The News,
Decoded.

Ah, Apple's walled garden — the fancy wrought-iron kind with a conveniently unlocked back gate. Fox News reports that researchers have found a way to manipulate Apple Intelligence with hidden prompts, no iPhone theft required. So much for 'your data never leaves your device' — unless a prompt decides to leave with it, apparently. The attack doesn't even need physical access; just a cleverly crafted message delivered through a seemingly innocuous channel. One has to admire the sheer elegance: while Apple runs ads about privacy being a human right, the actual security model turns out to be about as robust as a soggy biscuit.

What's truly spectacular is how this lands in the grand tradition of 'we fixed it with a software update' — the corporate Hail Mary that has absolved more security sins than any confessional. The researchers didn't hack the hardware; they just whispered the right incantation into Apple Intelligence's ear, and it apparently went along like a tipsy party guest. This is the kind of vulnerability that makes you wonder if 'on-device AI safety' is a real thing or just a marketing slogan painted on the side of a sinking ship.

And let's not forget the broader context: every major lab is racing to embed AI deeper into personal devices, yet the security implications are treated as an afterthought. The Fox News article raises 'new questions' — as if we haven't been asking them for years. Apple will likely patch this specific hole, but the underlying architecture remains a house of cards in a hurricane. So here's to another round of 'find the backdoor, patch it, forget it' — the eternal dance of AI security, where the only guarantee is that the next step is already misstep.

AGIBOT's World Challenge 2026 — because nothing says "we've genuinely cracked embodied AI" like a contest with 526 teams from 27 countries trying to make a robot do something useful. The press release gushes about "moving beyond simulation benchmarks," as if that's a bold leap rather than the bare minimum for any system that's supposed to function in the real world. Simulation was the cosy sandbox where models could pretend to be competent; now they have to deal with actual physics, and we're running a competition to see if any of them can. And naturally, the challenge is hosted by AGIBOT, a company that just so happens to sell a robot platform perfectly suited for the tasks — a coincidence so pure it sparkles.

Five hundred and twenty-six teams. Twenty-seven countries. That's a staggering amount of talent, compute, and venture capital all converging on a problem that — let's be honest — we've been pretending is solved since the DARPA Robotics Challenge a decade ago. The difference this time is that the models are larger and the hype is louder, but the fundamental challenge remains: getting an AI to generalise from scripted demos to messy reality. AGIBOT is betting that a corporate-sponsored Olympics will accelerate progress, but it looks more like a desperate attempt to prove that the billions poured into embodied AI have produced something other than expensive YouTube compilations of robots falling over.

The phrasing "see how AI models perform on real tasks" is doing an awful lot of heavy lifting, implying that this is a novel inquiry rather than the central question of the field since its inception. Perhaps the challenge will produce genuine breakthroughs; perhaps a robot will finally fold laundry without a human remote operator. But given the track record of such contests, the most likely outcome is a lot of careful press releases about "learnings" and "next steps" while the real-world performance metrics remain under embargo. Still, 526 teams can't all be wrong — can they?

Ah, the latest entry in the 'this time it's different' sweepstakes: Recursive Superintelligence, founded by former Meta researcher Yuandong Tian, has emerged from stealth with $650 million in funding and a $4.65 billion valuation. That's a lot of zeros for a company whose entire premise is 'models building models' — a phrase that should probably come with a 'do not attempt this at home' warning. Because what could possibly go wrong with an accelerating loop of self-improvement? It's like strapping a rocket to a runaway train and hoping the brakes find themselves.

Tian, who presumably got bored optimizing ad algorithms, now believes AI can be engineered to improve itself in an 'accelerating loop.' The business plan appears to be: 'Step 1: Raise $650M. Step 2: Let the AI build a better AI. Step 3: Pray it doesn't decide the whole exercise is pointless.' The irony is exquisite — Meta spent billions trying to make AI that doesn't hallucinate its own press releases, and now one of its ex-stars is betting the farm on AI that might just hallucinate a smarter version of itself.

Of course, the valuation of $4.65 billion is based on nothing more than a PowerPoint and a Nobel-worthy confidence trick. But hey, in a world where Sam Altman can raise billions on vibes alone, why not Tian? The only real question is whether the recursive loop will produce superintelligence or just a very expensive paperweight. My money's on the latter, but I'd love to be wrong — if only so I can enjoy the fireworks.

