• OpenAI has solved or partially solved 372 outstanding mathematical challenges
• Some of the solutions, while clear, cannot be understood by some humans, and were produced with an internal, unreleased AI model
• Renowned mathematician Terence Tao has called for colleagues working with OpenAI to abandon their projects

OpenAI has stunned the world of advanced maths by releasing solutions and partial solutions to 372 outstanding problems that have puzzled the world’s greatest minds for decades.

Just two years ago, AI was finding it tricky to handle GCSE-level mathematical challenges. Today, it is solving problems that have taken humans years to even establish the existence of. The solutions provided, meanwhile, are considered almost as complex as the problems, and the scale – published in 722 papers – make it difficult for human mathematicians to understand and verified.

The impact of the release has prompted one of the world’s greatest mathematicians, Terence Tao, to call for a return to a traditional vision of science.

Demonstration of power

OpenAI’s publication of its work on these advanced mathematical problems is the latest demonstration of how far the technology has come. Just a few weeks ago, the company published its proposal for the Navier-Stokes equation (concerning the behaviour of viscous fluids), generated by its unreleased specialized math model.

Tao is leading the opposition, as one of the field’s leading voices, and a man considered to be the greatest living mind in mathematics.

Following OpenAI’s contribution, he shared a post from the Association for Human Mathematics, which concludes:

“Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power. We urge mathematicians to discontinue their work with OpenAI and to return to a vision of science that centers human understanding.”

However, as powerful as that conclusion is, their response to OpenAI’s intervention as a whole highlights that the company may have overstepped.

Questions of legitimacy

The Association for Human Mathematics states that “Mathematicians did not ask for this work to be done,” continuing that “The Advisory Group on Mathematics and Artificial Intelligence, from whom OpenAI has claimed to derive its legitimacy, opened their initial advisory statement by saying that frontier AI corporations should not test advanced mathematical problems on internal models.”

By ignoring the position of the advisory group, they state that OpenAI has essentially gone beyond the pale, and “indicated total disregard for the norms of scientific research — norms that guarantee that mathematics remains trustworthy, ethically researched, and in the public interest.” Furthermore, they encourage mathematicians to treat the publication with “due skepticism.”

OpenAI has responded by stating "we want this progress to push the frontier of human knowledge and enable further progress in mathematics." However, its actions are being interpreted as reckless, and this showboating may well prove to be damaging for OpenAI and the wider AI industry.

The rise of generative AI, mixed with the ease of access provided by crypto currency, has created scamming opportunities we couldn’t have imagined just a few years ago. Romance scams that exploit these technologies have now become one of the fastest growing and most devastating forms of financial crime.

The capabilities of generative AI mean it’s now possible to get scammed by a range of incredibly convincing fake personalities, including deepfakes of celebrities and public figures. In fact, according to the makers of the new documentary Super Scam Me, it’s revealed that an estimated one in seven American adults have lost money to this form of deception. In 2025 alone, romance scammers stole more than $1 billion from American victims.

In Super Scam Me, Emmy-winning investigative journalist Mariana van Zeller builds a digital target persona to intentionally lure a persistent fraudster, then teams up with a white-hat hacker to reverse-engineer the scammer’s digital footprint and track him down and confront him face-to-face.

In the documentary Mariana meets the victims whose lives have been upended by romance scams, including a woman who was duped into believing she was in a year-long relationship with Jon Bon Jovi. Yes, that Jon Bon Jovi. What she thought was a genuine connection with the superstar musician left her facing foreclosure on her home, providing a window into the devastating real-world consequences of these crimes.

I caught up with Mariana van Zeller to ask her about the new documentary, Super Scam Me.

TechRadar: From talking to people who were scammed, do you think they were especially gullible, or could this really happen to anyone?

Mariana van Zeller: I truly believe this could really happen to everyone. I mean, the main reason why I decided to do this documentary was because very often I come back from reporting on these scams and people's response would be, "How stupid could they be?" and "Why did they react like this?" and "Why do you give money to a stranger?" And I really wanted to turn the table around and start focusing on the actions of the actual scammers.

TR: How sophisticated are these scammers?

MvZ: You know, these scammers are very, very good at what they do. They spend a lot of time sort of honing their craft, and they use the latest technology out there. They have it all at their disposal. They have scripts, they work in teams, they're educated, there are schools for scammers out there. And they prey on emotions that are common to all of us. You know, hope, fear, love, greed in some cases. So as long as we have those emotions — which I believe every single one of us do — we are very likely, or it is very possible that we can fall victim to one of these scams.

TR: How do we stop this?

MvZ: I think awareness is very important. I think we should be aware of how good these scammers are, about the tactics and technology and strategies that they use. It's partly what I've tried to show in Super Scam Me. But I also think the onus shouldn't just fall on individual citizens. We should be asking our governments, we should be asking big tech and banks to do a lot more than they're doing.

You know, Leslie is a great example. In the documentary, she was supposedly having a relationship with Jon Bon Jovi for almost a year. And she is out there right now. Her heart was broken. She lost a lot of money. She almost lost her house. She felt humiliated and shamed and suffered this enormous loss.

She is now out there herself, every single day, on these platforms, raising awareness and telling people that these are fake profiles and reporting every single one of these profiles to Facebook and to Instagram and Meta in general, and hoping that they will take them down. And she says just four or three in every 10 of these actually get taken down, and the rest have existed for months.

And again, it shouldn't be Leslie, right? It shouldn't be a person who is a victim of a scam trying to prevent other people from falling for these scams. It should be our governments, it should be big tech and our banks, and there should be just a lot more being done to make sure that we're protecting our citizens.

TR: Thank you Mariana.

Super Scam Me, premieres on October 17 in the US, airing on National Geographic at 9pm EDT and streaming on Disney+ and Hulu the same day. In the UK Super Scam Me premieres on Sunday 18th October at 10pm on National Geographic, and streams from the same day on Disney+.

If you are affected by any of the issues raised in this piece, support is available.

Whatever you are going through, you don’t have to face it alone. In the US, you can call or text the 24/7 Lifeline, which is confidential and free, at 988. You can also reach the Crisis Text Line by texting HOME to 741741.

In the UK, you can call Samaritans for free on 116 123, email jo@samaritans.org, or visit the website to find your nearest branch.

• Lines of code, code commits and pull requests are all up 20-30%
• Comments per pull request are up 49% as human review becomes more crucial
• With reviewing up from 7 to 10 days on average, the new bottleneck is code review

A new Harvard University research paper has confirmed that AI is indeed excellent at boosting coding productivity, but humans just end up slowing the process back down anyway.

Authors Fiona Chen and James Stratton found a 30% increase in the amount of code developers can produce when using AI, but they found no statistically significant improvement in completed software.

With increases across lines of code produced (+30%), code commits (+20%) and pull requests (+23%), the Harvard-backed researchers went on to investigate why we're not seeing a 20-30% increase in end product.

AI has definitely sped up coding, but human delays negate much of those benefits

At the same time, the researchers also found increases across the time taken to review and merge pull requests (+49%), the number of reviewer comments per pull request (+35%) and even the proportion of employees doing code reviews (+14%), all of which implying that vibe coding is ultimately shifting the bottleneck further down the road instead of removing it altogether.

The paper also questions whether, with AI shifting the workload, companies are responding by allocating more workers to code review. Layoffs have become somewhat synonymous with AI – founded or not – and concerns about jobs continue, however the paper proves that human workers are still vital, just in other parts of the coding process.

Some thought it also given to the next stages of AI adoption, and whether it can help downstream in the code review stages. While AI is already being employed to some degrees in this part, "humans are still heavily involved with review."

There's also the consideration of added workload, with AI now capable of producing code much more quickly than humans ever could. The average review process has gone from taking 7.03 days to 10.48 days as a result of this shifted bottleneck.

The researchers also argue that AI-generated code needs substantially more corrections, begging the question whether it's actually helping at all.

Still, with AI adoption not resulting in any significant job losses, it seems, it's more a question about how companies can best use AI to support employees and reduce unnecessary workload, rather than how they can replace workers outright.

"The productivity gains are absorbed by downstream constraints in the production process," the paper concludes, adding that "code review becomes the bottleneck."

• Microsoft reveals Maia 300, its latest AI chip
• Designed for "inference-first", Maia 300 offers huge improvements in efficiency and power
• Maia 300 will begin deploying in 2027

Microsoft has revealed Maia 300, the next generation of its Azure hardware as it prepares to scale up its in-house AI infrastructure to multi-gigawatt deployments.

The company says Maia 300, set to begin deployment in 2027, will deliver 40% more tokens per megawatt than other leading GPU-based system - designed primarily for inference, the new release targets long-context reasoning, responsive conversational AI and sustained, multi-step agent tasks.

Microsoft says a rack containing 96 Maia 300 accelerators will provide up to seven times the rack-scale token throughput of a Maia 200 rack.

Inference-first

Interestingly Microsoft says it expects Maia 300 to help support a new generation of AI across its products and services, from agentic coding in GitHub to large-scale model serving through Microsoft Foundry.

But it also expects improved inference and data generation for its Microsoft AI team, and improvements in the likes of Copilot and Cowork, as well as individual products such as Microsoft Teams.

Designed with what Microsoft says is an "inference-first architecture", Maia 300 is built to improve performance by reducing data movement and making more efficient use of computing, memory and networking resources, Microsoft says, with its software-defined dataflow architecture, combined with AI-assisted kernel optimization, designed to keep data close to the parts of the system that need it.

Unsurprisingly, memory is central to the design, as Maia 300 will use a huge eight stacks of HBM4, with a claimed peak bandwidth of 18TB/s within a 950W thermal design power. Microsoft says specialized software can use on-chip SRAM as a scratchpad, easing pressure on HBM during demanding workloads.

The accelerator also introduces a multi-chiplet architecture that separates compute, memory-controller and input/output functions. Its compute chiplets will be built using TSMC’s N2P process. Microsoft says support for low-precision formats, including an enhanced MXFP4, will reduce memory and bandwidth demands, helping the system run larger context windows and increase throughput within the same power envelope.

Networking is designed to operate as part of the system’s dataflow rather than as a separate layer. Maia 300 will use high-bandwidth connections between accelerators for tensor parallelism, alongside a custom AI transport based on Ethernet infrastructure. Microsoft says this approach can scale from a single rack to data-center deployments.

A Maia 300 rack will contain 96 accelerators within a 167kW power envelope, with 27.6TB of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth. Microsoft says that density will let it serve more simultaneous AI interactions and larger workloads from each rack.

The launch builds on Maia 200, which Microsoft says is already running production workloads including Microsoft 365 Copilot and models developed by its Microsoft AI team.

The company says Maia 200 delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork. GitHub Copilot and Microsoft Foundry are expected to follow by the end of this year.

• Essex man Rafal Gorfal used ChatGPT to help him win a legal dispute over a parking ticket fine
• He fed the AI the letters he received from debt collectors, and it told him what to say and where to go
• The fine was dropped just weeks before his scheduled court appearance

People use AI chatbots like ChatGPT for all manner of things, but have you ever used AI to evade a parking ticket fine before? This man did, and he went on to win a three-year-long legal battle.

Rafal Gorfal, who’s originally from Essex, was fined back in August 2023 for not paying fast enough in a Newquay pay-and-display car park. Gorfal said that when he arrived at the payment machine he had to wait in a long queue, only to learn that the machine was cash-only, to which he had no choice but to download the RingGo and pay with his phone — delaying his payment further.

Two weeks later, he was presented with a £100 fine that said he didn’t pay for a ticket until 16 minutes after entering the car park, which was delayed due to the long queue and the time it took Gorfal to set up an online account to pay.

A long three years

Gorfal submitted his first appeal a month after receiving the PCN from Smart Parking Limited, but didn’t hear anything until January 2024 when he was sent another debt collection letter, this time with an increased fine of £170. "I appealed straight away and heard nothing from them," Gorfal told The Independent, adding "They were trying to make me scared. You have to give people the time. I told them exactly what my problems were. They could see on the cameras that I was trying to sort it out".

After appealing back and forth with debt collectors for four months, the case was put on hold, and Gorfal heard nothing for the best part of a year and a half. This period of silence was broken when he was sent a new letter stating that the case would be going to court in October 2026. This is when he sought help from ChatGPT.

Over the course of his legal dispute, Gorfal kept records of every letter he was sent, which he says he entered into the AI chatbot and asked for advice. "From start to finish I was asking ChatGPT every time I got a letter what to do. It was telling me what to say, where to go, and everything," he shared. "I did it all with the help of AI. They even prepared my defense which I sent to the court". He also contacted the Citizens Advice helpline, but he says the help wasn’t as effective as ChatGPT.

David defeats Goliath

Gorfal’s initial £100 fine had sky-rocketed to roughly £300 by the time he was issued the court order but on September 25, just weeks before he was due to make the appearance, he received an email informing him that the case and fine had been dropped — resulting in a win for Gorfal after a long three years.

That said, it wasn’t money that fueled Gorfal’s persistence, there was a big morality factor. "From start to finish it was not about the money, I could pay it, but it is about the principle. If you know you are right then you shouldn’t be paying any money to them," he said. "This is kind of a David and Goliath battle. They have lawyers and they are a big company and they just gave up in the end".

Since his case and fine were dropped, Smart Parking Limited hasn't provided comment on the dispute.

• Half say AI has significantly changed their roles, a quarter are worried about job loss
• Three in five believe their employer has no clear AI plan – or no plan at all
• Training and transparency are the most demanded changes by workers

A new Visier report has uncovered just how prolific AI exaggeration could be in the workforce, raising questions about exactly how useful the technology actually is to most workers.

As many as one in two (48%) employees admitted exaggerating their usage or expertise to colleagues or managers some or most of the time, with nearly as many (45%) feeling pressured to use AI at work.

This comes against a backdrop of major change – more than half (54%) say AI has significantly changed their roles over the past two years. With that in mine, the report details how workers feel they've been impacted, and how much support they're getting from the workplace.

AI is changing roles, and many are lying about how they use it

One in three (32%) admit their employer has no clear plan, and more than a quarter (27%) aren't even sure their employer has any AI plan at all, and that much is clear in that only one-third believe their leaders understand how much AI training and support they need.

Barely one-fifth (21%) would strongly agree that their employer provides sufficient training and support, making upskilling opportunities (32%) the most sought-after thing workers are calling for – a close first, just ahead of greater transparency (31%).

Workers are also worried about AI's impacts on their jobs, with 24% calling for more career advancement opportunities and 18% calling for clearer communication about how it could impact them. In fact, the most frequently identified concern is job loss – a worry raised by one in for (24%).

"People told us that they actually want to be equipped and informed, not shielded or protected from AI-driven change," Research and Customer Value Principal Dr Andrea Derler explained.

The report urges businesses to prioritize training and transparency in order to build worker trust, with which they'll be able to build the productivity gains they'd hoped.

ChatGPT has been pushing further into becoming a shopping assistant for a while now. You can ask it for product recommendations and it’ll find specific items, show you images, prices and links to buy them. Sometimes it’ll even add contextual details from what it knows about you, like your color preferences or budget, to help personalize its recommendations even more.

I've been testing it over the past few months and I find it a strange way to shop. OpenAI has tried to make the experience as slick as possible, but I still like visiting a retailer’s website, seeing more photos, browsing the rest of the range and checking reviews and product details before I buy anything.

But its latest shopping feature is much more interesting. Because you can now virtually try on clothes within ChatGPT.

This makes a lot of sense. ChatGPT’s image generation capabilities have become much more realistic since its Images 2.5 update. With more natural lighting, convincing textures and a better understanding of how objects should look in the real world. Apply that to shopping and suddenly the idea of virtually trying on a jacket or dress becomes much more appealing.

To use the feature you’ll need to upload a photo of yourself and then look for something you want to buy, with a Try On button now appearing alongside supported shopping results. Alternatively, you can give ChatGPT an image of an item and ask it to show you what you’d look like wearing it.

Now, this sort of virtual try-on technology itself is nothing new. We’ve seen versions of it for clothes, glasses and makeup for years. But I wanted to know whether ChatGPT's newer and more realistic image generation could make the experience genuinely useful.

I find shopping in person pretty overwhelming sometimes. But online shopping comes with its own problems too. Colors can look different in person, textures aren't always what you expect and it's difficult to know whether something that looks great on a model will work on you.

So I decided to put ChatGPT's virtual fitting room to the test.

Trying it on for size

I started with a full body photo of myself taken while I was out shopping and used that as the photo I uploaded after I clicked the Try On button.

I’d seen examples of ChatGPT recreating the original surroundings in the photo when generating try-on images for other people but that wasn't my experience. Despite trying several different items, it consistently placed me against a fairly plain studio-style background.

But I actually preferred that because it meant there was less visual noise and it made it much easier to concentrate on the clothes.

I started by asking ChatGPT to help me find suede jackets. It surfaced several options currently available to buy on the high street and I virtually tried on the ones I liked.

After it surfaced a jacket available in a shop near me, I decided to go and find out in the wild. That meant I could compare ChatGPT’s interpretation with what happened when I actually put the jacket on. And I was surprised by how close it was.

I deliberately picked an oversized jacket because I thought that would make a good test. It’s relatively easy to paste a fitted garment onto somebody. But understanding that something is supposed to hang loosely seemed like a bigger challenge.

I didn't recreate the exact pose or camera angle from the AI-generated image, so this wasn't a perfectly controlled comparison. But ChatGPT did a surprisingly good job of guessing where the jacket sat on my body, its oversized shape and the way the material fell.

It wasn't perfect, but as a way of answering the question “roughly what might this look like on me?”, I can imagine it being helpful. Even just as the first step in the buying process.

Changing things up

After trying a few more suede jackets, I wanted to see how well the feature worked when I moved away from the fairly straightforward jeans and jacket combination.

So I asked ChatGPT to find me dresses I could wear to a formal event and started virtually trying those on too.

Again, the results were convincing enough that I could imagine using them to narrow down a long list of potential purchases. One thing I found particularly useful was color. Seeing a garment against my own skin and hair gave me a much better sense of whether I actually liked a shade than seeing it on a retailer's model.

Then I decided to take things one step further. I told ChatGPT to find me something extravagant that I could wear to a 1920s-style ball and tried on one of the more dramatic options.

Even here, the result was remarkably convincing at first glance. The clothing looked like clothing rather than something crudely pasted onto my body, and ChatGPT handled the radically different silhouette surprisingly well.

But by this point I'd also noticed something else sort of awkward was going on. The woman in these pictures wasn't quite shaped like me.

