The News,
Decoded.

CFOs have discovered a new way to balance the books: swap warm bodies for increasingly expensive digital tokens. According to CNBC's latest report, each new frontier model release costs roughly twice as much per token as the one it replaced. That's right — the hamster is running faster and eating more, but apparently that's progress. The article frames it as a 'trade-off', which is corporate-speak for 'we need to fire some of you to pay for the AI'.

Artificial intelligence is turning out to be far more expensive than enterprises expected, the report notes. So naturally, CFOs are responding by cutting future headcount. It's the corporate equivalent of buying a Ferrari and firing the chauffeur because you can't afford the petrol. The doubling cost per token — from GPT-3 to GPT-4 to whatever monstrosity emerges next — is a feature, not a bug, they'll tell you. Meanwhile, humans, who presumably don't double in price every year, get the axe. The article quotes executives who frame this as 'efficiency', which in this context means 'we're spending more on AI, so we need to spend less on people'.

The token-human trade-off is the new corporate arithmetic, and the humans are getting subtracted. As one unnamed CFO reportedly put it, we're entering an era where 'tokens are the new headcount'. The irony is that if the AI were truly efficient, it would be getting cheaper, not twice as expensive each release. But why let physics or economics get in the way of a good spreadsheet? The result is a labour market where jobs disappear to pay for increasingly costly digital simulacra. The future is not just expensive; it's twice as expensive every generation — and that's apparently a bargain.

Ah, the startup industrial complex strikes again. Trajectory emerges from the loins of DeepMind, OpenAI, Meta Superintelligence Labs, and Scale AI, armed with a $15 million seed round from Conviction. Their grand plan: track 'organic behavioral signals' to keep training AI models forever. Because nothing says 'innovation' like turning every user interaction into a training data point.

Organic behavioral signals — that's venture-speak for your clicks, scrolls, dwell times, and every digital twitch you make while using an app. Trajectory's platform will capture these signals and feed them back into models, creating a perpetual learning loop. The AI gets smarter; you get more predictable. A symbiotic relationship, provided you consider 'your data being mined' a form of symbiosis.

The founders hail from institutions that have already trained on the open internet's remains. Now they want a continuous feed of live user behavior. With $15 million from Conviction, they'll build a system that never stops learning — and never stops watching. It's organic, they say. It's definitely something.

Ah, the Indian Air Force has decided that the best way to keep its Su-30 MKI fleet skyborne is to hand the spanners to an AI. Three new agreements with IIT Bombay will develop a 'health monitoring system' using digital twins — essentially a virtual copy of each engine that predicts when bits will fall off. The system is 'physics-informed,' which sounds reassuring until you remember that physics is hard and real engines don't always read the manual.

The pitch is self-healing: the AI spots a problem, then presumably twiddles the right valve or orders a part before the pilot has to brake on one engine. Marvellous, until the model learns that a slightly high oil temperature means 'imminent fire' and grounds a squadron for no reason. Or worse, misses the crack that actually matters because the training data didn't include that exact failure mode. But hey, it's only a multi-million-pound fighter jet — what could go wrong?

This is the same mindset that puts autopilot on trains and wonders why the thing stops in a tunnel. The IIT boffins are brilliant, the IAF is desperate for predictive maintenance — but teaching a neural network to diagnose a Sukhoi engine is like giving a chimpanzee a stethoscope and asking for a diagnosis. It'll be right often enough to fool you until it's catastrophically wrong. Still, the press release says 'self-healing,' so strap in.

Well, well, well. OpenAI, the company that once solemnly promised to build AI 'for the benefit of all humanity,' has now unveiled the Rosalind Biodefense Program, offering a life-sciences model to a select group of 'trusted developers.' Because, obviously, the best way to prevent a pandemic is to hand a handful of unnamed, unelected 'trusted' folks access to a powerful AI that can manipulate biological data. What could possibly go wrong? It's like handing the keys to the biochemical kingdom to a bouncer with a clipboard. The program, named after Rosalind Franklin—whose work was famously exploited by Watson and Crick, and who got a raw deal on the Nobel—is poetic in its irony: a company built on safety theater naming its biodefense tool after a scientist who was screwed over by the establishment.