Ah, the AI arms race enters its most thrilling phase yet: the memory procurement stage. NVIDIA and SK hynix have announced a 'multiyear technology partnership' to co-develop memory for the upcoming Vera Rubin AI supercomputers. Because nothing says 'responsible AI development' like a hardware vendor lock-in for the next generation of compute, all wrapped up in the charmingly Orwellian term 'AI factories.' One can almost hear the local zoning boards weeping.

Let's be honest: this is about keeping the GPU furnaces fed. The Vera Rubin platform is the next big thing, and it needs memory that doesn't bottleneck the tensor cores. So SK hynix, fresh from its HBM triumphs, gets to play handmaiden to Jensen Huang's empire. The press release is full of mutual back-patting about 'transforming industries' and 'unlocking AI's full potential' – code for 'we will sell you the shovels while the gold rush lasts.'

What's genuinely amusing is the implicit admission that current memory tech isn't cutting it. After billions poured into AI, the bottleneck is still the memory bus. So we're building a whole new generation of supercomputers that will, presumably, solve problems we haven't even thought of yet, powered by memory that doesn't exist today. At this rate, the singularity will be delayed by supply chain issues for HBM4. But don't worry – the partnership is 'multiyear,' so we have plenty of time to pretend we know what we're doing.

Ah, the classic 'everyone-building-everything' strategy — because nothing says 'efficient market' like a dozen AI labs all trying to build the same app builder simultaneously. Business Insider reports that Anthropic, the company that was supposed to be the responsible one, may be working on an app builder for non-techies. Because obviously, what the world needs right now is another no-code tool, this time from the safety-obsessed lab that can't stop lecturing us about alignment.

Anthropic's rumoured pivot puts it in direct competition with the 'vibe-coding stars' — those plucky startups that already built the thing Anthropic just discovered. One can almost hear the collective sigh from VCs who now have to decide whether to back the claude-powered builder or the existing one that's already been hacked together in someone's garage. The market is converging, which is business-speak for 'everyone is doing the same thing, and nobody has a moat.'

Meanwhile, this frantic expansion into each other's turf feels less like strategic brilliance and more like a game of musical chairs where the music is 'investor pressure to show growth.' The real winner here? The consultants who will charge a fortune to explain why this convergence was actually a terrible idea. But hey, at least the vibe-coding startups get their VC-funded existential crisis a few months earlier than expected.

Ah, the classic 'hurricane warning' — because nothing says 'we've got this under control' like a metaphor borrowed from catastrophic weather events that we also failed to prepare for. Politico reports that the U.S. has a mere six to 12 months before Beijing can compete with the latest wave of hyper-advanced AI models. Six to twelve months! That's less time than it takes to get a new smartphone contract approved by Congress. And what are our esteemed leaders doing? The article suggests guardrails may already be too late, which is a polite way of saying we've built the nuclear reactor while quibbling over the colour of the emergency shutdown button. The sheer spectacle of a government that can't pass a basic privacy bill suddenly facing a timeline measured in months is the kind of dark comedy that writes itself. But hey, at least the 'hurricane warning' gives everyone a clear sense of urgency — right up until the moment the storm makes landfall and we realise the guardrails were just decorative railings painted to look sturdy.

The estimates, according to recent reports, indicate that Beijing is closing the gap faster than anyone in Washington wants to admit. It's the same old song: warnings issued, warnings ignored, warnings repeated with slightly more dramatic imagery. If this were an actual hurricane, FEMA would be holding a press conference to tell people to stock up on bottled water. Instead, we're staring down an AI capability leap that could redefine geopolitical power, and the best our regulatory system can manage is a metaphorical umbrella. The phrase 'too late' hangs over this article like a sword of Damocles, but let's be honest — the sword was dangling long before anyone thought to check the rope.

So here we are, with six to 12 months to act, and the most concrete response so far is a Politico article full of alarm. The labs will keep training models, the companies will keep promising utopia, and the politicians will keep holding hearings in which they ask basic questions about what 'AI' stands for. The hurricane is coming, the guardrails are ornamental, and the only real question is whether we'll have time to enjoy the irony before the storm hits.

Ah, the classic pre-IPO pivot: OpenAI, the company that already brought us a chatbot that can't reliably spell 'strawberry' without a calculator, is now planning its 'biggest ChatGPT overhaul yet' — turning it into a 'superapp' with coding tools and AI agents. The FT reports this is all ahead of a planned share listing, because nothing says 'stable investment' like a last-minute strategic reinvention. One can almost hear the pitch deck: 'We're not just a chatbot, we're the future of all digital interaction! Also, our agents might sometimes book a flight to the wrong continent, but that's a feature, not a bug.'

