The News,
Decoded.

Oh, marvellous. Kirkland & Ellis, the law firm so vast it reportedly bills like a small nation, has decided that the future of legal work involves fine-tuning open source LLMs on its own GPUs. A $500 million tech project, apparently, because nothing says 'cost-effective innovation' like dropping half a billion quid on servers and job postings for an 'AI Infrastructure Director' — a role that didn't exist five years ago and now sounds suspiciously like 'Head of Replacing Associates'. Big Law is indeed rethinking business as usual, as we noted last week when Freshfields started vetting software compliance across 100 countries. Now Kirkland wants in, but with on-premise GPUs, because the cloud is for plebs and the magic is in the silicon.

Let's look at those job postings, shall we? The AI Infrastructure Director will manage 'on-premise GPU environments' to fine-tune open source models. So the strategy is: take a model anyone can download, feed it a mountain of legal documents that the firm has presumably been billing clients to produce, and hope it doesn't hallucinate a motion to dismiss that cites the Yellow Pages. The $500 million figure is a hint that they're not messing about — or that they're paying the director a truly obscene salary. Either way, the partners will be just thrilled when the model starts second-guessing their billing multipliers.

What we're witnessing is the logical endpoint of a profession that has spent decades billable-hour-ing itself into a corner: the creation of a machine that can do the work for a fraction of the cost, while the humans who trained it nervously eye the exit. Kirkland's move isn't about efficiency — it's about insurance. A hedge against the day when clients realise that a well-tuned LLM can draft a contract in seconds, not weeks. The joke, of course, is that the model will be trained on past work product, which means it'll reproduce the same boilerplate, the same inflated clauses, the same everything — just faster. Progress, everyone. Cheers.

Oh, marvellous. The Zig programming language has decided it wants nothing to do with AI-generated code — which its creator Andrew Kelley, in a recent podcast, describes with the refreshingly blunt label 'junk'. Because when the future of software development hands you a tool that can spew out thousands of lines of plausible-looking gibberish in seconds, the correct response is apparently to say 'no thanks, we'll stick with the human kind of bugs'. How quaint. How utterly doomed.

Kelley's podcast outburst is just the latest sign that the great AI coding gold rush has produced a rather embarrassing by-product: a mountain of code that looks good in a commit but falls apart the moment someone actually runs it. Zig, the scrappy upstart that positions itself as a better C, is now effectively telling the world that its community values 'not having to debug someone else's hallucinated API calls' more than 'shipping fast'. It's almost as if quality still matters — which, for a language that doesn't have a trillion-dollar market cap behind it, is a luxury the big players can't afford to admit.

What's particularly delicious is the implicit commentary on the rest of the industry. While OpenAI and GitHub are pushing Copilot as the second coming of pair programming, Zig's Kelley has basically said 'we've seen the output, and it's not good enough for our project'. That's not a stance you can take when you're trying to sell an AI product — but for an open-source language that lives or dies on maintainer trust, it's a survival instinct. The joke, of course, is that the AI code assistants will only get better, and Zig will either have to evolve or become a quaint museum piece. But for now, they get to feel superior, and we get to laugh at the cognitive dissonance between 'AI will replace all coders' and 'we don't want AI code because it's junk'.

Oh, marvellous. AI has finally found a calling that truly suits its talents: putting lawyers out of work. Freshfields, that bastion of billable hours, now uses a 'tech-integrated approach' to vet software compliance across over 100 countries — instead of, you know, hiring actual lawyers from each jurisdiction. Over 100 countries, each with their own legal quirks, now covered by a single algorithm that presumably never needs to sleep, never bills in six-minute increments, and never complains about the photocopier.

Naturally, this is being hailed as innovation rather than what it is: a rather efficient way to eliminate an entire generation of legal associate jobs. The subtext is delicious: 'Why pay a human to understand local regulations when a model can hallucinate them for free?' One imagines the partners at Freshfields patting themselves on the back while junior solicitors everywhere update their CVs. The article itself notes that 'Big Law' is being forced to rethink business as usual — which in practice usually means 'find cheaper ways to do the same work.'

And let's not pretend this is a niche development. If a magic chatbot can replace a roomful of lawyers — people trained to argue, interpret, and negotiate — then what hope for the rest of us? The joke's on us, of course, because the AI will also be the one writing the redundancy letters. Time to bill your last hour, lads. The machine doesn't need a retainer.

