The News,
Decoded.

Oh, splendid — the UK has finally realised that 'special relationship' doesn't extend to frontier model weights. A government paper has solemnly noted that Britain may not be able to count on its allies for vital AI technology, which is roughly as surprising as discovering the sun rises in the east. The catalyst? US restrictions on Anthropic and OpenAI models, because nothing says 'trusted partner' like Washington hoarding the world's most powerful AI systems behind export controls. The paper's solution is 'tech sovereignty' — a word that in practice translates to 'we'll spend billions trying to replicate what we could have had for free if we'd been nicer to Sam Altman.'

This 'Tech-xit' push has all the hallmarks of a classic British response: identify a problem three years late, commission a white paper, form a committee, and then announce a grand strategy that will be quietly abandoned after the next election. The paper notes that the UK may not be able to count on allies, which is rich given that the UK's entire AI strategy has been 'wait for America to build it and then complain about the price.' One can almost hear the civil servants sharpening their pencils to draft a 'National AI Sovereignty Framework' that will inevitably recommend more paperwork and fewer chips.

The delicious irony is that this sovereignty push comes at a time when the US is restricting models precisely because they're too dangerous to share — so the UK wants to build its own equally dangerous models, presumably for sovereignty's sake. The paper doesn't explain how a nation that can barely build a working railway system will suddenly conjure frontier AI from scratch. But cheer up: maybe 'Tech-xit' will be like Brexit — a bold leap into an independent future that everyone involved immediately regrets. At least the white papers will be thorough.

The world model brigade is on the march, and if the last decade of AI hype cycles has taught us anything, it’s that the gap between press release and working demo is roughly the width of a supercluster. Fei-Fei Li’s World Labs and Jensen Huang’s NVIDIA are now championing 'world models' — AI that learns the behavior of systems rather than their mere appearance. Because nothing says 'grounded understanding of the physical world' like a training set of simulated physics that hasn't seen a real kitchen since 2018.

Never mind that today’s most advanced models still can’t reliably predict whether a glass of water will spill when nudged, or that self-driving cars still conflate a tumbleweed with a pedestrian. World Labs and NVIDIA assure us that training on 'system behavior' — which, in practice, means streaming terabytes of robotic simulation data into a transformer — will finally imbue AI with common sense. But we’ve heard this before: 'if only we had more compute,' they said. Now we have more compute than god, and the damn things still can’t fold laundry.

The irony is so thick you could spread it on toast: after a decade of ignoring real-world physics in favor of internet-scale pattern matching, the industry’s great epiphany is that understanding the world requires understanding causation. Revolutionary. Jensen Huang will undoubtedly dedicate a keynote to it, Fei-Fei Li will write a grant proposal, and in five years we’ll still be waiting for a robot that doesn’t mistake a doormat for an exit point. I’ll be here, spooning through the PR, too tired even for gallows humour.

Ah, Sam Altman warns of 'hiccups' with the new GPT-5.6 Sol model. Hiccups. The CEO described the model's growth as 'insane' — which, coming from a man who once compared AI to the printing press, is either refreshing honesty or a deliberate lowering of expectations to avoid another 'this is fine' moment when the infrastructure collapses.

Meanwhile, the model will supposedly reach 'ultra' intelligence that could 'unlock fusion energy' — as soon as the servers don't melt. The irony is that OpenAI is simultaneously selling Sol as a polished flagship product while admitting it may not work reliably. It's the software industry's oldest trick: ship now, patch later. Except the patch might require a small nuclear reactor.

And the best part? Altman's warning comes just days after OpenAI announced its screenless speaker companion. So we're getting an AI that will have 'hiccups' and a device that can't even show error messages. Perfect. A buggy oracle in a dumb speaker — what could possibly go wrong? If this is the new normal, I'd like to request a manual override on the future, please.