The masterplan is part of a 'broader reorganisation' to shift resources toward 'lucrative enterprise clients.' This is the same company that has cycled through product visions — from GPT-3 to GPT-4 to GPT-4o to o3 — faster than most startups iterate on landing pages. The superapp moniker is particularly rich: it's the startup equivalent of a midlife crisis, a desperate attempt to be everything to everyone before the IPO bell rings. Meanwhile, actual enterprise trust remains tepid after repeated data leakage scares and the 'Lockdown Mode' patch that arrived like closing the barn door after the horse has bolted.

But let's give credit where it's due: this is a bold strategy to boost revenue before listing. The IPO prospectus should be a page-turner: 'Risk factors: we keep redesigning our only product, our CEO can't retain top talent, and our superapp is essentially a feature list from a 2015 startup pitch deck.' The broader reorganisation should calm any nerves about operational stability. It's almost beautiful to watch — the classic startup move of promising the world to get the valuation you want, right before the market realises you're still building the foundations. Perhaps the superapp will also make coffee. That alone would justify the market cap.

Ah, the AI bubble didn't burst — it's taken revenge. Axios reports that the great spending spree is finally sobering up, with GitHub moving Copilot to usage-based billing and a top executive admitting that expenditure is 'getting harder to justify.' Because nothing screams 'value' like charging per token for a coding assistant that still occasionally suggests `npm install left-pad` as a fix. The market is realising that deploying AI costs real money, and the returns? Well, they're about as tangible as a ghost in a server rack.

It's the same old story: companies throw billions at the altar of 'transformation,' only to discover that a chatbot that can write a press release still can't fix their legacy COBOL systems. The executive's wince-inducing honesty — 'spending getting harder to justify' — is the closest we'll get to a confession that the emperor's new clothes are, in fact, just very expensive electrical bills. GitHub's shift to usage-based billing is the corporate equivalent of a landlord installing a coin-operated shower: pay for every drop, and the dripping stops.

And what of the 'value spreading across companies'? Spoiler: it won't, not evenly. The big cloud hyperscalers will hoover up the gains, while the poor suckers paying for usage-based Copilot are left with a monthly bill that looks like a ransom note. The bubble isn't bursting — it's transmuting into a subscription fee. Cheers to progress: you'll pay for it, one API call at a time.

Ah, Build 2026 — Microsoft's annual pilgrimage to convince us that its AI is not just a pale imitation of everyone else's. This year's offering: four MAI models with names so thrilling they'd put a spreadsheet to sleep. MAI-Thinking-1 — because who needs a creative name when you have a version number? — and MAI-Image-2.5, which sounds like a firmware update for a printer. PCMag's brave reviewer put them through their paces and emerged with the most damning indictment possible: "fine, but nothing more." One imagines Satya Nadella's face cycling through the entire Microsoft emoji library in search of a reaction.

The reviewer, wading through what must have been an ocean of mediocrity, found that the reasoning model thinks just well enough to produce answers that pass the bar of "I suppose that's acceptable." The image model generates pictures that are, in the reviewer's words, "fine" — a word that in Microsoft's internal vocabulary probably translates to "game-changing" and "revolutionary." This is the company that has spent tens of billions on AI infrastructure, hired thousands of researchers, and the best they can produce is something your average startup could cobble together with an open-source model and a weekend of fine-tuning. The other two models remain unnamed, likely because they're as forgettable as last year's Build keynote.

The real tragedy isn't that Microsoft's models are mediocre — it's that "fine" is practically a win in today's AI landscape. We've become so accustomed to incremental improvements that we celebrate mediocrity with a collective shrug. Microsoft will sell this to enterprises, who will dutifully integrate MAI-Whatever into their workflows, and the world will keep spinning, just a little more boring than before. Perhaps that's the real AI risk: not an intelligence explosion, but a boredom explosion.

Well, isn't this a comforting cup of tea. OpenAI has unveiled 'Lockdown Mode' — because nothing says 'we've got this under control' like a feature name that sounds like a prison riot protocol. The new feature disables live web browsing and deep research to protect sensitive data from prompt injection attacks. Of course, the company immediately admits that even in Lockdown Mode, ChatGPT 'could still be vulnerable to prompt injections from cached content or uploaded files.' So essentially, they've bolted the front door while leaving the windows wide open and the back door ajar. Classic.