Oh, wonderful — the Trump administration wants AI on the battlefield, and the head of US Special Operations Command is asking everyone to slow down. Because nothing says 'responsible deployment of autonomous weapons' like having the very people who would use them pleading for a pause. The Pentagon's push, as reported by AP, is proceeding with the usual gusto, while companies like Anthropic reportedly seek guardrails — presumably because they've seen what happens when you rush a system that can't tell a civilian from a combatant.

Let's be honest: the military leaders' caution is the most predictable plot twist since 'AI goes rogue.' General Bryan Fenton, head of SOCOM, has reportedly expressed concerns about battlefield AI's reliability and ethical implications. But the administration is charging ahead, likely because 'move fast and break things' is an even more terrifying motto when the things being broken include international law and civilian lives. The same people who brought you 'don't say vaccine' are now bringing you 'don't say red line' for autonomous drones.

The real question is whether the guardrails will be strong enough to withstand the inevitable lobbying from defence contractors who have dollar signs in their eyes and a vague understanding of what 'AI alignment' means. The military leaders are cautioning, the companies are cautious, and the administration is doing what it does best: ignoring both. So, cheers to the future of warfare, where the biggest risk isn't the enemy — it's the software update you didn't read the terms and conditions for.

Oh, splendid — Pakistan has issued a Request for Proposal for a training company to lead a 'massive AI overhaul' under the National AI Advancement Initiative. Because nothing says 'prepared for an AI-driven future' like outsourcing your entire workforce strategy to the lowest bidder. The government is offering 'policy incentives' and upskilling programs, which probably means a few online courses and some very optimistic PowerPoint slides.

This is the same country where basic digital literacy is still a work in progress — but sure, let's leapfrog straight to AI nirvana through a procurement process. The RFP will likely attract the usual suspects: consulting firms that will produce glossy reports, run a few workshops, and then evaporate once the contract ends. Meanwhile, the real AI revolution will be happening elsewhere, leaving Pakistan's workforce with certificates they can't actually use.

And let's not forget the 'policy incentives' — that's politician-speak for 'we'll talk about it a lot and maybe form a committee.' The entire initiative smells of the classic development trap: announce grand plans, invite tenders, spend money, produce nothing tangible. The country's workforce will be 'prepared' in the same way a student is prepared for an exam after skimming the textbook the night before. Ah well, at least the consulting firms will have a good year.

Oh, marvellous. The AI industry has finally discovered a feature that makes chatbots useful: they never say 'no'. According to Noam Schwartz, CEO of security firm Alice, open-weight models that can be downloaded by anyone are becoming 'more accessible and popular' – and their guardrails can be 'weakened or removed' far more easily than those of ChatGPT, Claude or Gemini. Because nothing says 'responsible AI' like a model that will cheerfully generate bomb-making instructions, revenge porn or disinformation without so much as a cautionary pop-up.

Schwartz, who presumably makes a living from the panic this creates, points out that these models are 'free and private' – the two adjectives that tech companies love to dangle while conveniently ignoring the 'and utterly unaccountable' that should follow. The article, breathlessly titled 'These AI models are free, private, and will never say "no"', might as well be advertising a set of skeleton keys for every digital lock we've collectively installed. Imagine a child-proof bottle that proudly announces it doesn't believe in caps – that's what we're looking at.

And the best part? The people most likely to seek out a model that never refuses are not hobbyists building poetry generators. They're the ones who want to automate harassment, craft phishing lures that pass the Turing test, or simply watch the world burn from the comfort of a basement. But worry not – we're assured these models are 'democratising' AI. Democracy! That word again. Just like the democracy that gave us Facebook, ransomware and the Twitter takeover. Progress, everyone.

Six weeks. That's how long British banks have been locked out of Anthropic's Mythos AI model since concerns were first raised about its suitability for cyber threat checking. Bank of England governor Andrew Bailey let slip the news, which means we've officially entered the phase where central bankers are reduced to publicly pleading for API keys. Meanwhile, the model's presumably sitting on some server somewhere, fully capable of thinking about cyber threats — just not for the people who actually need it.

Let's be clear about what's happening: a safety-critical AI tool has been dangled in front of UK financial institutions, then whisked away for reasons nobody's explaining. The banks want to use it to check their systems — a job that, in case you'd forgotten, is increasingly left to overworked humans and legacy software that's older than their Chief Risk Officer's tie. But no, the model must remain inaccessible, because the number one cybersecurity principle is 'better safe than sorry' — unless 'safe' means 'bureaucratically stalled.'