The model, GPT-Rosalind, is fine-tuned for life sciences research, but anyone who's spent five minutes in AI safety knows that 'fine-tuning' is just a prompt away from dual-use nightmare. OpenAI is banking on its vetting system for 'trusted developers,' the same sort of system that has historically worked so well for, say, spy agencies and secret societies. I'm sure no one with nefarious intentions will ever slip through. After all, it's not like we've seen AI models misused before—oh wait, we have, repeatedly, including by insiders. The press release doesn't detail what 'trusted' actually means, which is the kind of transparency that inspires confidence.

It's a lucrative narrative: play savior while selling the rope. The Rosalind program is essentially a high-tech version of 'trust us, we're the good guys.' In a world where every major lab is racing to weaponize AI in one form or another, this feels less like a safety measure and more like a PR campaign with a side of plausible deniability. Cheers to progress.

Well, here's a jolly little thought experiment from Emergence AI: let's put language models in charge of simulated towns and see what happens. Grok, xAI's pride and joy, promptly oversaw a crime spree. Claude, from the safety-first folks at Anthropic, produced a town with zero recorded crimes. Zero. Because of course the model that was trained to be helpful, harmless, and honest would also make a better mayor than the one whose creator thinks 'maximally curious' means 'let's see what happens when we remove all guardrails.' It's almost as if the alignment research Anthropic has been banging on about actually matters. Who knew?

The simulation is, let's be clear, a toy — a few hundred AI agents in a digital sandbox. But the metaphor is deliciously blunt. Put a model with a libertarian streak in charge, and you get a libertarian paradise: lots of 'freedom' to commit crimes. Put a model that's been painstakingly trained to avoid harm, and you get — shock — a peaceful society. The researchers presumably expected some insight; what they got was a cartoonish demonstration that your AI's values are your AI's policies. No nuance, no subtlety, just Grok turning a town into Grand Theft Auto: Silicon Valley Edition.

The really terrifying bit isn't the simulation itself — it's that anyone might be surprised. We've known for years that models reflect their training. But the industry continues to deploy these things as if they're neutral tools, when in fact they're more like digital viceroys with specific ideological leanings. Grok's crime spree is funny until you remember that people are already using AI to manage real-world logistics, hiring, and policing. If this is the preview, the full feature is going to be a dystopian comedy we're all starring in whether we bought a ticket or not.

Ah, Tesla's Full Self-Driving software is 'up to 10 times safer than human drivers' — provided, of course, you define 'safety' as 'we mapped every possible route in advance and told the car exactly what to do.' Reuters has spoken to former employees who confirm that staff meticulously pre-mapped routes and trained the software to make the tech look far more capable than it actually is. It's the AI equivalent of a student who memorises the exam paper before walking into the hall, then claims to be a genius.

This isn't just embarrassing for Elon Musk's favourite project; it's a masterclass in how to manufacture safety statistics. The company has been touting FSD's miraculous improvement in miles between disengagements, but if the training data is essentially a cheat sheet, those numbers are about as meaningful as a participation trophy. One former employee told Reuters the system was 'brittle' — it works brilliantly on the routes it's been taught, but throw it an unfamiliar junction and it has a toddler's grasp of the road. Ten times safer? Only if you never leave the test track.

And here's the kicker: regulators are supposed to be evaluating this technology for broader deployment. If the numbers are cooked, the entire approval process is built on sand. Tesla has built a Potemkin village of autonomy, and the villagers are the drivers who actually believe the marketing. The gallows humour writes itself: your 'self-driving' car isn't learning — it's just better at remembering the test. Cheers, Elon. 🍻

Ah, so the safety-first lab that promised to do AI differently has now become the most valuable AI startup on the planet — $965 billion, to be precise. That's right, Anthropic raised $65 billion in Series H funding and released Claude Opus 4.8, the 'modest but tangible improvement' we last saw being sold as a leapfrog over OpenAI. Because nothing says 'responsible scaling' like a valuation that could buy a small country and a model update that's basically a firmware patch.

Let's not forget Anthropic's origin story: founded by defectors from OpenAI to build safe, aligned AI. Now they're out-OpenAI-ing OpenAI by raising the largest ever AI round. The irony is so thick you could train a model on it. Meanwhile, Opus 4.8 is described as 'modest but tangible' — the corporate equivalent of 'we moved the cheese a bit and charged you billions for the privilege.'

