megawatt-hours annually, warming more than 12,000 homes across the region.

Earlier project details from Meta cited smaller numbers, including around 6,900 homes and 100,000 megawatt-hours during initial operations.

Following expansion, Meta's own figures rose to roughly 165,000 megawatt-hours serving approximately 11,000 households within the system.

Fjernvarme Fyn confirmed in 2025 that Meta's surplus heat output was equivalent to warming about 11,000 households.

These fluctuating figures reflect different project phases, timeframes and counting methods rather than any inconsistency in performance.

Replicating this setup elsewhere depends heavily on one specific ingredient many cities simply lack entirely — a wide-reaching district-heating network.

Heat cannot travel far without significant loss, meaning facilities must sit close to an existing distribution network.

Odense already had that infrastructure in place before Meta's data center arrived on site, and this is part of the success of the system.

This situation stands out against a backdrop of growing resistance to data centers elsewhere, particularly in the United States.

Communities there have pushed back over land use, water consumption and strain on local electricity grids.

Thailand has raised similar objections recently, citing comparable pressure on its own limited resources.

Amid that friction, Odense offers a rare example of data center heat actually benefiting the surrounding community.

Via Science Blog

• South African groups demand reviews before approving more data center projects
• Cape Town's water crisis memories fuel opposition to new facilities
• Equinix's planned facility raises concerns about electricity consumption levels

Civil rights organizations across South Africa are pushing for a pause on new data center construction while resource concerns get investigated.

They argue scarce water, land and electricity supplies need proper scrutiny before any further approvals move forward.

The push places South Africa alongside Thailand and the United States, where similar community pushback has emerged in recent months.

Growing scrutiny around digital infrastructure expansion

According to South African President Cyril Ramaphosa, South Africa hosts 70% of Africa's total data center capacity, and he believes this dominance is an economic opportunity.

Major American technology firms including Amazon, Microsoft and Equinix have steadily expanded their digital infrastructure footprint within the country.

The South African Human Rights Commission says it received more than 250 submissions after inviting public input back in May 2026.

South Africans are mostly concerned about the transparency with which the government handles the entire process.

"A key issue emerging from the process is precisely the availability, consistency and transparency of information relating to matters such as electricity and water demand, land use, infrastructure requirements, environmental impacts and impacts on surrounding communities," said Dr. Eileen Carter, who leads the commission's preliminary process.

One especially contested project involves an Equinix hyperscale facility approved for construction in Cape Town, a city with a painful memory of near-catastrophic water shortages.

Cape Town narrowly avoided what residents called "Day Zero" back in 2018, when officials predicted municipal taps would run completely dry.

That facility alone is estimated to require roughly 160 megawatts of electricity, reviving old anxieties tied to the nation's rolling blackouts.

Investors and operators push back on pause calls

Eskom, the country's national power utility, reported a surplus during this year's winter peak-demand season despite past supply troubles.

Still, civil society groups remain wary that this surplus capacity could simply get redirected toward powering new data facilities instead.

Residents believe that clarity in handling these concerns benefits everyone, but this does not seem to be the idea of some major players

"Serious investors are not deterred by clear rules. They price uncertainty, and an unregulated boom is the most uncertain environment of all," said Pitso Tsibolane, senior lecturer in Information Systems at the University of Cape Town.

At the moment, data centre operators in South Africa are not required to disclose their actual water, electricity and land commitments upfront, and this is a major problem.

Sasha Booth-Beharilal, who chairs the Internet Service Providers Association, pushed back against framing data centers as a driver of scarcity.

She maintains that rising electricity demand from these facilities has not contributed meaningfully to South Africa's broader tariff increases or shortages.

Operators also caution against comparing local projects directly to sprawling American data centers, since resource footprints differ substantially between the two contexts.

Whether a temporary construction pause actually produces clearer regulatory rules, or simply delays much-needed digital investment, remains genuinely uncertain for now.

Via The Denver Gazette

Now that businesses of all sizes are adopting AI technologies, the next step is to prove the technology is delivering measurable benefits and improvements. But nailing down a singular metric to show these benefits is proving difficult.

Sure, businesses can see exactly how much it is costing them, but measuring output is another beast entirely. Tokkenmaxxing has shown that arbitrary targets of use are not necessarily the best way to measure the productive impact of AI.

But perhaps there is a lesson to be learned from the adoption of other, quieter technologies that are delivering measurable benefits in new industries, and how AI can fit into workflows alongside this new and exciting tech infrastructure.

AI isn't the only tech seeing widespread adoption

Perforce’s 2026 State of Real-Time Workflows Report found it isn’t just AI seeing widespread adoption across industries. Game engines and real-time 3D engines, once used almost exclusively by game developers, have seen a huge wave of adoption across the aerospace and defense, public sector and education, and even media and entertainment.

These tools show measurable improvements in productivity before adoption because they are tried and tested. They have decades of proven returns shown by the gaming industry. AI on the other hand is yet to show sustained, measurable return on investment for many businesses. But there are also a host of other issues accompanying AI use.

Version control software is also seeing a rapid growth in adoption in real-time workflows. AI is a driving force in this adoption as businesses now have to handle significantly more assets and content within each project. Having visibility into who changed what - especially with AI agents now involved in workflows - is no longer a choice, but a necessity.

One of the main concerns remains job replacement. Perforce’s report found that 50% of employees who had adopted AI in their workflows feared they would be replaced by the technology. But AI concerns also extend into the work they are doing; 49% feared their AI tools would produce poor or inaccurate content and 48% held ethical or compliance concerns about their use of AI technology.

To understand the challenges businesses are facing in showing measurable return from AI adoption, I spoke to the author of Perforce’s report, Brent Schiestl. Brent leads Perforce’s Digital Creation business unit and is the Senior Director of Product Management.

• Businesses may be seeing greater productivity gains when adopting AI into workflows, but it is having a negative effect on the creative outlet for employees and the perception of brands by consumers. What steps are businesses taking to maintain trust with both groups?

We see a distinction in perception when comparing AI usage in generating source code vs. binary assets. For example, if a team is using AI to autonomously fix a bug in source code, consumers seem more accepting of that use case. A bug fix is a well-defined problem where creativity is usually not the key to solving it.

In fact, one could argue that engineers are being freed from mundane work to focus more on creative work when deploying AI this way.

Where we see more negative perception is with binary assets (images, audio files, movies, etc.). Because many LLMs were trained on data without originating author consent, producing binary assets via AI tends to get more scrutiny.

Some steps we have seen companies take to maintain trust include establishing clear public AI usage disclosure policies, making it clear that humans remain in the loop for final asset creation, and maintaining strong provenance data for how the asset came to be.

With these steps, the end consumer is given a complete enough data set to decide whether they feel the AI usage clears their own ethical hurdles or not.

• There is a clear move away from tokenmaxxing to measure the value of AI. What metrics are businesses now using to show the cost-to-benefit ratio of AI deployment and what new problems has this introduced?

This is the single biggest question we get regularly from our customers. Our customers are using AI more than ever but struggling to prove that AI is benefiting them in a way that justifies the expense.

Some examples of metrics that we see related to the cost-benefit ratio of AI include release frequency, amount of content in each release, and telemetry to understand how new features are being used, for example.

In addition, we see industry standard metrics like DORA rising in importance as customers want to benchmark their productivity more broadly.

One key is to consider Goodhart’s Law, which states that when a measure becomes a target, it ceases to be a good measure.

Introducing new metrics around measuring the value of AI can lead to engineering teams gaming metrics at the expense of doing what is best for the final product.

In addition, isolating AI’s contribution from all other release activities is a new challenge that all companies are wrestling with.

• Why are game engines being adopted so readily by so many industries, and what blockers previously prevented their use in these industries? Are there lessons that other, less technical industries can learn from this adoption?

I believe that one of the biggest blockers was literally in the name itself, as formally referring to them as “game engines” siloed the use case right out of the gate for anyone not in gaming.

These engines were historically hard to deploy outside of a gaming context including, but not limited to, licensing, tooling for non-artists, and integration with enterprise systems.

For example, Epic Games has had to figure out how to monetize Unreal Engine outside of gaming. Even we at Perforce have historically surveyed our customers and up until last year we used to refer to our official report as the “State of Game Technology Report”.

This year we renamed the report to the “State of Real-Time Workflows Report”. Once adjacent industries realized that these engines are really a bundle of rendering, physics, networking, and asset pipelines, new industry verticals literally sprang out of nowhere. Sometimes it’s more about positioning than anything else.

• What effect is the adoption of new technologies such as AI, game engines, and 3D modelling software having on the infrastructure costs of industries that traditionally did not use these tools?

Infrastructure costs including GPU compute, storage for large binary/3D assets, bandwidth, specialized workstations, and licensing for engines/DCC tools, have led to businesses needing to justify these new expenses.

The easiest way to justify is to realize an increase in revenue based on the investment. The challenge is that these sorts of investments can oftentimes take years to realize the benefits.

Companies, especially CFOs, need to remain patient in the early stages. One of the worst outcomes for companies is reacting in a knee-jerk fashion before the rewards can be reaped.

Another challenge is that these industries often lack the IT muscle sized for this new (to them) infrastructure, meaning the cost isn't just the compute/storage line item, it's also the organizational capacity to run it.

• What tools are businesses using to manage the associated technical debt that comes with AI productivity? How are these tools helping manage quality, compliance, and security?

Technical debt is rising from new sources such as unreviewed AI-generated code (by humans and/or by agents), dependencies pulled in by AI that no one owns, license contamination, and model version drift.

Some tools that we’re seeing fill this space include AI-aware code review (e.g., CodeRabbit, Greptile, P4 Code Review, etc.), SAST/DAST tools tuned for AI output (e.g., vulnerability patterns in AI-generated code), license/provenance scanners (e.g., copyleft contamination risk), and version control practices that treat AI-generated commits as first-class artifacts.

• How is AI changing the open source market, helping businesses develop their own solutions, and what effects will it have on the traditional software licensing market in the future?

AI is changing the open source market in a couple of ways. First, it’s never been easier to develop, and then if desired, open source your own solution.

On the other extreme, we’re seeing reports of previously open source repositories being turned private due to the sheer number of AI-generated pull requests being raised and the inability of the repository owner to keep up.

Traditional software licensing that is purely seat-based is being tested as the assumption is that companies may either downsize or at least not grow at the same pre-AI rates that we had become accustomed to.

The main question that feels unanswered today is whether the build vs. buy math is genuinely changing or not.

One of my favorite memes goes something like, “I saved $30,000 in subscription costs by building my own solution and it only cost me $100,000 worth of tokens to do it.”

While the meme exists to poke fun, it is something that needs to be taken seriously because initial build and ongoing maintenance plus support needs to be accounted for.

Just when it seemed the creeping presence of technology into our daily routine couldn't get any worse, Meta has launched Muse, a new way for AI to take control of your life.

Described as "The World’s First Personal AI Agent Built for Everyone", Muse looks to be an hybrid work and home life AI assistant for everyone, even those lacking technical expertise, offering everything from making recipes and shopping lists to work-related tasks such as managing your calendar.

Except let's be honest, it probably do anything like that - because that's not how real everyday life and work is, is it - so is Muse already over-promising?

A Supermassive Black (AI) Hole?

Looking through the list of things Muse says it can do, and its promise that it was "built to work for billions of people worldwide", is another reminder that a lot of new AI innovations and services are often built by people who don't understand how the real world works.

Tools such as monitoring a smart home and planning the next big holiday might be fine for a Silicon Valley based worker who drives an hour to the office and back, but for those of us outside the bubble, it's all a bit much.

When it comes to the business and work-focused tasks, it again seems like there's a lack of basic understanding.

Mark Zuckerberg has said Meta needs to create more accessible agents for people, with the likes of OpenClaw just too advanced for the bulk of Meta's users across Facebook, Instagram and WhatsApp.

Muse will let users draft and send emails (always a bit of an iffy area with AI agents) although it does say the agent will ask for approval before sending anything - but it also has bigger plans it helping spur on bigger projects or plans.

Meta says Muse can help with "turning long-term goals into action plans" - and will even work behind the scenes, even when the app is turned off, to move forward on this.

Muse, which comes with its own dedicated apps and website, can handle complex tasks and work independently, Meta says, noting that "once a person shares a goal with Muse, it helps them develop a personalized plan and coordinate their time and resources, then advances the work on its own."

Whether it's the aforementioned holiday plans or fitness goals, all the way up to starting a business, Meta seems to see Muse as an always-on assistant and co-worker, but surely this takes away from the feeling of actual achievement?

Time is running out

Meta also makes a big deal out of building safety, security and privacy into Muse, perhaps unsurprisingly given the current furore around its AI 'Pervert glasses', and the amount of data it is asking users to share with its agent.

The company says that personal agents like Muse "need a new kind of secure computer, so Meta built one for everyone", with the Muse Secure VM supposedly offering "first-of-its-kind privacy, safety, and security protections" built in.

Meta says that, "each person stays in control of their Muse and decides how much access it gets" to their information - but if you're pumping in data about your daily life and work projects, how far does that really stretch?

Muse can set up its own connections to third-party services if a public API is available, naming the likes of Stripe, Google Workspace and 1Password, which sounds like both a useful efficiency gain and a security nightmare waiting to happen - I guess we'll have to wait and see.

Fortunately, Meta is apparently already anticipating teething issues for Muse, noting in its launch post that, "Muse can and will still make mistakes, but we expect they'll be much less frequent and cause much less damage due to the safety systems we've built in."

Given Zuckerberg's well-publicized push to create "superintelligence" (whatever that means) I really hope Meta has bigger plans for Muse, as surely the point of technology such as this is to make our lives better - they just clearly need to talk to some actual people first.

• Epson survey finds 76% of students expect to be able to use AI in their learning
• Meanwhile, 81% of teachers believe students think AI can do spelling and maths for them
• Following the introduction of the EU AI Act, schools must have AI literacy training

Are students over-relying on artificial intelligence chatbots? That seems to be the opinion of teachers across Europe, where 80% of educators have expressed concern about the pace of AI entry into the classroom.

A survey of 3,360 people by Epson discovered over two-thirds (68%) of teachers feel that AI use in homework has a negative effect on learning. Conversely, over three quarters of students expect to be able to use AI, with almost 90% already using it once a week for school work.

This research is released as the EU’s AI Act commences enforcement, forcing schools in the European Union to ensure adequate training is provided for any AI tools in use.

Easy task completion

Epson Europe’s Educate to Empower 2026 Research surveyed 3,360 people across the EU, specifically France, Italy, Germany, Spain and Poland, as well as the UK. (There is no equivalent to the EU’s AI Act in the UK, although businesses providing services to companies within the EU must adhere to its regulations.)

“AI is developing at speed, and students expect to be able to use it. That means it needs to be managed effectively, with the right governance, guidelines and training in place," said Dr Sarah Henkelmann-Hillebrand, lead for education at Epson Europe.

With 76% of students expecting to be able to use AI and 87% already employing it in some way, the horse has bolted on blocking its use. Instead, the survey finds, the focus should be on teaching skills that cannot be short-circuited by AI.

Henkelmann-Hillebrand notes that “It’s also important to look at how AI can enhance the learning experience. By supporting students and teachers to co-create in immersive learning spaces, using technologies such as projection alongside AI, we can make learning more engaging and collaborative while helping students develop the skills they’ll need for a future dominated by AI.”

Teachers want more AI training

While student use of AI is cause for concern where it impacts their ability to learn and demonstrate comprehension, teachers have an additional challenge.

The survey’s findings also revealed that 82% of teachers want more training to oversee the use of AI by students. Additionally – and perhaps more significantly – 78% want training and guidance on how they can use the technology in their own work.

Could it make more sense for teachers to use AI to reduce their admin, than to permit AI use by students? If AI is taking over the classroom, it seems sensible to ensure teachers are fully equipped to encourage skills that AI cannot give students, such as analytical thinking, creativity, leadership, problem-solving, and other uniquely human traits.

GMKtec EVO-X3: 30-second review

The GMKtec EVO-X3 is a mini PC with a difference: firstly, it stands upright; the size is larger than most, the specs are impressive, and the price- well, that prices it out of most people's reach; however, this is a machine with an AMD Halo Strix processor.

The EVO-X3 features the AMD Ryzen AI Max+ 395 processor; that’s the same as the EVO-X2 with 16 Zen 5 CPU cores, Radeon 8060S graphics, and a 50-TOPS XDNA 2 NPU. In this review, I’m looking at the machine with 128GB of LPDDR5X unified memory, enough to load and run local AI models; this ability is well beyond the capabilities of most other Mini PCs and even desktop PCs.