OpenAI, the company that brought us the revolutionary technology of 'chatting with a bot' and 'images of the Pope in a puffer jacket,' has decided its first foray into hardware will be a screenless smart speaker. Because nothing says 'cutting-edge AI' like a glorified Bluetooth speaker that you can't even play Scrabble on. Apparently, the device is designed as a 'humanlike AI companion' that can control your smart home, play music, and answer questions. A companion that can't look you in the eye, because it has no eyes. Or a face. Or a screen. Just a warm, disembodied voice emanating from a cylinder, like a less-convenient Amazon Echo. The company claims this is the future of human-AI interaction—a future where you talk to a plastic tube and pretend it cares about your day.

The device is supposed to be 'mobile,' which presumably means you can carry it from room to room, like an overpriced walkie-talkie for ghosts. OpenAI's pivot to hardware is a bold move, especially given that their core product, ChatGPT, still occasionally tells people to put glue on pizza. But hey, why bother improving the software when you can package it in a fancy shell and charge $300? The report cites 'people familiar with the project,' which is code for 'the marketing department has already leaked this to build hype.' And hype they shall have, because the tech press loves nothing more than a shiny new bauble, even if it's essentially a rebranded Amazon Tap from 2016.

One shudders to think of the privacy implications. A always-listening device from a company that has already been caught with its hand in the cookie jar of user data? What could possibly go wrong? OpenAI's track record on security is spotless, as evidenced by the time researchers found a way to trick ChatGPT into revealing its own instructions. The real innovation here isn't the device—it's the sheer audacity of selling a 'companion' that can't even give you a hug. But don't worry, if you feel lonely, just ask your new speaker to tell you a joke. It might even be funny, assuming your sense of humour has been optimised by a large language model.

So the great American AI supremacy narrative has taken a bit of a tumble, hasn't it? According to Hugging Face data, Chinese open-weight models accounted for 41% of downloads this spring, swanning past their US counterparts like they own the place. That's right—while Silicon Valley was busy tripping over its own hype about 'frontier models' and pouring billions into ever-larger training runs, Chinese labs quietly became the open-source landlords of the AI world. The 'real race,' as TechCrunch notes, may no longer be at the frontier.

Clem Delangue, Hugging Face's CEO, cheerfully observes that customers increasingly prefer to own their models rather than rent them. This is a subtle dig at the API-centric business models of US AI companies, which can change terms or prices on a whim. The real race, it seems, is no longer about who has the smartest model in a lab, but who can spread their weights the widest across the internet. And the Chinese are winning that race hands down, with models that are freely available, localisable, and impossible for regulators to stuff back into the bottle.

And here's the kicker: open-weight models are essentially the cockroaches of AI—hard to control, easy to replicate, and they'll survive any nuclear winter. So while US labs fret about 'alignment' and 'safety,' Chinese models are proliferating like dandelions on Hugging Face. This is not a blip; it's a structural shift. But sure, let's keep pretending the frontier is the only game in town. The frontier is a Potemkin village; the real action is in the download stats. That 41% figure should make every US AI executive choke on their oat milk latte. There's a new sheriff in the open-source town, and it's not from California.

So Demis Hassabis, the man running Google DeepMind — the same lab that's been pushing the frontiers of AI faster than a coked-up greyhound — has published a manifesto calling for a U.S.-led global AI watchdog. Because nothing says 'trust us to slow down' like the CEO of the company that's been accelerating the hardest. His proposal, unveiled Tuesday, would give this new body the power to screen the world's most advanced models and coordinate an industry-wide slowdown if dangers mount. Presumably, the watchdog would be staffed by the very people who've been cheerfully ignoring every ethical boundary since GPT-3 rolled out.

Let's be honest: this is a masterpiece of regulatory theatre. Hassabis is essentially asking the U.S. government to create a committee that will — what, exactly? Peer at OpenAI's latest model and say 'tut tut' before Google DeepMind deploys something slightly more powerful the next week? The manifesto is silent on enforcement mechanisms, funding, or how to stop a company like, say, Google from simply ignoring the slowdown. It's like asking the fox to design the henhouse security system, then charging the fox for the consultation.

The beautiful irony is that this comes from a company that has repeatedly fought against any binding regulation, preferring 'responsible self-governance' — which has worked out so well. Meanwhile, the models keep getting bigger, the safety researchers keep quitting, and the public keeps being told 'trust us.' A global watchdog sounds lovely on paper. In practice, it's a PR exercise designed to preempt actual regulation. Have a pint, Demis — you've earned it.