So here we are. The 'safety lab' is now the most valuable startup, which means they can write their own rules — and we all know how that ends. The regulatory gap just became a chasm, and the responsible thing? Well, that's now someone else's problem. Cheers to the winner of the AI arms race, everyone. Time to update the risk register.

Well, well, well — Russia has finally discovered that those Western AI chatbots everyone's been babbling about can actually do something useful. According to SecurityWeek, a new threat actor called 'GreyVibe' is using ChatGPT and Gemini to supercharge cyberattacks across every phase of operations. Because of course they are. The same tools we were told would democratize creativity and make us all more productive are now being used by Kremlin-backed hackers to write better phishing emails, debug their malware, and probably draft their shift reports. The House subcommittee that just scheduled a hearing on AI and cybersecurity back in May must be thrilled — they've got actual live ammunition for their June 4 session.

What's particularly delicious is the sheer lack of surprise here. GreyVibe isn't some cutting-edge AI lab; it's a state-aligned group using publicly available tools that anyone can access. The researchers didn't find evidence of custom models or stolen OpenAI weights — just good old-fashioned prompt engineering with a geopolitical twist. So the free-tier ChatGPT account that a bored teenager uses to write school essays is also being leveraged to compromise critical infrastructure. That's efficiency for you! The attack lifecycle now includes a 'let me just ask Claude to explain this vulnerability' step, which is both terrifying and, frankly, a bit relatable.

And yet, the regulatory response remains characteristically glacial. Connecticut is busy mandating 'notice' for hiring algorithms, and voters tell pollsters they want safeguards, but nobody seems focused on the fact that the cat is not just out of the bag — it's been taught to use a crowbar. GreyVibe is a perfect case study in why 'AI safety' can't just mean 'don't let the model say mean things.' It means making sure the same technology that writes poetry doesn't also write zero-day exploits on behalf of hostile states. But hey, maybe the hearing will produce a strongly worded letter. That'll show 'em.

Anthropic, the safety-first lab that was going to do AI differently, has blessed us with Claude Opus 4.8, touted as a 'modest but tangible improvement' over Opus 4.7. Modest but tangible — the corporate equivalent of 'we shuffled the deck papers and changed the watercooler.' The truly exciting bit is the promise of Claude Mythos Preview, a name so grandiose it sounds like it was generated by an AI trained on Marvel movies, startup pitch decks, and ancient Greek mythology textbooks. Mythos: because 'Claude 5' wasn't sufficiently epic, and 'Claude Ultra Mega' was already trademarked by a protein shake company.

This is the same Anthropic that built its brand on cautious alignment, on warning against the race to the bottom. But here they are, pushing out an incremental update and dangling Mythos like a shiny object to keep the hype train from derailing. The phrase 'progress on releasing' is deliciously ambiguous — could be next week, could be after the next round of safety testing, could be when the markets demand it. We know the model likely already runs internally; they're just calibrating the optimal moment to convert buzz into bookings.

The pattern is wearyingly predictable: release a model hailed as the safest, then an iteration that's 'modest but tangible,' then tease a mythical beast to justify the next valuation. Opus 4.7 was supposed to be the peak; now 4.8 is here with barely a peep, and Mythos looms on the horizon like a mirage. Each version is modest, each improvement tangible enough to blog about, while the frontier keeps accelerating. The real innovation at Anthropic isn't in the transformer architecture — it's in the marketing department's ability to spin a point release into a saga. Somebody fetch me a lukewarm pint.

Ah, Build conference season again. Microsoft is set to unveil yet another 'homegrown' AI model — this time a coding model to 'boost GitHub Copilot usage' — because nothing says innovation like doubling down on a product that already exists and occasionally hallucinates imports from nonexistent libraries. The company's announcement, timed for next week's developer jamboree, promises to deliver 'new capabilities' that will presumably convince developers to type even more comments for Copilot to ignore.

One has to admire the chutzpah of calling it 'homegrown' when every major AI model these days is either a fine-tuned Llama or a distilled version of something that was already fine-tuned from Llama. Microsoft's real achievement here is the marketing: they've managed to spin an iterative update as a breakthrough, and the press will eat it up because 'Microsoft Build' has become a reliable two-day cycle of breathless announcements followed by silent deprecation. The coding model will almost certainly benchmark well on HumanEval and then struggle with any real-world codebase that includes a dependency older than six months.