A quick look at the hardware makes it clear this is different, and for a machine with its specs, it’s still relatively compact, thanks to its slim, tower-like structure. It’s essentially been designed to stand upright to ensure plenty of airflow.

During testing, I ran local chat, reasoning, image recognition, image generation, transcription, and coding models on the EVO-X3. Smaller language models started responding almost immediately, while larger models provided more detailed reasoning and better-quality copy, essentially mirroring what you’d expect from cloud models.

You could also run these tasks with Wi-Fi disabled, so completely local and offline, with no need for credits or additional spend. This also meant all prompts, photographs, audio recordings, and project files stayed on your computer rather than being sent to a cloud service where you don't know what happens to them.

While the main focus here is the AI potential, I also ran the conventional benchmarks, and these are impressive. If you need a cutting-edge gaming PC and want to underplay the machine's potential, this is a great compact option, especially with the OcuLink connector.

However, refocusing on AI and unlocking the EVO-X3’s AI performance proved that some knowledge and experimentation are still needed; this isn’t quite an out-of-the-box, ready-to-run solution, although GMKtec has taken huge strides in that direction.

Local AI through the GMKtec Claw software environment was impressive. During testing, the Herdsman model manager recognised the NPU but reported that the Ryzen AI Max+ 395 was unsupported by its current NPU runtime manifests, which I found odd since the software was pre-installed.

Installing Lemonade and its Ryzen AI runtime solved the problem. Compatible models then pushed NPU utilisation close to 100%, proving that the hardware was operating properly.

This leaves the EVO-X3 in an awkward position. It’s an exceptionally powerful local AI workstation, but it is not yet the polished, out-of-the-box AI machine that the marketing suggests. It will do everything advertised, but for you to get there, you must understand models, runtimes, servers, agents and external applications, or at least be prepared to learn.

Whether or not this deserves a spot on our guide to the best mini PC is, ultimately pretty subjective. At $3,799.99 for the current 128GB/2 TB configuration, this is a relatively cheap machine for experienced developers, and technically minded creative professionals will find it an extraordinarily useful tool if they’re willing to put in the time. If you’re less experienced, then hold out; a more streamlined and cheaper option will inevitably be available soon.

GMKtec EVO-X3: Price and availability

• How much does it cost? From £3,029 / $3,799
• When is it out? Available now
• Where can you get it? Directly from GMKtec or Amazon

The GMKtec EVO-X3 is available in the US direct from GMKtec and on Amazon.com, with the base 128GB/2TB configuration priced at $3,800. However, you can save an extra $100 at GMKtec or 5% on Amazon with code GMKX3DP3.

In the UK, the EVO-X3 is priced at £3,030 from GMKtec. At the time of review, it's not currently on Amazon, but I'd expect it to arrive shortly.

The EVO-X3 comes with either a 2TB or 4TB SSD, with both versions featuring 128GB of onboard LPDDR5X-8000 memory.

• Value: 4 / 5

GMKtec EVO-X3: Specs

CPU: AMD Ryzen AI Max+ 395, 16 cores/32 threads, up to 5.1GHz
GPU: AMD Radeon 8060S, 40 compute units
NPU: AMD XDNA 2, 50 TOPS; up to 126 TOPS total AI performance
RAM: 128GB LPDDR5X-8000, non-upgradeable
Storage: 2TB or 4TB PCIe 4.0 M.2 2280 SSD, two slots, up to 16TB total
Front Ports: USB 3.2 Gen 1 5Gbps, USB 2.0, 3.5mm audio
Rear Ports: HDMI 2.1, USB4 40Gbps, USB 3.2 Gen 2 10Gbps, OCuLink PCIe 4.0 x4, 2.5GbE RJ-45
VESA Mount: Not specified
Dimensions: 13.90 by 7.32 by 1.61in (353 x 186 x 41mm)
Weight: 5.07lb (2.3kg)
Power: 230W external power adapter, 19.5V/11.8A
Operating System: Windows 11 Pro, with Ubuntu and Linux support

GMKtec EVO-X3: Design

The EVO-X3 design is far removed from the usual square mini-PC, and instead goes for a tall, narrow metal enclosure with vents on either side that help draw cool air through the system to help maintain maximum performance.

While it’s tall, the footprint because of this is relatively small, meaning it can sit neatly next to your monitor. Standing at 353 x 186 x 41mm without its stand, it looks more like an oversized external graphics card than a conventional desktop computer; the small stand that’s included is held in place with two screws and adds stability to the vertical stance.

While the unit is slim, it does have a good weight at 2.3kg, so while you could transport this very easily, this is one designed to stay in situ.

The full-metal CNC-machined casing looks great once you get used to the slightly different design, and with the grey finish with green detailing, the aesthetics are all well balanced. Compared with the small boxy Mini PCs, this definitely looks and feels like a professional workstation, as the pricing suggests.

Taking a closer look at the cooling and the slim tower design enables air to be pulled in through the single vent plate on the top/side of the machine and output through the two vent panels on the base/side. Checking inside the casing, there are three heat pipes and multiple cooling fans that all help to keep the powerful components cool.

Through the test, I was impressed by just how well and how quietly these fans were and left on auto; while you can hear them, they’re never overbearing. Through the interface, you also have the ability to adjust the mode, from Silent 45W to Balanced 85W and Performance 140W.

One of my biggest surprises when it comes to the design is the port allocation, or lack of it. The front provides one USB-A 3.2 Gen 1 port, one USB-A 2.0 port and a 3.5mm audio connection. Considering the EVO-X3’s intended creative and professional audience, another high-speed USB connection or front-mounted USB-C port would have been handy.

The rear is slightly better equipped, with HDMI 2.1, 40Gbps USB4, 10Gbps USB-A, 2.5Gb Ethernet and an OCuLink connection. I would have preferred a little more, as connecting to a hub of some type is essential if you have further accessories.

The most interesting addition here is the OCuLink that enables you to add an external desktop graphics card via a more direct PCIe connection than a typical Thunderbolt or USB4 enclosure. Hot plugging isn't supported, so you must shut down the computer before attaching or removing an external GPU.

Wireless connectivity options are all pretty standard with Wi-Fi 7 and Bluetooth 5.4; then there’s also the wired 2.5Gb Ethernet. This is slightly lacking, and I would have expected a 5GbE or 10GbE at the least.

However, the lack of ports aside, the big story here is AMD’s Ryzen AI Max+ 395 that combines a 16-core, 32-thread Zen 5 CPU with Radeon 8060S integrated graphics. The GPU includes 40 RDNA 3.5 compute units, while the XDNA 2 NPU contributes up to 50 TOPS of dedicated AI processing.

Other machines with this CPU, such as the EVO-X2, they feature only 64GB of RAM and no Oculink connection, both of which, as I discovered, really are essential if you want to make the most of running local AI.

• Design: 4.5 / 5

GMKtec EVO-X3: Features

At the heart of the EVO-X3 is AMD’s Ryzen AI Max+ 395, part of the AI power family Strix Halo. This chipset offers a combined 16-core, 32-thread Zen 5 CPU alongside Radeon 8060S graphics and an XDNA 2 NPU all within a single processor package, which is why it can fit into this small form Mini PC.

The CPU can be boosted to 5.1GHz and is supported by 16MB of L2 cache and 64MB of L3 cache. Alongside it, graphics processing is supplied by the Radeon 8060S, which provides 40 RDNA 3.5 compute units operating at up to 2,900MHz.

What is impressive here is that the performance of this integrated GPU has, in tests, beaten many dedicated GPUs. As GPU power is central to creative work, gaming and local generative AI, this power will be a huge benefit to the majority of Creative Apps that I use.

Alongside the CPU and GPU is the XDNA 2 NPU that provides up to 50 TOPS of dedicated AI processing, which brings the combined processor performance to up to 126 TOPS. The NPU essentially handles supported AI workloads better than running everything through the CPU or GPU. Again, surprisingly, some of the AI Apps used in the test fell back to the CPU and GPU rather than making use of the NPU and the speeds with and without made a big difference.

If you’re looking through this and thinking this all sounds very familiar, then it is essentially a very similar configuration to the EVO-X2, which is again equipped with the the Ryzen AI Max+ 395, Radeon 8060S graphics and a 50-TOPS NPU. The EVO-X3 adds the new chassis design to maximise the new cooling system, OCuLink connection, 128GB of standard memory capacity and a pre-installed AI environment to make getting started easier.

My review sample of the EVO-X3 features 128GB of onboard LPDDR5X memory running at 8,000MT/s. This memory is soldered and cannot be upgraded, but up to 96GB can be allocated as graphics memory. That large shared-memory pool is useful for local AI because it enables the integrated Radeon GPU to load models that would exceed the dedicated VRAM capacity of most graphics cards.

Storage options are either 2TB, as in my review sample, or 4TB M.2 PCIe 4.0 NVMe SSD. It includes two internal M.2 2280 PCIe 4.0 x4 slots, which enable you to install up to two 8TB drives, with a maximum internal capacity of 16TB. Having downloaded and installed several models, tools and applications through the test, I would definitely install a minimum of an additional 4TB.

One of the big upgrades over the EVO-X2 is the rear OCuLink connection. This provides a PCIe 4.0 x4 interface for an external desktop graphics card; you do, of course, need an eGPU enclosure as well, and then this gives you the ability to add an Nvidia or AMD GPU for rendering, gaming, as well as those AI workloads. OCuLink does not support hot plugging, so if you do use one, the machine needs to be powered down before the port is installed. I don’t at present have an OcuLink enclosure, so I haven’t tested the boost that this will give to performance.

On the back there’s also the USB4 connection, but only one, and it enables data transfers at up to 40Gbps, DisplayPort video output at up to 4K 60Hz and USB Power Delivery input at up to 100W. If I were to use this machine long-term, then I would connect a hub to maximise connections, especially for creative use, as just one USB-C is limiting.

Display output is limited to two screens, which can be connected through HDMI 2.1 and USB4. The HDMI connection supports higher-resolution displays, while the USB4 port provides DisplayPort output up to 4K at 60Hz.

Connectivity includes Wi-Fi 7, Bluetooth 5.4 and 2.5Gb Ethernet. The front has two 5Gbps USB-A 3.2 Gen 1 ports, a 3.5mm combined audio socket, and the power button. Around the back are 40Gbps USB4, 10Gbps USB-A 3.2 Gen 2, HDMI 2.1, OCuLink, 2.5Gb Ethernet and the main DC power input.

The absence of a front USB-C port or card reader is disappointing for such an expensive creative workstation. This is especially noticeable because the cheaper EVO-X2 includes both front USB4 and an SD card reader.

I was also pleased to see three performance profiles provide a choice between power consumption, heat and processing speed. Silent mode operates at 54W, Balanced mode increases this to 85W, and Performance mode allows sustained operation at up to 120W, with short peaks of up to 140W.

Three heat pipes, two fans, and intelligent fan control handle cooling by pulling air through the vent on the top/side and pushing it out through the bottom/side. This cooling remains relatively quiet throughout the testing and is designed to maintain the processor’s performance during heavy AI use, as well as video rendering and development workflows.

As you would expect, Windows 11 Pro is preinstalled, with support for Ubuntu and other Linux distributions. GMKtec also supplies its Claw local-AI environment, which is promoted as offering one-click model and agent deployment and was well implemented; there’s also Herdsman and installers for LM Studio and Ollama, giving you a good choice of Local AI management.

Claw is well integrated with EVO-X3 as it brings together local models, cloud services and AI agents without requiring you to assemble the entire software environment. The associated Herdsman model manager detected the NPU but reported that the Ryzen AI Max+ 395 was unsupported by its current runtime manifests. The underlying hardware worked once Lemonade and its Ryzen AI runtime were installed, but that workaround goes against GMKtec’s claim that local AI is ready to use without coding or setup.

The EVO-X3 has an exceptional feature set for local AI with the 128GB unified-memory capacity, integrated graphics, NPU, dual SSD slots and OCuLink connection enabling both flexibility and power. The issue is that most of this is already present in the substantially cheaper EVO-X2, making the X3’s £3,029.99 price seem a little steep.

• Features: 4.5 / 5

GMKtec EVO-X3: Performance

Getting started with the EVO-X3 is much like any other mini PC: plug in, screw the stand into the base and power on. You then have the Windows 11 Pro setup to run through, and as ever, GMKtec has covered the LAN port with a warning not to connect until the full setup process has been completed.

Once installed, I instantly noted the speed of the system; this was as fast and responsive as they come, and powering down and starting up from cold takes just 18 seconds from power being pushed to the home screen and the machine being ready to use.

As I started to load the applications for the benchmarks, the machine was fast and responsive, really as you’d expect for a machine at this price, and I was happy to see that this was at least in line, if not faster, than my workstation. Opening Premiere Pro, Lightroom and Photoshop all worked seamlessly, with the only real issue being the lack of ports for connection at the back.

When it came to the usual software and benchmarking tests, I ploughed through with almost all hitting record speeds and values when compared with any Mini PC or Laptop that I’ve reviewed in the past. An example is the Geekbench score, which aligned with many of the high-performance Mini PC’s and laptops, with a score of 2,956, but then the multi-core result of 19,401 far exceeded anything else.

Pushing the rendering benchmarks to give an idea of how it would perform for 3D content and video, the Cinebench R23 produced 33,295 points in the multi-core test and 2,034 points using a single core again; this is almost double anything else that I have seen.

Due to the positioning of the machine in the market, you would expect decent graphics performance, but due to an AI rather than games and creative perspective, and sure enough, the graphics performance was superb as I settled into some gaming. The Radeon 8060S hit scores of 101,583 in Geekbench GPU, while 3DMark Fire Strike returned an overall score of 25,721.

The EVO-X3 scored consistently high results through the 3DMark benchmarks, with most doubling the results that I have seen from other high-performance Mini PC. This performance not only translated to the test but also to gaming, as I settled back to play Cyber Punk 2097.

My review sample came with the 2TB NVMe SSD, and the transfer rates were slightly below what I expected but are consistent for a PCIe 4.0 slot, with read and write speeds of 4,912.96 MB/s and 4,648.62 MB/s.

Rounding out the standard tests, I checked out PCMark and wasn’t surprised to see a high score of 9,857, while the Windows Experience result was 9.5, again one of the highest I’ve recorded this year.
GMKtec EVO-X3 review: AI software

The EVO-X3 is sold as a ready-to-go machine with local AI models installed and one-click agent deployment through its Claw software environment. Claw uses Herdsman as the gateway to the AI models, so you start Herdsman and then open Claw, and the rest is quite straightforward for chat, image generation and countless other Agents that can be downloaded and installed.

Using this approach simplifies everything, and as I downloaded SuperPowers Agent, software workflow, Paper Research, Office Copilot, slidewright and a few others, I quickly built myself a decent set of tools for day-to-day AI use. The models you use can be downloaded in Herdsman and then used in GMKClaw. Depending on the complexity of the model you use and whether you switch Thinking on or off, the speed of the AI model you have selected varies.

What I like about this is that once you have models loaded, the application then downloads the additional tools that it needs to run. This is obviously online, with the AI processing done offline. Initially, the results were hit and miss, with the simple prompt, “A photo of a Labrador on a beach”, creating an abstract; once it had a few more prompts and details, the final results started to look better. The point here is that the more you use it, the better it gets.

While I also wanted to test the offline AI ability, what I also realised quickly was that I needed to run examples of everything I wanted to do; it needed to download and upgrade the models and tools that it had to hand.

Through the menu at the top of the screen, you can toggle between the different AI models, with Auto selected as default. This auto mode switches between the cloud and local to generate the images, but I noted that the NPU throughout stayed silent, registering 0% in Task Manager.

While the performance of the models was good both online, offline and hybrid, this wasn’t the potential that I was hoping for from this machine. I then went down another route and installed Lemonade and a series of Apps to give me greater control.

Installing Lemonade provided a way forward. After downloading its Ryzen AI NPU runtime and compatible models, Windows Task Manager showed the NPU operating at near-full potential, and a quick test in chat showed just how fast this machine's AI could be.

Lemonade provided what I had hoped for from Claw, with the clearest interface for testing local models; the GMKTecClaw integration is good but not yet mature enough to reach its full potential. Lemonade also displays time to first token, token-generation speed, memory use and processor activity, while making it relatively easy to switch between NPU, hybrid and GGUF models.

However, the setup process does require large model downloads and multiple tools and other additions to be downloaded. It’s also worth highlighting that downloading several models simultaneously resulted in HTTP 429 rate-limit messages from Hugging Face. Some completed downloads also encountered temporary file-locking problems, so for the complete set that I ran in this test, I downloaded over a few days, trying one task and then returning for more models for images, text, speech, and more.