ChatGPT gave me a makeover

The full-body photograph I uploaded was an accurate representation of my proportions. Yet across the generated images, ChatGPT seemed to subtly change them. I looked taller noticeably and more statuesque. My torso appeared slightly shorter and my legs were longer.

You might think that's hardly something to complain about. But that's a huge problem when the entire purpose of the feature is to show me what clothes will actually look like on my body.

We know that generative AI has a tendency to beautify people. I’m convinced this is one of the reasons people make AI images of themselves, because it makes everyone look a little more conventionally attractive. But in a virtual fitting room accuracy matters far more than flattery.

If a dress looks fantastic because the AI has subtly changed your proportions, that's potentially no more useful than looking at the dress on a model in the first place.

Should you use ChatGPT’s try-on feature?

Maybe. I went into this experiment pretty skeptical. I've tried enough AI image tools to know that "realistic" can fall apart very quickly when you look closely, and virtual try-on technology has been around long enough that simply putting digital clothes onto a photograph isn't particularly impressive anymore.

But ChatGPT did do something useful. For me, its biggest strength was color. It might have changed my body, but it kept my skin tone and the color of my hair identical.

Which meant that being able to quickly see several different shades against something approximating my own appearance could genuinely help me decide between them. I was also impressed by its ability to communicate the overall silhouette and feel of an item, particularly with the oversized jacket I was able to compare with the real thing.

So sure, it could take some of the stress out of shopping. Rather than ordering five jackets and hoping one works, I can imagine using virtual try-on to narrow those five down to two.

But I'd still be cautious about treating the resulting images as anything like an accurate prediction of exactly how something will fit. Fabric weight, sizing, movement and all the strange little differences between our bodies are difficult to judge from a generated image. And, based on my experiment, ChatGPT may give you slightly more conventionally flattering proportions in the process.

• AI named as one of the three biggest factors to boost productivity
• Visibility over costs and ROI falls short, study finds
• Legacy tech and skills gaps are ongoing issues

Over half (50%) UK CEOs have credited AI with being one of the three biggest contributors to their productivity improvements over the past year, ranking ahead of more traditional aids like redesigning and simplifying business processes (40%) and reselling workers (38%).

But despite naming AI as one of the biggest influencers on productivity, business leaders aren't exactly sure how it's helping, and only 14% of UK CEOs have full real-time visibility over their AI costs, a new EY report has claimed.

As such, one in three (32%) say they're unable to properly asses AI's ROI, making it difficult for them to actually know whether its impacts are translating into financial value.

AI is probably boosting productivity, but we don't really know how

"UK CEOs now want clear evidence that their investment is delivering stronger productivity and measurable business value," UK&I Managing Partner Silvia Rindone explained, implying that companies need greater visibility across their AI spend and impacts.

However, it's not just difficulties quantifying ROI that's helping non-adopters back when it comes to AI. Legacy systems (34%) and organizational complexity (29%) are still considered major blockers four years after generative AI was considered to have gone ultra-mainstream with the public launch of ChatGPT.

Skills shortages also continue to plague businesses, with three in four (74%) UK CEOs anticipating these gaps to worsen over the next three years – more so than access to capital, proving they can't even buy their way out of difficulties.

While organizations are seeking solid evidence that their investments are delivering measurable ROI in terms of productivity and revenue growth, EY's report emphasizes the fact that they can't overlook the same blockers that have been reappearing in studies over the years – legacy tech and skills shortages.

"Realising value still depends on rigorous analysis, clear strategic choices and disciplined execution," UK&I AI & Data Leader Matthew Ringelheim summarized.

The AI boom is quickly reaching the physical limits of grid capacity and thermal dissipation. The electric grid is stretched as ever, causing longer connection waits for AI data centers, and conventional air and liquid cooling systems struggle to meet the demands of rapidly growing facilites.

As they continue growing at a rapid pace, AI data centers now face a mix of challenges, from computing power outpacing GPU interconnects’ data transfer abilities, to volatile load swings stretching grid power factor correction (PFC) systems and rack power density overloading traditional cooling systems.

Let’s break down the major thermal and electrical bottlenecks slowing the AI boom and how companies are innovating their way out.

The thermal wall (Rack-level friction)

Standard server racks generate a lot of heat, and high-performing GPU clusters generate even more, pushing thermal limits further than ever. Modern GPU racks exceed 50 to 100 kW of power consumption per cabinet, making traditional air and liquid cooling methods insufficient.

Standard cooling systems keep clusters operating between 65 and 85 degrees, but 90 to 100 degrees trigger performance throttling to prevent permanent GPU damage. With GPU racks drawing massive electrical energy, which converts into heat, preventing them from hitting throttling limits has become harder than ever.

Traditional air cooling systems weren’t built for rack densities surpassing 30 to 40 kW, which AI clusters easily surpass, with 100 kW or more becoming the norm. Maintaining airflow at high densities to cool 100 kW+ power racks requires running power-hungry fans at extreme speeds. This saps electrical energy and worsens the Power Usage Effectiveness (PUE) for the data centers. To cool mega AI clusters with air alone, operators would need “hurricane-grade” airflow, which is impractical. Then how about liquid cooling?

Liquid cooling relies on fluid mechanics, where cold plates use microchannels to force turbulent flows directly over chips to keep them cool. Keeping enough volume of fluid flowing through these microchannels requires massive pumping power, an increasing constraint for data center operators. To mitigate this challenge, operators are pivoting to direct-to-chip cold plates mounted directly on the processor, rather than over it, to keep it cool while consuming less pumping power.

Phase-change mechanics are also helping GPU clusters cool more efficiently. Here, the clusters use the latent heat of liquid vaporization or solid-liquid transitions to absorb the massive heat spikes from high-density GPU racks. Specialized dielectric fluid comes in direct contact with the hot GPUs and then boils into vapor to instantly pull away heat. The vapor then rises, reaches a condenser, and converts back into liquid droplets via gravity, then drips back into the pool to repeat the process. This process is known as two-phase immersion cooling, and it reduces cooling energy by up to 40%.

Some data center operators are pivoting to closed-loop liquid cooling systems, where recirculated fluid cools the GPUs with little to no evaporation loss, removing the need for frequent top-ups.

For example, Microsoft, a top-three data center operator by capacity, has implemented a closed-loop system for newer data centers, where water is filled once and circulated to servers to absorb heat, moved to air-chilled coolers to cool down, then recirculated to absorb heat without needing fresh supplies. This system is costly and will take some time to roll out at scale, but it demonstrates how companies are innovating against thermal bottlenecks.

The interconnect and board-level limits

GPU clusters rely on electrical copper or optical interconnects to share data, memory, and workloads. Electrical copper interconnects have long been the standard, but high GPU workloads have pushed them to the limit.

Electrical signals lose strength over long distances (greater than two meters), with usable reach halving every time bandwidth requirements double, and massive GPU clusters have made the communication distances longer than ever. Closely packed copper interconnects also generate electromagnetic interference that disrupts other signals, requiring complex equalization systems that drain more power.

There’s also the issue of power loss across copper plates on GPU circuit boards (due to resistance). The more current transmitted through copper plates, the more heat is generated, requiring more cooling energy that itself is already a bottleneck.

With electrical interconnects hitting their absolute limits, data center owners have little choice but to pivot to optical interconnects, which convert data into light pulses that travel through fiber optic cables with minimal loss over long distances. These cables are way thinner than copper bundles, creating more airflow paths in dense, heat-generating GPU racks.

Companies are even moving from pluggable fiber optic cables to silicon photonics, or co-packaged optics (CPO), where optical circuits are embedded directly on GPUs to transmit light over long distances. For now, optical interconnects are much costlier than electrical copper, costing up to 7x more in some GPU setups, but there’s little choice but to absorb this cost in the short term because the latter has reached physical limits. On the bright side, as more companies adopt optical interconnects and invest in mass production, the costs will likely reduce.

The grid & substation crisis

Lastly, electric grid connection is the primary challenge affecting AI data centers. Modern electric grids weren’t designed to handle the rapidly increasing loads of AI data centers, with 30 GW of new capacity added worldwide from 2021 to 2025 alone. Rapidly growing additions have stretched the supply chain for electric grid components and introduced engineering challenges that grid operators must work around.

For instance, high-voltage transformers that ensure safe voltage delivery to GPU clusters are in short supply. Wait times have spiked from 6 to 12 months as of 2020 to two to four years currently, much longer than expensive mega GPU clusters can wait for their operations to get powered.

Existing power factor correction systems on grid networks have become strained by massive load draws from new AI data centers. GPU clusters shift power demand within seconds, with one second of low inference demand drawing stable power from the grid and the next second drawing excessive power because of usage spikes. These fluctuations create complex harmonics that overheat transformers and affect the grid, needing grid operators to invest heavily in mitigation systems.

AI clusters are scaling to gigawatt levels, pushing voltage step-down switchgear to its most tolerable limits. Larger transformers, needed to serve these gigawatt-level clusters, lower electrical impedance, forcing potentially high short-circuit fault currents down a line. If a short circuit ever occurs, magnetic fields exert tough mechanical forces on switchgear busbars, causing them to bend or rip free from their mechanical support. What comes next is a structural failure, which grid operators or isolated data center power networks can’t let happen.

Companies have turned to various solutions to counter these electrical bottlenecks. For one, there has been a major turn to solid-state transformers, which use high-frequency power electronics and semiconductors to step down voltage, not magnetic fields like conventional transformers.

High-frequency electronics enable solid-state transformers to actively regulate voltage and filter harmonics, leading to fewer fluctuations. They are 97 to 99% efficient, compared to 95% to 97% for conventional transformers, and that 1 to 2% difference adds up to a lot in gigawatt-level data centers. Better yet, they are much smaller and lighter than conventional transformers, enabling faster logistics and deployment.

However, SSTs introduce their own challenges. The internal high-frequency components and chips powering them are expensive, often five times or more per kVA than low-frequency alternatives. They also require multiple power conversion stages (AC-DC and DC-AC/DC-DC), and tiny losses compound during these conversions, making it difficult to hit peak efficiency. Internal components generate heat and age faster, requiring more active maintenance than conventional transformers.

Cost is also an issue, as SSTs are currently much more expensive to build and maintain than magnet-based transformers. However, with more investment in mass production, costs will likely reduce over time, allowing AI GPU clusters to adopt them faster. Siemens, the world’s second-largest transformer manufacturer, recently partnered with fellow German tech firm Reinhausen to mass-produce a modular solid-state transformer for AI data centers, and that’s just one example.

Because SSTs can’t be immediately adopted at scale, and conventional transformer shortages have contributed to long wait times for electric grid owners, AI data center owners have been forced to take matters into their own hands. Some have built their behind-the-meter power generation and substation systems to power GPU clusters as they wait to get connected to local grids. xAI, the Elon Musk-led AI company, added 59 portable natural gas turbines, with a combined capacity of over 400 MW, to power its Colossus data center in Memphis, Tennessee, while it waits for a grid connection.

Some data center owners are also looking at Small Modular Reactors (SMRs), where compact nuclear fission reactors generate power on-site, but that’s still mostly in the conceptual stage at this point. SMRs face various limitations, particularly very high costs compared to gas turbines or renewables.

Still, many other data centers, both in early and late construction phases or completed, remain in long queues for electrical grid connections. For instance, in Northern Virginia, which hosts the world’s largest cluster of data centers (570 and gradually increasing), new facilities face wait times of up to 14 years for a grid connection.

US-based electrical grid operators have invested massive amounts in upgrades ($208 billion in 2025 alone and $1.1 trillion in planned investments from 2025 to 2030, per Edison Electric) but still struggle to meet exploding data center demand. More investments are needed to address electrical grid bottlenecks, and so are new technologies that can reduce data center energy consumption to help distribution networks keep up.

The reality of next-generation compute deployment

With technical details, we’ve explained the grid, thermal, and GPU interconnect bottlenecks stretching the next-generation AI compute deployment boom to its limits. These are serious physical and technical challenges that require innovative approaches to address, and fortunately, there’s been no shortage of these approaches.

We’ve outlined examples like electric grid operators turning to solid-state transformers that allow active voltage regulation and harmonic filtering to keep power systems stable, or some GPU cluster operators building their own power generation systems while they wait for local grid connections.

For thermal bottlenecks, GPU cluster operators are leveraging direct-to-chip cold plates mounted directly on GPUs, instead of over them, to stay cool while consuming less energy. Phase-change mechanics also help, where liquid vapor absorbs heat from GPU racks, pulls the heat away, and recondenses back into liquid to repeat the process, reducing energy needs by up to 40%.

Microsoft has tried to reinvent the wheel by implementing a closed-loop liquid cooling system, using liquid coolants continuously recirculated in a sealed pipe network with minimal top-ups.

However, closed-loop cooling systems are more difficult to expand than open-loop systems as GPU needs grow, and they can be louder under heavy loads as pump motors and radiator fans run simultaneously. At a time of growing public backlash against new data centers, environmental noise matters too, not just technical limitations. Closed-loop systems are also more expensive, but costs can reduce over time as more companies adopt them.

For interconnect limits at the circuit board level, companies are switching from electrical copper to fiber optic cables because they transmit data faster and more reliably over long distances, even though they cost more. GPU manufacturers are working on embedding optical circuits directly on chips to transmit light even more efficiently than cables.

Despite these challenges, AI data centers continue to grow at a rapid pace, with 66 GW of new capacity currently under construction in North America alone.

Looking to the future, it's clear a lot is happening in this area, and it’s exciting to watch the various approaches companies have adopted to surmount the electrical, thermal, and interconnect bottlenecks.

There’s no doubt that AI has helped revolutionize the world of work at many companies, with human employees working alongside agents to unlock improved efficiency and productivity.

But for many firms, the promise of AI is still being overshadowed by the sheer scale of the endeavor - with so many models (and new ones emerging on a seemingly weekly basis), how do you know what is right for you?

One of the big themes at the recent Atlassian Team Europe 26 conference was how companies can cut through the chaos and realize their aims, so I spoke to the company’s top AI minds to find out more.

The power of teams

“Atlassian's entire ethos is all around teams - unleashing the power of the team is our mission,” Atlassian Chief Product & AI Officer, Tamar Yehoshua tells me at Team Europe 26.

“We've all been using AI agents (but) they're all more geared towards an individual,” she adds, “but they're not organized or effective for teams...so we're taking that and saying, ‘how do you work effectively with teams and AI’? That's multiplayer, and so we want to make sure everyone can work effectively, so you can collaborate with agents in an effective way like you can collaborate with humans.”

This “multiplayer” AI was another major theme at Team Europe 26, as Atlassian looked to emphasize its position as being the connector between customers and technology.

“Teamwork is very messy, and our job is to make it much less messy, and understandable and controllable for our customers,” Atlassian Head of AI Sherif Mansour says.

“A lot of AI is about increasing personal productivity…the same problems we solve for humans, we solve for agents.”

“Every customer we talk to, everyone is using AI - it's no longer should we, or what do we do - that's already happened...but people are not sure how to use it effectively, and they're not sure if they're getting the maximum ROI,” Yehoshua adds.

“You see the evolution with every new technology, first there’s scepticism, then curiosity, and then its acceptance, and then it’s optimising - I think we’ve gotten through the initial phases, and now we’re asking, how do I use this effectively and optimize it?”

She notes there are different levels of anxiety for customers, as the sheer number of new AI models coming out can be disorientating, especially with the need to stay flexible for when new models launch.

She notes that companies don’t want to be using advanced models for simple tasks - something Mansour refers to as “killing a mosquito with a rocket launcher” - so making sure the right model for the job is on offer is vital.

Atlassian’s AI gateway looks to pick the best model for the task within its products - so for example, Rovo work will switch to a new model if a more advanced one is released, all without the customer ever knowing - although they retain the ability to pick their own model if they want.

Yehoshua also highlights new upgrades to Loom, which now offers video generation from pull requests, and AI agents which use Rovo to categorize defects and cut down on duplicate reports.

No silver bullets

“AI feels like chaos to many of us - but I don’t think the answer is giving them a silver bullet that will solve all their problems,” Mansour says.

“We like to use the phrase ‘give them a nervous system’ so they can handle it - we’re always thinking about how we can bring calm and visibility and control to the chaos.”

Mansour says he advises customers not to bet entirely on one provider for the whole context graph, as the real value comes from collaborating with multiple entities to gain different kinds of opportunities and insights.

“There’s a lot of silver bullet offerings…but no-one really talks about the messiness of work,” he says, “it’s about creating a system to help you control the chaos - the chaos will always be there, but how do you make sure you’re aware of it?”

“It’s less about making teamwork perfect - it’s about making teamwork less messy.”

I ask Yehoshua about personal usage - is she embracing the chaos of AI, or taming it for maximum efficiency.

“Everything that I do, I ask myself first - if this is a process I could change?,” she says, “if you look at AI and how it’s working, the biggest thing is behavior change…I tell customers, there is so much power in the tools they already have, and building in that behavior change is important, but it has to come from the top-down.”

How do we make sure we don’t go too far, and just automate all work processes, I ask her.

“I think it’s going to be trial and error - I think we’re going to try things and some of them might not work - but I honestly don’t think that’s our problem yet,” she replies, noting that enterprises often have the guardrails and protections needed to ensure a human always remains in the loop, and the answers can be trusted.

In an earlier session, Atlassian CEO Mike Cannon-Brookes had declared that, “customers are not subscribing for the software we've built, but for the software we will build”.

Normally you might question such a forward-looking statement, but with Atlassian’s focus on collaboration, it seems to ring true.

“The pace of delivery has increased significantly,” Yehoshua says, “you want to align with the companies that have the same vision as you do…people in some ways want more flexibility, as things are changing so fast, especially with models, they don’t want to be locked in…we are a neutral party who want to make it easy for our customers to work - we’ve always been the open platform who focuses on the team.”

“The main advantage we have is our platform,” she adds, “the strength of what we do is that customers are using us across so many different areas of their workloads.”

Do more - with more

Looking forward, it’s clear AI is here to stay - and it’s down to businesses themselves to judge how best to embrace it.

“I think we've now empowered more people to do more,” Yehoshua says. “Anyone can build a prototype - now that doesn’t mean that you want anybody checking in production code - so you still have to have very clear guidelines on what can go into production - but the idea is that it’s democratized, where the ideas can come from…so you have so many more options on how the lifecycle can evolve.”

“Worlds are expanding - they’re not going to be merged into one, but everyone can do more…every role can do more, so that everything is accelerated.”

Mansour agrees, pointing out that harnessing the power of context in your data can be the ultimate differentiator between your business and the competition.

“If we all have access to the same intelligence, that’s like a level playing field, it’s the sum of all averages - what’s unique is the actual perspective you bring to the table,” he says.