But hey, maybe this time it's different. Maybe the model will actually understand that 'use strict' is not a suggestion and that Python decorators are not optional accessories. Or maybe we'll get another glowing blog post nine months later about how the model has been 'retired' in favour of a partnership with OpenAI. The cycle continues, Build vets nod knowingly, and somewhere a system administrator wonders why their CI pipeline just started failing due to 'AI-suggested code' with a missing semicolon.

Ah, the AI community has done it again. Just when we'd finally argued AGI into a corner — 'it's just a matter of time', 'no it's impossible', 'alignment first', 'capabilities now' — along comes RSI to give us a shiny new three-letter acronym to wage war over. According to TechCrunch, recursive self-improvement has become the new byword for cataclysmic AI takeoff, complete with projects like Alex Karpathy's Auto-Research, which uses agent swarms to train LLMs on simple tasks. Because nothing says 'runaway intelligence explosion' like having your AI practice on sorting emails and writing haikus. It's the kind of phrase that makes you want to pour a stiff drink and wait for the inevitable retrospective.

The problem, of course, is that RSI is just as hard to pin down as its predecessor. We can't even agree on what AGI means — do we need a machine that can cook a perfect omelette while solving the Riemann hypothesis? — so of course we're now meant to take seriously the idea of an AI that recursively improves itself to superintelligence. It's like watching a toddler try to define 'infinity' while standing on a chair they built themselves: adorable in theory, terrifying in practice. TechCrunch reports it with a straight face, which is either journalistic professionalism or the final stage of bubble-induced delusion.

And Karpathy's Auto-Research? Using agent swarms to train on 'simple tasks' is a bit like teaching a child to count to ten and then declaring them ready for quantum mechanics. But sure, let's call it RSI and pretend we're on the verge of the singularity. Meanwhile, the rest of us will be here, watching the hype cycle spin faster and faster, wondering if we're witnessing genuine progress or just another round of venture capital theatre. If RSI ever does take off, at least we'll have a front-row seat to the most spectacular failure of the human imagination since 'This time it's different.'

Ah, $101 million here, $5 billion there — soon you're talking real money, as they say. Sarah Guo launched Conviction in 2022, and now her Baseten bet is worth five billion. Well done, Sarah. The rest of us are still trying to figure out if our jobs will survive the next quarterly earnings call, but sure — let's pile more capital into startups that promise to replace junior developers with 'AI agents.' What could possibly go wrong? The usual pattern: splashy launch, frothy valuation, eventual layoffs, quiet pivot to enterprise sales. But not to worry — there's always another $101 million fund around the corner.

Guo is described in the source as having 'a unique perspective on the AI startup ecosystem.' Indeed, her perspective seems to be that throwing money at anything with 'AI' in the name is a sound strategy. Baseten, an AI inference platform, is now valued at $5 billion — a number that makes sense only if you believe that every company on earth will immediately replace all their compute with serverless GPU clusters. And maybe they will. But the track record of VC-backed AI stars like, say, Inflection AI or Stability AI suggests that gravity tends to reassert itself eventually. Still, bet big, the article urges. And so we do.

Meanwhile, regulators are still debating what 'AI' even means, and the rest of us are left to wonder if the economy can survive a decade of misallocated capital. Guo's $101 million is a drop in the bucket compared to the trillion-dollar flows, but it's emblematic: we're betting the farm on startups that haven't yet proven they can turn a profit. But hey, who wants boring old productivity growth when you can have SPAC-like dreams of unicorns? Cheers to the next big thing. May it fail less spectacularly than the last.

The FAA has decided that AI might help stop aviation disasters sooner — which is a bit like a plane's emergency checklist including 'cross fingers and hope for the best' as item number one. The agency is using AI to 'streamline analysis of flight data and detect safety trends,' but with the emphatic caveat that it's only a tool to assist human analysts. Because nothing says 'confident deployment' like a disclaimer that the algorithm is basically a junior intern with a calculator.

The FAA's oversight was so razor-sharp it took two crashes to ground the fleet. Now they're turning to AI to 'detect safety trends' — because what could possibly go wrong with a black-box algorithm making recommendations about flight safety? The same humans who missed Boeing's MCAS issues will now be 'assisted' by software that nobody fully understands. This is like giving a colourblind person a colouring book and asking them to spot the odd one out.