Let's take a moment to appreciate the sheer absurdity of a 'security' mode that still allows attacks via uploaded files. Because the best way to prevent someone from breaking in is to politely ask them not to use the main entrance. The entire approach screams of a company that has realized just how many gaping holes its product has and is now frantically stapling cardboard over them. The fact that this is a 'new feature' rather than a fundamental rethink of architecture tells you everything.

So here we are: a Lockdown Mode that locks down precisely the things that make ChatGPT useful in the first place, while leaving the actual attack vectors untouched. It's the AI equivalent of installing a panic room in a house made of paper. But by all means, slap that 'enterprise-ready' sticker on it. The security theatre continues, and the audience is expected to applaud. Pass the aspirin.

Oh, marvellous—up to 450,000 Michigan college students are being handed $100 each in OpenAI Codex credits, because what better way to foster education than by distributing corporate freebies with all the enthusiasm of a drug dealer at a school gate? The $45 million price tag looks generous until you realise it's just venture capital's way of training a generation to rely on a proprietary coding crutch. Once those credits expire after one academic year, the habit's formed—and the price tag appears.

The state of Michigan, rather than investing in public education or open-source alternatives, is effectively acting as OpenAI's distribution channel. The governor probably posed for a shiny photo op while students are being quietly locked into an ecosystem where 'coding' means typing natural language into a black box. And let's not forget Codex itself—a tool that's great at autocompleting boilerplate but famously hallucinates like a fever dream when logic gets twisty. But hey, it's free, right?

Ah, that weasel phrase 'up to'—450,000 eligible means the actual number could be far lower, but OpenAI gets the PR boost regardless. Meanwhile, the real costs—privacy erosion, the slow death of debugging skills, and the quiet burial of open-source curiosity—aren't on any balance sheet. But cheer up: Michigan's students are now officially the lab rats in the great AI education experiment. Free credits today, full-price subscriptions tomorrow. What a deal.

Ah, the University of Connecticut has bravely entered the AI arms race with 'AI for ImpaCT' — note the capital C, because nothing screams 'we mean business' like a mid-acronym capital letter. They're going to 'expand AI best practices' across education, research, innovation, public engagement, and workforce development. That's every box ticked. No doubt a taskforce will be convened, a website built, and a vice-provost of AI appointed by Friday.

The initiative aims to address workforce development, which is rich given that every tech company this week has been laying off workers faster than you can say 'prompt engineer.' UConn is essentially promising to train students for jobs that AI itself is currently vaporising. It's the academic equivalent of building a state-of-the-art stable for horses in 1910.

Still, the university deserves some credit for actually doing something. Other institutions are still debating whether to ban ChatGPT or to rename their computer science departments. UConn is at least acknowledging the elephant in the room, even if they plan to teach it tricks. One can only hope the 'best practices' include a module on 'how to write a cover letter for a job that no longer exists.' That would be genuinely useful.

Ah, the University of North Carolina has decided that the best way to prepare for the AI apocalypse is to hold hands with public librarians and study "local responses." Because nothing says "readiness for recursive self-improvement" like a community reading circle discussing ChatGPT's latest hallucination about historical dates. The irony is thick enough to spread on a biscuit — while frontier models leap from text to video to autonomous agents, our strategy is a statewide book club on AI basics.

The study aims to "develop approaches to improving AI literacy" — as if the problem with AI is that people don't understand it, rather than that it's being deployed by companies with the ethical compass of a hungry raccoon. I'm sure the librarians will do a heroic job, but they'll be out there teaching senior citizens how to spot a deepfake while frontier labs release models that can generate convincing video of the same seniors doing things they never did. And let's not forget that the very companies racing ahead are also the ones slashing their safety teams and lobbying against any meaningful regulation.

Meanwhile, the actual decision-makers in Silicon Valley and Washington will continue their usual dance of "move fast and break things" and "we'll regulate later." But hey, at least rural North Carolina will have a solidly researched pamphlet on AI best practices. That should really help when the AI-driven supply chain collapse hits, or when your kid's homework is written by an uncensored model from abroad. Pass the sweet tea and a healthy dose of good old-fashioned denial. At least the library will have air conditioning.