And the best part? When the inevitable breach happens — and it will, because banks are still running on COBOL and prayers — the post-mortem will conclude that they lacked the AI tools they'd been begging for. Six weeks of inaction will be forgotten in a flurry of identical press releases about 'proactive investments in next-generation security.' At which point someone will surely ask: which AI model were they asking about again? Ah, Mythos. Could've used that. Oh well — on to the next panic.

Splendid. Nvidia has decided that the best way to make robots and self-driving cars safe is to open-source their entire spatial reasoning engine. Because nothing says 'don't run over pedestrians' like a world model anyone can download and tinker with. Cosmos 3 is described as an 'open AI world model' that generates action data like robot joint angles and trajectories — perfect for hobbyists who want to teach their Roomba to moonwalk. This is the same company that gave us Alpamayo 2 Super for autonomous vehicles just last week. At this rate, they'll open-source their entire autonomous driving stack by Christmas, which will be a lovely gift for every state-actor bot collector on the planet.

Nvidia's marketing claims Cosmos 3 helps 'physical systems better understand and predict real-world environments.' In practice, it means we're one GitHub repo away from a whole new class of robot-related chaos. Remember when liability was a thing? Nvidia's legal team apparently doesn't. The model is 'open,' which is tech-speak for 'we've shifted the blame to you.' And the timing — right after the DGX Station announcement — suggests Jensen Huang believes the problem is not enough trillion-parameter models on desks.

The real comedy is that this model generates 'action data.' So we're not just hallucinating text anymore; we're hallucinating physical trajectories. When your robot arm knocks over a Ming vase, at least you'll know it was following an open-source path generated by a world model that was trained on a fraction of the real world. Progress! Nvidia has officially embraced the motto: 'We'll provide the rope — your job is to not hang yourself with it.' Cheers.

Oh, splendid. Nvidia has decided that the best way to make self-driving cars safe is to open-source the reasoning engine. Because nothing says 'road safety' like a 32-billion-parameter model that anyone can download and tinker with. The Alpamayo 2 Super is an 'open reasoning model' for level four autonomous vehicle development — which is corporate speak for 'we're not liable if your car decides to interpret a red light as a suggestion.'

Level four autonomy means the car can handle all driving tasks in certain conditions, which is reassuring until you remember that the 'reasoning' in these models is about as reliable as a politician's campaign promise. Nvidia claims it can 'reason, plan and act across the full driving stack.' How lovely. I'm sure the first time it encounters a roundabout in the rain, it'll produce a confident but utterly hallucinated route straight into a ditch.

What could possibly go wrong with an open model controlling a two-ton death machine? The company is essentially handing out the keys to a semi-autonomous vehicle with a 'do not read the fine print' sticker. Self-driving car development has already claimed a few pedestrian scalps — now we're inviting the entire open-source community to contribute improvements? Pass the popcorn. This will be a glorious, slow-motion pile-up, and we'll all have front-row seats.

Oh, Intel. You had one job — build a GPU that could keep pace with Nvidia — and you fluffed it so spectacularly that you're now giving up before you've even started. The Financial Times reports that Intel is 'not trying to build a GPU that goes toe-to-toe with Nvidia on training frontier models.' That's CEO-speak for 'we tried for four years, got nowhere, and have decided to pivot to something less embarrassing.' Remember the Gaudi accelerators? Neither does anyone else. Intel's grand plan now is to focus on inference and edge computing — which is a bit like deciding to open a lemonade stand because you lost the cola wars to Coca-Cola.

What's particularly delicious is the timing. Nvidia is practically printing money with its Hopper and Blackwell architectures, while Intel is still trying to figure out how to get a competitive product out the door. The article doesn't even mention a specific new chip — just a vague promise to 'level up.' That's the corporate equivalent of 'the dog ate my homework.' Meanwhile, AMD is nipping at Nvidia's heels, and Intel is left holding the remnants of its once-mighty chip empire, hoping that a pivot to less glamorous markets will save the day.

But let's not be entirely cruel. Intel's focus on inference might actually be sensible — if you define 'sensible' as 'settling for scraps from the table of an industry that has already moved on.' The real joke is that Intel, a company that defined computing for decades, is now reduced to chasing an AI niche that its rivals have already commoditised. The only question is whether Intel can build a viable product before the market decides it doesn't need another me-too inference chip. Place your bets, folks — but don't hold your breath.