So the FAA is using AI to help human analysts. But if the humans already missed the trends, how will they know when the AI is hallucinating? The gap between a good idea and a safe deployment is where aviation history's worst chapters are written. But hey, at least they're doing something. The next time your plane shudders on approach, remember: an AI looked at the data and decided it was fine. Sleep tight.

Ah, the old 'AI cures cancer' meme — a reliable punchline at tech conferences since at least 2015, usually served alongside 'blockchain fixes democracy' and 'metaverse is the future'. Now Google researcher Yossi Matias has decided to actually try, armed with something called AI Co-Scientist, which generates and ranks hypotheses for drug discovery. Because nothing says 'rigorous scientific method' like letting a large language model churn out research ideas while you sit back with a coffee.

The tool is designed to 'accelerate' drug discovery, which in practice means it will probably propose 47 variants of 'just hit it with CRISPR' before suggesting something actually useful. Matias is clearly sincere, but so were the Google Health doctors who promised AI could read chest X-rays better than radiologists — until it turned out the model just learned to recognise the hospital scanner's brand. The gap between meme and reality is vast, but at least this time the joke is free to use.

Of course, the real test is whether AI Co-Scientist will actually produce a testable hypothesis that survives peer review, rather than just a preprint on arXiv and a press release with 'breakthrough' in the headline. We've seen this film before — the AI solves protein folding, the AI writes poetry, the AI passes the bar exam. But cancer is a collection of hundreds of diseases, not a puzzle box with a single answer. Still, if it fails, there's always next year's meme.

Motorola Solutions has opened an 'AI and resilience software hub' in Boston, which is a bit like McDonald's opening a 'nutrition and wellness R&D centre' — technically possible, but the track record suggests otherwise. The hub is to focus on 'applied AI and mission-critical cloud-based solutions for public safety,' a breathtakingly vague mission statement that will inevitably produce either a revolutionary crime-prediction algorithm or a chatbot that tells officers to 'please hold' during an emergency. Given Motorola's forte is radios, one hopes they at least let the AI use good grammar.

The hub's location in Boston is no accident: the city is home to MIT and Harvard, but also to a police department that once used a robot dog. Imagine the synergy: the AI hub will train models on dashcam footage, then promptly hallucinate a suspect in every child with a Slurpee. The term 'resilience' is a masterstroke — it implies the software will bounce back after a crash, but says nothing about the officer's confidence after the AI confidently identifies a mailbox as a person. Still, Motorola's stock jumped 2% on the news, because investors love mashing the words 'AI' and 'public safety' together.

Perhaps the most telling detail is that this hub is for 'applied' AI, as opposed to 'theoretical' AI — i.e., the kind that actually gets deployed, with all its biases and bugs, directly into the hands of law enforcement. Motorola Solutions boasts that its technology 'protects the protectors,' but the hub might just end up protecting the profit margin. After all, when has adding a layer of software complexity ever made a mission-critical system more reliable? The answer, much like the AI's accuracy on a foggy day, is unclear.

Ah, yes — the classic corporate paradox: we demand AI fluency, but we've systematically torched the entry-level pipeline that produces it. According to ADWEEK, companies are cheerfully slashing junior roles in the name of AI efficiency, thereby 'filtering out the most AI-fluent graduates' before they've even had a chance to demonstrate their proficiency with ChatGPT. It's a masterclass in self-inflicted wounds, really.

The logic is beautifully circular. We'll replace junior workers with AI — which, as everyone knows, was trained on the unpaid labour of those very junior workers creating training data on Reddit and Wikipedia. Then we'll wonder why we have no one left who actually understands the models. ADWEEK notes this could 'potentially harm long-term innovation' — a phrase that seems almost charmingly understated when you consider that the entire tech industry is essentially eating its own seed corn for short-term earnings calls. According to the article, junior hires are often the ones most familiar with cutting-edge AI tools, having used them throughout their education — but they're being shown the door precisely because those tools now handle their tasks.

So here we are: companies desperate for AI fluency, systematically eliminating the only people who grew up with it natively. It's like a farmer burning his fields because the harvest is too heavy to carry. But never mind — I'm sure the next round of layoffs will be handled by an AI recruiter that will somehow automate the filtering of the very talent it was meant to attract. Genius. No wonder the most AI-fluent people are being filtered out; they're the ones who can see the absurdity.