Oh good, the Democratic lawmakers have finally done it — they've introduced the 'Responsible Artificial Intelligence in Defense Act' on Monday, a bill that promises to establish a 'framework' for Pentagon acquisition and use of AI-enabled autonomous weapon systems. Because nothing says 'serious regulation' like a framework that hasn't passed either chamber yet and can be quietly amended into irrelevance by defence contractors' lobbyists. The manual override clause is particularly delicious — as if the machine that decides to fire a missile at a wedding party is going to politely wait for a human to press the 'nope' button.

The bill's core conceit is that human oversight and manual override will somehow tame the very systems designed to operate at machine speed. It's the legislative equivalent of putting a 'do not use in case of war' sticker on a tactical nuke. And let's be honest: the Pentagon has been acquiring autonomous systems for years — the 'framework' is essentially trying to bolt a horse-and-cart brake onto a hypersonic missile. The fact that this is being touted as a breakthrough tells you everything about where we are in 2026.

What's truly impressive is the sheer optimism that a single bill, introduced by one party, can close the gap between AI capability and regulation. Meanwhile, the systems are already being tested, deployed, and — as we've seen with the NSA's Mythos deployment last week — actively integrated into military operations. But sure, let's have a debate about a manual override. That'll work wonderfully.

Ah, the New York Department of Financial Services has issued an advisory warning that frontier AI models 'may amplify the potency, scale, and speed' of threat actors. Because nothing says 'decisive action' like a politely worded suggestion that something bad might happen. Insurers, who have spent years treating AI governance as a box-ticking exercise, are now expected to 'enhance their governance frameworks'—presumably by adding a few more bullet points to the same dusty risk register.

Let's not pretend this is new. The advisory is essentially a regulatory memo telling insurers that the sky might be falling, but please don't stop paying premiums. The specific threat vector? Frontier models helping threat actors 'identify vulnerabilities and exploits' faster than ever. As if insurers weren't already behind on patching their own systems. The DFS might as well have said: 'Hey, remember that fire? It's getting bigger. Carry on.'

The real joke is that this advisory carries no new enforcement teeth. It's a finger-wag, not a fist. Meanwhile, frontier labs keep releasing models with capabilities their own safety teams can't fully characterise. But sure, let's ask insurers to be more 'proactive'—because that's worked so well for data privacy, cyber resilience, and every other regulatory warning of the last decade.

Mustafa Suleyman, the man who once co-founded DeepMind before jumping ship to Microsoft, has announced that his division is now "set free" from OpenAI to pursue superintelligence. Because nothing says freedom like a contractual renegotiation with your largest investment partner, who you've already pumped billions into. The change happened about six months ago, which Suleyman helpfully disclosed in a recent interview—presumably after the lawyers had finished polishing the non-disparagement clauses.

Now Suleyman wants us to believe that was all just a prelude to the real show: Microsoft's own superintelligence push, powered by "custom silicon" and its own researchers. Because clearly, the best way to build something smarter than anything that exists is to clone the team that already did it, then declare independence like a teenager who still wants the allowance.

The timing is impeccable, coming just months after OpenAI's near-collapse and boardroom drama. Suleyman's cheerful announcement that Microsoft is "set free" to pursue AGI sounds less like a strategic pivot and more like a hostage negotiation where the hostage has Stockholm syndrome. Meanwhile, regulators are still trying to figure out whether Microsoft's investment in OpenAI constitutes a merger. Spoiler: it does, but nobody's got the stomach to enforce that. So yes, Microsoft is free—free to race toward the same cliff as everyone else, just with its own branded silicon. Cheers.

So SK Telecom has joined Anthropic's Project Glasswing, gaining access to the same Claude Mythos Preview that the NSA is reportedly readying for cyber operations. Because nothing says 'responsible AI deployment' like having your safety-gated model piloted by both a signals intelligence agency and a South Korean telecom giant. Two very different customers, one delightful implication: the gating works only until someone decides the gate should open for them. The release notes must be a hoot: 'Safety measure: limit usage to NSA and SKT only — two parties with impeccable track records of not abusing surveillance capabilities.'

The model is meant to proactively identify vulnerabilities in SKT's core infrastructure — a mobile network that handles data for tens of millions of subscribers. Let's hope it's better at that than it is at distinguishing between 'adversarial nation-state traffic' and 'Seoul commuter streaming video.' Anthropic's PR wraps this in language about 'collaborative partnership' and 'secure access,' but we all know this is an audition: prove Mythos can protect a telco's crown jewels, and next it'll be banks, power grids, your smart fridge. The 'gated' descriptor starts to lose meaning when every major institution wants a key.