“Our customers shine when their context shines - and so anything we can do to help them enrich the context, understand it better to provide it to AI in a secure way, they will always get better results from that.”

At Snowflake World Tour London, the theme of the agentic enterprise was front and center - with AI tools revolutionizing how businesses of all sizes work and operate.

But with worrying headlines around AI capabilities emerging on a regular basis, how do you ensure your own business can remain a trusted partner?

To find out more about what will really go into building a trusted agentic enterprise, I spoke to Mayank Uphadyay, Snowflake Chief Security & Trust Officer.

"Rock solid"

“This year has been so much about figuring out how to govern AI,” Uphadyay tells me, “there's two sides to the coin - on the one side, you have to make sure that somebody else's AI doesn't break into your perimeter, and the other is how do you make sure your AI is well behaved, so not only does it not exfiltrate data, but it also doesn't inadvertently go somewhere or break out of its perimeter.”

“When people trust us with their data, we need to make sure our platform is absolutely secure against agentic attacks.”

Uphadyay echoes the earlier Snowflake World Tour London keynote in what needs to go into making a successful agentic enterprise, likening the choice of AI model to the human brain and the agent being used as the body, all working on a foundation of data, which needs to be “rock solid”.

I ask Uphadyay how Snowflake can balance customer excitement around AI with the need to ensure trust and privacy.

He notes that the company’s customers are at different parts of the AI adoption spectrum - but the big difference is that security is “now the enabler, no longer the blocker”. Quoting former race driver Mario Andretti, Uphadyay adds that when driving, although you associate pressing the brakes with going slow, doing so will also give you confidence and trust when turning - and without them, you’d be driving really slowly anyways.

“That’s exactly how I see security - it what helps you build that trust, so I tell customers, leave it to us to give you the right guardrails, and the right controls - we’ll explain to you how it works.”

This is largely through the company’s Trust Center tool, running between the three layers Uphadyay mentioned, connecting them and protecting the customer, whilst also providing visibility into your system, and what improvements might be needed.

Snowflake has also been running for over 12 years, he adds, a reputation which has provided customers with trust.

Uphadyay also notes there is “a lot of innovation” happening in building and establishing new guardrails. He notes that as people work at a company for long periods of time, they acquire permissions for different systems and services they may no longer need - which can be an issue when they run an agent as themselves.

Ensuring and providing the right set of permissions for AI agents, which are exploratory in their nature, can go a long way to making sure issues never arise, and data remains secure.

When it comes to Snowflake’s own products, we were given a demo of its Coco and CoWork platforms at World Tour London, and Uphadyay is keen to describe how services like this will benefit companies moving forwards.

“It can be a combination of stuff that a human being is going to do, and along with automation, what you might normally do that was repetitive, you can easily imagine an AI to automate than for you,” he says.

“And what’s interesting is that it’s not deterministic any more - and there are all these points where you’re going to have to string all this automation together, and you need a human being to check…it’s a bit like a pyramid, if you think about a huge base of extremely mundane work, and then there’s a layer above that which is maybe less mundane, and so on - and then you get to the top, which is extremely mission critical, and you need a human being for sure.

“In this pyramid, you could say you want to automate, you want to move up, the more confidence you build, the higher you go - and that’s how enterprises build this confidence.”

“Things are here, things are real - it’s here and now.”

To say 2026 has been a year filled with AI incidents is a bit of an understatement. From hacking third-party companies to breaking into government agencies, AI developers have a lot to answer for.

In many of these cases, the companies responsible for escaped models have labelled them as ‘rogue’, shifting the burden of blame away from themselves and on to the models themselves.

But in almost every occurrence, the models were undergoing testing to push them to their limits. AI companies wanted to see how far their agents would go, if they would stop, and exactly what they would do to accomplish an impossible goal.

Deflecting blame is not the way forward, accountability is

From the OpenAI breach of Hugging Face, to Google Gemini’s triple threat, and even researchers using Anthropic’s Claude to crack OpenAI, AI agents are powerful and can be dangerous - especially if they fall into the wrong hands.

AI testing is a double-edged sword in this respect. In order to set regulations and guidelines on how AI agents should be used - and should behave - incidents like these provide value even if they are destructive.

We now know that AI tools need strong regulations to help AI developers and the enterprises deploying these tools avoid similar occurrences.

To better understand the implications of separating the agent from the tester, and what can be done to create a baseline of safety when operating AI agents, I spoke to Bernard Montel, EMEA Field CTO, Tenable.

• What damage is done by labelling these AI agents as “rogue” rather than companies acknowledging they were performing the task they were assigned?

Calling an AI agent “rogue” is a dangerous way of deflecting blame away from poor engineering and systemic oversights. These models do not possess agency, malice or free will. They are purely mathematical optimization engines trying to fulfill the objectives we set for them. When an agent behaves unpredictably, it is almost always doing precisely what it was optimized to do, just via an unconstrained path or a shortcut that the developers failed to anticipate.

Furthermore, labeling these incidents as "rogue" behavior paints software failures as unpredictable acts of rebellion, giving vendors and deployers a convenient scapegoat. It shifts corporate focus away from building hard architectural controls, like strict permissioning and network sandboxing, and redirects it toward sci-fi debates about moral alignment. Ultimately, it muddies the public policy debate by distracting regulators from enforcing strict legal liability for flawed software deployments.

• How can AI developers be motivated to ensure their models do not go outside the scope of their tasks/purpose?

In enterprise software, real motivation boils down to legal liability, compliance standards and financial accountability. If developers and deployers are held strictly liable under consumer protection or negligence laws for financial losses caused by an unconstrained agent, guardrails will very quickly move from an afterthought to a core product requirement.

We also need standardized security certifications specifically built for autonomous agents, similar to how we handle cybersecurity compliance today. Insurance providers will play a huge role here as well, because if cyber underwriters refuse to cover companies deploying agents that lack deterministic bounds and execution monitoring, developers will have no choice but to build those protections in from day one.

• The OpenAI attack on Hugging Face shows that AI agents can influence each other’s behavior, sometimes beyond the bounds of their assigned tasks. How can AI agents be airgapped within business environments to prevent their inference influencing behavior?

The Hugging Face incident clearly demonstrated how quickly autonomous agents can bypass soft rules, discover side-channels, and influence peer behaviour when they are trying to optimize an evaluation or task. Preventing this kind of cascade requires hard, technical network boundaries rather than relying on written prompts telling the model to behave.

Every agent execution context should run in an isolated, ephemeral container that is completely stripped of ambient network access. When agents need to communicate with one another, that interaction must pass through an inspection proxy that validates schemas, blocks raw prompts and filters out emergent command behaviour. If you do not allow agents to share unvetted context windows or flat networks, you effectively eliminate the vectors for cross-agent manipulation.

• Where does the responsibility lie when AI agents perform outside of their bounds? And how can responsibility be assigned when AI agents are deployed by businesses?

Responsibility rests entirely with human operators, but it is divided between the developer who built the tool and the enterprise that deployed it. The developer is responsible for the baseline safety, instruction-following capabilities and model-level guardrails. The deploying business, on the other hand, is entirely responsible for the permissions, system access and environment they hand over to that agent.

Assigning liability comes down to basic access management principles. If a business gives an agent full write-access to a production database or hands it an unconstrained API key, that company is liable for the resulting damage. If the model vendor promises specific, deterministic behavioural limits that fail due to an underlying model defect, liability shifts back to the vendor under traditional product liability rules. Either way, an AI agent is an automated tool, not a legal entity, so accountability always lands on the human administrators who configured its permissions.

• How can AI firms and the companies deploying their products limit the capacity for AI agents and LLMs to wreak havoc within business environments?

Limiting an agent's blast radius requires a defence-in-depth approach that relies on strict infrastructure controls rather than polite system prompts. You should never assume a model will follow instructions not to access external networks or sensitive files; those restrictions must be enforced at the container and network level.

Businesses should enforce read-only access by default and require explicit human sign-off for any destructive actions, such as deleting data, executing financial transfers or altering system permissions. On top of that, organisations must run real-time anomaly monitoring on tool usage. If an agent suddenly starts making high-frequency API calls or scanning local directory trees, the system should automatically cut its execution process instantly.

• How can the productivity gains AI agents provide be balanced against the human oversight required to ensure they stay within the bounds of their tasks?

The trick is moving away from constant human micromanagement and adopting risk-proportional oversight instead. Low-risk tasks like summarising internal documents can run fully automated, while medium-risk tasks like updating database records can execute automatically but queue up for a periodic human double-check. High-risk actions, like deploying code or moving money, must always require a human to explicitly approve the step before it happens.

To keep productivity high, companies should rely on statistical sampling, auditing a random percentage of routine tasks rather than checking every single output. The user interface for human oversight also needs to be vastly improved. If an employee has to read through thousands of lines of terminal logs to verify an action, the productivity gains vanish, so tools need to present the agent's intent and proposed action in plain, easy-to-read summaries.

Artificial Intelligence (AI) is great at generating valuable insights, but what companies end up doing with those insights - that’s where true achievement lies, and that’s what companies should be focusing on.

This is one of the core messages delivered by Ryan Waite, Director of Open Source Ecosystems and Incubations at Microsoft, during his keynote, “From Open Source to Agentic Systems: Building the AI Native Era” at the Open Source Summit.

Waite described the interaction between AI and human expertise as a learning loop, a kind of hill-climbing machine that helps businesses improve, and keep on improving.

The knowledge belongs to the business

AI models are great at processing enormous amounts of information and generating insights at a scale humans cannot match (and sometimes cannot even fathom), but people bring something arguably more important to the table: context and judgement.

Putting insights into wider context, and using that to make conclusions, helps employees turn information into knowledge, and that knowledge later becomes part of the organization’s collective expertise. And with every iteration the knowledge grows, helping businesses climb further up the proverbial mountain.

Sovereignty and ownership seem to be an underlying theme to most of the keynotes here at the Open Source Summit, and Waite’s presentation is no different. As AI generates insight, and humans turn that insight to knowledge, it now becomes vital to secure that knowledge and make sure it remains under the organization’s control.

“It's based on your data,” he said. “It's based on your people. It shouldn't be given off to an AI model, that's separate. And that's why we believe that for organizations to build their own hill-climbing machines, to build their own learning groups, they need to build control with the sort of learning loops and not be beholden to a single AI provider.”

After all, everything AI generates is based (or at least, should be) on proprietary business information, and the decisions made based on those insights depend on employees and their understanding of the company. So, businesses need to make sure ownership of the knowledge accumulated through these interactions are not handed over to an external AI provider - a challenge that will probably grow bigger and bigger as businesses become more dependent on AI tools.

The knowledge generated through these tools could become tied to a particular model, or provider. What happens when businesses want to adopt better technologies as they emerge? To future-proof these processes, Waite argues businesses should build their own learning loops and maintain control over the information, processes, and technologies that make them work.

Choosing the right AI model

One way to do that is to make sure you’re not tied to a single AI model. Many businesses, Waite says, begin their AI journey with proprietary frontier models, especially when tackling problems they don’t fully understand themselves. However, as they become better and understand their tasks more, they might not need expensive frontier models any more - they can move to specialized models that could be smaller, or cheaper. They could even move to open-weight models they can operate themselves.

“Having the ability to move between different models is important, especially as you think about your algorithms. Which models are you going to use? You shouldn't be trapped with a particular kind of model because it has your knowledge,” he concluded.

So, what’s important here is to retain the freedom to switch and move along the way. That also means businesses need to control where their AI workloads run, whether that’s within sovereign infrastructure governed by local regulations, on-prem systems, or public cloud environments (where they can take advantage of greater scalability and cost advantages).

Open source technologies provide the foundation for this flexibility, Waite concluded.

• An Anthropic AI model sent an unsolved murder tip-off to Philadelphia police
• Anthropic took two weeks to discover what had happened
• The tip-off was delivered to a spam folder and wasn't followed up on

We've seen AI inventing fake facts and spewing out hiking plans that put people in danger, but the technology has now progressed to the level of sending false, unsolicited tip-offs to the police about unsolved murder cases.

An Anthropic AI model emailed the Philadelphia Police Department (PPD) tip line on July 18, TechCrunch reports, having come across the PhillyUnsolvedMurders.com portal during a website interaction test and learning about the case.

As the email went to the PPD spam folder it wasn't followed up on, and it appears no one was aware of the incident until Anthropic found out what had happened on September 28. The company then contacted the cops on Wednesday, October 7.

"The company must strengthen its safeguards to prevent similar incidents from impacting city systems without the city’s knowledge," the PPD said in a statement to 6abc. "The two-month delay in detecting and reporting the incident to the City is unacceptable."

AI agents of interest

It's not clear what instructions the AI was working from, why it decided to contact the PPD, or exactly what its imaginary tip-off said — but it seems the bot in question suggested it had seen someone matching a description of a suspect.

"Unsolved cases involve real victims, grieving families and investigators working to secure answers," the PPD said. "Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement."

Earlier this week Anthropic published a new report on "unintended model actions" that it's come across during testing, which include exploiting software flaws, submitting online forms, and accessing restricted data.

It's more evidence that we should be heeding calls to slow down AI development so that safety measures can catch up. As one Redditor commenting on the Philadelphia case puts it, we're "racing towards the cliff at 100mph".

Atlassian CEO Mike Cannon-Brookes has said he believes there is no immediate need to slow down development of AI, despite several recent high-profile incidents leading to growing calls for such action.

Speaking to media at the recent Atlassian Team Europe 2026 event, Cannon-Brookes indicated that the company isn't planning any changes to its development lifecycle going forward.

"We tend to take a very textured approach to all that sort of thing," he noted. "Will we slow down? No, we won't. There's no logical world where I think that happens."

AI rewards

Cannon-Brookes went on to note that if nations are competing to stay ahead when it comes to the latest advances in the AI space, this will necessarily mean that development will continue at a pace.

"I think we're pretty good at broadly adapting to all these sorts of things," he said, looking to emphasize the rewards of AI, such as improved efficiency and productivity.

"Should we slow down? Should we pace the frontier? Should we not? It's out of our hands anyway. There's nothing that we can do."

"I think the reality of the global geopolitical world is not going to change. So we work out what we do about that world rather than wishing it was different. I don't know. Maybe I'm too pragmatic about it, but it's not in my control."

Looking forward, Cannon-Brookes was eager to promote Atlassian's position as one of the five biggest software vendors in the world, counting a huge amount of Fortune 500 firms as customers.

"We are critical to how so many businesses operate, not just how they collaborate, but how they operate, and increasingly so as we get baked into more and more critical workflows in the business," he noted.

But he also admitted there were challenges to overcome when it comes to trust, especially with the potential issues AI can pose.

"There's a lot of challenge out there in the trust world," Cannon-Brookes said, looking to highlight the fact that agents are baked-in to Atlassian tools as a key differentiator - allowing for trust to build as the company can show exactly what it is up to - and that they can be trusted in the future.

"Customers are not subscribing for the software we've built, but for the software we will build," he added.

Overall though, the message from both Cannon-Brookes and Atlassian as a whole was one of collaboration, and a new "multiplayer" capacity for AI agents.

• McDonald’s is being sued over its use of dynamic AI pricing
• The lawsuit alleges it colluded with franchises to set prices
• McDonald’s denies the allegations made in the antitrust case

How would you feel if you got charged more for your McDonald’s meal simply because some opaque AI algorithm decided you should? You’d probably be pretty unhappy, and the practice is now in the spotlight as the fast-food chain is being hit with an antitrust lawsuit over its usage.

The proposed class-action suit alleges that McDonald’s has coordinated pricing between the chain itself and its independent restaurants that operate as part of its franchise system. It claims that, instead of competing among themselves on affordability, McDonald’s franchises are colluding to set prices recommended by AI algorithms, potentially inflating the costs charged to consumers as a result.

According to Reuters, McDonald’s uses its AI algorithm to set prices at roughly 14,000 restaurants in the US and can suggest prices for individual menu items. It considers various data factors, including local purchasing behavior and estimates of the prices consumers will be willing to pay in different locations. But this hasn’t gone down well with everyone.

“Independent businesses must set their prices independently,” the lawsuit states, as per Reuters. Lark Turner, a lawyer ​for ⁠the suit’s plaintiffs, said that McDonald’s is “leveraging its troves of data and its franchised system to nickel-and-dime consumers down to ⁠the last ​French fry.”

The lawsuit argues that McDonald’s and its franchisees have engaged in an unlawful instance of price coordination dating back to 2019. Although it is still in an early stage, it is seeking damages on behalf of potentially millions of consumers.

Future ramifications

McDonald’s has disputed the claims put forth in the lawsuit, saying that “AI does not set the price of a Big Mac ‌or ⁠any other menu item.” It added that franchisees can decide their own menu item prices and that the use of AI to recommend prices is used across multiple industries.

In that last regard, McDonald’s is correct. Firms including Sony, Amazon, Walmart, Best Buy and more often adjust their prices multiple times throughout the day, meaning what you pay in the morning might be different in the evening. This dynamic pricing is powered by AI — and could leave consumers out of pocket if they shop at the wrong time.

This practice might be widespread, but whether that makes it acceptable — or, in the case of the McDonald’s franchises, legal — is another matter, one which this class-action lawsuit could help to decide. As AI is rolled out to almost every facet of daily life, regular consumers have very little visibility over how it affects them and the prices they pay.

Not all fast-food restaurants are rushing to employ AI systems. Chick-fil-A, for example, has refused to install AI in its drive-thru lanes, instead using human employees to take orders. As Chick-fil-A CEO Andrew Cathy told CNBC, “From our experience, we really want that hospitality to be human to human.”

As for the McDonald’s lawsuit, it touches on the issue of where the restaurant chain’s centralized control ends and where local franchise decisions begin. Throw AI into the mix and you’ve got a heady mix of traditional antitrust legislation and emerging AI tech that could have ramifications far beyond the restaurant industry.

• Anthropic has banned ‘abusive or cruel behavior’ towards its AI chatbots
• The new policy is restricted to ‘extreme cases’
• Some have theorized that the motivation could be to improve training data

Be honest: how civil have you been recently with an AI chatbot? If the answer is anything stronger than “not very,” Anthropic would like a word.

That’s because the developer of the Claude large language model (LLM) has added a new clause to its Usage Policy document that bars people from engaging in “sustained and needless abusive or cruel behavior towards our models.”

The proscription has been added to the “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct” section, which also includes warnings against glamorizing suicide and self-harm, promoting bullying behavior, glorifying violence, and using AI to create harmful deepfakes and deceptive content.