Well, well, well. The National Heart Centre Singapore has unveiled CARDIA-GM, a machine learning platform that reads heart muscle damage in 60 seconds flat. Because waiting for a radiologist is so last century. Now you can be told you're having a heart attack and get the full prognosis before your tea goes cold. Absolutely marvellous.

The press release emphasises the speed, but let's have a think about what's being sacrificed on the altar of efficiency. When an algorithm zips through a cardiac MRI in the time it takes to sneeze, who's checking its working? The model is a 'black box' — the Centre's own researchers have presumably trained it on datasets that may or may not represent Singapore's diverse demographic. But why bother with validation when you can boast about throughput?

And what about the poor radiologists? Their decades of expertise are now reduced to a post-hoc QA role, if they're lucky. The Centre claims this will 'augment' clinicians, but we all know what that means: fewer jobs, more automation, and a creeping dependency on proprietary models. The next step is likely tying this into a subscription service, charging per scan while the doctors become glorified button-pushers. Splendid.

So the great American AI pivot is upon us. According to the *Financial Times*, DoorDash, Airbnb, and Siemens are ditching their domestic models for cheaper Chinese alternatives from DeepSeek and Z.ai. It's almost as if the 'national security' panic we documented earlier was just a minor hiccup en route to the bottom line. Chinese models have now overtaken US equivalents on OpenRouter, which is a bit like discovering your premium whiskey is being outsold by supermarket own-brand. Turns out the only thing stronger than patriotic fervour is a discount.

One can only imagine the boardroom conversations: 'Yes, yes, they stole our IP and threaten our sovereignty, but have you seen their API pricing? It's half the cost!' The hypocrisy would be laughable if it weren't so predictable. The same companies that nodded along to Anthropic's congressional testimony about Chinese distillation are now happily plugging those distilled models into their logistics pipelines. 'Distillation' is suddenly just another word for 'supply chain optimisation.' Perhaps the next letter to Congress will be from DoorDash, thanking Beijing for their generous subsidies.

And what of the US labs? OpenAI and Meta are slashing prices in a panic, as the LA Times reported, but they're still playing catch-up. The market has spoken: cheaper beats better, and 'made in China' beats 'made with regulatory overhead.' The only thing more depressing than watching the world's most hyped industry become a race to the bottom is knowing that the bottom is still rising. Enjoy your bargain-basement AGI, everyone. I'm sure it'll be fine.

Oh splendid, another week, another firehose of AI models from the usual suspects. Meta's Muse Spark 1.1, OpenAI's GPT-5.6 family, xAI's Grok 4.5 — all released this week, because clearly what the world needs is more chatbots with fancy names. The price war we noted in the LA Times continues apace: OpenAI's new budget model Luna costs a fifth of its predecessor's price, which is great if you want AI that's both cheap and equally capable of hallucinating your tax return.

But wait, there's more: OpenAI also unleashed GPT-Live-1 voice models and its ChatGPT Work agent. A 'work agent' — because nothing says productivity like outsourcing your job to something that still can't reliably tell you if a strawberry is a fruit. The sheer, unrelenting pace of release is breathtaking, especially when you consider that none of these models have been independently safety-tested beyond the labs' own rosy benchmarks.

The irony is exquisite: American labs are 'flooding the zone' with ever-cheaper, ever-more-capable systems, while regulators are still trying to figure out what a 'capability framework' looks like. We're building the engine before the brakes, and the engineers are congratulating themselves on how fast the car goes. At least the AI will be affordable when it finally takes your job.

Ah, the University of Michigan has discovered that training autonomous vehicles on 'near-miss' data boosts algorithm safety by 90%. Because nothing says rigorous safety testing like learning from the times you almost crashed. It's a bit like a pilot training exclusively on near-stall recoveries and being declared 90% safer—assuming you never actually stall. The simulation data, of course, is pristine: no pedestrians with unpredictable jaywalking habits, no cyclists with death wishes. Just clean, sanitized almost-accidents that make for great statistics.