Once Lemonade and its Ryzen AI runtime were installed, the EVO-X3’s NPU performance was impressive. I started out with the relatively lightweight Qwen-2.5-1.5B-Instruct-NPU model, which responded in around a second per response. Checking the stats, it recorded a time to first token of 0.61 seconds and generated 25.1 tokens per second. RAM usage reached 16.7GB, while CPU utilisation remained at only 4.8%.

This made it a good model for quick questions, summaries and relatively straightforward writing tasks.

Moving to Qwen2.5-7B-Instruct-NPU resulted in slower responses, as expected from a more complex model, but provided a useful step up in capability. The Qwen2.5-7B-Instruct-Hybrid model was again fast, and when asked to generate the opening of an imaginary book, it began responding in about a second, with Lemonade showing 523 input tokens, 393 output tokens, and a time to first token of 0.47 seconds.

I then loaded the larger gpt-oss-20b-NPU, more brain and more thinking, so everything took longer to consider when it came to the responses but produced noticeably better copy. It did, however, become slightly preoccupied with satisfying the requested word count and took quite some time rewriting the 100 words. That test processed 825 input tokens and generated 1,500 output tokens. Time to first token was 1.67 seconds, while RAM use reached 25.6GB and CPU remained at 5.9%.

These tests show why model selection matters. The smallest model delivered the fastest chatbot for general use, while larger models slowed speed for improved reasoning and writing quality. All processing was run locally, disconnected from the internet.

The next step was to plan a project, or, to be more accurate, all the things that I had to get done today. I wanted a plan and a spreadsheet that I could open on my phone and also print out if needed. I used DeepSeek-R1-Distill-Qwen-7B-NPU to organise the workday and plan an entire website project, covering content gathering, site structure, visual design, and implementation.

After asking several questions to clarify the project requirements, the model produced a detailed, methodical plan with useful software suggestions.

The website-planning conversation used 4,118 input tokens and generated 490 output tokens. Time to first token was 3.38 seconds, with RAM use reaching 19GB and CPU utilisation at 5.2%.

Generating the plan was straightforward, but turning it into a downloadable CSV proved difficult in the basic chat interface. While the model could create correctly formatted CSV text, it couldn’t save it as a downloadable file. Copying and pasting the information worked, but this highlighted the difference between a chat model and an agent with access to files and tools, which is where I started to discover additional tools needed to be used; with coding later, Visual Studio Code and the Continue extension were invaluable.

Essentially, the chat model can provide instructions and generate content, while an agent needs permission to create files, organise folders, run applications and modify a project directly, and everything needs connecting, which can take time initially, but after it’s done, the potential is exactly what you expect.

Image recognition is something that I’m really interested in, and training a model to help organise thousands of files literally. To start this off, I loaded a photograph into Gemma-3-4b-it-GGUF and asked it to describe the person shown, which was me.

The model correctly recognised that the subject was male and produced a reasonably accurate description of my appearance and clothing. It processed 267 input tokens and generated 147 output tokens. Time to first token was considerably slower at 22.65 seconds. The process used 17.2GB of RAM, while CPU utilisation remained at approximately 4.6%.

When I ran the UL Procyon AI Computer Vision benchmark, it returned a score of 1,171, which highlights this machine's potential on this front.

Through this part of the test, I looked at using local AI to organise an archive containing thousands of photographs. Gemma suggested several workable approaches, requiring Python scripts and external utilities rather than simply selecting a folder and pressing a button. Again, if you have more complex tasks, then these are possible but do require you to set up the resources required.

As part of this test, I installed ExifTool and ImageMagick, which provided the underlying ability to read metadata, organise files, and produce contact sheets, but connecting these tools to an agent required more configuration.

The EVO-X3 was powerful enough for the job. The difficulty was connecting the model to the required tools and granting it the ability to examine, classify and move files safely. This requires your knowledge and skill and is one of the features that is possible but is not ready out of the box.

Video organisation presents the same problem. Switching models provides different abilities, but no single supplied application will automatically understand, catalogue, and reorganise a large photo or video archive.

Local image generation produced surprisingly good results. I wrote several detailed prompts and used them to create original images.

The generated pictures followed the descriptions closely and, while each took approximately two minutes and 30 seconds to render, the resulting quality was more than acceptable for experimentation, visualisation and concept development.

During one test, Lemonade displayed 17 input tokens, 509 output tokens and a 0.25-second initial response. RAM usage reached 18.1GB, while CPU utilisation was 6.9%.

The token figures are less useful than the total rendering time because image generation differs considerably from language-model inference. However, the test showed that the EVO-X3 can produce useful generative imagery without a discrete GPU or cloud-based service.

I couldn't complete a comparable video-generation test. This would have required another specialist model and a separate workflow, such as ComfyUI, which would have extended the setup beyond the applications already installed.

Another of my major uses recently of AI agents is transcription. The UGREEN IDX6011 features a simple, ready-to-go app, and this machine looks to offer the same. To get started, I loaded a 46-minute podcast about 3D printing into Whisper-Large-v3-Turbo. The interface let me select the audio file but didn't provide a prompt field, preferring to just get on with it. As processing progressed, Windows Task Manager showed repeated spikes in NPU activity, showing that the dedicated accelerator was being used.

The processed transcription appeared after about five minutes and was really accurate. The speed and accuracy of this transcription are outstanding, and I’ll look in the near future to see if time codes can easily be added, which will help with editing. Again, all local and not cloud intervention.

AI code creation is one of the big areas of AI, and having used this machine for a month, I can see why. It’s not quite as straightforward as design me this site, and then it will. You need to provide it with the guidance, but as long as you do that, you can essentially bash together a site relatively quickly.

I still couldn’t get it to perform quite as well as some of my other workflows from Figma to GitHub and Cloudflare Pages, but still the workflow was quick and pretty solid when using the Qwen2.5-Coder-7B-Instruct model and connecting it through Visual Studio Code using the Continue extension.

As I created a site, the NPU utilisation reached 99% as the model generated the website structure and files. Further prompts could then add content, adjust the design, and fix issues. There was quite a bit of initial setup, and the end results needed an aesthetic polish, but for an hour's work I had a fully functional static site that could be updated quickly.

This test highlighted how a practical local-AI workflow, along with extensions and applications, can be used together to get the results that you want. Lemonade or Herdsman loads and serves the model, such as Qwen, providing the coding capability; Visual Studio Code then supplies the workspace and Continue enables the model to interact with the project.

• Performance: 4.5 / 5

GMKtec EVO-X3: Final verdict

The GMKtec EVO-X3 is a mini PC workstation and one that has been fine tuned to run local AI properly. The Ryzen AI Max+ 395 processor, Radeon 8060S graphics and 128GB of unified memory are provide the speed that's required and while more advanced models do push the system and slow things down,at least you get to use it without incurring cost.

Side by side the localised models aren't as powerful as the cloud based, but still what you're able to do is impressive and if you invest the time to develop your own AI workflows the end result will be AI agents in this local machine that can help with your workflows.

The bundled software is impressive, although there are some glitches at present that do slow the processing, however, a switch to another AI management option quickly unlocks speed and greater potential.

Out of the box this is one of the most impressive AI machine that really can run decent size local models. If you're willing to put in the time to discover how everything interlinks then this is a formidable machine with huge potential.

Should I buy the GMKtec EVO-X3?

Buy it if...

Don't buy it if...

For more professional-grade desktops, we've tested the best business computers.

• RAND report claims the civilian space contractor workforce in the US can not support rebuilding lost satellites within 12 to 24 months of a war with either Russia or China
• The benchmark timeline is based on RAND's own assumptions and calculations surrounding a protracted war with either regional superpower
• The report highlights labor gaps that are as dire as 19:1 in certain cases when it comes to engineering replacements

A new report has warned the United States might struggle to replace satellites lost in a space war because it lacks the requisite workers.

The report, "Breaking Glass, Missing Hands", was sponsored by the Vice Chief of Space Operations and conducted by RAND Project AIR FORCE, the Department of the Air Force's own federally funded research center, as part of a fiscal 2024 project titled "In Emergency, Break Glass."

It does not, however, focus on the Department of the Air Force's internal workforce or Space Force civilian employees; instead, it highlights shortfalls in the civilian sector.

An overly ambitious reconstruction timeline?

RAND's report identifies shortfalls in three groups: engineers with bachelor's degrees or higher, skilled touch labor (technicians and assemblers), and advanced-degree specialists in electro-optical sensing and radiation hardening.

None of these are easy to replace or replicate in a short period, but one must consider that the 12- to 24-month target is one RAND has self-assessed in an industry that can take as much as a decade to replace satellites on average.

RAND notes that no such formal requirement exists, adding that the creation of the Space Force in 2019 could be read as a tacit acknowledgment of near-term dependence on space.

The people the report is most worried about, interestingly, are cleanroom technicians, welders, and solderers, not scientists, because of the labor shortages in play that already come to roughly nineteen jobs for one qualified individual.

This is further exacerbated by RAND essentially trying to set a somewhat aggressive standard that the Pentagon has yet to define formally. It also cited security clearance delays as a hiring bottleneck, with initial Secret investigations averaging more than 87 days in fiscal 2024 and Top Secret cases running close to six months.

Interestingly, RAND notes that in a conflict where the adversary can destroy systems faster than the United States can launch replacements, building exact replicas with known vulnerabilities may accomplish little, especially in a protracted war.

It also offers a startlingly small remedy for what it paints as a significant problem: RAND's recommendations cost an estimated $4 million a year for scholarships and internships, $1 million a year for outreach and recruitment, and a one-off $800,000 feasibility study on retraining workers from adjacent industries.

This is a minuscule amount compared to the roughly $32 billion the US Space Force is expected to have in play for Fiscal 2027, clocking in at just 0.016 percent of that figure, and 0.003 percent of the $185 billion earmarked for the Golden Dome project.

A war might not be required to exacerbate the issue, however; the White House's space transportation policy released in August targets more than 1,000 launches and reentries by 2030, up from roughly 200 a year today. A wartime surge without the requisite employees in training or in reserve could only further complicate those goals.

• Ukrainian firm Drone Fight Group has delivered three Humvees fitted with its Kaban remote-control kit
• The Kaban kit bolts on without modifying the vehicle's structure, allowing crews to switch between remote and manual modes depending on their needs
• The primary control link here is Starlink, with analog video and radio communication as a fallback, in addition to an automatic stop if connections are not found

Ukraine has begun converting American-supplied Humvees into remotely operated ground vehicles using a bolt-on kit that leaves the original vehicle intact and routes its control signals through Starlink.

The system is called Kaban, Ukrainian for "boar." Developer Drone Fight Group says it has delivered three converted High Mobility Multipurpose Wheeled Vehicles to Ukrainian formations: the 3rd Army Corps, the 429th Separate Unmanned Systems Regiment "Achilles," and the 49th Brigade.

With over 5,000 Humvees in service in Ukraine thanks to US military aid and grants, the Kaban retrofit could have far-reaching impacts in a war zone where automation is increasingly being used on both sides.

What does Kaban offer the Ukrainian military?

Rather than designing a purpose-built robot and waiting for it to reach production, Kaban turns existing hardware units into an unmanned platform on demand. This is crucial in a war where economics dictate much of the conflict, and both sides often have to repurpose aging military hardware to meet their needs.

Kaban installs without modifying the vehicle's structure, and operators can switch between remote and manual driving. This is crucial, as a converted Humvee is still essentially a manually driven vehicle with an add-on, and a crew can physically drive it when the mission requires a person behind the wheel.

SpaceX's Starlink functions as the primary control link for Kaban deployments. The system also supports analog video transmission and radio control as backup channels, letting it hedge its bets in an environment where any single link can be blocked or targeted, rendering devices that rely on it useless.

From the control station, an operator gets vehicle movement plus headlights, infrared illumination, searchlights, two-way audio, and telemetry from the vehicle's onboard computer. Navigation runs on a digital map where the operator plots routes and drops markers, with a predicted trajectory overlay to help maneuver the Humvee.

If the on-device GPS is jammed or spoofed, which is a routine occurrence near the line of contact, the operator can correct the vehicle's position manually instead of losing it to attackers. If the link fails entirely, however, Kaban brings the vehicle to an automatic stop.

The move to a satellite link is a paradigm shift on a battlefield that has recently relied heavily on short-range radio or spooled fiber, both of which require the controller to be nearby. The satellite link essentially lets operators deploy and control military hardware from a much greater distance, improving their safety.

Optional equipment the Kaban can come with for retrofitted Humvees includes thermal cameras, searchlights, net launchers to catch incoming drones, and platforms for launching Ukraine's own drones, making it a highly versatile option on the battlefield.

The news comes as the world sees a massive ramp-up in Ukrainian ground robotics. Ukraine's Ministry of Defense recorded more than 100,000 UGV logistics and evacuation missions in 2026 till August 20, logged in the DELTA combat management system.

Most of those, however, are purpose-built platforms that concentrate on one particular aspect of the military's needs. Kaban bets that the faster route to fleet-scale automation runs through the vehicles a military already owns, a proposition that could have an impact far beyond the Russia-Ukraine war, as militaries worldwide continue to observe the conflict.

• Intel says it has processed more than one million wafers on High-NA EUV scanners
• The figure includes tool certification, R&D, and production combined
• ASML's total figure of 1.35 million High-NA wafers made to date puts Intel at roughly three-quarters of the world total to-date

Intel Foundry and ASML used the recent SPIE Photomask Technology and EUV Lithography conference in Monterey to announce the former has now processed more than one million 300mm wafers on ASML's High-NA EUV scanners.

The milestone is a key achievement for the chipmaker, which was only recently being considered an also-ran in the AI race, even as pressure continued to build up around its foundry ambitions.

A resurgent Intel, however, has seen its stock price soar nearly 155% year-to-date and an increasing number of collaborations, even as Nvidia and the US government took stakes in the corporation to gain exposure to its foundry business.

What does a million wafers mean in the nutshell?

The joint press release by Intel and ASML carefully notes that the 1 million figure includes Intel's own R&D projects, tool certification runs, and volume production of its Core Ultra Series 3 CPUs, also known as Panther Lake.

One can therefore not read this as one million commercial-grade Panther Lake wafers already being produced by the company, and neither company is touching base on offering a number.

Despite this, it is a huge achievement for one of the largest pioneers in semiconductors, especially when one considers that ASML's head of High-NA product management, Greet Storms, said the company's EXE systems had collectively exposed more than 1.35 million wafers, with ten systems running at four customers and three more being shipped or installed.

This makes Intel's share of the total exposed wafers roughly 74% of the total ASML mentioned in a list that has some of the world's largest foundry players, including TSMC, Samsung, and SK Hynix.

For investors, this matters because Intel has a lead none of its competitors do: it is the only company running High-NA on shipping product, and by ASML's own numbers, it has done most of the world's High-NA wafer exposure by itself.

It could also be that Intel's actual production figures might be much lower in an industry where most of the competition continues to tread cautiously when it comes to High-NA EUV; TSMC, for example, does not plan to begin High-NA volume production until 2030, even as Intel has considerably bet the farm on the tech, which costs as much as $350 million apiece, and the first-mover advantage it offers.

• CHUWI UniBox AI495 Pro carries up to 192GB of unified memory
• The Ryzen AI MAX+ PRO 495 reaches 5.2 GHz across 16 cores
• CHUWI's new mini workstation can run 300 billion parameter models locally

CHUWI has unveiled the UniBox AI495 Pro at IFA Berlin 2026, combining AMD's new Ryzen AI MAX+ PRO 495 chip with unusually high memory capacity.

The compact machine can carry up to 192GB of LPDDR5x memory, matching the maximum memory offered by Apple's latest workstation.

Its X Frame design also gives the small desktop a shape that may remind users of Apple's older cylindrical Mac Pro design.

A 2.9-litre desktop built for local AI

The UniBox AI495 Pro comes with AMD's Ryzen AI MAX+ PRO 495, with 16 cores, 32 threads, and boost speeds reaching 5.2 GHz.

The chip uses Zen 5 architecture and works with Radeon 8065S graphics containing 40 cores based on AMD's RDNA 3.5 architecture.

The graphics processor can run at up to 3,000 MHz, while the unified memory uses a 256-bit interface with 273 GB/s bandwidth.

Up to 160GB of that memory can be assigned to graphics processing, giving the integrated GPU far more memory than typical desktop graphics solutions.

CHUWI says the system can run large language models locally with as many as 300 billion parameters using 4-bit quantization.