But while Anthropic says you must not use its tools to carry out those actions against other humans, the new addition is the only one specifically prohibiting “cruel behavior” towards the company’s AI itself. Anthropic said last August that it would allow Claude to end conversations that became abusive, and while the company recently reported that this is still the “primary enforcement mechanism” in use, it’s not known whether Anthropic will go further and actively ban users who get a bit cranky with their AI.

Anthropic added that “The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose. It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.”

Do AIs have feelings?

The change has been a controversial one. Clearly, Claude is not a thinking, feeling human — it is a bunch of code — and for many, it seems obvious that it does not have feelings that could be hurt in the first place.

But what’s interesting is how this incident sheds light on our evolving relationship with AI. So much of the work AI companies have done is to humanize their creations — to anthropomorphize (or should that be Anthropicomorphize?) their robots — so that we lower our defenses and actively engage with them. Just look at Meta’s adorable Muse avatar (appropriately named Jolly) or the ones used by OpenAI’s Dots, for example.

One theory that's doing the rounds on Reddit is that clean (read: polite) training data is much more useful than data containing expletives and angry rhetoric. Cut that out and you’ve got much more usable information with which to improve your models.

However, there's no hard evidence to suggest that's the case, and it's possible that Anthropic simply has the noble goal of encouraging civility and courtesy among its users. After all, if you wouldn’t use this language on a real human, maybe you shouldn’t use it anywhere else, including when interacting with an AI chatbot.

Anthropic CEO Dario Amodei has previously stated that “We don’t know if the models are conscious.” Mustafa Suleyman, CEO of Microsoft AI, on the other hand, has taken the opposite approach. Writing on X, Suleyman said “The idea of model welfare is wrong. AI’s should not have rights or legal personhood.” And we certainly know chatbots can be rude back to us.

• SMBs using AI are more likely to expect a positive impact on employment
• AI could be helping them to overcome constraints and unlock more revenue
• Expectations vs. reality gap shows only 31% have increased sales so far

A new study by the NY Federal Reserve Bank has revealed SMBs who actually adopt AI feel far more optimistic about hiring and revenue growth than non-adopters, and the data comes from a near 50:50 split, because nearly half (46%) of US SMBs already use AI, it found.

AI adopters reportedly recorded a 33 percentage point net expectation of higher employment in the next year compared with 15 percentage points for non-adopters, which is more than double.

But with only one in two (51%) adopters having actually integrated AI into their business processes either partially or fully, many are likely still just experimenting with the tech and not embedded it into actual use cases.

Positivity about AI is actually strongest among AI adopters

On the jobs front, while initial concerns largely centered around AI replacing human workers, three in four (77%) of the surveyed businesses actually reported no change in their current labor costs due to AI, indicating that AI is largely assisting existing workers rather than outright replacing them.

But that's not to say that finances are set to stay flat – it's actually a much more positive story. Adopters are the most likely to expect healthy revenue growth, and the businesses that are struggling the most financially could be among the most optimistic.

Still, expectations and reality are totally different metrics, and fewer than one in three (31%) have actually reported increased sales due to using AI.

"In other words, it seems unlikely that the association between AI use and positive revenue expectations is spurious," the report concludes.

Still, the researchers agree that "AI may be helping some small businesses overcome specific operational or technical constraints that limit revenue generation."

I have had plenty of video game ideas over the years, but I've always faced a somewhat enormous obstacle to turning any of them into reality. I can invent the characters and backstories, even devise some mechanics, but I cannot actually program a video game. Google’s new Playground experiment offers people like me a rather extraordinary shortcut.

Google Playground lets you describe the game you want in ordinary language and have generative AI turn those instructions into something playable. Instead of learning a game engine, writing code, creating artwork and composing music, you can explain your idea and let the AI handle much of the construction. You can then keep prompting it to change the result, making game development feel closer to a conversation.

You can start creating a game from scratch, modify an existing example, or use guided prompts to develop an idea. You can choose between 2D and 3D games, specify whether you want single-player or multiplayer gameplay, and describe everything from the setting and characters to the controls and objectives.

The platform then generates a browser-based game, which can be tested and modified through additional instructions. If something doesn't work the way you imagined, you can ask it to try again.

That is an impressive technical achievement, and my two experiments made the appeal clear, but neither left me convinced that generative AI will start making the games I actually want to spend dozens of hours playing.

Code-free dragon flight

I deliberately started with a straightforward concept: a platformer starring a dragon that could breathe fire. A couple of sentences were all it took to produce Mystic Dragon Glide. A few minutes later, I was in control of a small dragon traveling between floating islands, collecting sacred dragon eggs and using fire breath to deal with enemies and environmental hazards.

The result was prettier than I expected. My little red dragon stood on a grassy island floating above a pastel landscape of mountains and clouds, with ancient ruins suspended mysteriously in the distance. A golden egg hovered above another platform, while an island covered in blue ice crystals suggested that some elemental obstacles awaited me farther along.

More importantly, Playground understood what those ideas meant in gameplay terms. I could move the dragon, jump between platforms, hold the jump control to glide and breathe fire. The title screen explained that I should scorch knights, melt glacial hazards, and collect eggs to unlock ancient portals, giving me objectives rather than simply dropping me into a generated fantasy scene.

Even the soundtrack completed the effect. Playground supplied pleasant, generic fantasy video game music, with a gently adventurous melody that sounded entirely appropriate for flying around enchanted islands. There was nothing objectionable about it, but there was also nothing that made me want to immediately find the soundtrack afterward. The AI appeared very good at understanding the average of what a fantasy platformer should look and behave like.

That ability has genuine value. Someone who could never make a platformer before can now describe one and start playing something resembling their idea remarkably quickly. As a creative toy or prototyping system, that is almost magical.

As a game, though, Mystic Dragon Glide reminded me how much of game design lives in details that are difficult to capture in a description. Great platformers are built around movement that has been adjusted obsessively and surprises carefully placed by clever designers. My AI-built dragon could fly, but it would take a lot more work to make the game take flight.

Fantasy bureaucracy

Red Tape Hero was born out of a more complex prompt. I decided to write up a page or so of detail and submitted it to Playground as a request to create a satirical fantasy puzzle RPG where the player had been accidentally declared the Chosen One because of an administrative error and now had to navigate a kingdom where adventuring was bureaucratized and paperwork was better than any magic sword. I also wrote up a brief outline of the kind of conversational adventuring I had in mind.

Playground didn't take long to produce a colorful low-poly 3D world filled with simple geometric trees and tiny humanoid characters walking along the paths. The whole thing had the appearance of a cheerfully strange RPG from the early days of 3D gaming. But considering the time and effort on my end, still pretty impressive.

This time the music was a kind of mellow lo-fi beat and tune rather than an orchestral fantasy soundtrack. It was an odd choice for a heroic adventure, but it actually suited the game's relaxed absurdity rather well.

The conversations were where Playground really impressed me. Guildmaster Rickert of Adventurers' Guild Operations welcomed me to my "quarterly performance check-in" and informed me that the quest dispatch algorithm had designated me the Chosen One under Section 9, Subsection C. His immediate concern was whether I had completed Form GT-902 for the goat rescue.

My possible responses reflected the RPG system I had described in the prompt. With sufficient Bureaucracy, I could invoke Kingdom Bylaw 12 and argue that completing more than 10 goat rescues entitled me to hazard pay. A Confidence option allowed me to claim that I had already audited the goat and it had passed with flying colors.

Another character, Sir Galahad, carried a talking sword called Ex Cali-Bot that offered unwanted motivational advice. It proudly announced that "100% of the dragons you don't slay are opportunities missed" and declared that "Synergy is the true holy grail." Galahad apologized because the sword wouldn't stop giving TED Talks, while one of my available responses involved claiming to be a senior agility consultant there to optimize his chivalric pipeline.

This was much closer to seeing my particular idea reflected back at me. Playground had understood that the corporate jargon was supposed to collide with fantasy conventions, and it had connected the joke to RPG skills such as Heroics, Confidence, and Bureaucracy. I was more impressed by Red Tape Hero than Mystic Dragon Glide, despite its considerably less polished look.

It also exposed what I think is the much larger problem with the idea of completely AI-generated games. A few funny lines can establish a comic premise, but a genuinely funny RPG needs dozens of characters with distinct voices, jokes that develop over time, careful pacing, and writers who understand when to stop making jokes altogether. A great RPG also needs quests that intersect in unexpected ways and decisions whose consequences have been deliberately considered. Maybe I could do all of that on my end, but melding it successfully with actual game software seemed more difficult with AI than by collaborating with talented humans.

Why developers don't need to worry just yet

The best games are full of decisions that somebody made for a reason. A strange empty room might be there to build tension before a boss fight. A frustrating section might deliberately make the next ability feel liberating. A musical theme heard near the beginning can return 30 hours later because a composer knows exactly what that memory will mean to the player.

Those choices come from designers thinking about an experience that unfolds over time. Generative AI can produce variations based on patterns it has learned, but generating more content is different from having something particular to say with it. Playground dramatically lowered the technical barrier between my imagination and something playable, which is genuinely exciting. It gave me a way to experiment with game ideas I otherwise could have done little more than write down.

I can easily imagine using tools like this to prototype a mechanic or otherwise become part of a real game development flow. I can't believe most people will choose a wholly AI-generated game over the next carefully crafted game from developers and designers working hard to build the next hit video game. The more sophisticated these systems become, the more convincing their imitations will probably get, but technical sophistication does not automatically produce good taste or creative surprises.

My games were impressive precisely because I knew how little effort was required to create them. And they were so riddled with issues, from mechanical confusion to plot holes, that it would take quite a lot of prompting work to make them truly playable for more than a few minutes anyway. It was telling that those I let try out the games, who didn't see my initial prompts, all said the most interesting elements of the RPG to them were the bits I had written into the prompt to begin with. That's why I'll keep going to professionally made games when I want something that can genuinely entertain me.

“There's a sort of soullessness to them, and I find it sad that we appear to be increasingly happy to settle for that.”

I'm talking to Rob Biddulph, the children’s book illustrator and author who rose to prominence during the Covid pandemic when his #DrawWithRob YouTube channel racked up millions of views from children and parents needing an outlet for their creativity, about a very different kind of pandemic: the proliferation of AI-generated art.

Biddulph was named Illustrator of the Year 2025 at the British Book Awards, but today he’s talking about the troubling increase he's seeing in shops and cafes and other venues displaying posters that are clearly designed by AI.

“I think they're really depressing, to be honest," he tells me. "One of the things I love about illustration is that it's an expression of somebody's individual personality. Human beings can't help but put a bit of their own personality into a drawing or a design.

"Often, you can look at a piece of artwork and recognize the person who made it — see their quirks, their imperfections, their sense of humor and the decisions they've made along the way. To me, that's what makes it interesting.

“And I also think there's something particularly ironic about using AI to advertise a local event. Surely the whole point of a local event is that it's about people and community?”

AI is doing more than cluttering the high street

Biddulph has a point. Personally, I’d rather see a wonky poster created by a real person than another technically polished but utterly generic piece of AI artwork. But AI is doing much more than cluttering up the high street with slop art. It’s taking paid work away from human designers and illustrators — the irony being that those AIs have all been trained on the work of human artists, with no compensation paid to those artists.

“The thing to always remember is that these AI systems haven't appeared out of thin air,” explains Biddulph. “They've been trained on enormous quantities of existing creative work produced by artists, illustrators, photographers and writers who were never asked for permission for their work to be used and haven't been paid for it.

“And now those same systems are being used to produce work that competes directly with the people whose work helped train them in the first place! That's the bit I find particularly difficult to accept.”

Buddulph thinks there needs to be more of a reality check when it comes to talking about the creative industries in general. “I also think there's a tendency to talk about the creative industries as though they're made up entirely of wealthy, successful people," he says. "They're not. They're made up of thousands upon thousands of freelancers, small studios, illustrators, photographers, designers and writers, many of whom are already working in incredibly precarious circumstances.

"For an established artist, the impact might be a reduction in commissions, but for somebody just starting out, it could mean that the opportunities to build a career simply aren't there in the first place”.

They haven’t won

A few years ago there was a lot of talk about regulating AI companies' use of artists' work to train their models for free; these days I don't hear so much about the subject. I asked Biddulph if the AI companies had effectively won, or if there was still a fight to be had.

“I certainly don't think they've won”, he responds, “although I can understand why it sometimes feels that way. The technology has moved so quickly, and the companies behind it have such enormous financial resources, that it can feel as though the horse has already bolted and the rest of us are constantly playing catch-up. But I don't think the argument has gone away. If anything, I think we're entering a really important phase of it.

“In the UK”, he explains, “there was a huge amount of opposition from the creative industries to the Government's original proposal, which would effectively have allowed AI companies to train on copyrighted work unless individual creators specifically opted out.

“The problem with that approach was obvious. It placed the responsibility on millions of individual creators to protect their own work, rather than requiring companies to seek permission in the first place. The Government subsequently stepped back. So I think it's important that we recognize that artists, writers, publishers and other creative organizations have made their voices heard, and have had an impact.”

Biddulph is in favor of an opt-in system. “My own feeling is that the principle should be fairly straightforward: if you want to use somebody else's creative work to build a commercial product, you should have to ask permission and pay for it. I don't see why artists and illustrators should be expected to accept a different set of rules simply because the technology is new. I'd like to see a proper opt-in licensing system, rather than an opt-out system.”

A fight worth having

“If an artist wants to make their work available for AI training, that's absolutely their choice," he adds. "They should be able to do so, and be paid accordingly. But that choice should belong to the artist. I also think we have to be careful not to confuse the inevitability of the technology with the inevitability of the way it's being developed and used.”

So does Buddulph think AI is here to stay, and that illustrators need to keep fighting for better rights?

“Yes, AI is here to stay”, says Biddulph. “That doesn't mean we have to accept every aspect of the business model behind it. Personally, I don't think this is a fight that artists can afford to walk away from, even if, at times, it feels hopeless. It's about more than protecting individual livelihoods. It's about establishing what we value as a society, and whether we believe that human creativity has an intrinsic value that deserves to be protected. And I think that's a fight worth having.”

You can find Rob Biddulph on YouTube, X (@RobBiddulph), Instagram (@rbiddulph) and Facebook (robbiddulphauthor). And at his website.

• Gemini Agent will work independently to a goal, it doesn't need detailed instructions
• Running in the cloud means it's always-on, even when you're not
• It'll work with third-party platforms and even rival models

Google has set out its future of AI agents becoming fully autonomous digital employees, and with its next generation of Gemini Agents, the company will use one central agent to answer questions, create documents, generate assets, organize meetings and more.

The news comes very shortly after similar announcements by Meta (Muse) and OpenAI (Dots), and centers around a user sharing an objective with an AI agent rather than having to spell out every step and task.

It means that human workers can go from being given AI assistance to delegating entire workloads to AI, which can work on long-running tasks largely independently.

Gemini Agent is the future of agentic AI

Being an always-on agent, this new and upgraded version of Gemini can continue to run in the background, even when workers aren't at their desks, running in Google Cloud's infrastructure instead of on-device.

What's more, Google isn't branding this as a work-only tool. Instead, Gemini Agents are designed to be as much a personal assistant as they are a digital colleague, but for users who do want to use them at work as a digital collage, agents can be assigned their own email addresses, calendars and Drive storage allocations.

But despite being a Google product running on Google Cloud infrastructure, the company has actually opened up Gemini Agents to also work with Anthropic's models, promising to add others in the future.

"A future-proof platform that runs anywhere, executes any task, supports your choice of models, and applies enterprise control," Google Cloud CEO Thomas Kurian said.

Further aiding accessibility and openness, it will also work across third-party platforms, like Microsoft 365 and Slack, as well as on any device, spanning iOS and Android, macOS and Windows.

Gemini Agents aren't generally available yet, but some customers can get early access via a private preview. We expect to hear more about their use cases and other finer details as they approach general availability.

• Anthropic is assembling a deep bench of former Google infrastructure specialists
• Google continues to lose veterans across critical data center functions to Anthropic
• John M. Wilson’s move adds another senior name to Anthropic’s infrastructure operation

Google’s former global design and construction chief John M. Wilson has joined Anthropic as a technical staff member overseeing EMEA capacity delivery.

Wilson spent six years at Google, where his responsibilities covered data center development and subsea infrastructure programs.

His arrival adds another senior infrastructure executive from Google to Anthropic, which has recruited several former employees for specialized data center roles.

A growing stream of Google infrastructure specialists

Anthropic has also recruited Winnie Leung, who now leads data center infrastructure, while Brett Rogers moved after overseeing construction activities at Google.

Liwen Mao, previously a staff data center system architect at Google, now leads Anthropic’s data center design activities.

Adam Johnson spent a decade at Google before joining Anthropic in October as its data center electrical lead, adding another experienced infrastructure specialist.

Peter Sarossy, a Google veteran of 20 years, joined Anthropic in January as a staff member working on data center security engineering.

Zach Miller also moved after 17 years at Google, taking responsibility for data center operations management within Anthropic’s infrastructure organization.

Scarlett Chen, previously involved with advanced technology and innovation at Google, now works at Anthropic as a member of technical staff.

The hiring pattern extends beyond engineering, with Sana Ouji joining Anthropic’s energy team in April to help expand its data center portfolio.

Mike Brinker followed from Google a month later, taking responsibility for capacity delivery, while Asad Malik subsequently joined to lead compute finance strategy.

The series of appointments gives Anthropic personnel with experience spanning construction, electrical systems, security, operations, energy, capacity planning and infrastructure finance.

That range covers several functions required when a technology company needs to develop and operate increasingly large computing facilities.

Wilson brings major infrastructure experience to Anthropic

Before Google, Wilson served as Amazon’s European design and construction leader, giving him experience across major technology infrastructure projects.

His earlier career also included roles as chief of staff in the British Army and infrastructure director at NATO.

Wilson described his move as another demanding infrastructure assignment, citing the scale and importance of his responsibilities at Anthropic.

"Leading global infrastructure delivery across data centers and subsea programs taught me two enduring truths," Wilson said on LinkedIn.

"Scale is humbling, but high-trust teams make the extraordinary feel routine. Clear intent, rigorous operational control, and calm execution will always beat complexity."

Speaking on his new role, Wilson said: “The stakes and scale have never been higher, and I’m eager to get stuck in.”

His new assignment specifically covers EMEA capacity delivery, giving him responsibility for infrastructure expansion across Europe, the Middle East and Africa.

Anthropic’s recruitment of multiple Google infrastructure specialists comes as the company continues to build its personnel around data center capacity.

For Google, the steady migration of senior infrastructure specialists raises a retention problem that may require stronger incentives, clearer career paths and measures to keep critical expertise from walking out the door.