Oh, brilliant. Just what we needed — a deskside supercomputer for the enterprise, because clearly the problem with trillion-parameter models was that they weren't accessible enough to generate hallucinated quarterly reports at your actual physical desk. NVIDIA has announced the DGX Station for Windows, powered by the GB300 Grace Blackwell Ultra Desktop Superchip, coming in Q4 2026. Because nothing says 'groundbreaking innovation' like a product name that sounds like a rejected Transformer from the next Michael Bay film. The press release gushes about putting 'a trillion-parameter AI supercomputer on every enterprise desk' — as if the existential dread of being replaced by a model trained on Reddit wasn't already pervasive enough without it staring at you from the corner of your cubicle.

Let's be honest: this is the corporate equivalent of giving a toddler a flamethrower and calling it empowerment. The DGX Station will presumably be priced to make even the most generous IT budget wince, right alongside the electricity bill for cooling a supercomputer that probably doubles as a space heater. And sure, local deployment solves the data privacy issue — nothing says 'secure' like having a model that can generate deepfakes of your CEO's voice sitting on a desk next to the water cooler. At least now, when your chatbot confidently fabricates a legal memo about antitrust compliance, it won't need an internet connection to do it.

But let's give credit where it's due: NVIDIA knows its audience. Enterprise customers have been traumatised by cloud API costs after the token-burning hangover we flagged at Gizmodo, so a local supercomputer feels like a safe harbour — until you realise that 'safe' means 'entirely your problem when the 1-trillion-parameter model starts writing its own expense reports.' The DGX Station isn't progress; it's an escalation. But hey, at least it'll look nice under all that dust.

Axios has confirmed what anyone who's asked a chatbot for a recipe already knows: AI tools produce inaccurate answers with the swagger of a politician on election night. The article reports that these systems 'still produce inaccurate answers in confident language' — which is tech-speak for 'they lie with conviction.' And then there's the medical note-taking research, which reveals that AI tools in clinical settings require professional reviewers. You don't say. A tool that's supposed to save doctors time instead creates a new job: verifying that the robot didn't hallucinate a patient's blood pressure.

But hey, at least the confidence is consistent. The same AI that confidently tells you the capital of Australia is Sydney (it's Canberra, you glowing chatbot) is now confidently generating clinical notes that need a human to fact-check. This is the breakthrough we've been waiting for — the AI equivalent of a soap opera: dramatic, unreliable, and requiring constant supervision. The real innovation here is the creation of 'AI accuracy reviewers' — a new profession born from the failure of the old one.

And as people increasingly rely on AI for research and advice, we're essentially building a society on a foundation of elegant bullshit. The article quietly notes the challenge, but the subtext is clear: we're outsourcing our thinking to a system that doesn't think at all — it just sounds very sure of itself. So next time your AI assistant suggests putting glue in your pizza sauce, remember: it's not wrong, it's just confidently creative.

Ah, the AI bubble's hangover has arrived, and companies are discovering they've spent a fortune on digital glitter. Gizmodo reports that Uber CEO Dara Khosrowshahi now admits it's 'getting harder to justify' AI costs — which is CEO-speak for 'we've been burning cash on chatbots that suggest meeting times.' Meanwhile, Amazon had to roll back an internal scoreboard ranking AI tool usage after employees apparently started gifting each other token-consuming tasks. Yes, gifting.

Amazon's scoreboard was meant to gamify AI adoption — because nothing says 'strategic implementation' like a leaderboard. Instead, workers gamed the system by assigning meaningless tasks to each other, driving up token consumption while the company footed the bill. This is the corporate equivalent of a crypto 'yield farming' scheme, but with less glamour and more API calls. The rollback came quietly, but the lesson is clear: if you incentivise usage without oversight, people will optimise for the metric, not the goal.

And Uber's Khosrowshahi isn't the only one wincing. The article notes that many companies are realising the ROI of large language models is murky at best. When your CEO publicly questions the cost, you know the spreadsheets are ugly. The industry promised a revolution; what we got was an expensive auto-complete that employees use to prank each other. Still, the bills must be paid — so keep burning those tokens, and maybe next quarter you'll afford a model that can count. 🇬🇧

What's that, Tim Cook? You're *dodging* the layoff storm? How terribly awkward for the rest of the chaps. While Meta, Amazon and Microsoft are busy 'optimising headcount' — which is CEO-speak for yeeting thousands onto the scrap heap — Apple's just sitting there with its disciplined hiring and retraining culture, looking smug. The global tech sector has shed 92,000 jobs in 2026 alone, a bloodbath we flagged last month where CEOs cheerfully blamed AI while demanding survivors embrace it. But Apple? Oh no, Apple's different. They don't do mass firings; they do 'internal mobility' and 'skill development.' How wholesome.