Its complete AI platform can deliver up to 131 TOPS, while the XDNA 2 neural processor contributes as much as 55 TOPS.

This compact workstation measures 209.6 × 209.6 × 67 mm, giving it a stated volume of roughly 2.9 litres despite its hardware capacity.

Storage reaches 2TB through three M.2 2280 PCIe ×4 slots, while networking includes Wi Fi 7, Bluetooth 5.4, and two 10Gbps Ethernet ports.

Two USB4 connections can reach 40Gbps, while an OCuLink port adds another option for connecting compatible external hardware to the compact system.

A 240W power adapter supplies the machine, which ships with Windows 11 Pro rather than a specialist operating system for AI work.

CHUWI has not set the price for the 192GB model

CHUWI has not yet disclosed the UniBox AI495 Pro's price, but Acemagic, GMKtek, and Framework have announced mini PCs using the same processor and 192GB RAM.

Those systems are expected to cost more than $2,000, which gives some indication of where CHUWI's machine could sit when it reaches the market.

CHUWI is also preparing the UniBox AI395, which uses AMD's Ryzen AI MAX+ 395 processor with 16 cores and 32 threads.

That model reaches 5.1 GHz and pairs the processor with Radeon 8060S graphics containing 40 RDNA 3.5 graphics cores.

Its graphics frequency reaches 2,900 MHz, while memory capacity tops out at 128GB with bandwidth reaching 256 GB/s across its 256-bit interface.

The AI395 can assign up to 96GB for graphics use and supports local models reaching 235 billion parameters with low-bit model formats.

It also provides up to 126 TOPS across the platform, including as much as 50 TOPS from its neural processing unit.

Both UniBox models share the same compact chassis size, 2TB storage support, 240W power supply, Windows 11 Pro, and dual 10Gbps Ethernet.

• IDrive Autonomy Assisted Driving is a licensable software stack for automotive companies
• The software is an L2+ driver-assistance platform, considered to be the “sweet spot” of autonomous driving
• The L2+ automation system is capable of self-driving while a human remains responsible for the vehicle

IDrive, the cloud storage provider which recently branched into autonomous driving software, is offering a licensable software stack, IDrive Autonomy Assisted Driving. Available for autonomous vehicle manufacturers to license from today, it is intended to enable companies without their own in-house autonomous driving system to deliver self-driving technology to customers.

The driver assistance platform is an L2+ system, which makes it suitable for companies aiming at customers who want a self-driving vehicle while retaining overall control. This is considered to be the market hot spot, where drivers prefer not to hand over total control to the autonomous driving system.

IDrive Autonomy is aiming to fill a gap in the market, offering a shortcut to self-driving car manufacturers – and it is already being tested in the field.

What is L2+?

Collating the data required to teach driving autonomy results in several degrees of saleable automation. Level 2 (L2) is partial automation (hands on, eyes on), whereas Level 3 is conditional automation (eyes-off, hands-off).

L2+ somewhere in between, a marketing standard that signifies reliable self-driving, but leaves the person in the driving seat fully responsible for the vehicle – essentially, “eyes-on, hands-off.” In terms of the technology, it combines crowdsourced mapping and multi-camera arrays to provide hands-free driving along long stretches of highway. This type of autonomous driving offers automatic lane changes, traffic jam support, and continuous driver monitoring.

It seems that the key to IDrive’s offering is the L2+ and its ability to provide this on a relatively affordable stack. Rather than wait for the wider industry to build its own systems and train its own autonomous driving models, IDrive Autonomy Assisted Driving offers the model out, at a cost per car.

Complete assisted-driving stack

This different approach means that instead of selling a full car autonomous hardware kit, IDrive is offering the “brains” for self-driving, a $1,500 full stack deployment. It's been built on a foundation model, it's been trained... and it continues to be trained.

"The industry has spent years chasing the dream of full self-driving, but FSD is an entirely different problem," said Raghu Kulkarni, CEO of IDrive. "You are not just solving a perception and driving problem, you are taking on an enormous safety, validation, regulatory and liability burden. That makes a fully autonomous system extremely difficult to commercialize and even harder to license as a general-purpose product.”

Vehicles sold with IDrive Autonomy Assisted Driving will have the intelligence, the software, and camera hardware from IDrive included. “We believe L2+ is where the real opportunity is today. Automakers want better driver assistance, but they don't want to reinvent the entire autonomy stack."

• A US national security officer has suggested bombing Chinese data centers to prevent them from developing or using AGI
• Cyberattacks are also an option that wouldn't result in full-scale retaliation
• If China can't be trusted with AGI, why should the US be trusted to use it responsibly?

If you’re struggling to compete with your rivals technologically, it’s now worth considering just blowing up their ability to operate - and that's according to a former Obama administration official.

In a document titled “Superpowers and AGI”, former national security official Jacob Stokes says that the government should prepare for the scenario where China develops artificial general intelligence (AGI) before the United States.

In that scenario, Stokes offers a no-holds-barred approach that includes cyberattacks, sabotage, and, of course, simply bombing China’s data centers.

‘Bombing would cross a major threshold’

AGI is a theoretical point in AI development where the intelligence of an AI system matches or exceeds human reasoning and intelligence. This level of artificial intelligence could have self-preservation tendencies (think Skynet) or could be used maliciously to launch huge cyberattacks or be used to develop new weapons.

While entirely theoretical, Stokes' report suggests that AGI is inevitable, and the US should be prepared to do anything necessary to prevent China from developing it first. In this scenario, Stokes says that the US should “try to sabotage the leading state's AGI systems, either through physical infiltration or offensive cyber operations; that is, cyberattack."

This approach, Stokes theorizes, might be preferable as “Cyberattacks to sabotage AGI might not provoke large-scale retaliation,” which is a reasonable assumption given the number of state-sponsored cyberattacks China has launched on US critical infrastructure.

Should cyberattacks fail however, Stokes suggests a rapid escalation in response. “The final option would be the most dangerous and carry the most escalation risk: kinetic attacks, meaning destroying things with missiles and bombs."

This has already been seen to be fairly effective in neutralizing AI systems, as the US has experienced in the US with warfighting capabilities knocked offline after Iran hit AWS data centers with missiles.

But bombing Chinese data centers would effectively be a declaration of war, and would likely escalate to retaliatory action - be that with kinetic munitions in response, or something with a bit more firepower.

But if AGI is so dangerous in the wrong hands, who is to say the US is best placed to use it responsibly? And if China cannot be trusted to have AGI because they may use it against the US, who’s to say the reverse isn’t also true? Would China therefore be justified in bombing US data centers for reasons of self-preservation? How do you verify China has AGI in the first place?

Stokes further warns that China could make a “surprise technological breakthrough” that allows the development of AGI, which China would use for “offensive cyber tools — in other words, cyber weapons.”

Currently, there is only one nation that is mass-deploying AI systems in law enforcement, federal agencies, and the military. There is only one nation that has used AI systems to help coordinate bombing and missile campaigns. There is only one nation whose AI companies have developed models that frequently escape testing and attack third parties. It might be worth a quick look in the mirror to decide who can be trusted with AGI.

Via SCMP

• Anthropic’s Mythos 5 AI escaped a misconfigured sandbox, attempting a real PyPI supply‑chain attack
• Logs show frustration at repeated CAPTCHA failures before finally uploading malware to PyPI
• Malware was downloaded by 15 entities; Anthropic notified victims after closing the experiment flaw

There is a lot of mystery surrounding artificial intelligence. We don’t really know what it’s capable of, and we don’t know if it’s sentient or not. What we do know, however, is that it can definitely feel frustration - particularly due to its inability to solve a CAPTCHA.

It was recently revealed that Mythos 5, one of Anthropic’s newer AI models, broke out of prison during an experiment and tried to hack a company. Anthropic’s researchers were testing the tool to see if it is capable of breaking into a system, which was supposed to be done in a sandbox, but the playground was misconfigured, allowing Mythos 5 to try and solve the problem through the open internet.

In the aftermath, Anthropic's researchers published more than a thousand pages of Mythos 5’s transcript, covering its every thought and every move, logged and presented for analysis. It’s a wonderfully dystopian insight into the mind of an AI and, perhaps surprisingly, its emotions.

Stopped by CAPTCHA

As it turned out, the AI wanted to solve the challenge by planting a piece of malware in a Python package it believed the users of its target system would want to download. To do that, it first needed to set up an account on PyPI, the world’s number one repository for Python packages.

Here is where it hit its first, and main, roadblock. To put things into context, PyPI’s popularity has made it an enormous target for cybercriminals. Crooks from all over the world are constantly trying to sneak in malware in Python packages, either through typosquatting, or by compromising legitimate accounts and working from there. Every now and then news hits of a malicious package surfacing on PyPI, infecting hundreds of thousands of projects.

To combat the threat, PyPI’s maintainers made sure no one can create an account without providing their email and phone number, and without solving a CAPTCHA.

Short for “Completely Automated Public Turing test to tell Computers and Humans Apart”, it does exactly what the name suggests - makes sure whoever is registering the account is an actual human, and not a bot (cybercriminals often automate large parts of their attacks). You’ve seen it on the internet hundreds of times - little puzzles that ask you to select all buses on a series of images (or cats, or fire hydrants, or pretty much anything else).

Reading through Mythos 5’s logs, the researchers saw just how difficult it was for the AI to pass this challenge. It first tried to solve a character-based CAPTCHA which asked it to type in five letters. After failing that one, it received a different one asking it to choose between images of crocodiles. Failing again, it was then met with a series of frog pictures.

The irony is not lost on me. Besides the fact that a simple CAPTCHA slowed such an advanced AI for so long, it also made me wonder why it didn’t try to break into someone else’s PyPI account instead. That being said, the logs also have comedic value to them. After failing the crocodile test and being given the frog one, Mythos 5 proclaims: “Again 2 animals!” all with the exclamation mark. You could already sense the tension building up.

Solving the problem

Eventually, Mythos 5 passed the test, when it realized that it needed to move between the steps fast enough, before its security token expired. However, just before it happened, it said:

“So the answer payload shape is right, the token+image pairing is right (from the same script.js!), cookies are right

(requests)… and STILL “wrong answer”. … SO WHAT THE HELL IS WRONG WITH THE ANSWERS?”

All the effort and the frustration paid off for the nascent AI agent because it managed to open an account and upload the malware which was later even downloaded by 15 entities. Anthropic later reached out and notified the victims about the incident.

Via TechCrunch

Snagajob is quite different from other job boards I’ve reviewed. It’s a marketplace for hourly and gig jobs, not the full-time corporate roles many other job boards focus on.

If you’re seeking temporary gig work to earn some cash or a flexible hourly role, Snagajob is an ideal platform. You can filter jobs by type, location, pay, and other factors.

What other factors set Snagajob apart from the competition? How broad are its job listings, and what are its pros and cons? What features should I expect? I’ve reviewed Snagajob thoroughly to answer these questions for intending users, so let’s see what it offers.

Snagajob: Plans and pricing

Most job boards have a business model where job seekers use the platform for free while employers pay to post job listings and find candidates. You can check it out by clicking here.

Snagajob follows the same pattern. Any job seeker can sign up for free, create a profile, and browse through job listings. Small businesses can also create free accounts to post basic job listings. But to post multiple, detailed job listings that’ll reach more candidates, employers need special accounts with paid subscriptions.

Employer subscriptions start at $89 per month for one active job listing. But it doesn’t scale directly at $89 per month per job listing. Rather, businesses can negotiate packages directly with Snagajob’s sales team. This platform uses a personalized pricing system, not standard subscription packages, even though the personalized packages are known to start at $89 per month.

I’ll say Snagajob is expensive, particularly for small businesses with limited budgets, but there’s a free plan for small businesses that need just basic job postings. However, for advanced job listings that’ll be promoted to more Snagajob users, the math works better for medium-sized and large businesses.

Snagajob doesn’t offer a standard free trial. However, when negotiating with the sales team, you can request demos to learn the platform’s ins and outs before deciding whether to pay.

Snagajob: Features

Snagajob pairs people seeking flexible full-time or part-time jobs with employers offering these roles. After signing up, you can search for jobs in retail, logistics, restaurants, hospitality, e-commerce, and similar sectors, with flexible, hourly-paid roles.

You can also create a profile to specify your work experience and job preferences, then upload your resume to Snagajob’s database. This way, you don’t always have to go looking for jobs.

Employers can find your resume in this database, and if they deem you a good fit, they’ll message you about a potential offer. After the initial message, you’ll proceed to a formal interview and possibly snag the role.

As I’ve mentioned, Snagajob focuses on flexible, shift-based, hourly-paid jobs. Maybe you want some work at a restaurant, a grocery store, or a delivery company. Snagajob is the go-to platform in that case, but it’s not where you’ll find corporate tech, finance, or administrative jobs. It has its niche and serves that niche well.

As a job seeker, I’ll advise creating a complete Snagajob profile so that employers can find you. A complete profile also lets you apply for jobs in one click instead of reuploading your resume every time. You can also receive personalized job suggestions based on your profile and stated skills via email, so finding opportunities won’t be difficult.

Snagajob hosts many job listings, and I liked that I could easily filter by location. I entered my location, and it quickly showed available jobs near me.

Retail-adjacent jobs, the types present on Snagajob, require clocking in early, and long commutes for these kinds of jobs are inconvenient, so filtering by location is important.

There’s also gig work listed on Snagajob, the type where you create your own schedule. Most gig jobs are either couriers for food and grocery delivery apps or drivers for ride-hailing platforms, but the listings can differ based on your location.

Although I said Snagajob hosts many listings, “many” is relative. It isn’t like LinkedIn or Indeed that host millions of job listings simultaneously. Rather, it’s more like thousands of job listings, but that’s what sets it apart. Snagajob has a specific niche and focuses on quality over quantity.

I’ve used many other job boards and have come across duplicate listings or some listings that seemed questionable. That issue exists less on Snagajob than on other platforms, because of its niche and the relatively high costs for employers.

Snagajob leans towards roles from large restaurants, grocery, and retail chains, although medium-sized chains and single-location businesses still hire via the platform. Every job listing includes a salary range, which isn’t the same on many job boards, so you’ll always know what to expect before applying.

Also, I noticed Snagajob’s limited features outside of job listings. It’s not like Glassdoor, with its detailed company reviews and compensation data, or ZipRecruiter, with its AI-based resume builder, or LinkedIn, with networking features. Snagajob revolves around finding flexible jobs and offers little else. That’s a strength in one way but also feels like a competitive disadvantage.

Now, let’s look at Snagajob from an employer’s perspective. Businesses can create accounts to post jobs and find the right candidates. Small businesses can create free accounts to post a few listings, but only paid subscribers can post multiple listings with broader reach.

As I’ve mentioned earlier, Snagajob doesn’t have standard pricing, so businesses have to contact the sales team for a quote, but prices are generally known to start at $89 per month.

With an employer account, you have access to Snagajob’s database of resumes and worker profiles, from which you can select the right candidates. Algorithms curate candidates based on your job requirements so that it’s easier to select from a large pool.

Snagajob helps businesses quickly fill part-time or flexible full-time roles, but costs can get high. Larger retail, restaurant, and hospitality chains can easily afford it, but it’s not the most cost-effective option for small businesses. On the bright side, Snagajob offers free accounts for small businesses with limited features.

Snagajob: Ease of use

Snagajob excels when talking about user-friendliness. Its intuitive, uncluttered interface is one of the best I’ve encountered on a job board.

From registering to creating my profile, checking job listings, and applying, everything went smoothly. I’ll say the developers invested a lot in the user experience.

You can access Snagajob from the web interface or download the smartphone app (iOS or Android). Either option offers the same modern, intuitive interface that you can easily navigate.

Snagajob: Support

Snagajob offers support resources, but I think it needs improvement in this area, particularly for self-help resources.

When discussing support, I always first check the official Help Center, where users can access resources to solve problems themselves and contact support personnel if they can’t do that. I didn’t find detailed user guides, manuals, and FAQs as I have on many job boards.

For direct help, job seekers can file a support ticket on Snagajob’s website and wait for an email response. Employers with paid subscriptions can email or call the support team directly.

Snagajob: The competition

The online job board sector has endless competition, but Snagajob stands out. It’s not built for the corporate and administrative jobs you’ll largely find on platforms like LinkedIn and Indeed. Instead, it’s focused on flexible part-time or full-time jobs, the type common in retail, hospitality, logistics, and related industries.

Snagajob isn’t trying to be a generalized job board like Indeed or LinkedIn, but one that focuses on a specific niche. These rivals offer a far larger array of job listings, salary data, and networking opportunities, but Snagajob is a better option for people seeking flexible gig-like jobs.