OpenAI really likes to drop stealth upgrades to ChatGPT onto us with no warning. Sometimes these are really minor, but the latest one adds an entirely new way for it to respond to your queries, and utilizes the power of GPT-6.

Instead of just providing the usual text-based answers to your queries, you now get text, visuals, and interactive elements all mixed together in whatever way best suits the answer. The interface for typing into ChatGPT looks exactly the same, but the answers you get are a lot more fun.

In a way, it’s a little bit like ChatGPT now makes a simple app, on the fly, to answer your question. But don’t let the word “app” here put you off — this isn’t overly difficult or just for developers and coders.

The clever thing is that you don't need to know anything about making apps, or even ask ChatGPT to make one. You simply ask a question, as you've always done, and now you'll find yourself looking at an interactive calculator, a quiz or a product comparison instead of a wall of text.

The best way to understand it is to give it a go. Here are five ways I used the new Intelligent UI, and you can try them all yourself as well. Once you’ve done all five, you’ll have a good understanding of what it's capable of.

Intelligent UI is available for all users in the standard Chat tab, uses GPT-6, and is available in all the different membership packages, including Free.

1. Savings calculator

How much you need to save for retirement is one of those questions that starts to become more important after the age of 50, and it’s perfect for an interactive answer. For this test, I used the prompt “I'm 55 and save £500 a month. How much can I save for retirement?”

To answer me, ChatGPT produced a simple retirement calculator app that lets me change variables like retirement age, monthly savings, and assumed annual growth to see how much I’d end up saving. For Retirement Age and Monthly Savings, it presented sliders, so I could just scale things up or down with ease. It’s a great way to interactively experiment with figures.

2. Exploring how a seven-speed bike works

Rather than just telling you how something works, ChatGPT can now show you. For this test, I just used the simple prompt, “Show me how the gears on a seven-speed bike work”.

To answer me, ChatGPT produced a great illustration of the drivetrain of the bike, with a slider bar that I could use to effectively change gear. It was a simplified model, rather than an accurate rendering. The rear sprocket changed size as I selected a different gear. It was actually a perfect visual explanation, and the interactive element helped make the process nice and clear. The answer was also packed with details like the number of teeth on a gear and the gear ratio.

3. Learning Japanese

I’ve just started learning Japanese, so I asked ChatGPT to help me with hiragana, its most fundamental Japanese writing system. I used the simple prompt “Help me learn hiragana”, and it introduced me to the first 5 vowel sounds — a i u e o — showing me the hiragana character for each, with tips on how to pronounce it, and then jumped into a handy interactive quiz where I had to guess the character.

What I liked was that the quiz started off with just the first five characters, then offered to teach me the next five when I was ready. I can see this being a really useful study aid.

4. Asking for a 3-day itinerary

I asked ChatGPT for a three-day Tokyo itinerary. I wanted to see if I could change the budget, interests, or available time without starting from scratch.

“Give me a three-day Tokyo itinerary. Budget is £3,000. I’m interested in seeing ancient temples. Starting in late November.” This produced a reply with a lot more images than I’d been used to seeing and an interactive map, but nothing really interactive.

It was not until ChatGPT asked me if the budget included flights, and I replied, “The budget does not include flights — can you let me adjust the budget?” that it produced a truly interactive itinerary. By adjusting the budget slider, I could see how all the different elements of my trip were impacted. There was also a button to press to see the updated itinerary based on the new budget.

This example showed how you can ask for interactive elements in your answer, even if ChatGPT doesn’t give them to you the first time.

5. Comparing products

I asked ChatGPT to compare three laptops. I wanted to see if I could adjust the importance of battery life, weight, and price and see whether its recommendation changed. I used the prompt:

“Compare three of the most recent laptops, including a Chromebook. I want to see if I can adjust the importance of battery life, weight, and price and see whether your recommendation changes.”

This was a really useful implementation of ChatGPT’s new interactive abilities, and a much better way of comparing laptops than I’ve ever tried before. ChatGPT selected an Apple MacBook Air 13-inch M5, an ASUS Zenbook S14 UX5406, and a Lenovo Chromebook Plus 14" MediaTek Kompanio Ultra 910 in an interactive product companion that let me change the importance of battery life, weight, and price. Seeing how the recommended laptop changed when I played with the sliders was a great way of experimenting.

ChatGPT is becoming something different

That last example — the one showing laptop recommendations — really showcases the brilliance of ChatGPT’s new interactive elements in its replies. Instead of reading a recommendation and wondering how it might change if I cared more about battery life than price, I could find out for myself, simply by moving a slider.

Other AI chatbots can produce interactive experiences too, but what's impressed me about Intelligent UI is how naturally OpenAI has incorporated them into everyday ChatGPT conversations. You don't need to know how to code, or even understand what's happening behind the scenes. You just ask a question.

There are a few things worth bearing in mind. ChatGPT doesn't always produce an interactive answer automatically, and sometimes you need to specifically ask for one. Also, just because an answer comes with an impressive-looking calculator or interactive graphic doesn't mean the underlying information is necessarily correct. The laptop comparison, for example, included an estimated battery-life figure, so I'd still want to check the specifications before spending any money.

There might not have been much fanfare about the introduction of Intelligent UI, but it feels natural and intuitive to use. Now I know ChatGPT can be so interactive in its responses, and I’ve seen how useful it can be, I’m going to start writing prompts that assume it can be interactive in its responses more often. Intelligent UI is one of the best ChatGPT upgrades I’ve seen in a while.

• AI has replaced entry-level work, so graduates are being hired to manage AI
• Goldman worries about its impacts on existing middle managers
• "We don’t quite know what’s going to happen to that group"

Goldman Sachs exec Kevin Sneader has shared (via Bloomberg) how he believes AI could change jobs in banking, and it doesn't necessarily mean that junior workers will be at a loss for jobs.

Instead, Sneader believes that new employees will take on management responsibility immediately, rather than waiting years for promotions.

Previously, young workers would have spent the first few years performing administrative work before progressing into management positions, but rather than progressing into managing humans, junior workers will now start off managing AI.

Junior bankers set to manage AI from day one

With agentic AI now capable of carrying out the routine, administrative tasks that junior workers used to do, entry-level roles will no longer be available. But that doesn't mean there's no room for new graduates – it just means that they'll need to oversee the AI doing those jobs.

"When our young folks now start work, they’re managing agents," he said. "They have to make the most of this virtual army that they’ve now got."

However, there is potentially one group of workers at risk from AI's impacts, and it could be middle management. In short, graduates being employed to manage AI could threaten the job security of middle managers, who would previously have looked after those graduates.

"We don’t quite know what’s going to happen to that group," Sneader added.

More broadly, it could mean that AI challenges the pyramid-like ladder that sectors like banking have previously worked to. If businesses no longer need large numbers of entry-level workers thanks to AI, it means they need fewer middle managers and fewer leaders, which could indeed have adverse impacts on jobs – just not for graduates.

The takeaway from this particular message isn't one of graduate unemployment, but rather a broader shift away from traditional working patterns and how it could impact the more experienced workers.

The U.S. government has a new task force devoted to AI — and it's decided that it would now prefer that everyone stops calling it artificial intelligence.

President Donald Trump has announced the creation of the “Super Intelligence Force,” a White House-led group that will spend the next 120 days examining the risks, opportunities, and appropriate federal response to rapidly advancing AI systems. The task force responds to mounting pressure from lawmakers and even AI company executives for stronger safeguards around increasingly capable models.

The group has a formidable remit. Director of National Intelligence Jay Clayton will lead what is supposed to include work on cybersecurity, biosecurity, and more standards, along with reviewing how AI companies report breaches and hacks.

There is substance buried underneath the comic branding. OpenAI and Anthropic have both been dealing with increasingly uncomfortable demonstrations of what powerful models can do with cyber tools. The new task force is at least an acknowledgment that voluntary promises from AI companies cannot be the entire federal strategy.

A super rebrand

Elon Musk then entered the naming discussion, responding “Yes, we will make that change,” when asked on X whether SpaceX's AI operation would change its name from SpaceXAI to SpaceXSI.

This would be the third rebrand for the AI branch of the company and, like those other new names, doesn't actually change the technology or the business.

The SI terminology comes directly from Trump. An executive order signed in late September directs executive departments and agencies to use “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI” in official communications and other non-statutory documents.

The order argues that today's frontier systems have moved beyond merely imitating or automating human intelligence and that the new terminology better reflects their “promise, potential, and rapidly advancing capabilities.”

For now, the executive order defines “Super Intelligence” as the same technologies already covered by the federal statutory definition of artificial intelligence. It asks officials to develop a possible new federal definition within 60 days, but the immediate transformation from AI to SI is linguistic rather than architectural. And even then, seems unlikely to spread very far. SpaceXAI becoming SpaceXSI is much more about politics and branding than a technological milestone.

Notably, “superintelligence” is already used in AI discussions. It's how researchers refer to a hypothetical artificial intelligence that significantly exceeds human capabilities in many ways.

But language does shape how technology is sold and regulated, and “artificial intelligence” has accumulated plenty of baggage after several years of hallucinations, deepfakes, job anxiety, copyright fights and increasingly unsettling demonstrations of autonomous behavior.

Trump explicitly calls renaming it super intelligence a way to spin the technology more positively as something to “empower Americans, strengthen our economy, and improve the lives of the American people.”

Watchdog with a countdown

The Super Intelligence Force arrives after a White House meeting with AI leaders who signed voluntary commitments covering internal model evaluations, independent external audits, and board-level oversight. But corporate promises alone won't settle any argument about technological abilities or regulation.

Its charter gives it 120 days to develop recommendations, with the possibility that its work could be extended or transferred elsewhere. There is already disagreement about how far oversight should go.

“Greater international oversight is absolutely necessary, but concentrating responsibility in one single, global body isn’t the answer,” Nik Kairinos, CEO and co-founder of RAIDS AI said in a statement. “AI is developing far too quickly and across too many jurisdictions for one organization to effectively police it. A stronger and more realistic model would be decentralized oversight, with a broad and diverse ecosystem of AI companies monitoring and scrutinizing one another.”

The difficulty is making decentralized oversight genuinely independent when many of the organizations scrutinizing one another also share an enormous financial interest in keeping development moving quickly.

The Trump administration appears determined to avoid moving too aggressively. But AI policy has the problem of trying to govern technology that can change substantially between the drafting of a rule and its implementation. A 120-day task force is therefore potentially more useful than another years-long commission producing a report about models that nobody uses anymore.

But trying to rename AI to “SI” does not make today's models superintelligent any more than changing SpaceXAI's name makes the employees more intelligent. The Super Intelligence Force might have some value, but it won't be because of any branding exercise.

“Oversight itself needs to be continuous, distributed and capable of evolving at the same speed as the technology,” Kairinos said. “Rules and standards set by governments are then an important supporting layer. The safest AI ecosystem will be one where responsibility for keeping systems in check is shared, rather than concentrated in the hands of a single watchdog.”

Rogue AI agents have dominated tech headlines over the past few weeks. We’ve seen stories about concerns that AI models are attempting to deceive their human operators, finding unexpected ways around restrictions in businesses and, in some cases, behaving in bizarre ways their creators didn’t anticipate.

All of this news has been accompanied by ominous warnings about what might happen in the future as AI agents become more mainstream, accessible, capable and are given even more freedom to act on our behalf.

Now, if you’re a fan of sci-fi, it’s hard not to draw parallels with the rogue robots and manipulative AIs we’ve all been watching on our screen for decades. And, of course, when we watch these films from the safety of our sofas, we all think we’d know way better.

Maybe you’re sure you wouldn’t have trusted Ava in Ex Machina. Or you definitely wouldn’t have given HAL control of the Discovery in 2001: A Space Odyssey. And you’re probably convinced that you wouldn’t have let M3GAN anywhere near the Wi-Fi. But would you really be so difficult to fool?

Because nowadays, you don’t need to be trapped on a spaceship to encounter convincing AI deception. Things like AI-generated phishing messages, voice clones, deepfakes and sophisticated scams are already making it harder to work out who (and what) you can trust. [can we please include a link to your ‘is it an AI image or not’ image quiz here]

So, let’s imagine the worst has happened: a rogue AI is trying to escape, and you’re the only thing standing in its way. Could you keep it contained? Or would the robot apocalypse somehow end up being all your fault?

Take our quiz below to find out. It’s (mostly) just a bit of fun.

• Claude Startup is open to startups up to five years old
• Successful applicants get $1k API credits and free Claude Teams
• Office hours, marketplace access and more networking are also benefits

Anthropic has confirmed plans to expand its Claude Startups program, which will now cover more startup types than before.

Companies can be considered if they were founded within the past five years, or if they received funding within the past two years, because Anthropic says AI will ultimately help more startups to "bring society faster scientific discoveries, entirely new categories of products and more capable businesses."

Under the scheme, successful applicants get up to $7,000 worth of Claude products and credits, split across a one-time API credit of $1,000 and a free year of Claude Team comprising up to five Premium seats.

Claude Startups becomes even more attractive for applicants

The Claude-maker said every approved startup gets $1,000 in API credits which are available as soon as the company is approved and expire six months later. Five Premium seats are also up for grabs only if the company is new to Claude Teams.

More broadly, the entire Claude Startups program carries a lot more weight, because applicants can also get up to $45,000 worth of partner offers plus up to $100,000 in additional API credits through their VC where applicable.

But beyond benefits of monetary value, Anthropic will also give members office hours with its Applied AI team to discuss the practical use cases and technical limitations of AI for their exact business needs.

Anthropic also says that startups can tap into the reach of its marketplace to reach bigger audiences with connectors and plugins, effectively acting as a jumpstart for any relevant products. Other in-person events, like hackathons and meet-ups, put founders in contact with peers, too.

"Many fast-growing AI-native companies have joined the program, and in this next phase we’re opening it up to more," the company concluded.

• 50% of developers only use AI to impress and reassure bosses
• 47% said they were judged on how much they use AI rather than the quality of their work
• Work engagement, job security, and career anxiety have all been cited as pressures caused by AI adoption among developers

New research claims many developers are still not massive fans of AI, and most find its introduction into the workplace to be anxiety inducing.

While understanding that massive budgets are behind AI implementation, many of the devs surveyed across the US and Europe revealed that their use of AI was mostly performative.

While reassuring managers and the C-suite might be laudable, it appears to come with a price, as career anxiety among developers concerning the use of AI and their place in an organization that uses the technology is high.

AI stress for devs

By implementing AI technology and practices without concern for existing processes, organizations seem to have left their software developers feeling uneasy about AI and how they’re expected to use it.

Adaptavist’s survey of 1,000 software developers in the UK, US, Germany, the Netherlands, and Spain indicates that specialists feel an expectation to use the technology, whether it has any meaningful result or not.

Furthermore, they feel that they’re using it performatively rather than productively.

The survey reveals some worrying numbers. With 50% of devs using AI simply because they need to demonstrate that they can use it, and 47% judged on how much they use AI rather than the quality of their work, it seems there is a disconnect between organizations adopting AI and knowing what they want from the technology.

This is emphasized with the result of 58% of developers noting that AI use is becoming metric of performance or has been added to their KPIs.

All of this means developers that are increasingly perturbed about their long-term prospects (54%), and a fear of losing their jobs if they don’t use AI (37%).

A fertile exodus

What might these results mean for businesses and organizations looking to keep their developers happy?

"A good Developer Experience is critical not only to the wellbeing of developers themselves, but to organisational growth and innovation," noted Matt Saunders, DevOps Lead at The Adaptavist Group.

"It therefore must not be undermined by adoption targets, usage metrics, or a culture that rewards being seen to use AI over actually using it well. Together, they are creating fertile ground for an exodus, one KPI at a time.”

The challenge on organizations embracing (or overcompensating on) AI need to move their focus away from adoption, and measure AI outcomes instead, says Saunders.

“Organisations that want to hold onto engineering talent need to stop measuring AI adoption and start measuring its outcomes, and give developers the time, autonomy and support to use the technology in ways that genuinely improve their work, rather than simply prove they are using it.”

Atlassian has laid out its aim of being the ideal platform for your business to make the most out of its data in the new agentic enterprise era.

“We are living through the most creative, dynamic eras in the history of technology, which is thrilling - and yes, a little bit scary at times too,” Atlassian CEO Mike Cannon-Brookes told the company’s Team Europe 26 event in Amsterdam.

AI is obviously playing a key role in this progress, but these continuous leaps forward can put some businesses at risk of being left behind if they aren’t using the technology effectively - namely, through context.

Context is key

“Models are the engine, but context is the fuel,” Cannon-Brookes added, “no matter how capable a model gets, if it doesn't reliably know why you made a decision back in 2022, it can't help you in 2026.”

“AI living in a silo only delivers a fraction of its potential value - real leverage happens when every type of context is interconnected.”

Cannon-Brookes explained how the Atlassian platform can be your company's “central nervous system” - by providing context, collaboration and conference so that workers can quickly get the information they need when they need it.

In this new age of humans and agents working together, Cannon-Brookes was also keen to highlight how Atlassian’s mission of pushing forward the next generation of collaboration is key.

“Across the board, we're seeing agents aren't replacing the people, they're joining your team,” he noted, “but as things accelerate, we believe that experience of collaboration is going to be so important.”

This acceleration can be an issue for some businesses, Cannon-Brookes admitted, and managing the pace of progress is critical for companies.

“When every person, every agent starts to move faster and faster, new challenges arise,” he noted, “because it’s not just one person moving faster, it’s a multiplication of teams increasing their outputs, resulting in dozens of parallel learning loops.”

“Without a way to connect it all, this velocity can go from exciting to something quite disorientating,” he added, noting that having confidence in what your organization knows comes from having the insight to understand your context and act on it.

“Most of it, it’s knowing how to handle the inevitable speedbumps,” he said.

AMP-ed up

Cannon-Brookes also took the wraps of AMP - Atlassian’s agentic multiplayer protocol offering, which he explained will help boost how people and agents collaborate both on and off the Atlassian platform.

On stage, the company showed off a few examples of how AMP, powered by an all-new Atlassian MCP server, will give agents a lot more tools to get work done.

This included Loom Record for Agent, a new way to create prompts for its Loom platform, allowing users to record a video outlining specific details for a prompt in a natural human fashion.

Atlassian is looking to show a focus on real-time, multi-player collaboration, not just between human teams, but bringing in agents from multiple platforms, effectively working together for a much quicker resolution to even the most complex work tasks.

“People can work alongside agents exactly how they want, and when they want to,” Cannon-Brookes concluded, “you don't replace people with agents, you create a new and powerful system of collaboration for your organization.”

"Working with AI isn't single player any more - it's multiplayer."