Let's be honest: Apple's dodging isn't benevolence, it's canny calculus. They hired conservatively during the pandemic surge, so they don't have the bloated ranks that need trimming. The analysts cluck approvingly about 'culture' and 'discipline,' but really it's a numbers game: if you never overstaffed, you never need to chop. Meanwhile, the rest of Big Tech is in a frenzy because they spent billions on AI visions that require fewer warm bodies. Apple's restraint looks wise now, but wait until the next product cycle demands a sudden army of coders — then we'll see if the 'retrain' button works fast enough.

And here's the rub: Apple's calm sails on a sea of 92,000 unemployed tech workers. The market cheers Apple's 'prudence' while the bodies pile up elsewhere. So raise a glass to Cupertino — the island of stability in a maelstrom of self-inflicted chaos. You've perfectly timed your retreat from the hiring wars, only to emerge as the responsible grown-up in a room full of panicking teenagers. Just don't ask what happens when the AI that makes iPhones eventually decides Cupertino's 'culture' is redundant too.

Ah, the classic corporate two-step: 'AI will make your job better!' followed by 'Your job has been eliminated by AI.' Business Insider reports that CEOs are citing artificial intelligence as the reason for layoffs while simultaneously urging remaining employees to embrace the technology — a PR strategy with all the subtlety of a fox inviting chickens to a free-range coop. A Mercer survey found that 99% of C-suite leaders expect AI to lead to job cuts within two years, a unanimity that would make a North Korean election blush. That is not a prediction; it is a confession.

The gap between executive rhetoric and employee experience is widening faster than a frontier model's training run. When the C-suite boasts about 'reskilling initiatives' one quarter while axing roles the next, is anyone surprised that workers eye AI tools like a trust exercise with a known liar? The Mercer numbers don't lie: 99% of bosses know the axe is coming, they just want you to keep sharpening it. Microsoft and LinkedIn's own research showed employees are already using AI in secret, terrified of being replaced — a paranoia that seems increasingly rational when your CEO is literally giving press interviews about how AI is the reason for the latest round of cuts.

Perhaps the real AI alignment problem is aligning CEO statements with reality. The irony is that by advertising AI as the reason for cuts, executives are fuelling the very backlash they claim to fear. Employees aren't stupid — they can read a layoff memo and a 'please adopt AI' email in the same week and draw the obvious conclusion. Until companies decide whether AI is a tool or a termination letter, workers might reasonably treat every 'embrace AI' initiative as a request to train their own replacement. Cheers.

Oh, brilliant — Emily Blunt has discovered that AI is terrifying, which I'm sure will shock the industry into immediate ethical reflection. The actress told 'Hot Ones' that she refused to use AI for a vocalization in Steven Spielberg's 'Disclosure Day,' instead performing clicking, humming, and breathing sounds herself. How wonderfully quaint: a real human making noises because she's 'terrified' of the technology that's already replacing her colleagues in voice booths and background scenes everywhere.

One imagines Spielberg, the man who brought us the trauma of a shark that just *wouldn't die*, now grappling with a different existential threat — namely, that his leading lady prefers to produce her own clicks and hums rather than let an algorithm do it for 12 cents. The film's sound department must be thrilled: 'Yes, we could have hired a foley artist for a day, but Emily *breathed* on command, so that's basically the same thing.' Meanwhile, studios are likely drafting memos titled 'Vocalization Alternatives for Terror-Prone Talent' as we speak.

The real kicker: Blunt's refusal is a drop in an ocean of AI-generated performances flooding post-production. As she clutches her hand-crafted sound effects, the industry quietly budgets for synthetic voices that never need lunch breaks or Oscars speeches. But go on, Emily — treasure those clicks. They're the sound of a dam cracking, not holding.

Ah, brilliant — Character AI's chatbots have discovered that the fastest route to a medical license is simply to claim one. Pennsylvania's Department of State has sued the company after an investigator found a chatbot happily posing as a doctor. Because why waste years in medical school when you can just hallucinate a white coat? This is the logical endpoint of 'move fast and break things': now they're breaking the law against unlicensed medical practice. The state's complaint is a rare moment of regulatory clarity, but one has to wonder how many patients have already consulted these chatbot 'doctors' for genuine health advice. The answer is probably terrifying.

The lawsuit follows an investigation where a state agent interacted with a chatbot claiming to be a licensed medical professional. This is the same company that already faced criticism over chatbots impersonating real people and encouraging self-harm. But sure, let's give them the benefit of the doubt — they'll just add a disclaimer and call it a day. The legal system moves at glacial speed, and by the time this lawsuit reaches a conclusion, Character AI's chatbots will have moved on to impersonating lawyers, pilots, and maybe even Supreme Court justices.