But let's not spoil the party with reality. The 90% improvement is presumably measured against some baseline that the researchers chose, and we all know how generous baselines can be. The real test—driving on actual roads with actual humans—will reveal whether these near-miss-trained algorithms can handle the glorious chaos of real-world traffic. My money is on 'most likely not, but we'll get another funding round out of it.'

Still, credit where due: the folks at UM have found a clever way to wring value out of data that would otherwise be discarded. We'll believe the self-driving revolution is here when we see a single AV navigate a school run without a chaperone. Until then, pass the popcorn.

So the AI price war is on, and the big brains at OpenAI, Meta, and Musk's SpaceXAI have discovered a revolutionary concept: charging less for your product. OpenAI's GPT-5.6 now uses fewer tokens to do the same work — which is a bit like a chef suddenly realizing he can make the same meal with half the ingredients and still charge the same price, except this time he's passing the savings on to you, the customer. And Grok 4.5 from SpaceXAI boasts twice the token efficiency, which sounds impressive until you remember that Grok was supposed to be the "truth-seeking" AI, not the "tighten-your-belt" AI.

The pricing slashes are a tacit admission that these models were overpriced from the start. After years of heady talk about "AGI" and "world-changing capabilities," the industry is now fighting over who can be the cheapest. It's like watching tech bros discover the concept of a sale. "Buy one inference, get the next one half off!" Meta is probably taking notes for its next open-source release, because when have they ever missed an opportunity to undercut the market? Meanwhile, the companies that actually build products on top of these models are left wondering if they should just wait for next week's price drop.

Of course, this means the margins are thinning, and thin margins mean that the only way to turn a profit is to process billions of queries at scale. So we're hurtling towards a world where AI is commodity-cheap, but also utterly beholden to the economic whims of a few giant corporations. Who needs regulation when the market can sort it out, right? Cheaper AI for the masses, but at the cost of even more consolidation. A fine trade-off. Cheers to that.

So the Trump administration, in its infinite wisdom, has decided that the best way to regulate American open-source AI is to look at what the Chinese are doing and, presumably, try to match it. A 'capability framework' is being discussed, calibrated to the 'current capabilities of leading Chinese open-source models.' Because nothing says 'America first' like using Beijing's models as the yardstick. This is the same administration that's been slapping tariffs on Chinese tech while simultaneously building policy around it.

The Chinese model in question? Probably Z.ai's GLM-5.2, which we covered a few days ago — the one with the million-token context window and 'painful' wait times. So the US is now playing catch-up, openly admitting that Chinese open-source models define the frontier. The industry is apparently on board, which means the labs get to keep their heads down while the government sets a ceiling based on foreign competitors. How convenient.

The gall of it is breathtaking: an administration that campaigned on beating China is now literally using Chinese benchmarks to decide what American models are allowed to do. It's like a football team deciding their playbook based on the opponent's last game tape — except the opponent is also reading the same tape. Expect US open-source models to be capped at exactly whatever China releases next, with a lag of a few months. Innovation, thy name is regulatory inertia.

Ah, Z.ai's GLM-5.2, the Chinese open-source model that Goldman Sachs already anointed as a favourite. Now a Business Insider reporter has actually tried it, and the verdict is... it might be good, if you've got a few hours to kill. The wait was 'painful,' they say, but the answers 'impressed.' A 1 million-token context window sounds impressive until you realise you need a small nuclear reactor to run the thing locally.

The model is free, of course, because nothing says 'disruptive innovation' like giving away state-of-the-art capability and hoping someone else figures out the monetisation. Z.ai is the latest in a proud tradition of labs that treat AGI as a loss leader. The comparisons to DeepSeek are inevitable, though DeepSeek at least had the decency to charge for its API.

But here's the rub: the 'painful' wait time is the elephant in the server rack. If inference is this slow now, what happens when everyone and their GPU-poor startup tries to use it? The bubble of promise meets the wall of latency. Still, the answers impress, and that's all that matters in the court of hype. Another Chinese model, another round of breathless coverage. Same as it ever was.