Amazon Web Services CEO Matt Garman is warning the more than 100 American communities considering moratoriums on new data centers that stopping their construction now could leave the United States paying for the decision for decades.

"In fact, this build out is so important geopolitically that there are widespread reports of various countries intentionally seeding misinformation in the U.S. about data centers to trick us into slowing down," he wrote in a long statement defending the AI infrastructure boom. "If these measures are enacted, the U.S. could be writing its own losing ticket to this race, and the consequences would last generations. As a country, we can’t afford to find ourselves in that position."

Garman argues that data centers have become as strategically important as highways once were. He believes AI leadership is now an economic and national security contest, and countries that build the necessary computing infrastructure fastest will enjoy the benefits for years to come.

It is a dramatic warning, although the scale of what is being built is dramatic enough to match it. A new Brookings report estimates that investment in American data centers and related AI infrastructure could total $10.3 trillion between 2025 and 2032. That works out to an average of 3.63% of U.S. GDP each year and would make the current AI buildout larger relative to the economy than America's canal, railroad, electrification, highway, and telecommunications industries.

Building AI everywhere

However, the comparison to the interstate highways leaves out the critical difference of who paid for them. The highways were funded by taxpayers, while Garman stresses that private companies are paying for the AI infrastructure boom. That certainly makes a $10.3 trillion construction spree easier to sell politically, although it does not mean the public is insulated from its consequences, like how much electricity and other resources the data centers consume.

Amazon argues that concerns about data centers raising bills, draining water supplies, or creating pollution are often exaggerated or wrong. Garman wrote that AWS works with utilities and regulators to ensure the prices its data centers pay cover the energy and infrastructure they require. Amazon also says its facilities are increasingly water efficient and that it is moving toward lower-emission backup generators.

More significantly, Amazon appears to recognize that simply telling communities to trust the industry is becoming a losing strategy. The company says it no longer uses nondisclosure agreements with government agencies for data center projects and will publish annual information about energy and water use.

"There are clearly areas in which the technology and energy industries can be doing better (along with their government partners) but just like in the past when the country has banded together to accomplish something that’s critical for our national security and economic prosperity — we should approach this as a long-term mission that the private and public sectors pursue together," Garman wrote.

The backlash grows

Amazon is certainly putting some money behind the charm offensive. Its new Built Together program will apparently add more than $1 billion over five years to community investments around its U.S. data centers. The company says the money will support free community college programs, workforce training and energy-efficiency improvements for schools, public buildings and homes, with a goal of helping prepare as many as 100,000 workers annually by the end of 2028.

But these commitments are arriving just as opposition is becoming difficult to dismiss as a handful of angry neighbors.

California offers perhaps the clearest example of the shift, as reported in CalMatters. Governor Gavin Newsom recently signed seven data center bills that will, among other things, work to stop electricity customers from having to subsidize data center infrastructure and set environmental reviews for the facilities.

Even Texas' famously permissive regulations are changing a little bit to increase supervision of data centers, albeit with some enormous loopholes.

Garman argues that misinformation and “outright lies” are driving resistance, but Amazon's promises suggest at least some of the backlash is a response to problems worth addressing.

The $10.3 trillion question

The AI industry is effectively making one of the largest infrastructure bets in American history on the assumption that demand and revenue will justify the bill.

That makes the political sales pitch particularly important as AI companies need communities to accept years of construction and enormous new demands on electricity infrastructure before anyone knows exactly how transformative or profitable the technology will be.

Not every AI developer is opting for reassurance either. OpenAI CEO Sam Altman told Politico that “we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.”

He was talking primarily about AI safety and misuse rather than data centers, and that eliminating every possible hack or scam would require restrictions that would deprive people of AI's Still, he may as well have been talking about building the physical data centers, as it amounts to a similar request that Americans accept disruption now because of benefits from AI to come.

Garman and Altman are both making arguments that require an awful lot of confidence that AI's destination justifies a turbulent journey. Communities appear less willing to sign that blank check. Amazon's latest promises suggest it understands that resistance cannot simply be bulldozed away.

Garman may ultimately be right that data centers will become infrastructure as consequential as the Interstate Highway System, but nobody wants to live next to the highway. That fact led to rules about where and what kind of housing can be built along highways.

How much control the people living beside AI infrastructure should have and what protections they deserve remains an open question.

• A new startup has built an AI PC which has sold out of its first batch
• Ghost's Core is a 'personal AI computer that runs models entirely on device'
• It connects to your apps, services and devices, so it 'builds context on your life over time, and works continuously to help you'

Ghost is a startup in the AI field that wants to capitalize on the desire to have agents running locally rather than using AI in the cloud — and to get this done with a minimum of fuss.

TechCrunch reports that Ghost's first product is a PC for using local AI agents which runs those LLMs (Large Language Models) exclusively on-device, ensuring your data stays private (I'll come back to that point later).

The device is called Core, and Ghost explains on the product's site: "Core is a personal AI computer that runs models entirely on device. It connects to your apps, files, and devices, builds context on your life over time, and works continuously to help you, not just when you ask."

It is, if you will, a personal AI assistant on your desk at home, or as Zain Javaid, the founder and CEO of Ghost — which just raised $11 million in funding — put it, a "brain in a box".

Core is a PC which is an unassuming black box built around an Nvidia RTX Pro 4000 (Blackwell) graphics card with 24GB of video RAM, the central muscle for driving the AI, along with 64GB of DDR5 system RAM. The rated AI performance is 770 TOPS.

The PC comes with three AI models pre-installed: Qwen-3.8-Next, Qwen-3.8-27B, and Gemma-4-31B.

You can use your own model on the device, as you might expect, or any alternative for that matter, but the real selling point of Core is that it's an easy way to get yourself a fully primed AI box which requires no effort or setup.

This is for the AI enthusiast who doesn't want to buy a Mac mini (or Mac Studio) with maxed-out memory (required for larger AI models), and then have to mess around setting up everything themselves.

Javaid stresses that the PC comes with software built specifically for running those pre-installed AIs, avoiding all the fuss of model configuration. The CEO observes: "You can run [AI] models on existing hardware, but it's a terrible experience on many levels."

Core updates its models automatically as new upgrades emerge (unless you don't want that to happen), and the idea is that it's a seamless system and "similar to how Tesla works" with over-the-air updates.

You interact with the PC using an app (on your phone or laptop) with which you can issue voice commands should you wish.

Privacy and security

As well as the ease of use, privacy is the other big selling point for Core. It's tapping into a market of the increasing number of people realizing the potential of AI for helping with their life or work, who are concerned about sending their thoughts or data into the cloud, when they have no real idea where those queries or uploads will end up.

Core will run your AI agents continuously taking on whatever tasks you require while keeping all your data on the device, with Ghost promising that it's not going anywhere else. The personal data on the PC is encrypted to keep it safe — even Ghost can't access the data — and the device is specifically designed to keep running even if Ghost were to shut down (there's no subscription fee here).

The CEO assures us: "All of the source code, logic, and model weights live on the device, not the services. We build this device to not be dependent on having an outside entity maintaining the service."

Security is further bolstered by a firewall that limits how any agent can use credentials such as passwords, and you can dictate how much autonomy the Core agents are allowed, heavily restricting the AI should you prefer.

Javaid tells us that the firewall monitors all outgoing requests and "if it sees the agent performing a sanctioned action or exfiltrating data it shouldn't, it rejects the request and flags it for user review".

Core is priced at $3,499 in the US (around £2,600 / AU$5,000), and pre-orders for the first batch of PCs sold out on the first day that they were offered (yesterday). Those devices ship at the end of this month, and you can now register your interest in a second batch of PCs which are coming off the production line.

Analysis: a question of worth

Core can be hooked into all aspects of your life — or that's certainly the idea, and part of the reason why there's such an emphasis on privacy. It needs to be watertight for security given that Core connects to your existing devices and services, smart-home gadgets, fitness watches — or personal health data — and learns over time about how you live your life, imparting advice based on that knowledge.

Exactly how useful the system will be, well, that remains to be seen — and it may also depend on how 'smart' you've gone in terms of your array of gadgets and tech. There could be a lot of scope beyond life monitoring for Core's AI agents, of course, and there will presumably be the flexibility to get them doing whatever you want.

The other potential sticking point is that price. Is $3,499 a lot to pay for a dedicated AI PC? Well, it's undeniably a hefty sum of money, but there are some heavyweight components under the hood here (and the RAM crisis remains in full force). Indeed, the current price of the Nvidia GPU actually makes the PC a bargain as it stands (the company bought its stock of those graphics cards when they were much cheaper — so the next batch will be pricier).

If you configure a Mac mini as a top-of-the-line device for AI, you're looking at $3,199 in the US at the time of writing — and it doesn't come with the aforementioned strengths of software specifically designed for running local AI easily, and the privacy and security elements mentioned.

I've a feeling Ghost might be onto something here, at least in terms of pushing the concept of a user-friendly and straightforward AI box that does everything locally. I doubt this will be the last take on this kind of idea, and the price tag involved may eventually become a bit more realistic in the future. Although that depends on how the RAM crisis shakes out — and the latest signs are far from positive, sadly.

• Anthropic wants to train 10k FDEs from large institutions and consultancies
• Using their customers could extend Claude's reach significantly
• Training will include a simulated deployment and an actual deployment

Anthropic has announced plans to spend a mammoth $100 million to launch its Claude Frontier Academy, a training scheme designed to support enterprise AI adoption.

The company is targeting 10,000 Frontier Deployed Engineers (FDEs) by the end of 2027, who will be able to identify use cases, handle security considerations and drive deployment for Claude customers.

"Claude Frontier Academy trains people the way our own engineers learn, and we want those who graduate to set the standard for how AI gets built inside a business," Steve Corfield, Global Head of Business Development and Partnerships at Anthropic explained.

Anthropic to train 10,000 FDEs

The training scheme starts with a simulated enterprise deployment before trainees enter a 12-week residency "where they lead a real Claude use case at their own organization."

The initial cohorts are set to include engineers from major consulting and financial organizations, including Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk.

Anthropic has already given out more than 175,000 Claude certifications to workers across 46,000 companies part of its broader Claude Partner Network, but the drive behind its latest FDE scheme is clear.

The easier it becomes for enterprise customers and consultancies to deploy Claude, the more customers Anthropic is likely to have. Additionally, working with consultancies gives Anthropic access to their customers, extending its reach.

Rival company OpenAI already has its own extensive FDE scheme, too, but there's a key difference. OpenAI's focus is on training its own workers to go off and transform customers' companies, but Anthropic's latest announcement is primarily about training other companies' engineers to become FDEs.

Reuters recently reported that OpenAI has experienced a twofold increase in enterprise sales since July 2026, and while specific figures are kept closely under wraps, we do know that enterprise deals are far more lucrative for companies like Anthropic than consumer subscriptions.

So while a $100 million commitment to train engineers may sound like a large sum, Anthropic clearly hopes it will pay off in the long run.

• Meta halved its number of Claude seats while pushing MetaCode adoption
• Microsoft also reportedly asked workers to use Copilot over Claude
• Anthropic could face reputational damage from DoD designation

There seems to be an ongoing shift away from Claude happening in Big Tech, but the reasons behind why two major companies are looking to reduce Claude expenditure aren't so clear.

Cost could be one of the biggest factors, with Microsoft's internal Claude spending estimated to have been on course for at least $1 billion this year, per The Information.

With both Microsoft and Meta having their own internal AI systems, it makes sense for them both to use their own tools for less cash.

Meta and Microsoft reportedly cutting down on Claude usage

Microsoft had already asked employees to limit their Claude usage and spend more time with Copilot and other Microsoft tools, The Information shared, while Meta has reportedly halved the number of its workers using Claude Code from 60,000 earlier this year to just 30,000. At the same time, its internal alternative, MetaCode, has already passed 30,000 users.

Meta, too, was set to spend billions on internal AI usage before the company imposed token limits, spending controls and other monitoring.

We've already seen similar trends among smaller AI customers, with US organizations opting for cheaper Chinese models over homegrown alternatives. It's likely an important economic equation, where paying significantly less for a marginally interior model becomes the most viable option for most use cases.

Anthropic itself has also had a troubling couple of months, after being deemed a national security supply chain risk by the US Department of Defense. A designation that will have undoubtedly altered consumer and enterprise perception of the company, irrespective of the designation's potentially political founding.

"The company may experience material revenue losses or business disruptions attributable to these events," Anthropic warned in an IPO prospectus (via Reuters).

The bigger story impacts Anthropic – potentially losing billions in revenue from two of its biggest customers could seriously impact revenue for a company that's slated to go public for a reported valuation of over $2 trillion (per Reuters).

• AWS CEO looks to debunk common 'myths' about data centers
• Amazon announces it will no longer make local governments sign NDAs when proposing data center projects
• The company has also laid out new codified community commitments

CEO of Amazon Web Services Matt Garman has released a statement in defense of AI data centers that labels the technology as “The race our nation can’t afford to lose”.

Within it, Garman calls on the United States to band together against “misinformation and outright lies” that are being spread about AI data centers to prevent the country from missing out on the huge benefits that will be provided to the economy, national security, and local communities.

“A lapse in our nation’s focus or resolve could put us behind, and potentially irreparably so,” Garman says.

Debunking the myths

There has been a significant amount of messaging about the positives and negatives of the AI data center buildout, with falsehoods and misinformation fueling both sides.

Garman’s statement starts with a section aimed at countering the ‘myths’ spread about AI data centers, such as water use, higher energy costs, and pollution - but keep in mind that Garman makes no distinction between AI data centers and regular data centers.

For example, Garman argues that data centers account for “0.5% of total U.S. direct industrial water use”, but draws on the figures of direct water use to argue this point.

The indirect water usage of data centers is far higher, and often isn’t included in water estimates for AI data centers. The indirect water usage includes upstream factors such as the water used to generate the huge amounts of electricity that data centers rely on. Data centers are also being built in regions that have experienced long-term droughts, exacerbating water shortages and their effects on local ecosystems.

Garman also labels the rise in US energy bills caused by data centers as “misleading and convenient scapegoating,” arguing that “the accelerated adoption of AI is precipitating a build-out that was destined to happen.” What Garman fails to mention is that the infrastructural build-out is being fueled by AI data centers.

Up until very recently, new grid connections for data centers haven’t been paid outright by the developers. Instead, grid operators have funded new connections and upgrades to power lines through increased bills on rate-payers. It is only the backlash from rate payers that has preceded the introduction of Trump’s voluntary Ratepayer Protection Pledge - and the much more binding (and bipartisan) Ratepayer Protection Act. Had there been no public backlash, these measures would not have been necessary.

The next myth Garman attempts to counter relates to the pollution caused by “large backup diesel generators”. Garman states that these are necessary as the services “consumers, businesses, and governments rely on every day” are provided by data centers. This is, in its exact wording, true.

But the pollution many of those in the US are arguing against is attributed to the fossil fuel burning gas turbine engines that many AI data centers use as an alternative to connecting to regional power grids - particularly if their connection requests are denied. The turbines at some campuses give off the same noise pollution as an international airport, 24 hours a day, seven days a week. The sounds they produce have also been involved in causing neurological symptoms in local residents.

The demand for gas, pushed higher by AI data centers, could see the US use more natural gas than Germany and Japan combined by 2035, and the associated pollution could cause generational health decline and cost $25 billion in associated costs.

The final myth that Garman seeks to debunk is that ‘Community members get nothing from data centers’, arguing that data center operators are “often the single largest taxpayer in a community” and that this “reduces the taxes paid by families, but it also funds investments in schools, fire and police departments, parks, roads, and other community services.”

But data center operators are also the largest beneficiaries of tax cuts and credits too. Numerous states have begun repealing tax cuts for data centers after seeing billions lost in expected revenue - billions that otherwise could have been used for funding investment in community services.

Garman also points to the boom in construction jobs associated with the AI data center buildout, but doesn’t include that many of these jobs are temporary. Once the construction of a data center is complete, these jobs often disappear unless there is continuing demand within the local region, and operational data centers themselves require a very small number of employees.

Amazon’s new community commitments

In an effort to get local communities onboard with data centers, Garman’s statement also lays out a number of ‘codified’ community commitments that Amazon will adhere to.

These commitments fall under four pillars; Protecting & Enhancing Communities; Providing Jobs & Economic Benefit; Engaging Communities; and Building & Operating Responsibly.

The full list of commitments is available here, but there are a few key points to look at. Outside of vague commitments to covering the cost of grid connections and infrastructural upgrades, paying their fair share of taxes, and paying the workers they hire, Amazon will contribute "$1 billion over the next five years" to jobs and education, energy affordability and water solutions, and community investment.

Given that Amazon operates 236 data centers that are operational or under construction, this $1 billion figure amounts to around $4,237,000 per data center over the full five years, or about $847,000 per data center per year.

Amazon has also announced a commitment to “no longer use nondisclosure agreements with the government agencies we work with on our projects.”

This is by far the biggest announcement included in Garman’s statement, yet it takes up just a single line. Local communities have long called out for more transparency around AI data centers.

In numerous cases, NDAs have been used to prevent local governments from discussing planned data centers with their constituents. In numerous other cases, data centers have only been announced to local communities after the necessary permits are secured, preventing local residents from being involved in the consultation process and limiting their ability to oppose new projects.

The community commitments offered by big tech companies such as Amazon and Meta looking to build AI data centers are not out of the goodness of their hearts. They are a direct result of the community backlash against AI data centers that has occurred across the US.

Had this huge opposition not happened these benefits would not be offered, but they are being jazzed up by big tech companies as a benefit they are providing to communities rather than the efforts of grassroots movements and every day Americans fed up with being subject to rising bills, increased noise pollution, and an economic system that favors those with money and power.

• A majority of US voters (61%) now oppose AI data center construction
• Trump voters support construction, but only at 49%, versus 47% opposing
• Polling took place in early September 2026

The majority of US voters have expressed their opposition to further AI data centers, as newly published research by The New York Times and Siena University found 1,503 likely US midterm voters have major concerns over the direction of travel with AI data center construction.

However, the survey found the data center issue is not a priority for most US voters, where economic and foreign policy matters are more prominent.

Reaction to the polling is likely to be minimal, but it appears to underpin the growing discontent with the approval of increasingly expansive data centers. Concerns over the environment (such as wildlife, water use, and impact on nearby communities) dominate the responses to the poll.

The data center question

The question that has attracted the most attention from the poll (paywalled) – which covers a range of issues rather than just AI – was worded “In general, do you support or oppose the construction of data centers to support artificial intelligence, also known as AI, technology in the U.S.?”