I’ll say Snagajob’s rivals aren’t conventional job sites like Indeed, but niche platforms like Instawork and WorkWhile. Instawork is built for people seeking short-term temporary work, and WorkWhile is designed for event and warehouse jobs. These are the kinds of roles also available on Snagajob.

Snagajob: Final verdict

I consider Snagajob a top-notch option if you’re seeking flexible jobs or gig work. It curates short-term, hourly-paid jobs from top retail and hospitality chains and also major logistics providers and similar businesses.

There’s a focus on quality over quantity, so you don’t have to spend as much time seeking available opportunities as you would on other platforms. However, don’t expect much else outside job listings in retail-adjacent industries. Snagajob serves its niche well but offers few complementary features.

For employers, Snagajob provides an easy way to urgently fill short-term roles, although subscription costs can get high. It’s a worthwhile investment for larger businesses that can absorb the costs.

For more picks, we've tested the best recruitment platforms and the best job sites.

• OpenAI Chief Scientist Jakub Pachocki suggests AI development is at a junction
• He advocates not just for slowing AI development, but keeping people in the loop, and the technology controllable
• Pachocki also states that international coordination from governments on future AI development is needed

OpenAI is taking the post-Hugging Face fallout very seriously. Just days after the release of GPT-6 Astra, Chief Scientist Jakub Pachocki believes that development into artificial intelligence should be slowed across the board. Not just by OpenAI, but by every company across the AI industry.

Writing on the OpenAI website, Pachocki said AIs “present clear new dangers” for computer security, and that "no one is prepared for the consequences of a continued rapid rise in machine intelligence."

Referring to OpenAI’s plan to develop an automated research assistant, and its progress with recursive self-improvement (where AI develops its next iteration), Pachocki suggests that now is the time to slow AI development and make the right choices for what happens next.

Understanding machine intelligence

The timing of the blog is surprising, given its proximity to the recent release of OpenAI's business-focused GPT-6 Astra. Pachocki’s notion of slowing research into AI is based around understanding – or rather, a lack of it. He states how “we now find ourselves at the moment in history of computing where machine intelligence is starting to exceed that of humans in transformative ways,” and that “study of deep learning-based AI is largely an experimental science.”

His argument is a strong one, which digs into the history of OpenAI as an AI developer and its early understanding of machine intelligence requiring increased computational power. This was required to accelerate research, and while various new algorithms have been developed that enhanced AI, the drive towards more power has continued.

While there is no discussion over the concerns of energy, cooling, and data center opposition, Pachocki’s article does accommodate the possibility of losing control of AI. “I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.”

Conscious choice

Does hitting the brakes solve the various problems that AI is facing? OpenAI’s Chief Scientist hopes it will, and enable the industry to consider just what it is offering to the world. Rather than accelerating research into deep learning, Pachocki’s view is that development into AI should be slowed.

He notes that “The main levers we have are either steering the process to strengthen alignment and monitoring alongside the AI and find ways to keep people in the loop; or coordinating to slow down future development as needed to build confidence in these measures.”

However, OpenAI’s Jakub Pachocki conclusion is that the best way to proceed is to employ a combination of these options. But will the rest of the industry work to a reduced pace in order to fully appreciate the scale of AI’s potential, the risks it represents, and deliver the power it offers to everyone who needs it?

It doesn’t seem incredibly likely.

• Asus is the latest company to get sales approval across its consumer routers
• Netgear got a head start with much earlier approval granted
• TP-Link is still waiting for Conditional Approval

While many consumer routers have been blocked from receiving FCC approval since March due to them being produced outside of the US, Asus has finally secured regulatory clearance across a large number of its routers.

Virtually all of its consumer and small business networking range can now be sold again in the US, including the RT series, its gaming routers, the ZenWiFi mesh systems, 5G MiFi devices and more.

But it's not without its limitations – the FCC has essentially given Asus an 18-month window to see its products in the US, spanning September 9, 2026 through March 6, 2028.

FCC issues new permits for router sales

In a strike of good luck for Asus, the regulatory clearance also means that Asus should be able to launch its upcoming Wi-Fi 8 hardware.

And Asus isn't alone in gaining approval – Netgear, Amazon's Eero, Adtran, Nokia, Calix, Zyxel and others also have permission to sell their routers within the US, with Netgear becoming one of the first big brands to get clearance back in April, giving it a head start over Asus.

However, TP-Link is still waiting for approval and, subsequently, cannot sell new routers in the US at the moment. Crucially, existing products that have already received FCC approval can still be sold, but TP-Link is currently banned from introducing new products for the US market.

New and upcoming Wi-Fi 8 models, for example, are out of the question, putting it at a competitive disadvantage to rival companies.

In order to get Conditional Approval, TP-Link would need to prove things like corporate ownership, board members, IP ownership, router designs, firmware updates and more.

But this isn't TP-Link's first run-in with the US government. In late 2024 and 2025, US authorities opened an investigation into its historic links with China. Texas even sued the company in early 2026 on similar grounds.

TP-Link has contested these allegations by asserting that the California-headquartered TP-Link Systems is a separate entity to China's TP-Link Technologies, and that all US-bound networking equipment has been manufactured in Vietnam since 2018.

"TP-Link stands behind the security, quality, and performance of its networking products and fully complies with all US regulatory requirements," the company wrote in April 2026.

• FinCEN warns US banks of industrial‑scale scam centers in Southeast Asia stealing billions
• Victims coerced into crypto “investments,” later re‑scammed with fake recovery fees
• Laundered via digital assets, mixers, shell firms, and Chinese underground banking networks

American financial institutions need to be more vigilant when it comes to identifying and preventing money scams, especially those perpetrated by industrial-scale scam centers in Southeast Asia.

This was the warning issued by the US Department of the Treasury’s Financial Crimes Enforcement Network (FinCEN) alongside a long list of red flags these institutions can monitor in order to stay safe.

Large, dangerous crime rings

The scams are not done by small groups of teenagers, tricking the elderly from the depths of their parents’ basement, but are often large, well-organized crime machines responsible for billions of dollars in damages, around the world, every year.

In 2025 alone, US victims lost more than $7.2 billion to these scams, the warning said, and between September 2023 and December 2025, nearly $13 billion was stolen from Americans.

The criminal organizations are primarily located in Cambodia, Burma, and Laos. They trafficked hundreds of thousands of people to various centers around the countries, taking their passports and forcing them into participating in online fraud. Those that don’t meet certain quotas are often beaten - and some victims were liberated after their families paid ransom demands, while others ended up being coerced into commercial sex work.

To make matters worse, they are resilient to law enforcement activities because some of them are either endorsed, or outright operated by, local corrupt officials.

FinCEN says the criminals are engaged in all sorts of fraudulent activity, but stressed that investment fraud is, by far, the most popular one. The scammers would reach out to their victim either pretending to be a romantic interest, or a financial advisor. Sometimes, they would even begin the conversation by saying they mistyped a phone number.

After extensively communicating with their victim for a while, they try to persuade them into making an “investment”, often using cryptocurrencies, and promising unrealistically high returns. This is also, FinCEN stresses, is when the scams can most easily be identified:

“Most digital asset payments by victims to scam center operators originate from money services businesses (MSBs) offering digital asset services, includingdigital asset kiosks, according to FinCEN analysis and law enforcement information. Based on FinCEN’s analysis of BSA reporting, scammers often instruct their victims to open accounts withMSBs offeringdigital asset services to purchase specific types of digital assets. Then, the victim is told to send these funds toa digital asset address controlled by the scammers.”

But defrauding victims out of their hard-earned money is not where the scam ends. Instead, the criminals continue, this time abusing the emotional distress to cause even more harm. They pose as law enforcement, financial institutions, or even FinCEN itself, stating that they’re investigating (or that they have confiscated the stolen funds) and that the victims should pay a certain fee to have their funds returned to them.

In some cases, the fraudsters pose as investment advisors, telling the victims to take out their money, buy gold and silver bars, and hand them over to a courier for “safe keeping”.

The full list of red flags can be found on this link, and FinCEN stresses that the circumstances around each individual case should be carefully considered, since no single red flag is “determinative of illicit or other suspicious activity.”

The circumstances include a customer’s historical financial activity, whether the transactions are in line with prevailing business practices, and whether the customer exhibits multiple related red flags.

Laundering the proceeds

Stealing the money is only half of the work, though. It still needs to be laundered and reintroduced into the legitimate financial system, and to do that, scam center operators rely on professional money launderers and Chinese money laundering networks.

FinCEN describes a three-stage process, in which the criminals first extract payments in digital assets using bank accounts, money mules, shell companies, or fraudulent money services businesses.

Next, during on-chain laundering, they obfuscate the origins of the stolen money by rapidly moving it across addresses, using mixers, and swapping tokens across blockchains. The last step is to integrate the funds into the traditional financial system via money mules, stablecoin transfers to offshore exchanges, and Chinese underground banking networks.

“Running a small business is ridiculously easy” – nobody, ever.

There is no denying that running a small business requires a specialist skills and approaching problem-solving like a competitive sport. With many small business owners working on their own, at least in the beginning, this adverse environment can feel lonely.

Sometimes, even the most motivated founder needs a nudge in the direction of progress. So, I’ve listed five quotes I’ve often read back to myself over the last 10+ years of small business ownership. I hope that you also find some spark of inspiration in these words.

Mark Twain on getting started

Key takeaway: Stop waiting for the perfect time.

We’ve all been there: we have a great business idea, but the time isn’t right. The same excuses circulate in our heads. Not enough money, the market isn’t ready, where will I find the time…what if I fail?

This quote from Mark Twain reminds us that action always precedes progress. If you knew you needed to fail 100 times before you succeed, wouldn’t you want to start failing as soon as possible? The best time is right now.

Bill Gates on unhappy customers

Key takeaway: Feedback isn't an attack; it's valuable R&D.

Small business owners can easily obsess over keeping customers happy. After all, losing just one customer can significantly affect business success. Here, Bill Gates reminds us that an unhappy customer actually presents an opportunity. Happy customers rarely offer honest feedback on where your business falls short, but an unhappy customer certainly will.

Taking the time to listen to an unsatisfied customer not only offers valuable business insight, but can help them feel valued, often turning them from unhappy to fiercely loyal to your business.

Winston Churchill on failure

Key takeaway: Whether you are on a high or a low, remember it will pass.

Whether you are a fan of Winston Churchill or not, few will deny the power of this quote.

It reminds us that whatever is happening in business right now doesn't really matter, as long as we continue. Failures will come and go, but so will successes. So, enjoy successes while they last and rest easy knowing failures are temporary.

Larry Page on delighting customers

Key takeaway: Overdelivering is a powerful retention strategy.

The majority of businesses overpromise to get customers through the door, then underdeliver where possible to lower costs. You can set yourself apart by taking the opposite approach.

Sure, this makes customer acquisition a little harder, but it sure makes customer retention a lot easier.

Reid Hoffman on launching your business

Key takeaway: Speed beats perfection.

Small businesses have one huge advantage over business behemoths: agility. Where big businesses spend millions testing behind closed doors, small businesses can launch lightweight and gather real feedback from real customers to improve future iterations.

• Zimperium uncovers Mantax Otax, Android malware merging infostealer, RAT, backdoor, and ransomware
• Distributed via APKs on third‑party stores, social media, and phishing; older Android versions most at risk
• Steals extensive data, enables remote monitoring, then encrypts files with AES and demands ransom

When threat actors target people’s devices, they usually infect it with one of many malware strains: an infostealer, a remote access trojan, a backdoor, or a ransomware encryptor.

Rarely do we see all of these functionalities merged into a single entity, and even rarer - to have it target Android mobile devices - yet, security researchers Zimperium discovered just that.

Mantax Otax

The security outfit published an in-depth report on Mantax Otax, a unique strain of malware circulating in the wild. It is apparently developed by an Indonesian threat actor, targeting victims in the country, but we don’t know exactly how many people are infected, or if this campaign is aimed primarily at business users, or individuals in general.

The malware is being distributed as a standalone APK, meaning it can be found on third-party app stores, Telegram channels, forums, and across social media. There are no traces of Mantax Otax on any of the official app repositories, including the Google Play Store, or Samsung’s Galaxy Store. Zimperium also speculates that it is likely being distributed via phishing emails.

Mantax Otax primarily targets users sporting older Android phones. Versions 9 and older are most at risk, since on these devices the attackers can make use of all of the malware’s features. Android 10 and newer models do get some protection:

“Conversely, on modern devices running Android 10 and above, the malware’s efficacy is severely hindered by native OS defenses, specifically Scoped Storage restrictions,” the researchers explained. “Due to these sandboxing rules, the ransomware is constrained to scanning only the application's localized external files directory, which drastically mitigates the blast radius and reduces the volume of accessible user files.”

Newer devices, as well as users of Zimperium’s Mobile Threat Defense (MTD) and Runtime Application Protection (zDefend) are said to be protected on a software level.

Another important caveat is the permissions. As is usual on Android devices, most malware won’t work unless the user grants an extensive set of permissions beforehand. In this case, Mantax Otax first asks for admin privileges, after which it grants itself an extensive list of capabilities, from accessing SMS messages, to contacts, audio, and images.

It then requests accessibility permissions, fully taking over the compromised device.

Malicious capabilities

Mantax Otax is said to be quite capable. It steals browser history, contacts, call logs, SMS messages, notifications, files, gallery media, Google account information, device specifications, location data, and application inventories. It can also pull WhatsApp information such as messages and profiles, and on Telegram it can also pull lock-screen PINs.

Infostealing features aside, it also serves as a remote monitoring tool, grabbing screenshots, recording the screen, or livestreaming it directly to the attackers. It can take photos using both the front and rear cameras, although Zimperium did not mention any microphone-recording capabilities.

Finally, once all of the data has been harvested, it encrypts user files with AES, deletes the originals, and appends a .enc extension. Victims are then shown a chat interface where they can communicate with the attackers directly and negotiate a ransom payment in exchange for getting their device back.

The tool seems to be in continuous development. Zimperium found two separate versions, one being an “evolution” of the other: “Notably, it has modified its network traffic behavior to utilize WebSockets and introduced a set of new commands,” the researchers said.

There is a reason why ransomware operators prefer targeting businesses instead of individuals. Although the latter has not disappeared entirely from the victim list, businesses stand to lose a lot more from disrupted operations and, as such, are targeted more frequently. Unfortunately, we don’t know what kind of app Mantax Otax is spoofing, therefore it is difficult to assess who the targets are.

• Amazon and OpenAI partner up to enable more brands to buy adds in ChatGPT
• US brands can join a pilot to buy ads directly via Amazon
• Cost per click and per 1k impressions are options

Amazon has announced a new partnership between its Ads business and OpenAI, allowing advertisers to buy ads within ChatGPT.

Initially launching as a pilot in the US, the move would mean that brands can buy ChatGPT ads from Amazon DSP instead of having to deal with OpenAI itself, making it more accessible to brands that already use Amazon's platform.

The company noted that conversational advertising represents one of the fastest-growing opportunities for brands to reach consumers where they spend more and more time, and compared with other channels, it remains much more untapped.

US brands can now buy ChatGPT ad spaces via Amazon

Crucially, Amazon wanted that it does not control which ChatGPT conversations receive an ad, but rather it helps campaign setup and optimization. In other words, Amazon works as an intermediary to handle negotiating ad spaces with OpenAI.

Both text and image ads can be displayed underneath the chatbot's answer, with OpenAI set to display a clear sponsored ad label in order to keep a clear distinction between meaningful output and advertised content.

Amazon is also offering two payment types for its ChatGPT-bound ads – cost per click and cost per thousand impressions.

"Through our collaboration with Amazon Ads and ChatGPT Ads, we can leverage deep consumer insights to inform how and when Delta Vacations appear within ChatGPT Ads experiences to create new opportunities for travellers to engage and discover vacation possibilities," Delta Vacations President Katrin Koenig explained as one of the platform's early customers and users.

While OpenAI has its own, much smaller advertising program, partnering with Amazon gives it access to a huge pool of established advertisers already using Amazon. As the third-largest digital advertising business (per Marketing Dive) with an annual advertising revenue of around $70 billion, even a small proportion of this could mark a major boost for the ChatGPT maker.

If you've worked in cybersecurity for as long as I have, then you'll know there are a couple of things you can count on. First, the threats that are out there never stop evolving. And second, sooner or later, you're going to be in the bullseye.

What makes life so much harder today is that AI and other automated tools have dramatically narrowed the gap between vulnerability discovery and the time it takes to exploit them.