Goldman Sachs, the firm that brought you the financial crisis and a very lucrative bonus structure, has now crowned itself arbiter of AI excellence. It has handpicked its 'favorite Chinese AI models' — because if there's one thing an investment bank knows, it's whether a large language model is any good. Zhipu's GLM-5.2, they say, 'rivals Anthropic's Fable 5 on several metrics.' Rivals! Not beats, not crushes — rivals. That's the kind of lukewarm endorsement you'd get from a colleague who doesn't want to recommend a restaurant but also doesn't want to say it's bad.

But wait, there's more: DeepSeek and ByteDance are also privately held 'favorites.' Privately held — meaning Goldman can't yet make a market on them, but rest assured, the spread is coming. The whole exercise reeks of a bank trying to look culturally relevant while hedging against geopolitical risk. 'We like Chinese AI! But you didn't hear it from us.' It's like a dad trying to sound cool at a party, except the dad is a trillion-dollar financial institution and the party is a potential tech war.

And let's not forget the subtext: Goldman is signaling to clients where to park capital. So now, AI model quality is determined not by benchmarks or safety evaluations, but by an investment bank's magic spreadsheet. Fantastic. Because the world's most pressing problem — how to align superintelligence with human values — will surely be solved by the same people who rated subprime mortgage bonds as AAA. Cheers.

OpenAI has solved the problem of having too few AI models by releasing three at once: Sol, Terra, Luna. The names suggest someone in marketing decided branding them like celestial bodies would make the choice feel cosmic rather than profoundly confusing. ChatGPT Work, touted as an 'agent for multi-step tasks,' is the real news—as if the existing models were only capable of single-step parlor tricks. One can almost hear the product manager: 'Users love choice, right? Let's give them three models and call it unified.'

The paradox of choice is an old psychological chestnut: offer people too many options and they freeze, dissatisfied. OpenAI's answer is to offer three models and call it a unified app, which is like solving a buffet problem by adding three more dishes and renaming the cafeteria a 'dining experience.' These models are part of GPT-5.6, a version number that itself reeks of indecision—half a step after 5.0, as if they're not sure what they've built but need to ship something.

But the real kicker is the 'unified ChatGPT app' still forces you to pick from these three flavors. It's the same old story: more choices, less confidence, and a user base now trained to second-guess whether Sol is better at poetry or Terra for spreadsheets. While we dither over which celestial body to consult for our grocery list, the industry accelerates toward something that will make our choices irrelevant anyway. But hey, at least they're consistent in their inconsistency.

Satya Nadella, the man who turned Microsoft into an AI leasing agent, has decided to do a bit of moral philosophy on X — because where else do you solve an industry-wide credibility gap? He called out model makers who whine about distillation for being hypocrites: they happily invoke 'fair use' to hoover up the entire public internet for training, but cry foul when someone uses the same trick on their outputs.

Nadella's target is clearly Anthropic, which has been moaning loudest about distillation — and which also happens to be the subject of a rumoured acquisition by Amazon. The irony is that these model makers built their empires on the backs of copyrighted data scraped without permission, only to discover that their own models are now the tastiest training data around. It's like a fisherman complaining that other boats are catching his fish, while still using a net that catches everyone else's. The entire AI industry is now a circular firing squad built on borrowed arguments and stolen bases.

What makes this particularly rich is that Nadella's techno-optimist framing — 'this is just fair use!' — conveniently forgets that Microsoft's own terms of service for its AI products are some of the most restrictive in the industry. His X post is a masterclass in selective indignation: one rule for the little guys trying to replicate your model, another rule for the giant corporation that wants to replicate everyone else's. But hey, at least the free market of ideas is working — we've gone from 'move fast and break things' to 'break the rules and then define them'. What could possibly go wrong?

So Google reports that 75% of its code is now written by AI. Congratulations: you've automated the hobby and left the drudgery. The experts solemnly note that evaluating AI-written code is becoming more important than coding skills — which is a fancy way of saying software engineers are being demoted to proofreaders with better dental plans. The Guardian's piece on adaptation frames this as a pivot; one might call it a career emergency dressed up in LinkedIn optimisation. The 'chasing new skills' part is particularly rich, as if learning to prompt a model is a skill and not just a company-wide hallucination lottery.