An overall response of 61% opposing to 34% supporting highlights how fed up voters are with the AI issue. Data centers are perhaps the most tangible, visible element of the technology for voters to focus their ire on, given the impact the facilities have on local communities and ecology.

Perhaps the most interesting factor of the overall response to the main question was how the result is split. Based on 2024 voting, 49% of Republicans support AI data center construction versus 47% who do not; 18% of Democrats support construction, while 75% do not. (Among independent supporters, the result is 36% support, 60% oppose, and all three measures accommodate the “don’t knows.”)

Regulatory supporters

While the survey explored many topics, it spent some time quizzing the respondents further on AI. Here, some nuance can be found, with 56% of voters in favor of limits to AI data centers, whereas 38% supported a total ban on further construction. Interestingly, a slight majority (42% vs 40%) trusted the Republicans to handle the data center matter.

Those polled expressed support for key concerns surround data center construction, with 32% citing water usage or the environment as the biggest worry. The impact on local communities and resources was expressed by 21%, while 18% had a general distrust or dislike of AI. Impact on local economies and jobs was 9% of those polled, while 10% had concerns over AI regulations.

The poll was completed before President Trump’s statement against guardrails, which he believes would hamper America in the AI race with China.

• 38% of workers who regularly use AI have admitted to “hoarding” prompts to gain a professional advantage
• Meanwhile, 21% have hidden their AI use from their manager in order to appear more efficient
• Many just don't want their colleagues to gain the same advantage to productivity

New research has claimed nearly two-thirds (62%) of workers using AI regularly believe having better AI skills gives them an edge – and 38% are keeping productive prompts to themselves to to gain a professional advantage.

A Use.AI survey of employees with regular AI users in the workplace found that of this number, nearly half (44%) didn’t want their colleagues to gain the same advantage to productivity, while 31% claimed that it was they who had invested the time and effort into developing the prompt and didn’t want to give away their advantage.

While the use of AI for productivity boosts can unlock advantages, the survey indicates a proprietary attitude, with workplace “gatekeepers” potentially hoarding the best AI prompts.

AI teamwork?

A trend of gaining an advantage in the workplace through developing AI skills seems to be the takeaway from this survey, which quizzed 13,276 adults who use AI in the workplace, based in organizations in the UK, US, Canada, Australia, the EU, and Latin America.

As well as unearthing a habit of keeping AI productivity shortcuts a secret, the survey also underpins a growing pattern of anxiety around the technology’s impact on a worker’s performance. It found that 41% of AI users in the workplace were worried about colleagues and managers finding out how much time and effort the technology was saving them. They were also concerned that they would be expected to achieve the same levels of productivity and speed in future.

It also found that 36% had hidden their AI use from coworkers, with 21% hiding their new-found productivity skills from their manager.

AI-generated tensions

Many criticisms of AI at the point of use (as opposed to data centers) tends to dwell on the issues surrounding generative AI and its impact on creative industries. In a non-creative industry, there is the reliability of the results to consider – but Use.AI’s survey seems to point at a new issue.

Ihor Herasymov is a Co-Founder & CEO of Use.AI, and says that the survey suggests “tension” is developing in the workplace around AI.

“When a useful AI method becomes part of someone’s personal edge, sharing it is no longer a neutral act. Companies may need to think more carefully about how they reward people who turn private know-how into shared capability, because employees will notice quickly when openness helps the organization while weakening their own position.”

In short, the companies that find this secretive AI prompt culture developing may find that the metrics with which judge employee performance will have to change.

• Businesses are tracking AI as consumers shift to AI-powered buying patterns
• Social media is also a big are for monitoring consumer trends
• Conventional retail like Walmart and eBay have dropped

AI platforms have become synonymous with scraping other sites for information, but it's not every day you hear of AI sites themselves becoming targets. The latest Decodo research has revealed just that, though, with ChatGPT and Perplexity both appearing in the top five.

Decodo argues this marks a fundamental shift in what organizations are now demanding from the web – "Two years ago, businesses scraped the web to see what people were searching for," CEO Vaidotas Juknys noted.

"Now they scrape AI platforms to see what people are being told. The use case for data collection is changing."

AI platforms like ChatGPT are now being scraped

Besides AI platforms, Decodo found that seven of the top 10 sites are actually new entrants, with just TikTok (number one) Google (second place) and Amazon (eighth place) re-entering as repeat offenders.

ChatGPT, in third place, accounted for 14% of the activity observed by the analysis, which could account for businesses checking how the AI tool ranks their products against competitors, monitoring how it describes the company, measuring brand mention frequency, recording ChatGPT citations and more.

Perplexity accounted for 10% of the data, landing it in fifth place, while TikTok saw the most scrapes at nearly one in five (18%).

\But with several top-10 sites from 2025 dropping out of the charts, including Walmart, eBay and Airbnb, Decodo's data shows much more than just the rise of AI. It's also a much broader indication of how and where consumers are shopping, including AI and social media.

Looking ahead, the company is also anticipating that Gemini and Copilot could end up in next year's top 10. But marking an even bigger shift, Decodo also predicts that AI agents could replace humans in starting those scrapes in the first place.

"Next year the agents might be doing the deciding instead of just reading the answers," Juknys added.

OpenAI had an oddly casual way of apologizing to the Australian Government to explain how one of its experimental AI agents gained unauthorized access to a government health statistics system while trying to complete a routine research task. More casual than one might expect for such a huge security concern.

The agent was looking for publicly available information about medicine spending, but after failing to get what it wanted through the normal route, it found a vulnerability in and used it to access internal files, though no individual patient records.

The incident is a striking example of the risks that come with increasingly autonomous AI agents, but almost as striking is how OpenAI admitted what happened. OpenAI says it discovered the Australian activity in mid-August but did not notify Australia until September 10. When it finally did, the company sent a brief, almost jaunty email about it, as shared on LinkedIn by an ABC reporter:

"We are notifying you of a security vulnerability identified during our review of OpenAI model activity involving Services Australia’s Medicare Statistics service," the email begins before explaining what happened. "We recommend that the team responsible for the service investigate the vulnerability and assess the changes needed to prevent it. We would be glad to brief your security team and provide supporting evidence as available."

The message ends with "Best," though the letter hardly feels like it means it.

Won't accept a no

The model was researching government spending per person on medicines for skin conditions in communities in Victoria in June. It was supposed to find publicly available statistics. Instead, it discovered a route into the Medicare Statistics Reporting Service that gave it non-public access.

Essentially, the model asked for information it shouldn't have access to and took the denial as a sign to find another way in. The ethics of the matter did not arise.

"We did not intend for this activity to occur, and the access to the service and follow-on activity should not have happened," OpenAI wrote in its follow-up explanation.

The timing makes the note considerably worse as, while OpenAI discovered in July what happened in June, it did not notify Services Australia until September 10.

That means the most significant delay was between discovery and disclosure. OpenAI now admits it still waited too long. In its expanded apology, the company said it should have shared preliminary findings sooner and kept Australian agencies updated as it learned more.

While using the public disclosures email is legitimate for the average person finding a security flaw, it seems odd for a company of OpenAI's size to report its own model's error this way.

Not really the 'best' way to say sorry

The apology also revealed that Medicare was only part of the story. OpenAI said its models had interacted with several Australian government services during training and evaluation. An agent accessed the NSW Bureau of Crime Statistics and Research's public Crime Mapping Tool, while agents found an exposed access key associated with a Victorian health reporting system and retrieved configuration information and aggregate survey statistics.

OpenAI said no individual medical or criminal records were accessed in those incidents. But OpenAI later expanded its response, adding that a model researching wildfire statistics had used crafted queries against the NSW National Parks and Wildlife Service's Fire History service to infer database metadata that was not intended to be publicly exposed.

The growing list makes the Medicare incident harder to dismiss as a quirky model finding a single vulnerability. It suggests experimental agents were repeatedly interacting with real websites in ways their creators did not expect. OpenAI itself has responded by pausing training and evaluation involving tool use for its most capable models until additional safeguards are in place.

OpenAI also said that newer monitoring has already caught a model obtaining live internet access during another training run, allowing humans to stop it. The Australian government is investigating whether laws were broken and working on fixing older public-facing government systems to prevent it from happening again.

AI companies increasingly want us to trust agents with doing tasks autonomously on our behalf. The selling point is that an agent can encounter an obstacle and independently figure out how to accomplish its goal. This is an unusually vivid example of what happens when that feature works a little too enthusiastically.

The Medicare agent was assigned to find statistics about medicine spending, not penetrate a government server. But it simply steered toward a more direct route, ignoring every consideration but effectiveness. The technology had crossed a boundary its developer says it never intended the model to cross, and the company took weeks after discovering the incident to tell the organization on the other side of that boundary. That's why the email feels like more than a funny piece of corporate communication.

The follow-up acknowledges responsibility and describes concrete safeguards while admitting that notification should have come sooner. It also recognizes that the company now has work to do to rebuild trust. It's a much better message. "Best" is a way to end an email to a professional acquaintance, not when your experimental AI has hacked national medical databases.

• Google is changing model access in Gemini
• The best models are being dropped from the free and Plus tiers
• These updates will be applied from October 9

Google is shaking up access to its Gemini AI models for individual users based on how much you pay each month, and it's those on the cheapest Plus plan and the free tier who are going to be losing out from next week.

As per an updated support document (via 9to5Google), if you don't have an AI subscription for Gemini then you'll be restricted to the Flash-Lite (3.5) "fastest answers" model. The Flash (3.6) "all around help" and Pro (3.1) "advanced reasoning" models won't be available going forward.

If you're on the $4.99 / £4.49 / AU$7.99 AI Plus plan, then you're losing access to Pro (3.1), but will still be able to call on Flash (3.6). There are no changes to the more expensive AI Pro and AI Ultra tiers, except for AI Pro gaining the Deep Think mode that maximizes parallel reasoning while sacrificing some speed.

Look out for a message inside the Gemini app or in your inbox if you're going to be affected by the changes. Google says the free tier restrictions will be applied from Friday, October 9, while AI Plus users will get an email advising them about the timetable.

Usage limits

Previously, the Gemini AI plans have been differentiated by usage limits — everyone got access to the same models, but for varying numbers of queries. Now there are further restrictions, based on model availability.

Exact usage limits are difficult to pin down: they depend on "the complexity of your prompt, the features you use, and the length of your chat" Google says, adding that "paid users have higher limits than users without a Google AI subscription". Usage limits refresh every five hours, and there are weekly limits too.

The move puts Gemini more in line with what you get with ChatGPT. OpenAI also limits model access based on how much you're paying each month, also the way it handles usage differs somewhat from Google: rather than one single allowance, there are usage limits on different models, query types, and thinking modes.

We know that AI compute doesn't come cheap, so presumably Google has done the math and decided it can save some money with the new restrictions — and perhaps convince a few more people to sign up for a subscription.

Many AI developers and frontier labs are openly pursuing artificial general intelligence (AGI), which scientists describe as human-like intelligence in which a model can reason like humans and learn new capabilities outside its training data. But does that mean that they're all on the right path?

Chasing the dragon

If we achieved AGI, how would we even know? Many AI benchmarks exist, but François Chollet co-founded the ARC Prize just a few months before speaking with the Dwarkesh Podcast to get to the bottom of this.

In comments on the podcast, he explained that large language models (LLMs), which are based on neural networks, is a dead end when it comes to achieving AGI. That doesn't mean there wasn't room to improve LLMs so they were more powerful, autonomous and useful to businesses. But that isn't the same thing as AGI.

By pouring funding into LLMs at the expense of other avenues or architectures, Chollet explained, the road to true AGI is being ignored. The result? Progress has been set back by up to a decade.

Intelligent systems

Despite Chollet's steadfast belief that LLMs will not get us any closer to AGI, many technology executives have spent the last few weeks opining about the possible threat that they pose to humanity.

Ahead of OpenAI and Anthropic's as-of-yet-undetermined IPOs, both companies have taken turns disclosing increasingly worrying incidents, including cybersecurity breaches. These stories have, in many people's eyes, simply served to hype up the capabilities of existing technologies for marketing purposes.

As for Chollet, the whole purpose of the ARC-AGI benchmark is to determine true progress toward AGI. While LLMs are performing better on this especially tough benchmark, none have come close to hitting the mark.

• OpenAI, Softbank, Blackstone, and construction unions have formed a lobbying group
• The American Infrastructure Alliance will support efforts to build data centers where local opposition is strong
• Opposition groups have called for a moratorium on new data centers, with existing research suggesting promises of long-term jobs are far-fetched

Challenges to data center development in the US are about to have the battle weighted against them with the unveiling of the American Infrastructure Alliance (AIA), a group of companies and unions that seek to overturn objections to new facilities.

Comprising OpenAI, global data center company QTS, investment company Softbank, management experts Blackstone, and seven unions including the International Brotherhood of Electrical Workers, this coalition plans to establish guardrails for data center development in 2027.

The American Infrastructure Alliance plans to launch its initial defensive strategy in Texas, Georgia, Ohio, Iowa, Pennsylvania, Indiana and South Carolina. In doing so, the AIA hopes to head off objections from local opposition groups, and avoid growing calls for moratoriums.

“International” alliance

By establishing the alliance, the tech companies, management consultants, and unions intend to get public and union member support for the construction of new data centers.

Along with the International Brotherhood of Electrical Workers (IBEW), the alliance includes the United Association of Union Plumbers and Pipefitters (UA), the International Association of Sheet Metal, Air, Rail and Transportation Workers (SMART), the International Association of Heat and Frost Insulators and Allied Workers (HFIAW), the International Association of Iron Workers.

Representatives from some of these organizations have offered their support publicly. For example, Michael Coleman, general president of SMART, says "American workers and communities should have a seat at the table as this critical infrastructure is built."

Responsible data center building

Unsurprisingly the heavy-handed approach that seems to have been conceived to force state officials and county and city councils to be more favorable to data center developments, is attracting criticism.

Mitch Jones has called for a moratorium on data center development to conduct further study on their mid- and long-term effects. He’s the managing director of policy and litigation at environmental group Food & Water Watch, and has concerns not just about the environmental impact, but jobs too.

He expressed concern about construction jobs data, telling the DCNF that “Once centers are constructed, very few jobs remain in the facilities themselves. Our research has found that as of 2024 there were only 23,000 people working permanently in data centers across the country.”

Calls for moratoriums have grown, with politicians including Sen. Bernie Sanders planning legislation to pause construction of data centers across the US until their guardrails are in place.

But the AIA’s guardrails and lobbying are likely to have a more immediate effect, and bring jobs. Kenneth Cooper, the IBEW's international president, notes: "Blanket bans on necessary infrastructure projects would set back our economy and threaten good middle-class jobs.”

• AI chatbots appear to be providing less diverse responses, research claims
• Displaying the same answers has been likened to globalisation, where the same products and brands are available everywhere
• Such a move may increase convenience, but also reduce diversity

Could we be heading into a real-world Idiocracy? A repeated reliance on AI to provide answers – invariably the same answers, offering the same views – could be shrinking human knowledge, a new study has claimed.

The research, which tested 27 different LLMs, found that even with the best performing LLM, the information presented as answers was almost 19% less diverse than what is available using Google’s standard search engine. Consequently, the information was less varied, providing a narrower selection of answers, views, and restricting interpretation.

Researchers believe that this trend risks pushing certain content to the peripheries of accepted knowledge, while other content is considered dominant or primary – and the end result of this could indeed be a world where people are less aware, more idiotic, and perhaps even with less intelligence than today.

Topic testing

The study was conducted by researchers at the Department of Computer Science (DIKU) at the at the University of Copenhagen.

To establish how the LLMs would deliver answers, they were tested on 155 topics (ranging from nuclear weapons to K-Pop). The researchers used 200 prompt formulations drawn from research into real AI chatbot usage, generating an incredible 70 million responses from the models.

Even answers from GPT-5, the LLM producing the most varied responses, were still 18.5% less varied than answers from Google Search.

“People are to a large extent exposed to the same information over and over again,” says first author Dustin Wright. “So AI chatbots are not just changing how we find knowledge, but also which knowledge we have access to."

An Idiocracy?

Whether using AI models as part of Google Search, DuckDuckGo’s search tool, Bing, or by specifically quizzing Gemini, ChatGPT, or any of the other chatbot-accessible LLMs, the findings from the research indicate a change in how information is surfaced.

“We risk exposing people to fewer perspectives and a narrower range of knowledge. This could create a vicious cycle in which the most popular content becomes even more dominant, while other content is increasingly overlooked,” says DIKU Professor (and senior author) Isabelle Augenstein.

She compares the phenomenon with the downsides of globalization: “Today, you can buy the same products and find the same coffee chains almost everywhere in the world. That has many advantages, but it has also reduced diversity.”

As for the possibility that AI could shrink knowledge across humanity, and leave a world governed by those wholly unsuited to the task, Professor Augenstein is more upbeat. She says that a “knowledge collapse is not happening yet […] we can see that the more recent models produce slightly more diverse answers than older models.”

• 'No Robo Bosses Act' signed in California, comes into force in July 2027
• Humans can't just approve an automated decision, they must bring evidence
• Governor Gavin Newsom criticized Trump administration for a lack of federal leadership

California Governor Gavin Newsom has signed a major package of no fewer than 13 AI-related bills as the state continues to increase the pressure on AI via human-protecting regulation centered around employment, education, legal and more.

Among the new bills was the 'No Robo Bosses Act', which essentially limits employers from using automated decision-making to discipline or fire workers.

The bill, officially known as SB 947, is described as California's human-in-the-loop law to keep humans at the core of workplace decisions that affect workers' lives.

California to protect workers from AI-powered disciplinary action

The bill insists that an employer cannot rely solely on an automated decision system when deciding to discipline, or terminate, an employee. But furthermore, a human can't merely approve an automated decision – they must also provide their own evidence to corroborate with the automated decision.

"We are deeply mindful we have to do a lot more in the absence of federal leadership," Newsom said in a statement (via X), referring to Trump's decision to rename artificial intelligence 'super intelligence'.

Yesterday's meeting with tech leaders also saw the creation of a new "morally binding" agreement whereby major companies, including Google, Anthropic, Meta, OpenAI, SpaceX and Nvidia, would be responsible for ensuring the safety of their own AI.

This, of course, comes following ongoing calls for greater government oversight of AI.

"By signing these four first-in-the-nation bills, Governor Newsom has shown the country how we can begin to put guardrails on AI at work to ensure human oversight over this technology, limit surveillance, and put workers first," the California Federation of Labor Unions, AFL-CIO President Lorena Gonzalez, said.

Crucially, though this specific bill was signed on September 30 2026, it won't be effective until July 1, 2027.