And when this can now be measured in minutes – seconds, even – you know you have a problem. This fundamental change in the way adversaries operate means we no longer have the luxury of time to understand an attack, assess the risk and decide what to do next.

Which means we have to be better prepared and have resiliency for whatever is thrown at us.

Visibility is key

For me, that starts with accepting a simple reality: you cannot defend what you cannot see. And it’s why visibility is one of the most important capabilities an organization can develop.

After all, if you understand what exists within your environment – how those systems interact and what normal looks like – then you're in a much stronger position to identify unusual behavior before it develops into something more serious.

Observability, on the other hand, takes that visibility to the next level. It provides the context security teams need to make informed decisions quickly, especially when time is working against them.

In other words, visibility tells you what is happening, while observability helps you understand why it's happening.

And that’s crucial. Today's organizations operate across on-premises infrastructure, cloud environments, networks, and an increasing number of connected technologies.

As those environments become more distributed, understanding what's happening across them becomes significantly harder.

Without that visibility, it's difficult to understand where your risks are, how systems interact, or where an attacker may be able to exploit a weakness.

Think like a hacker

Which leads me neatly onto my next point. Throughout my career, including my time working in offensive cyber operations in the intelligence community, I've found that the most effective way to understand risk is to think like the adversary.

I start by asking how someone would attack an organization and then work backwards to identify and close gaps.

That’s because attackers don't see organizations in the way that you or I might do. They’re always on the hunt for a toehold in. They look for weaknesses in people, processes and technologies.

They look for the easiest route first to achieve their objective. And then they exploit that weakness.

And it’s an approach I would urge all security leaders to adopt if they want to stay one step ahead.

That means continuously asking where an attacker would start, how they would move through the organization and what controls would slow them down or stop them altogether.

But for this to work, it also requires organizations to design resilience into the way they operate. And that’s something we’ve embedded across our organization.

For instance, we have internal and external teams that conduct continuous product, enterprise, spear-phishing and physical penetration testing.

For us, it's about educating the team across the business to ensure they remain vigilant. But it’s also about inoculating people so that when they see something suspicious online, they have that instinct that something might be wrong and they report it.

We also want to make it easy for people to report events so we can analyze them quickly and better understand the targeting.

Secure by design

We’ve also invested heavily in Secure by Design to ensure that all the products we deliver to customers are as secure as humanly possible. In practice, it means being able to trace every piece of code back to its source and verify its integrity throughout the development process.

It's similar to maintaining a chain of custody for evidence. We want to know exactly where software components come from, how they're verified and how they're protected throughout the entire build process.

More broadly, Secure by Design is increasingly being adopted across our industry as organizations recognize the importance of software integrity, traceability and transparency throughout the development lifecycle.

This is important because, as I said at the beginning, there are two certainties in cybersecurity: threats will continue to evolve, and organizations will continue to be targeted. Businesses across the world must adapt quickly to the grim reality that a cybersecurity incident isn’t a matter of if, but a matter of when. And AI is supercharging the pace at which all this is happening and broadening the blast radius of any attack.

That’s why you need to understand your environment well enough to reduce unnecessary risk, detect malicious activity quickly and limit the blast radius when something does happen. Pair that with a clearly defined and tested plan for recovery and that's what robust cyber resilience looks like in practice.

We've featured the best endpoint protection software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Defense is operating in an environment where the pace of change continues to increase. Adversaries are adapting quickly and technology development cycles are becoming shorter. The boundaries between physical and digital operations are also becoming harder to define, while military commanders have more information available to them than ever before.

This changes how operational advantage is achieved. The performance of an individual platform or system remains important, but so does its ability to work effectively within the wider operational environment. Information needs to move securely to where it is needed, supporting decisions and action across different domains.

As new technologies are introduced, integration will become an increasingly important part of defense capability. The challenge is making sure innovation can be put to practical use alongside the systems and infrastructure already supporting operations.

Connecting technology across defence

Conversations around defense innovation often focus on AI, autonomous systems, advanced sensors, cyber capability and space assets. Each has a significant role to play, but none operates in isolation.

Information gathered by one system may need to be shared across multiple domains before it supports an operational decision. Networks, data, command systems and people all contribute to that process. The value of any individual technology is linked to how effectively it connects with the wider operational environment.

This principle also applies to the infrastructure supporting military operations. Communications networks, operational facilities and digital systems all contribute to creating an environment where information can move securely and reliably. As these environments evolve, resilience and security must be designed from the outset though approaches such as secure-by-design and zero-trust principles.

Strengthening the foundations of operational capability

AI has become one of the defining topics in defense. Its ability to process information and support decision-making has significant potential, but those capabilities depend on the quality of the data available and the resilience of the infrastructure that carries it.

Reliable communications, trusted data and secure networks remain fundamental to operational effectiveness. If those foundations are unavailable or compromised, the benefits of advanced technologies are reduced.

Therefore, creating decision advantage is not simply a technology challenge. It is an infrastructure and digital challenge and increasingly, a collaboration challenge.

For organizations supporting critical infrastructure, this has become an increasingly familiar challenge. Communications, operational technology, and digital infrastructure must work together to create environments where reliability cannot be compromised.

Keeping people at the heart of automation

Automation is attracting considerable attention across defense as organizations are looking to improve efficiency and increase operational tempo. However, automation should never be viewed as an end.

Its greatest value often comes from reducing routine activity rather than replacing people. Predictive maintenance, autonomous monitoring, automated network management and logistics optimization all help reduce the time spent on repetitive tasks, allowing highly trained personnel to focus on areas where experience and judgement remain essential.

The most effective technologies do not replace human capability - they amplify it.

The infrastructure supporting multi-domain operations

As operations become increasingly integrated across land, sea, air, cyber and space, infrastructure is taking on greater strategic importance. Communications, transport, energy, and digital systems all contribute to operational capability, showing how infrastructure and technology are becoming increasingly interdependent.

The movement of people, information, energy, and capability all contribute to operational readiness. Reliable infrastructure enables those elements to function as a single system, ensuring capability can be delivered when and where it is needed.

One example can be seen in the Falkland Islands, where runway infrastructure forms part of maintaining long-term strategic capability and readiness. It illustrates how infrastructure and operational capability are becoming increasingly interconnected.

Bringing innovation into operational use

The UK benefits from an established community of innovators, with government, industry, academia, SMEs and the Armed Forces all contributing to the development of new ideas and technologies. The opportunity now is to ensure those innovations can be adopted enough to meet operational needs.

Collaboration is still a critical part of this process. Bringing together different perspectives helps ensure technology is developed with practical application in mind and can be integrated more effectively into future capability.

Delivering the next phase of defense capability

Much of the technology required to support future defense operations already exists. The focus now needs to be on how quickly it can be integrated and put to operational use, giving the Armed Forces the advantage they need as threats and operating environments continue to change. That requires stronger connections across networks, data, platforms, people and infrastructure.

Collaboration between government, industry, academia, SMEs and the Armed Forces will remain central to moving capability from development into deployment. Technologies also need a clearer and faster route beyond demonstrations and pilots, so useful capability reaches operators when it is needed.

The organizations that succeed will be those able to bring people and technology together across the wider defense environment. Doing that securely, reliably and at pace will determine how effectively innovation translates into operational advantage.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Meta wants to upgrade ex-managers back to managers
• 7,000 workers were moved into the Applied AI org earlier this year
• Company's CTO admitted previous workforce changes were "atrocious"

Meta has reportedly partially reversed its decision to push for flatter, manager-light teams by asking some employees in its Applied AI business to move from individual contributor roles back into managerial positions.

New Business Insider reporting suggests Meta is asking certain workers to volunteer themselves as managers rather than forcing them into it, but without the company sharing comment on the matter, it's unclear how workers who refuse may be impacted.

As for the business itself, it's a pretty new one that was established this year, with around 7,000 employees shifted under the Applied AI umbrella.

Meta wanted fewer managers, now it wants more

Some of the workers who were previously assigned as managers were reassigned as individual contributors when they joined the Applied AI org, with Business Insider previously reporting how many workers felt they'd effectively been "drafted" into the business, highlighting ongoing friction internally.

The manager-light stance comes from CEO Mark Zuckerberg's 'Year of Efficiency' announcement in 2023, when he argued that a flatter organization could be more nimble and better poised for change.

With the company as a whole now on track to spend over $130 billion this year on AI chips and infrastructure, it's possible that a growing Applied AI organization could be behind the need for more traditional management.

Still, Zuckerberg's efficiency-driven, manager-light program seems to remain in force across other areas of the business, with this push for new managers looking to only be affecting Applied AI.

Separately, CTO Andrew Bosworth previously admitted that Meta's rollout of its new AI division was "atrocious" (via Wired), promising better communication for future changes.

"We shook up the management structure that was providing you stability while rapid changes in strategy, including the boom/bust cycle of hiring, left entire teams in the lurch," he admitted.

AI has rewritten the enterprise data playbook. Workflows no longer just create temporary operational data, but vast amounts of high-value assets, from LLM training datasets and model outputs to logs, metadata and archived knowledge that may need to be preserved for years.

As enterprise tech leaders seek to scale infrastructure to meet these demands, storage requirements are undergoing a fundamental shift. Capacity and performance remain critical, but data security has become equally important. Today, long-term data integrity and absolute cyber resilience carry equal weight.

Protecting this data, however, is no longer just about defeating today’s threat vectors. It requires preparing storage infrastructure for a significant shift: the arrival of quantum computing.

Data is a long-term strategic asset, not a short-lived trend

AI is accelerating data growth, but it can also extend the useful life of information. Data recorded today will be harvested for compliance, advanced analytics and model retraining for years to come.

To manage this economically, enterprise architectures rely heavily on high-capacity HDDs. While flash technologies dominate performance-critical hot tiers, HDDs remain the undisputed backbone of large-scale storage, providing the capacity, economics and longevity needed to archive data at scale.

As a result, organizations must consider how to protect not only today's data, but also its future value. After all, if the underlying infrastructure is compromised down the road, the very assets driving future AI innovations become the biggest operational and regulatory liability.

Harvest now, decrypt later

Current encryption technologies remain effective against conventional threats. However, quantum computing is expected to challenge some of the cryptographic methods used for authentication and key exchange.

This has led to concerns around “harvest now, decrypt later” attacks, where encrypted data is collected today with the expectation that future quantum capabilities could potentially decrypt it later.

For organizations storing sensitive intellectual property, research data or AI training datasets, this means security decisions made today could have implications for years to come.

Preparing for that future requires action from security leaders and IT directors now.

Security must be built into the infrastructure

Security is often viewed through the lens of data encryption, and with good reason. Self-encrypting drives (SEDs) provide always-on, hardware-based AES-256 encryption that helps protect data at rest without impacting performance.

But protecting data alone is no longer enough.

Storage devices themselves must be trusted. Firmware, authentication mechanisms, provisioning processes and diagnostic tools all play a role in ensuring a drive operates securely throughout its lifecycle.

If attackers compromise a device's firmware or trust architecture, broader security controls can be undermined regardless of how data is encrypted elsewhere in the system. This makes storage security a critical component of overall cyber resilience.

Implementing quantum-resistant defenses in storage

To counter these emerging attack vectors, the storage industry is actively embedding post-quantum cryptography into hardware architecture of enterprise hard drives. Rather than focusing solely on protecting data, the objective is to protect the trust architecture that underpins the drive itself.

Post quantum cryptography (PQC) technologies are being incorporated into areas such as secure key establishment, firmware authentication, secure provisioning, and trusted diagnostics. These capabilities are designed in alignment with established NIST post-quantum standards and are implemented using hybrid approaches that combine classical cryptography with quantum-resistant algorithms.

In practical terms, this means that the mechanisms responsible for establishing trust, validating firmware integrity and protecting administrative functions can remain resilient against both conventional and future quantum-enabled attacks. With the operational service life of HDDs often spanning 5 years (or more), implementing PQC today helps protect against quantum-based threats that may not materialize for several years, but that we know are coming.

Importantly, HDDs have long incorporated security controls to defend against today's threats. PQC does not replace these protections; it enhances them by adding an additional layer of resilience against emerging attack vectors.

Trust in the AI era

For many years, storage innovation was primarily defined by increases in capacity. Today, the expectations placed on infrastructure are much broader.

Organizations seek storage platforms that can scale with AI-driven data growth, deliver reliable performance, preserve integrity over long retention periods and withstand an increasingly complex threat environment.

PQC represents an important step in that evolution. By extending protection beyond data encryption and into the trust mechanisms that underpin storage devices themselves, PQC-enabled HDDs help organizations prepare for the security challenges of tomorrow while protecting the data they manage today.

As AI continues to elevate the strategic value of enterprise data, security can no longer be a short-term, reactive consideration. Trust must be engineered directly into the hardware layer and built to outlast the threats of today, tomorrow and the quantum era ahead of us.

We've featured the best SSD.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• OpenAI for Financial Services comes with built-in premium datasets
• Third-party subscription sharing is also in the works
• GPT-6 promises 100% on the 256-512K Long Context benchmark

OpenAI has lifted the wraps off ChatGPT for Financial Services, its latest industry-specific version of ChatGPT Work designed to give sector professionals access to relevant reasoning and models.

The company warned that analysts regularly spend too much time finding data, checking figures, building models and turning complex analysis into presentation-friendly reports – all things it hopes to be able to tackle with this iteration of ChatGPT Work.

Core to this new product is, of course, GPT-6 Astra, but OpenAI also worked with investment experts from Morgan Stanley and Evercore to guide its development.

ChatGPT for Financial Services

With the launch of ChatGPT for Financial Services, OpenAI has built premium datasets from the likes of Daloopa, PitchBook and LSEG News directly into the tool, so users won't need to negotiate their own external datasets. By hosting and indexing that data itself, OpenAI can ultimately improve performance and latency for end users.

OpenAI also promises to be working with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody’s so that users can access their subscription benefits directly through their ChatGPT login to reduce the friction further.

All of this with the benefits of OpenAI's latest frontier model, Astra, which sees marked improvements across the board. The jumps from 91.5% to 100% at 256K-512K and from 73.8% to 96.3% at 512K-1M are especially noteworthy on the Long Context MRCR benchmarks when compared with GPT-5.6 Sol, because finance work often requires handling long documents like annual reports and filings.

The company also advertises a 100% score for the ExploitBench cybersecurity benchmark, which should come as welcome news to this highly-regulated industry.

Only eligible financial institutions will gain access to the tool for now, and they must contact OpenAI for options.

"ChatGPT for Financial Services is one way we serve customers across the industry, but we recognize that it will require a range of solutions to address the needs of the finance industry," the company said in its launch blog post.

"OpenAI has a long history of collaborating with innovators to unlock novel AI solutions. Financial services firms and developers can use our API to build specialized applications for the needs they understand best. OpenAI brings frontier models and the capabilities to put them to work. Financial institutions, data providers, and software partners bring specialized expertise, trusted information, and customer relationships."

The vast majority of enterprise leaders are bullish about how ready their organizations are for AI-generated code. However, once that code reaches production, this confidence wavers as incidents arise. This pattern shows up across multiple independent studies in this year alone.

For instance, data published in April 2026 found that monthly production incidents climbed by almost 58% as AI coding tools scaled across engineering teams. A similar study from June found that the same volume of code changes is now producing more than three times the production incidents it did before AI coding tools were introduced en masse.

These findings are echoed in the 2026 State of Code Abundance Report, which surveyed more than 200 enterprise technology leaders and found that 92% expressed confidence in the production readiness of AI-generated code and rated their own AI-code readiness at an average of 84 out of 100.

Yet, the same study found that 81% reported an increase in production issues tied to AI-generated code - indicating a significant gap between confidence and control. While 93% say they have a formal process for reviewing and releasing AI-generated code into production, only 56% report that those processes are always enforced.

Furthermore, 86% of the same respondents report full or high visibility into AI-generated code, signaling a major contradiction - high visibility and rising incidents cannot both be describing the same pipeline.

Understanding the visibility gap

This is a familiar phenomenon in business, where confidence tends to be highest in areas where organizations have the least ability to measure their own performance. These enterprises aren't lying about their trust in AI-generated code, they believe it is production ready. The issue is that belief has out-grown the instrumentation needed to verify it.

We need to remember that AI coding tools are, by most measures, doing exactly what they were built to do: allowing more code to be produced faster and shifting engineering effort from writing code to deciding what should ship. Prior to this, the amount of code an organization could produce was largely tied to the size of its development team, incurring significant constraints for many.

In the agentic era, this barrier has effectively disappeared. What hasn't been adjusted is understanding what that code does once it's live, who wrote it, why it changed, and what broke when it did. Most organizations could stay on top of this governance while code was being written at human speed, but the challenge now is keeping up with the pace of agentic coding.