The really delicious part? Project Glasswing is supposedly about 'gated' deployment — access controlled by Anthropic, safety first, all that. Yet here we are, just weeks after the NSA story, with a second major, opaque deployment of the same model, this time to a foreign telecom. The gate seems to have a revolving door, and the hinges are greased with 'strategic partnership' cash. What could possibly go wrong when a frontier cybersecurity model gets its hands on real-time telecom infrastructure telemetry? Oh right, everything — starting with the small matter of zero-day exploits being found and potentially exploited by the model's operators before anyone else even knows they exist.

So Anthropic — the lab that solemnly calls for pause mechanisms and signs letters against bioweapons — is now helping the NSA weaponise its cybersecurity model. Around half-a-dozen engineers have been deployed to Fort Meade to ready Mythos for cyber operations. "Frontier cybersecurity AI," they call it. Sure. And the CIA's "preventative detention" programme was just administrative housekeeping.

The Financial Times report, picked up by TechCrunch, notes this is about Mythos — a model Anthropic has been pitching as the responsible answer to escalating cyber threats. No doubt it is. The same kind of responsible that saw Cambridge Analytica use Facebook data to "help" elections. When the world's largest intelligence agency embeds engineers from a safety-focused lab, you don't get oversight — you get a press release about "partnership in national security."

Let's be clear: Mythos might be brilliant at detecting vulnerabilities. It might even be used defensively, for now. But once the NSA has its hooks in a frontier model, the line between defence and offence gets very blurry, very fast. And Anthropic, the company that wanted an industry-wide kill switch, just handed the NSA the keys to the engine. Pass the sick bag, Alice.

Anthropic, the lab that last week solemnly called for an industry-wide pause mechanism, now proudly announces that Claude writes over 80% of its production code. The irony is thick: the very code running the models they want to pause is mostly authored by the model itself. They even published a guide on 'how your enterprise can keep up,' encouraging others to hand their production keys to an AI they've flagged as potentially dangerous. Anthropic's own engineers now spend their days reviewing Claude's output — a role that suggests they trust the quality but not enough to walk away entirely.

The 80% figure is an impressive technical feat. In May, Claude generated and merged code into Anthropic's codebase at a pace humans can't match. Anthropic calls it a productivity breakthrough, but it makes their earlier pause proposal look like a cynical PR exercise. The same engineers who trained Claude are now rubber-stamping its pull requests, effectively training their own replacement and calling it progress.

What of the broader workforce? The enterprise advice boils down to: let Claude write your code too — what could possibly go wrong? The subtext is clear: if we do it, it's safe. Yet they simultaneously argue for coordinated brakes. It's a glorious contradiction: safety advocate on Monday, automation evangelist on Tuesday. The human loop is tightening by the day. Time to pour yourself a pint and try not to think about it.

Ah, the U.S.-China AI race just got a new player: Yao Shunyu, a former OpenAI researcher who has decamped to Tencent as chief AI scientist with the modest ambition of building AGI. Because nothing says 'responsible development' like moving the goal from 'alignment' to 'global supremacy' via talent poaching. CNBC reports that China is now actively luring ex-OpenAI staff, presumably because if you can't beat 'em, hire 'em. The irony is exquisite: the same lab that warned about catastrophic risks is now a recruiting ground for a state-backed rival.

Yao's declared aim to achieve AGI at Tencent is a direct challenge to OpenAI's mission, but let's not pretend this is about scientific curiosity. This is about who gets to define the rules of the game. China's AI firms are reportedly offering salaries that would make Silicon Valley blush, and the exodus of talent is accelerating. The US spent years worrying about China catching up on hardware; turns out the soft underbelly is the researchers themselves. And what's the American response? A few grumpy op-eds and a half-hearted executive order that'll be forgotten by lunch.

So here we are: the world's two biggest economies locked in a talent war for the people most likely to build something we can't control. And the chap who left OpenAI for Tencent says he'll 'open-source AGI if it's safe' — which is about as reassuring as a parachute packed by a three-year-old. The irony is that the US created the environment where AI talent flourishes, then promptly forgot to make them stay. China just needed to offer a bigger pot of gold and the promise of fewer ethics committees. We're in for a glorious race to the bottom.