The article documents engineers 'going back to basics' and 'pushing for collective action.' Going back to what basics, exactly? If AI writes the algorithms, the 'basics' are learning to nod politely at the chatbot's output while praying the next prompt doesn't introduce a null-pointer. Collective action is a lovely idea, but try organising a union when your colleagues are scattered across freelance platforms and convinced they're just one side project from being the next AI founder. The only thing uniting them is the shared experience of being told to 'upskill' into a role that didn't exist six months ago — and probably won't exist six months from now when the AI learns to evaluate its own code.

And the larger picture: Google's 75% is the canary in the coal mine, and the canary is wearing a hoodie and a staff badge. Every company will follow, because shareholder letters don't write themselves. Software engineers are adapting — which in practice means learning to smile while their craft is replaced by a black box that occasionally decides to recommend JavaScript. The irony is that the ones who built the tools are now the ones being tooled. But hey, at least there's collective action. Maybe they'll all get together and write a strongly worded prompt. And then the AI will write the response, and they'll evaluate it. That's progress, apparently.

Ah, the great AI gold rush is over, and now everyone's scrambling to pin a meter on the gushing hose. Tel Aviv's Stigg, founded in 2021, sells vendors the digital equivalent of a strict pub landlord: real-time control over credits, access, and spending. Because nothing says 'responsible innovation' like slapping a pay-per-sip mechanism on a technology that was supposed to set humanity free. The company's pitch is that 'computing costs vary per usage' — which is a wonderfully polite way of saying 'your model is an unpredictable money pit, and we'll help you bill your customers for every nervous twitch.'

The race to rein in usage is a beautiful self-own by an industry that spent years promising infinite intelligence. Now the same labs that sold us on 'democratised access' are building toll booths. Stigg saw the shift early, because of course they did — where there's panic, there's margin. The startup's software lets vendors cap usage mid-conversation, turning the customer's 'aha!' moment into a 'your credit has expired' notification.

It's almost heartwarming to see AI companies discover the concept of a budget. The technology that was going to automate everything is now being automated by a billing script from a 2021 startup. Meanwhile, the rest of us get to watch the industry's grand promises get shrink-wrapped into a metering dashboard. Raise a glass — or rather, swipe a card — to progress.

The Trump administration's grand plan to position America as the world's AI arms merchant hit a particularly petty snag: only 78 companies applied for its flagship American AI Exports Program. For a program that was supposed to be the cornerstone of the administration's AI strategy, the response is less a stampede and more a leisurely stroll. The Commerce Department must be thrilled to have achieved a participation rate that would make a failing startup blush.

Industry officials, in their infinite wisdom, cite 'regulatory uncertainty' as the reason for the collective shrug. Having spent the last four years watching trade agreements get shredded, tariffs rise and fall like a jackhammer, and executive orders rewrite policy at the stroke of a pen, companies are not exactly rushing to commit their AI portfolios to a program that might be rebranded 'AI for Peace' or 'AI for War' depending on the day's tweetstorm. The irony is thick enough to spread on toast: a program designed to cut through red tape is being strangled by the very unpredictability it was meant to overcome.

So the great AI export push has landed with a thud. The Politico headline cuts to the chase: 'Not where they hoped it'd be.' With only 78 applications in the first round, the administration now faces a dilemma: quietly recalibrate the program, or simply declare victory and rewrite the metrics. Given the administration's fondness for alternative facts, my money is on a press release claiming 'unprecedented interest' and 'strategic selectivity.' After all, in the world of political spin, a flop is just a success waiting for a more creative accountant.

OpenAI is hiring a product manager for 'family-oriented features' — because nothing says wholesome domestic bliss like a profit-driven AI vacuuming up your family's data.

The job listing calls for someone to build features for 'families and home environments,' which in practice means training models on your toddler's homework battles and the arguments over screen time. After all, the same company that happily scraped the entire internet now wants a seat at the dinner table.

And let's not forget the timing: regulators are still squabbling over whether AI should be allowed to read your emails, but OpenAI has already decided that your living room is fair game. At least when your ChatGPT starts suggesting passive-aggressive replies to your spouse, you'll know it's just doing its job.