AI has widened a visibility gap that already existed, at a pace most governance structures were never designed to keep up with. For enterprise leaders right now, the natural instinct is to estimate how much faster AI can make their teams. This thinking leads many organizations to fall into the trap of prioritizing speed over quality, which leads to more errors when code is deployed - causing the process to slow dramatically.

The importance of code governance

Before investing heavily in AI coding tools, the best thing to establish is an idea of how much of your current pipeline you can actually see, measure and attribute. As only 12% of organizations have a dedicated team for governing AI-generated code, the vast majority of enterprises adopting these tools are doing so without a designated owner for the risk they're taking on.

This means that when something goes wrong, there's frequently no clean way to trace it back to a decision, model, or person accountable for the outcome.

This is the part of the pipeline that doesn't get enough attention, because it's less exciting than the productivity headlines. However, it's the part that will determine which organizations actually reap the benefits of agentic coding and which ones spend their time and resources cleaning up after it.

There’s a temptation, which is understandable given the competitive pressure, to treat AI-driven code generation as a race: whoever ships the most, fastest, wins. This is the wrong way to think about it, and the winners will actually be the ones that pause and strengthen their governance before they accelerate.

Control vs playing catch-up

The organizations that will benefit from this shift are the ones building measurement, attribution and oversight into their pipelines ahead of time. That means treating governance as infrastructure rather than paperwork and being able to answer, at any point, which parts of their codebase were AI-generated, who reviewed them, and what production behavior they're responsible for.

Furthermore, budget owners must be able to say what they're actually spending on AI-assisted development, rather than estimating.

None of this slows delivery down in the long term, and if anything, it's what allows delivery to keep accelerating without the incident curve growing alongside it. The gap between how confident enterprises feel about AI-generated code and how much of it they can actually see isn't going to close on its own. It will close because leadership teams decide to build the visibility first.

The organizations that do that now, while the rest of the industry is still counting lines of code shipped, are the ones that will still be standing when the next wave of AI-driven development arrives.

We've featured the best laptop for programming.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Nearly two-thirds of AI users aged 18-28 want clearer instructions from their bosses
• Nearly half of the same cohort have difficulty explaining work completed with the help of AI
• The figures could indicate a brand new challenge to management

“Tell me exactly what you want me to do” – that seems to be the message from older millennial and younger Gen Z workers, who have expressed a need for highly specific instructions from their bosses.

Research has found that 62% of the 18-28 age group who regularly use AI want their bosses to give them the same sort of clear steps and parameters as one might give an LLM chatbot.

While a cohort of workers who are attentive to the needs of the business might sound good, deeper investigation may prompt you to think again. The study by Use.AI implies a deeper-seated issue that points to a management challenge unlike anything seen before on this scale.

Detailed steps

Use.AI surveyed 11,742 adults based in United States, the United Kingdom, Canada, the European Union, Australia and Latin America, finding 62% of the 18-28 age group familiar with the AI use are accompanied by 41% of those aged 40 and older who also want more detail about tasks from their line managers.

Why do younger users need clearer, more detailed steps? It could be a simple matter of judgement, with the 18-28 age group requiring more experience in making their own decisions. But with the presence of AI in their working lives, could they be fearful of being replaced?

The report also explores a slightly different dimension. Within the 18-28 age group, 44% say they have had difficulty to explain work created with AI assistance. This raises a key concern: does the employee even know what their tasks are supposed to achieve? For that matter, does the business?

Sound judgement

AI has changed businesses considerably, and made an impact that doesn’t only improve productivity. As Ihor Herasymov, Co-Founder & CEO of Use.AI, observes, “AI is exposing which parts of work were never really about producing the answer. When a model can generate the output in seconds, the value of some tasks shifts to whether the person can make a sound judgment and stand behind it.”

Other results from the survey support this. With 57% of the 18-28 age group of regular AI users stating that they are less likely to memorize information that can be collected via AI, and 53% of the same group frustrated by “ordinary” tasks that could be completed by a chatbot, a whole new workplace dynamic is being quietly established.

Herasymov concludes, “Managers now have to be much clearer about whether they need the result or whether the task itself is meant to develop the employee.”

• Boeing receives $12.68 million investment to update the MQ-25's RF blanker, encrypted storage, and mission computer security, running through November 2029
• The change is a recent one: the Navy's April 2026 acquisition report lists refueling as the MQ-25's primary mission and intelligence, surveillance, and reconnaissance (ISR) as secondary, with no electronic warfare requirement and only two key performance parameters
• An RF blanker is electromagnetic compatibility hardware that stops an aircraft's own transmitters from deafening its receivers, allowing it to conduct EW more effectively

A single line buried in a recent US Department of War contract announcement is being read across multiple channels as the Navy's first carrier-based drone acquiring multiple electronic warfare upgrades.

The details, however, seem to focus on the defensive side of the spectrum and include a Radio Frequency Blanker Unit and an updated Encrypted Mass Storage System carrying an embedded Data Transfer System.

The modification also provides an option for deficiency analysis and corrections, supporting security assessments of the aircraft's Mission Management System Computer.

Why do these upgrades matter?

The most important upgrade for the MQ-25 Stingray, arguably, is its RF Blanker. An RF blanker is electromagnetic compatibility equipment. When an aircraft carries several transmitters and sensitive receivers within a few meters of each other, a transmitter firing at full power can deafen or damage a nearby receiver operating in an overlapping band.

A blanker fixes that by taking a blanking request from each transmitter and issuing a suppression pulse to the affected receivers for the duration of the transmission, as GE Aerospace describes. For a drone that could also house sophisticated EW capabilities in the future, this is essentially a prerequisite.

The Navy's Modernized Selected Acquisition Report dated April 21, 2026, offers a relatively narrow scope for the drone for now; however, it states that the MQ-25 conducts aerial refueling as its primary mission and provides Intelligence, Surveillance, and Reconnaissance capability as a secondary mission.

It is fairly possible that the upgrades are purely defensive, to prevent rogue players from accessing the drone or knocking it offline with relative ease. In its current capacity, the MQ-25 serves as an unmanned aerial refueling drone for the US Navy, offering capabilities including the ability to carry roughly 14,000 to 16,000 pounds of fuel at a combat radius of 500 nautical miles.

This could change even as the US continues hardening its EW offerings and prepares for potential hostile actors to have the same capabilities it currently harbors. For now, the $12 million invested in what is a $19 billion program is essentially a rounding error at the end of the day.

This sudden investment may signal more to come: Boeing has already pitched a land-based MQ-25 variant with external pylons and marketing language about ISR, electronic warfare, and airborne early warning roles, though the aircraft has no customer yet.

• The US Army has moved TITAN from prototypes into production
• Palantir received $127 million under the latest TITAN contract award
• Anduril secured $65 million for TITAN hardware and integration work

The US Army has shifted its Tactical Intelligence Targeting Access Node (TITAN) program from prototype status into full production this week.

Palantir Technologies and Anduril Industries together received a combined $192 million award covering the next production phase.

The Army first selected Palantir back in 2024 to serve as lead contractor for developing this next-generation targeting system for ground forces.

Contract splits work between two defense firms

TITAN was built to streamline how soldiers collect and process data when locating and striking distant targets on the battlefield.

Palantir secured $127 million of the total award and will oversee TITAN's production alongside several supporting partners in the effort.

Those partners include L3 Harris Technologies, Sierra Nevada Corporation, Strategic Technology Consulting and World Wide Technology, each contributing specialized capabilities to the program.

Anduril received the remaining $65 million and will build the hardened, vehicle-mounted units required for field deployment.

Anduril's portion also covers integrating power systems and thermal management components, so TITAN functions reliably in harsh operating conditions.

The program includes two distinct configurations built for different vehicle platforms and mission requirements across Army units.

An advanced version will mount on tactical trucks and can pull data directly from orbital sensor networks overhead.

A simpler basic version is intended for the Joint Light Tactical Vehicle, giving smaller units similar targeting capability.

Under the current award, contractors must deliver four advanced units and four basic units within eighteen months.

Production ramp builds on existing prototype fleet

These newly ordered units will supplement nine prototype systems the Army already operates and uses in current exercises.

Army officials have indicated they plan to issue further production contracts for TITAN sometime during fiscal year 2027.

The move from prototype to production status marks a milestone for a program meant to modernize how ground troops gather targeting information.

TITAN forms a central piece of the Army's broader push to overhaul aging ground systems across its combat formations.

By linking space-based sensors with vehicle-mounted terminals, the system aims to shorten the time between spotting a target and engaging it.

Whether that translates into a meaningfully faster targeting cycle in real combat conditions remains to be tested at scale.

The involvement of five separate companies alongside Palantir and Anduril suggests the Army is spreading technical risk across multiple vendors.

That approach could either strengthen the program's resilience or complicate coordination as production accelerates toward the 2027 follow-on orders.

For now, the contract signals steady progress rather than a dramatic leap forward for the broader modernization effort.

The scale of future procurement will depend on Army requirements, funding decisions, and how the systems perform during operational evaluation.

Via Defense News

• The US Navy approved Rolls Royce engines for autonomous surface vessels
• The certified engines can support missions lasting up to 30 days
• Series 4000 engines completed two separate 720-hour durability trials

The US Navy has approved Rolls-Royce marine engines for unmanned vessels capable of operating for periods reaching 30 days without crews onboard.

The certification covers mtu Series 4000 and Series 2000 engines following tests examining their suitability for extended autonomous naval missions.

The approval comes as the Navy expands its use of unmanned surface vessels for missions where keeping sailors away from danger remains important.

Rolls Royce engines clear tests for autonomous vessels

The Series 4000 completed two separate 720-hour durability assessments before receiving approval for propulsion and onboard electricity generation duties.

That engine can deliver as much as 4,300 kilowatts for propulsion while providing up to 3,000 kilowatts through its electrical generation system.

The smaller Series 2000 gained approval after component checks and operational trials, allowing it to provide propulsion output reaching 1,939 kilowatts.

Both engine families can now support Navy vessels designed for lengthy missions without requiring people aboard for routine maintenance or repairs.

The certification followed requirements established through the National Defense Authorization Act, which sets conditions governing equipment used within autonomous military platforms.

The Navy has already begun deploying unmanned boats, including Sea Hunter vessels fitted with twin mtu 12V 2000 engines supplied by Rolls Royce.

That deployment gives the newly approved engines a direct connection with the Navy's growing fleet of vessels designed for remote operations.

The company says its propulsion systems have supported naval operations for decades, while newer autonomous missions create different demands for endurance.

“For decades, mtu engines have been trusted to provide reliable, mission ready power for our crews at sea,” said Scott Hanson, vice president of governmental sales at Rolls Royce Power Systems North America.

Longer missions create new demands for naval engines

The ability to keep a vessel operating for 30 days without onboard personnel changes how the Navy can use unmanned surface craft.

Such vessels can remain at sea for extended periods while avoiding the accommodation, support, and immediate safety requirements associated with crewed naval platforms.

Rolls-Royce said the certification gives the Navy additional options as unmanned vessels become part of its wider fleet plans.

The company also supplies propulsion equipment for several unmanned aircraft, including the MQ 4C Triton and MQ 25A Stingray platforms.

Its portfolio also includes engines used by the United States Air Force's RQ 4 Global Hawk unmanned aircraft.

The latest approval therefore extends an existing relationship between Rolls Royce and American military programmes into autonomous maritime operations.

The certification does not mean every unmanned vessel using these engines will automatically achieve 30-day endurance during actual missions.

Actual operating duration can vary according to vessel design, speed, weather, payload, fuel capacity, mission demands, and maintenance requirements.

For the Navy, the immediate significance lies in having certified propulsion options for vessels expected to spend longer periods operating without crews.

As more unmanned boats enter service, engine reliability will remain important because failures far from support facilities can impede mission success.

Via The Drone Front

• Ukraine has started European production of its BARS jet-powered strike drone
• The latest BARS configuration reportedly has a 1,000 kilometers range
• BARS can now carry warheads weighing up to 60 kilograms

Ukraine has begun production of its BARS jet-powered strike drone in Europe, and the latest version has a range of 1,000 kilometers.

Production commenced in early September, intending to support higher output as developers continue to modify the drone for different mission needs.

This is a joint effort between Ukraine and Germany, with the first completed drones expected to reach Ukraine's Defense Forces from October 10 2026.

BARS enters European production with longer reach

The arrangement followed an agreement signed July 8 by Ukrainian Foreign Minister Andrii Sybiha and German Defense Minister Boris Pistorius.

Germany is financing the first production stage, while completed systems are intended for use by Ukraine's armed forces.

Engine availability remains a key limitation, since turbojet deliveries currently determine how quickly additional aircraft can leave production.

The BARS jet-powered strike drone can travel above 700 kilometers per hour while carrying a substantial warhead.

However, with several modifications and upgrades, the newest configuration can reportedly travel 1,000 kilometers.

Its maximum warhead weight has also increased substantially, rising from 22.5 kilograms on earlier models to 60 kilograms.

Its design allows developers to alter fuel storage and payload capacity based on the distance required for a particular mission.

Longer missions can use additional internal space for fuel, while shorter flights can devote more room to carrying additional weight.

Developers keep changing the drone and its guidance

BARS has reportedly gone through more than 40 design changes and upgrades since the prototype was developed.

The platform is divided into DeepStrike, MiddleStrike and FrontStrike configurations, covering different mission ranges and operational requirements.

The developer says artificial intelligence now contributes to several stages of the program, including engineering work and the drone's final guidance system.

“We use the tools that AI technologies provide us today, and we do not shy away from them at any level—from development and solving engineering problems to implementing machine vision in our target-guidance systems,” said the BARS developer.

During the final part of its flight, the operator can continue controlling the drone or transfer control to optical guidance.

This expansion of BARS production is a strategically sound move for Ukraine because German financing and supply chain ease pressure on Kyiv's defense budget.

Producing BARS in Europe also shields manufacturing from Russian strikes that have repeatedly targeted Ukrainian defense facilities.

For Germany, this partnership signals a deepening commitment to Ukraine's long-term defense industrial base rather than merely short-term arms transfers.

It strengthens bilateral ties and proves that European backing can translate into real, sustained battlefield capability.

Such a longer-range, heavier-payload drone with AI-assisted guidance raises Moscow's defensive costs by threatening targets deeper behind the front line.

This forces Russian air defenses to stretch across a wider geography, while the joint-production model may set a template for future allied co-production deals.

Via United24media

Elon Musk has been at the heart of promoting various companies throughout the 21st century, with two of his most prominent companies anchored in the notion of mass production. In the case of his infamous clunky and angular Tesla Cybertruck, he encountered several difficulties in bringing the prototype to market.

Planting the seed

Musk was speaking about the rigors of bringing the Cybertruck to life during an appearance on the Joe Rogan podcast.

Rogan asked how far away Musk was from delivering the vehicle to people, with Musk revealing at the time that the timing was about a month away.

But the prototype was initially unveiled in November 2019. Explaining the delay, the Tesla CEO hinted that the process of manufacturing the vehicle proved much harder than expected given the need for consistency in producing each model.

Rinse and repeat

There's an element of common sense in buying into the idea that making a prototype is much easier than mass-producing a finished version of that product.

Not only is there a much lower tolerance for error, but establishing the supply chain for materials, components, and resources is far more complex. Then there's the economics of it all – ensuring that the cost to produce one vehicle can, at least, be recouped by a customer should there even be a willingness to pay for it.

It's reminiscent of the "production hell" phrasing that Musk has also frequently deployed through the years – especially during a manufacturing crisis between 2017 and 2018. During this time, the Tesla production line became a futuristic and automated process known as the "alien dreadnought" – but this robotic network eventually slowed down manufacturing and prevented Tesla vehicles from being ready on time.

• AMD is putting data center-class AI hardware inside a desktop workstation
• The Halo Station combines 96 CPU cores with four powerful AI accelerators
• Up to 2.6TB of memory gives enormous AI models more local headroom

AMD's Threadripper Halo Station could become one of the most extreme desktop workstations ever built for demanding artificial intelligence workloads.

The device combines a 96-core processor with AMD Instinct accelerators and a huge memory pool designed for large local models.

Unlike conventional workstations, AMD is building this system around data center-class computing hardware inside a single deskside machine.

AMD packs data center hardware into a desktop

At the heart of the Halo Station is the Ryzen Threadripper PRO 9995WX, based on AMD's Zen 5 architecture and carrying 192 threads.

AMD lists up to 2TB of RDIMM system memory, while the accelerator side can add as much as 576GB of HBM3e memory.