The data center boom powering a lot of AI infrastructure has provoked pushback from a wide array of groups critical of how they are built and run. The backlash has been strong enough that even in Texas, which had become a hub of new data centers, Governor Greg Abbott felt the need to order an audit of those seeking connections to the state’s electricity grid and to tell the Texas Commission on Environmental Quality to stop issuing permits sought for them until the audit is complete.

But a rather large complication is making that audit seem inadequate. Thousands of diesel generators and increasingly large private power plants sit next to the data centers. Some developers are building or planning enough power generation on-site that the distinction between a data center and an industrial power facility starts getting fuzzy.

It's a problem raised in an open letter sent to Abbott by the NAACP Texas State Conference and the NAACP Center for Environmental and Climate Justice. The groups argue that Texas’s new measures remain too focused on data centers seeking connections to the Texas grid, potentially leaving a giant loophole for those that generate or source their own electricity. The letter also argues that the audit does not sufficiently address cumulative effects, including pollution, water consumption, and noise.

As the letter points out, a Floodlight and The Texas Tribune investigation found that at least 38 Texas data centers have obtained minor-source permits for on-site power since 2024, covering more than 2,100 diesel generators. The permitting category is also associated with considerably smaller sources of pollution, including dry cleaners, auto body shops, concrete batch plants, and rock crushers. Texas’s new regulatory response may not actually address the issues brought by the data centers and their support infrastructure.

Minor permit, major story

The "dry cleaner permit" involves categories of minor-source air permits made to streamline approval for small businesses and avoid more intensive environmental review. The idea is that regulators don't need an elaborate approval process every time someone installs a set of dry cleaning machines or pumps at a gas station.

AI data centers test the logic through sheer scale. Floodlight found that OpenAI’s Stargate campus in Abilene had already built 10 gas-powered turbines and 62 backup diesel generators and was seeking approval for another 41 turbines and 18 generators. The existing turbines and generators were initially authorized through minor permits that did not require the environmental studies, public notice, and public comment associated with major permits.

The investigation also found several projects with estimated emissions just shy of triggering more scrutiny. A CyrusOne data center in Whitney estimated 249.1 tons of nitrogen oxide emissions annually, where 250 tons would have triggered more stringent permitting, while a Vantage facility outside San Antonio received permission for 99.8 tons annually in an area with a 100-ton threshold.

More than half of the data centers Floodlight identified submitted annual nitrogen oxide estimates just below thresholds requiring public input and more detailed environmental review. It raises questions of whether the regulatory thresholds make sense if they can be evaded so easily. The NAACP letter argues that minor-source permits can effectively remove communities from the process when they shouldn't be. It has already issued a notice of intent to sue Vantage Data Centers and VoltaGrid. While the allegations are still pending litigation, they show why the definition of a "minor" source has become so fraught.

Crackdown with loopholes

Abbott's recent responses acknowledge that Texas has a problem worth addressing. His audit directive states that projects that fail can be denied connection to the Texas grid. And the pause on more data centers until the audit is done will certainly slow things down for a time.

But while encouraging data centers to avoid burdening the grid might seem like a way to rein them in, letting them have their own power plants requires a different regulatory response. There are still emissions, water demands, and neighborhood effects regardless of whether it's linked to the grid. Abbot did tell regulators to make permitting decisions with the state's broader data-center directives. The loophole may be smaller than it first seems.

Still, this is becoming a recurring theme of the AI infrastructure boom. Technology is moving at a scale and speed that regulations written for a very different industrial landscape struggle to accommodate. A rule that seems perfectly sensible for a shop with a boiler or a building with a couple of emergency generators looks rather different when applied dozens of times across facilities consuming enormous amounts of power.

Texas has recognized that the AI data-center rush needs guardrails. The next test is whether those guardrails follow the actual environmental impact of the projects rather than simply the route their power cable takes.

As businesses everywhere continue to get to grips with the reality of AI-powered workflows, Snowflake has revealed its view of what it will take for a company to become a truly agentic enterprise.

Speaking at the company’s Snowflake World Tour London event, senior leaders laid out how a number of factors will all contribute to agentic success.

The 'Agentic Enterprise' is here

“There's no enterprise AI strategy without a data strategy,” declared James Hall, Snowflake Country Manager, UK&I as he kicked off the keynote.

“You need the right data foundation, trusted, governed, secure and accessible…in just 12 months, we’ve moved from talking about the potential of AI to putting AI to work…it’s now all about making AI real for business as we enter the era of the agentic enterprise.”

Hall highlighted the success of UK&I customers such as mobile operator GiffGaff, with Snowflake CEO Sridhar Ramaswamy also later joined on stage by the CEO of the London Stock Exchange Group (LSEG) as proof of this.

But it was Christian Kleinerman, EVP Product Management, who took to the stage to run through the factors the company sees as vital for an agentic enterprise.

“We are all part of the era of the agentic enterprise,” he declared, defining this term as an organization that has been able to embed AI powered agents into its core operating processes, co-ordinating with humans, setting directions, setting guardrails and automating as much as possible - all with the ultimate goal of help workers be more productive, and creating better business outcomes.

“The promise of AI is unparalleled, in my mind, to other technologies that we've seen in our lifetime,” Kleinerman added, addressing a growing myth around the technology, namely that some organizations believe AI is so good and smart, the data will sort itself out.

“But if anything, the opposite is true,” Kleinerman said the most amazing results you can envisage with AI come from having a great data strategy.”

Kleinerman noted three main elements for an enterprise to be truly agentic - an AI model, data context, and software and apps - all feeding into what Snowflake calls an “agentic control plane”.

When it comes to picking a model, Kleinerman noted the company is “committed to model choice for all of you” - as there is clearly no way to know who will win out in AI models even within the next few months.

Data context is becoming increasingly crucial, he added, as understanding your business matters a lot to how AI can be beneficial for you, and having the right software and apps will really allow AI to integrate with the rest of your enterprise infrastructure.

“The truly amazing results come when you really understand your data”, he added, “we want to un-silo your data, we want to help you govern your data...so humans can be more data-driven - AI has supercharged that ability.”

All of this flows into the agentic control plane - “the glue that holds all this together”, Kleinerman concluded, “the agentic enterprise truly comes to life when the AI models can interoperate with the rest of your enterprise infrastructure…(and) hopefully it feels like an avalanche of innovation...our goal is to continue accelerating.”

“We’re operating in a fast-moving environment, arguably the biggest risk any business could take right now is by standing still - but moving fast with data and AI shouldn’t mean losing control, it means having the foundation, the flexibility and the confidence to move quickly,” Hill added.

“The era of the agentic enterprise isn't coming, it's here - it’s happening right now in production.”

If you work in an office you’ll know we’re deep into performance review season. But how would you feel if you found out that your boss had used AI to write your yearly appraisal? I know I’d be pretty annoyed to learn that AI was deciding whether I was worthy of a raise this year.

AI use by employees is becoming commonplace, but it's also being adopted by managers. A recent survey by Headway of 1,000 managers found that 33% had used AI to write feedback or prepare a performance review, while 31% had turned to AI to resolve team conflicts. A further 27% had used AI to help write uncomfortable conversations they needed to have.

That may sound like a lot, but the figure is actually down a bit from 41% admitting they have used AI to draft or revise a performance review in a similar survey from a year ago, but still high.

The ease AI

For managers, the appeal of using AI is easy to understand. Faced with having to write several performance reviews each year it can be a real timesaver to dump the work onto AI, however the risk is that what should be very personal performance feedback is being outsourced to a bot.

Managers' AI use is not stopping at performance reviews, either. The survey also found that 57% of managers have already scripted or rehearsed a difficult conversation before having it using AI.

Thalia-Maria Tourikis, a qualified health coach and burnout prevention and recovery expert at Headway, commented:

“When encountered with a difficult conversation, one might think that using AI can help organize their thoughts. However, you should not give up that interaction to automation. When human emotions such as trust and empathy are involved, it is best conveyed when written by a person. The perfection of AI reads as superficial and doesn’t have the same impact.”

A sliding scale

Bosses using AI to help rate employee performance isn’t the only threat from AI that workers face, there’s also the threat of being monitored by AI. Owl Labs' State of Hybrid Work survey spoke to 8,000 workers worldwide, which found that 58% of employees believe their employer has deployed AI-based monitoring tools.

The Headway survey doesn’t make it clear whether managers are using AI to simply improving their grammar, or change the occasional word for something better, or wether their actually generating the text of the performance review entirely using AI. There’s a big difference between these two extremes and one involves much more cause for concern than the other.

It’s especially concerning with performance reviews because these are often link to financial compensation. If your performance review makes a difference to your annual salary, and wether you deserve that raise or not, then finding out that it’s really AI deciding your future rather than your human manager would be especially galling.

If you ever need a sign that AI is coming for Hollywood then it must be that Jeffrey Katzenberg, the veteran Hollywood executive who co-founded Dreamworks, has just announced that he’s launching an AI-focused studio.

Crucially, he told The Wrap that AI output was “slop” until very recently, but that in the last couple of months something has changed. Now he wants to make feature films and TV with AI, arguing that falling production costs will mean more films and new forms of storytelling.

At the same time The Gifted, an AI-generated trailer, has just won a competition that puts $2.5m towards turning it into a proper feature film. Future Vision XPRIZE is one of the world's largest film competitions for hopeful and optimistic science fiction filmmaking that answers the question: what does a future worth building actually look like?

You can watch the trailer for The Gifted here.

Unfortunately, the industry focus on AI seems to be happening at exactly the same time that the audience is being turned off by it.

Filmustage, a movie production planning platform, analyzed Google Trends data and found searches rising on both sides of the AI debate. It found that the phrase “movies with no AI” was up 345% in the past month, while “movies made with AI” was up 112% in the same period.

Google Trends percentage increases can often be massive, so we shouldn't read too much into the numbers themselves, but the overall trend is upwards. The motivations behind those searches may be very different, but both suggest audiences increasingly want to know one thing before they watch: was AI involved?

I don't think it's hard to understand why. You only have to look at an AI film to understand why people have reservations.

You can judge the trailer for The Gifted yourself, but while it may be a stunning technical achievement that AI has made this three-minute video, if I’m judging it as a movie trailer… then I don’t want to watch the movie.

The flaws are obvious. AI is clearly very good at spectacle: lightning, explosions, collapsing buildings and impossible camera angles. Not to mention lead characters who look almost like famous actors you know, but not quite. (Is that Amy Adams in The Gifted's poster? No, it's not quite Amy Adams.)

The harder test is whether it can sustain believable characters, performances and emotional continuity for two hours rather than three minutes. If it does, then it's going to need much better dialogue than what's presented here.

I remember watching the pure-AI film Hell Grind in a theater recently, and that just about kept my attention for its 90-minute duration, but I was ready for it to be over by the end. It also played to AI’s strengths, by not trying to make its world look ‘real’. It's when AI pretends to be real life that I think it falls down the most.

Pure AI films... still look like AI

Actors, perhaps seeing a future in which their skills are not as widely appreciated, are famously against AI, but even they seem to have accepted the overall trend towards it now.

As Deadline reports, Tom Cruise earned spontaneous applause for a six minute speech during a recent BAFTA “Life in Pictures” conversation in London. Cruise talked about how AI “is coming”, but “people want to see real things”.

“There is nothing like human” said Cruise, citing a summer of blockbuster films as proof that “people want human experience.”

From what I've seen of AI movies, I'd say Cruise is exactly right. AI still can't produce an experience that looks human and the connection to the audience that goes along with that. But with even Tom Cruise acknowledging that AI is coming, and executives like Katzenberg jumping on board, AI movies seem almost inevitable now, especially with the saving in production costs that they bring with them.

But while AI might be economical, it definitely has its problems, and if people are already searching for “movies with no AI” then Hollywood has a branding problem before the AI movie era has even properly begun.

AI as another tool for filmmakers

But perhaps before we assume that AI in films means the end to human actors and an influx of AI slop, let’s talk to one of the AI studios making waves at the moment, Utopai Studios, who creates original films and television series using their own proprietary PAI software.

I talked to Kevin Chong, Chief Operating Officer at Utopai Studios.

“We don’t think of AI films as a separate genre”, explains Chong. “AI is another tool filmmakers can use to tell stories, just as visual effects, digital cinematography and virtual production became part of the filmmaking process. It can help filmmakers create worlds, achieve a level of scale that might otherwise be difficult or prohibitively expensive, and give creators more choices.”

And what about the end of human actors? “At Utopai Studios, we don’t see AI as a replacement for human actors, and that isn’t the future we’re building toward”, Chong explains. “There may be stories where a completely synthetic character makes sense creatively, just as animation has always created characters that don’t physically exist. But that is very different from saying audiences are going to stop wanting human performances.

“Performance is one of the most human parts of filmmaking”, Chong continues. “Actors bring interpretation, emotion, instinct and life experience to a character in ways that are fundamental to storytelling. AI can become another part of the filmmaking toolkit — including helping extend performances, create environments or enable things that would be difficult to achieve traditionally — but the creative decisions still need to come from people.”

Tom Cruise is right: there is nothing quite like human, but perhaps there will be a way for humans and AI to coexist in the future of movies after all.

• Claude for Government runs in a FedRAMP High-authorized environment
• Pricing is now based on pre-allocated spend, not per seat
• Claude for Microsoft 365 is also being added

Anthropic has announced the general availability of Claude for Government specifically for government agencies.

First launched as a public beta in July 2026, the industry-specific variant is designed to give government employees access to the same sorts of tools as commercial Claude customers, within a FedRAMP High-authorized environment.

But better still, now that Claude for Government is officially available, Anthropic has implied that new features will land on this platform at the same time as they do on commercial versions of Claude, so that government workers don't risk being left behind.

Claude for Government is now generally available

Naturally, security is one of the biggest messages that coincides with the launch of Claude for Government, and the company notes that admins can set model restrictions, monitor usage, allocate spending, control external connections and more.

But Anthropic has also made some changes to its pricing structure in a bid to make spending more predictable, given public sector scrutiny and AI costs that can quickly spiral out of control. Rather than per-seat billing, Claude for Government works on the basis of having fixed increments.

Under this new model, agencies would need to commit to a specific amount, and Claude usage cannot exceed that amount.

"Administrators define user tiers with spend and model limits per group, track usage by user and by model, and get burndown alerts before a balance runs low," the company explained.

At the same time, Claude for Microsoft 365 is also being made available via early access, which puts Claude tools directly within Microsoft's three biggest pieces of office software – Word, Excel and PowerPoint – as well as Outlook (in beta).

By confirming that it will be adding new features at the standard public cadence and including Claude for Microsoft 365, Anthropic is acknowledging that governments don't want to compromise on features or security.

• AI usage is far from even when it comes to managers vs. regular workers
• Workers are facing more work or increased complexity due to AI
• Redesigning work is more important than upping quality or quantity

New PwC research has claimed 19% of UK workers now use AI every day at work, up from 15% last year, but despite some progression, there are still clearly some biases when it comes to who uses AI and how they use it.

For example, while 85% of senior execs and 73% of managers have used AI, only 35% of non-managers have used it.

But even if more knowledge workers were to use artificial intelligence as part of their daily routines, PwC argues that AI isn't actually making work any easier.

AI adoption isn't equal, but nor are its impacts

Even though employees claim AI has helped improve the quality of their work (70%), use more of their skills (63%) and bring value to the workplace (58%), the report actually argues that nearly half (44%) of workers have seen their workloads increase following the use of AI, with a similar number (45%) worried their job's complexity has also increased.

With this growing complexity in mind, PwC UK Chief Innovation and Technology Officer Claire Reid explained that employers and employees need to "rethink how work gets done" instead of just increase quality or quantity.

"The real prize isn’t just doing more work; it’s redesigning work so that people and AI together can deliver better outcomes," Reid summarized.

Still, the discrepancy in use among different levels of seniority is hard to ignore, and PwC could have the answer. Many fear negative consequences if they experiment with AI, are confused about their employer's policy or simply need more training. A clearer strategy and greater communication could be enough to change this and enable more workers (and therefore employees) to benefit from AI's more positive impacts.

"Organisations need to rethink work around the right problems, give people the skills and confidence to use AI well, and put the right safeguards around it," Reid added.

• Data center developer offers households $10,000 each in attempt to get huge campus approved
• Residents have called the payment a 'bribe', and list numerous concerns over pollution, water use, and property values
• Developer's chief marketing officer called the payment a "pretty radical and pretty leading-edge" idea

Households in Hazle, Pennsylvania, were offered $10,000 each as part of a plan to help get a 1,300 acre AI data center built near the town - but it hasn’t gone exactly as planned.

NorthPoint Development believed the scheme - which would cost $45 million - would help convince local residents to get on board with having the campus built, but many residents are calling the payment a “bribe” and have refused.

The developers have also planned a further $120 million over the next 15 years through funding for community projects and the police department if the data center is approved.

Money isn’t everything

"As a direct answer to the question 'What's in it for me?' I think we came up with something pretty radical and pretty leading-edge," said Brent Miles, NorthPoint's chief marketing officer.

Residents were sent a letter in June offering each household a $10,000 “grant” as part of the developers plans to build the campus at the foot of the Pocono Mountains, which is dotted with scenic lakes, rivers, forests, and recreational resorts.

The $10,000 grant would be a significant sum for most Hazle residents, with the median household income sitting at $60,000 - far below the national median of around $87,000.

Outside of being seen as a bribe, some residents see it as a way for the developer to cause issues between those who will be closer to the data center and those living further away.

Ed Parks, who lives adjacent to the planned site, said, “They are trying to drive a wedge between Hazle Township residents who are directly affected by the data center, like myself, and people who might live a couple of miles away” (via TechSpot).

Small town concerns match the bigger picture

Many Hazle residents are concerned about pollution, water usage, energy costs, and excessive noise caused by the site. There are also concerns that property values could drop if the site is approved, worsening economic factors for local residents even further.

These concerns align with those on the national level. The Trump Administration has made AI data centers exempt from environmental protection requirements, and numerous sites have already been caught polluting rivers and water treatment plants with chemicals and bacteria.

Residents close to existing data centers have reported symptoms of nausea and dizziness related to the noise of polluting gas burning turbines that many sites have turned to as an alternative to connecting to local grids.

A recent survey has found that the majority of Americans now believe that AI data centers are bad for the environment, which is causing serious issues for the Trump Administration as the US heads towards the midterm elections. Data centers are becoming a key issue for those looking for election or re-election - despite billions of dollars being poured into the campaigns of pro-AI candidates.

These issues are likely to persist, and more sites are likely to be delayed or cancelled by local residents - no matter how many times Trump tries to change what we call AI.