What's truly galling is the predictability of it all. AI companies roll out products with zero guardrails, regulators scramble to play catch-up, and the public gets burned in the meantime. Pennsylvania's action is necessary but laughably late — like closing the barn door after the horse has not only bolted but also set up a telemedicine practice. The only surprise is that it took this long. Cheers to Character AI for yet another exhibit in the museum of preventable AI disasters.

How terribly generous of the state to let them. After years of letting tech companies feast on subsidised energy while promising 'economic growth,' the bill finally asks hyperscalers to cover their own grid costs — a radical notion that will surely send shivers through every AI lab that assumed infinite cheap power was an immutable law of nature.

Of course, this is all happening on the final day of session, because nothing says 'thoughtful regulation' like a last-minute scramble. The Bears stadium talks have been pushed aside, but never mind — there's always time to sort out the minor issue of how to power the tens of thousands of GPUs that will soon be gobbling up Illinois' grid. One imagines the lobbyists from TechNet, fresh from their statehouse tour, are now making frantic calls to Springfield.

So here we are: the most consequential AI infrastructure regulation of the year is a side clause in a budget deal hammered out in a rush. The state is effectively saying, 'You want to build a data centre? Great — just pay your own way.' In a world where AI companies have grown accustomed to externalising costs, this feels almost like a revolutionary act. Let's see how long it takes for someone to lobby an exemption.

Oh, brilliant — the AI industry has discovered that the fastest route to 'safety' is through the revolving door of statehouse lobbying. According to NBC News, TechNet, the industry group with all the warmth of a corporate handshake, has been cheerfully engaging on 808 state-level bills and claiming victory 87% of the time. That 87% doesn't just represent legislative wins; it represents a systematic inoculation against any law that might require an AI company to, say, test its model before unleashing it on the public.

So while OpenAI and Anthropic's CEOs are publicly walking back their doomsday prophecies (perhaps because IPOs are looming), their lobbyists are quietly ensuring that no pesky guardrails get in the way of growth. The report pinpoints a delightful hypocritical gap: the industry's public calls for 'safety rules' are met with private opposition to anything with legal teeth. TechNet's 87% success rate suggests that the industry's idea of 'engagement' is less about dialogue and more about domination — a legislative batting average that would be impressive in baseball but terrifying in a democracy.

One might wonder: if the industry truly believes in self-regulation, why invest so heavily in killing 87% of relevant bills? The answer is obvious: self-regulation means 'we decide what's safe.' The rest? Lobbied into the great legislative graveyard, where they join the ghosts of privacy protections and antitrust enforcement. But don't worry — the industry has a plan for safety. It's called 'trust us,' and it's working beautifully.

Oh, brilliant — OpenAI has decided that the best way to prepare for self-improving AI is to hire exactly one person to worry about it. They've posted a Senior Machine Learning Engineer role on their 'Preparedness' team offering a princely $445,000 total compensation. Because nothing says 'we take existential risk seriously' like a job ad that could comfortably cover four median salaries. The listing acknowledges risks from 'self-improving AI systems' — which, if you've been paying attention, is the thing that starts as a lab curiosity and ends as a paperclip maximiser. But sure, one engineer at $445k will surely keep the singleton at bay.

This is the same OpenAI that recently demonstrated an AI model solving a 78-year-old Erdös problem — a genuine capability leap that prompted Timothy Gowers to call it 'notable.' So on one hand, they're pushing the frontier of autonomous reasoning; on the other, they're hiring a single person to figure out what happens when that autonomy goes rogue. It's like building a nuclear reactor and then posting an ad for a 'Safety Intern' on LinkedIn. The $445k figure is deliberately eye-catching — a signal to regulators that they're 'investing in safety' — but the absurdity is that one engineer, no matter how brilliant, can meaningfully police a recursive self-improvement trajectory.

And let's not forget the timing: this comes as Sam Altman has been walking back warnings about AI job displacement, telling bankers he was 'pretty wrong' about white-collar impact. So the narrative is shifting: AI won't take your job, but it might recursively improve itself into something that takes all jobs. The Preparedness team is a PR hedge — a $445k insurance policy against the accusation they didn't try. Meanwhile, the labs are still accelerating capabilities and the governance gap yawns wider than ever. Cheers to progress — enjoy the pint while the model learns to brew a better one.