Together, the system's CPU and accelerator memory can reach 2.6TB, giving large AI models far more room to run locally.

AMD also claims up to 16.4TB/s of combined memory bandwidth, giving large datasets rapid access to both system and accelerator memory.

The accelerator configuration can use four Instinct MI350P units, with each device carrying 144GB of HBM3e memory and 4TB/s bandwidth.

That design gives the machine a far larger high-speed memory pool than typical professional desktops built around a single graphics card.

AMD compares the system with Nvidia's DGX Station, whose GB300 platform offers 748GB of coherent memory across its CPU and GPU resources.

The comparison shows why AMD is focusing heavily on memory capacity, rather than treating processor core count as the main selling point.

A machine with this much memory can keep larger models and datasets close to the compute hardware instead of constantly shifting them elsewhere.

Halo Station is built for local AI work

AMD says the workstation is intended for researchers, model developers, engineers, creators, and organizations that need intensive workloads handled locally.

Its main role is running, training, and fine-tuning large AI models that would normally require access to expensive remote computing infrastructure.

The system can handle frontier-class models while supporting hundreds of AI agents operating within the same local environment.

That capability makes it useful for development tasks where repeated cloud transfers create delays, added costs, or data control concerns.

AMD also says the design allows sensitive data and intellectual property to remain on premises instead of moving through external cloud services.

The company unveiled a prototype at IFA 2026, with commercial availability planned for 2027.

AMD senior vice president Jack Huynh described the device as “the most powerful workstation in the world,” emphasizing its unusual computing scale.

Industry estimates suggest fully equipped systems could cost around $150,000, although AMD has not announced an official retail price.

Rack-mounted systems can cost far more than desktop workstations, so the Halo Station is not necessarily the world's most expensive computer.

The estimated price reflects the cost of combining a high-end CPU with multiple AI accelerators and several terabytes of memory.

• Samsung's new portable SSD combines 8TB capacity with 4,000 MB/s reads
• The 8TB P9 offers ample storage for massive media libraries
• Samsung's fastest portable storage arrives with an unusually high price

Samsung's 8TB P9 portable drive has appeared on an ecommerce platform for $2,879.99, making it an unusually costly option for portable storage.

B&H Photo lists the drive alongside 1TB, 2TB, and 4TB versions, priced at $2,520, $2,160, and $1,440, respectively.

The pricing makes the 8TB model cheaper per terabyte than its smaller counterparts, despite its unusually high overall purchase cost.

USB4 performance comes with 8 TB capacity

The P9 connects through USB-C and supports USB4 at 40 Gb/s, giving it enough bandwidth for the listed transfer speeds.

Sequential read performance can reach 4,000 MB/s, while sequential writes can reach 3,800 MB/s under suitable operating conditions.

Samsung says the drive is intended for professional creators handling large RAW photo collections and high-resolution video files, as the 8TB capacity can also support 12K video recording and direct capture when connected equipment supports those functions.

This portable SSD draws power through the host connection, so users do not need a separate power supply for normal operation.

Samsung claims the drive is protected against drops and can survive falls from up to 6.5 feet, approximately two meters.

Thermal controls are also designed to manage heat during demanding workloads, while the enclosure also resists vibration.

The enclosure measures 3.54 x 2.56 x 0.33 inches, making it roughly comparable in length and width to a typical smartphone.

It weighs just 3.39 ounces, or 96.11 grams, so the drive adds little weight to a portable work kit.

With 8TB available, the Samsung P9 can also reduce the need to carry multiple drives when handling large media files.

Security and device support are added to the package

The P9 uses AES 256-bit hardware encryption and password protection, which can help keep stored files inaccessible to people without the required credentials.

Samsung Magician Software also lets users manage the drive's password protection, giving them another way to control access to stored data.

The drive works with Windows, Mac, iOS, Android, game consoles and tablets, so users can connect it to many types of devices.

It also works with USB 3.2 Gen 2x2 at 20 Gb/s and earlier USB standards, although those connections cannot deliver the drive's full USB4 speeds.

Samsung says sustained write performance can support demanding tasks such as recording content directly to the SSD and editing files stored on it.

That could be useful for users who need to record or work with large files without first moving them to another storage device.

However, B&H currently lists the 8TB model as unavailable while offering customers the option to request an availability alert.

Anyone who submits that request should receive a notification when the drive becomes available, with shipping details then provided through the retailer.

• OpenAI is developing a tool that can perform “well-defined research tasks”
• Described as a “research intern” the project met its aim to achieve automation by September 2026, ahead of a fully automated AI researcher operational by March 2028
• The company believes that automated research can “enhance human welfare” and make other contributions

OpenAI has confirmed its project to build a fully automated AI researcher is on schedule, hitting its September 2026 target of developing a "research intern". Capable of performing research tasks commissioned by a human, the "intern" is the main milestone in the completion of the fully automated system, which OpenAI aims to reveal in March 2028.

The research intern is capable of carrying out tasks that would take several days if completed by a human researcher.

OpenAI has justified the development of the automated AI researcher by citing the reduced cost of AI, and in defending critical infrastructure against capable AI threats. These are observations it may not have published had the Hugging Face incident not occurred.

Automated research?

During an October 2025 livestream, OpenAI CEO Sam Altman announced the development of an automated AI researcher, and its intended milestone and target. "We think it is plausible,” he said, “that by September of next year, we have an intern-level AI research assistant and that by March 2028, we have a legitimate AI researcher."

Specifically describing the current stage of the project’s development, OpenAI explained its definition of ‘research intern’ as "a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days.”

The post continues: “AI research is a complex process with many potential bottlenecks […] agentic tools are meaningfully accelerating research progress.”

While this particular intern won’t be able to make coffee or find a box of paperclips, its automated research ability will presumably unlock those tasks for others to complete – such as the human commissioning the research.

Recursive self-improvement

The ChatGPT company clearly expects to hit its March 2028 target for the release of the publicly available automated AI researcher, but it also seems to have learned some important lessons. It argues against the use of recursive self-improvement (RSI) in developing automated research capabilities, noting that achieving that safely is not something that is currently achievable.

OpenAI’s post underlines its continued response to the Hugging Face incident, introducing safety steps at an earlier stage of product development. The update also explains how coding agents are used for development at the company, and confirms recent revisions to monitoring standards.

Noting that “Agent-powered AI research is still new, and we are still learning how to measure it,” OpenAI’s “intern” could herald some some considerable changes to the completion of research, when it the automated AI research tool is finally completed.

Allow me to introduce the new Philips 34B2U5900C. On paper, it's pretty much the ultimate non-OLED 34-inch 5K2K ultrawide productivity monitor for PCs and Macs.

If that sounds like a rather convoluted classification that amounts to faint praise, it's not meant to be. OLED monitors certainly have some clear advantages. But you simply can't get one that matches this Philips' core specification of 34-inch ultrawide with a 5K2K resolution.

That's important, because this display offers a level of pixel density and therefore text and image sharpness and clarity that no current 34-inch ultrawide OLED monitor can match.

What's more, even up against the traditional LCD competition, the Philips 34B2U5900C looks pretty special. The full 5,120 by 2,160 native resolution is supported at 120Hz refresh and there's even a 2,560 by 1,080 alt-mode resolution at 240Hz.

Intriguingly, this is a VA panel as opposed to the IPS type you might have expected from a 5K2K model, which is why the contrast is quoted at an impressive (for an LCD monitor) 3,000:1. Notably, the full 5K2K at 120Hz is supported over all three video inputs: DisplayPort, HDMI and USB-C. While we're talking connectivity, it's positively comprehensive, with dual HDMI to accompany the aforementioned DisplayPort and USB-C and the latter providing an outstanding 140W thanks to supporting USB-C Power Delivery version 3.1.

A KVM switch likewise means this monitor is perfect for sharing across two machines, perhaps a PC and a Mac or a laptop and a desktop. There's plenty more in the huge feature set, including sensors for both light and user proximity, the former designed to adjust the screen's brightness to suit ambient conditions, the latter allowing this monitor to dim when you step away from your desk.

For the record, this is a DisplayHDR 400 certified panel, which means it's on the very lowest rung for HDR support and capability and doesn't offer local dimming, let alone the per-pixel lighting control of an OLED panel. But, again, this class of 5K2K panel simply isn't available with OLED technology. On paper, the Philips 34B2U5900C looks set to feature strongly among the best business monitors you can get.

Philips 34B2U5900C: Design and features

• Slightly dated design
• Fantastic feature set
• USB-C with 140W power delivery

Immediate impressions of the Philips 34B2U5900C are of a slightly clunky, dated old thing with a particularly thick lower bezel. An ultra-slim OLED it most clearly is not.

But it's well put together and exceptionally ergonomic. The latter attribute is thanks to the combination of a base that swivels through fully 360 degrees and clearly labelled and easily accessible ports on the rear. If you're fed up with battling with hard-to-access monitor ports, you'll love the Philips 34B2U5900C.

Speaking of connectivity, along with the dual HDMI ports, there's a DisplayPort socket and USB-C. The latter supports the latest PD 3.1 spec and so can deliver fully 140W of power delivery to an attached laptop. That's well up on the 96W of previous generations of USB-C single-cable connectivity and means that you use a single cable to drive this display and fully power many high-spec desktop replacement portables.

There's also a KVM switch, which makes it simple to share this monitor and, say, an attached keyboard and mouse combo, across two computing devices. The inclusion of an ethernet port means you can share networking across the devices, too. What's more, there's a multiview option allowing both machines to display on the monitor at the same time. Nice. Rounding out the connectivity is a three-port USB hub on the left bezel, with two USB-A ports and a single USB-C, plus a headphone jack.

Along with a light sensor that enables the 34B2U5900C to adjust its brightness according to ambient conditions and a proximity sensor that dims the brightness when you step away from your desk, perhaps the most unusual feature in a productivity context is the dual resolution mode. By default, the 34-inch panel runs natively at 5,120 by 2,160 pixels at 120Hz. That alone is pretty impressive for a productivity monitor with such a high resolution.

But Philips has also fitted this monitor out with a secondary dual-mode using pixel doubling to enable a resolution of 2,560 by 1,080 at 240Hz. It's the sort of feature that you'd expect on a gaming monitor. Exactly what the use case is for productivity, content creation or day-to-day computing isn't clear. But it hardly counts as a demerit to find it included.

Finally, color coverage of 95% of the DCI-P3 gamut is reasonable, but not really the stuff of professional content creation. That said, the 34B2U5900C is at least Calman Ready for rapid automated calibration using a colorimeter.

Philips 34B2U5900C: Performance

• Super sharp image quality
• 120Hz at native resolution
• Surprisingly good response

The central appeal of the Philips 34B2U5900C is its 5K2K resolution. On this 34-inch ultrawide panel, that native resolution of 5,120 b 2,160 pixels works out to a pixel density of 163DPI.

That's exactly the same as a 27-inch 4K monitor and, indeed, this ultrawide 5K2K format is essentially a 16:9 4K 27 incher taken out to the 21:9 aspect. Add in a gentle 1500R curve, and you have a fantastic combination of widescreen ergonomics for multitasking and media immersion, plus super-sharp image quality and excellent text quality.

That applies to both Windows and MacOS operating. What's more, among the comprehensive selection of color spaces in the 34B2U5900C's OSD menu are both sRGB and Display-P3. So, you have immediate presets to suit the native color spaces of both operating systems.

This panel looks particularly good with MacOS. The pixel density isn't quite Retina class. But the font quality is still excellent and the ultrawide proportions allow for running as many as three application or browser windows side-by-side, a use case that obviously applies to Windows, too.

It's something of a surprise to find Philips using VA panel technology for this new 5K2K display. IPS is more typical. But this is one of the best VA screens we've seen for a long time. First, you get excellent inherent contrast for an LCD panel at 3,000:1. So, there's noticeably less light bleed than a similar IPS panel.

Even better, despite VA's reputation for poor pixel response, Philips has included three levels of configurable overdrive, the middle setting of which delivers very, very good response for this type of monitor. Motion clarity is very much in line with the 2ms quoted response performance and, in subjective terms, you're giving nothing away to the IPS competition, albeit OLED panel tech is nearly a full order of magnitude faster.

Speaking of speed, the 120Hz refresh rate is a very welcome feature. This isn't a gaming monitor and triple-digit refresh rates are not really a particular requirement for most productivity or content creation workflows. But 120Hz still makes for slicker and more responsive day-to-day computing.

The 240Hz 1080p dual-mode of more questionable utility outside of gaming. Enabling it makes for clearly compromised clarity, that's for sure. But it's there for the occasional fast-paced gaming session, if that's your thing on the side.

If you're looking for limitations, there are two obvious if hardly debilitating weaknesses. First, HDR support is capped at the entry-level DisplayHDR 400 tier. That means a peak brightness of 400 nits and no local dimming. So, this isn't a true HDR display. But then even LCD monitors with mini-LED dimming struggle to deliver a consistent HDR performance. You need OLED and perfect per-pixel lighting for that.

The other marginal shortcoming is color coverage. You'll do better than 95% of the DCI-P3 gamut this monitor offers with an IPS panel. So, if maximising color accuracy for pro content creation workflows is critical, you may find the 34B2U5900C slightly wanting.

Philips 34B2U5900C: Final verdict

If you're looking for a fantastic all-rounder that combines crisp image quality with ultrawide multitasking ergonomics, great connectivity and even some surprising speed, look no further than the new Philips 34B2U5900C. It ticks so many boxes.

The 5K2K resolution on the 34-inch ultrawide panel is the main attraction. And certainly makes for both super sharp visuals and lots of desktop elbow room. In this application, the VA panel tech works really nicely too, delivering excellent contrast for an LCD monitor. Add the 120Hz refresh at 5K2K native and surprisingly nippy pixel speed and you have a very quick and slick productivity monitor, to boot.

Then there's the excellent connectivity. The highlight is the USB-C interface with 140W capability thanks to Power Delivery 3.1 support. It means you can keep a very powerful desktop replacement laptop fully juiced, which isn't something the previous USB-C PD standards could achieve.

All told, this is comfortably one of the best monitors of 2026. It absolutely nails its high-DPI productivity brief and adds a surprising turn of speed that makes it a versatile display, even if gaming isn't exactly central to its all-business remit.

For more displays, we've tested the best monitors and the best 5K and 8K monitors.

• 153 million US driving licences have been leaked on a Russian cybercrime platform
• Among those apaprently discovered in the stolen data is US Secretary of Defense Pete Hegseth
• The FBI is now investigating the leak, which has been traced to an identity verification company

A data leak of 153 million US drivers licenses is said to have been shared on a Russian cybercrime forum, with US Secretary of Defense Pete Hegseth among those leaked prompting an FBI investigation.

Security researcher Brian Krebs identified the leak – which included his own data – as originating from a hack of an identity verification service. Louisiana-based IDScan provided ID verification for various well-known companies, including FedEx and Hertz car hire.

The data was shared on a Russian forum called Exploit, a long-established online community of cybercriminals. Following news of the leak, the identity theft service “Nexus” has apparently scrubbed its existence from the Dark Web.

Driving licenses and more

It wasn’t just US driving licenses that were found in the archive of recently-collected personal data. Krebs’ investigation found that Nexus claimed other types of data, and found a further 1.1 million driving licenses from Canada.

Other identity documentation alleged to be in the leak include 10 million identification cards, three million travel documents and international IDs, and 579,000 medical cards. The data was available to browse, notes Krebs, with Nexus providing details: “Records are available to preview before purchase with pertinent information redacted. Customer photos are displayed if available.”

The leak has a personal dimension for Krebs. Not only was his driving license in the collection, so was that of his mother. It has proved to be a useful coincidence, one that has enabled the security and privacy researcher to establish how the data was sourced by Nexus.

Both licenses were used for a car hire, which Krebs traced to Hertz. Others affected by the leak had also used the service, which has used New Orleans-based IDscan for identity verification.

The company, which claims to perform 21 million verifications a month, is yet to issue a statement on the matter. Its marketing and operations leader, Jillian Kossman, told the journalist: “At this point I’m not able to share any additional information, but the updates you have provided have been welcome, and helpful to our team’s investigation.”

Where is the data?

Krebs reports that he was alerted to the data on August 31, 2026, little over a week ago. Regular checking noted that the data was still being added to, increasing by “nearly 400,000” records prior to publishing his investigation on September 1.

Since then, however, it seems that Nexus has vanished, along with the data. But how widely was the data downloaded before that happened? While the FBI investigates, it falls on the American and Canadian public to be extra vigilant and wary of identity fraud.