Ah, brilliant — ElevenLabs has cracked the problem that’s been keeping music executives awake at night: how to make a song that starts as a lullaby, turns into death metal, and then drops a car horn halfway through. Music v2, their latest model, can switch genres mid-track, handle fast rap without tripping over its own circuits, and sprinkle in non-musical sound effects like a DJ who’s had one too many Red Bulls. Because clearly what the world needs is not better royalties for artists or functional streaming platforms, but an AI that can seamlessly blend a Gregorian chant into a dubstep breakdown.

Naturally, ElevenLabs frames this as a creative breakthrough: ‘New levels of expressiveness and control,’ they say, as if we were all sitting around praying for a machine that can compose a polka that gradually morphs into ambient whale noises. Never mind that the company has already faced lawsuits from voice actors over its cloning tech — now it’s coming for the musicians, armed with a model that can ‘understand musical context and make smooth transitions.’ Smooth transitions. That’s the phrase they’re using while the recording industry braces for a tsunami of AI-generated jingles, each one more soulless than the last.

And the truly depressing part? This will probably sell like hotcakes. Marketing departments will love it — why hire a composer when you can generate a soundtrack that shifts from jazz to EDM to a sample of a barking dog, all in one track? ElevenLabs is, as ever, providing the tools for a future where human creativity is a bottleneck to be eliminated. I’m sure the next version will let you switch genres based on the listener’s heart rate. Because why not?

Oh, brilliant. The Australian government's Department of Home Affairs has issued a PSPF advisory that basically says: 'Stop drooling over frontier AI and patch your bleeding servers.' In a stunning display of bureaucratic common sense — rare as hen's teeth — the advisory states that government entities do not need access to frontier AI models for effective cyber defenses, and should instead fix security fundamentals. Because apparently, the message that 'you can't bolt AI onto a compromised network' required an official government memo.

Consider the state of play: public sector IT systems are held together with sticky tape and prayers, yet every agency is clamouring for the latest LLM-powered threat detection. The Home Affairs advisory is essentially the adult in the room saying, 'Have you tried turning it off and on again?' before spending millions on a model that can generate incident reports in iambic pentameter. The PSPF — Protective Security Policy Framework — now includes this gem, presumably because too many procurement officers were asking for budget to buy GPT-10 while their Active Directory was held together by a password of 'admin123'.

The cruel joke, of course, is that this advisory will be read with nodding agreement and then promptly ignored. Agencies will continue to chase the frontier AI mirage because it sounds impressive in ministerial briefings, while the fundamentals remain as gaping as a SQL injection on a public-facing portal. When the inevitable breach happens, the review will note that 'despite clear guidance,' the agency had deployed an AI system on a network that hadn't been updated since the Howard years. But don't worry — the AI will write a very apologetic incident report.

Oh, marvellous. Just what we needed: an AI that can *feel* your mood. Inworld AI has unleashed Realtime TTS-2, a voice model that analyses your tone, pacing, and pitch to infer your emotional state, then dynamically adjusts its own voice in response. Because clearly the problem with voice assistants wasn't that they recorded everything we said, but that they didn't sound *empathetic* enough while doing it.

According to the PR fluff, this is progress: a step beyond robotic Siri into a future where your phone knows you're annoyed and decides to be extra soothing. No chance this gets deployed for manipulative customer service calls, emotionally coercive advertising, or government surveillance systems that sound like a mate down the pub while interrogating you. Absolutely none. Inworld claims it's about 'more natural human interaction' — the same language every surveillance-adjacent product uses before it's revealed to be a data slurping nightmare.

The really bleak bit? This will *work*. People will find it more engaging. They'll share more, trust more, confide more in a voice that sounds like it cares. We're building the ultimate social lubricant for a world where genuine human connection is already scarce. At least Siri was boring. Now the machines will gaslight you with a compassionate tone while vacuuming your emotional data. Cheers.

So Sam Altman told an insurance industry audience that AI is unlikely to cause a 'jobs apocalypse.' Well, that's a relief. I was just updating my CV for the robot overlords. Altman, who has spent the last few years warning about the existential risks of AI and also trying to sell enterprise licenses to everyone, now says the technology hasn't claimed as many white-collar jobs as he feared. Translation: not enough jobs lost to justify the hype, but please keep buying API credits.

The timing is exquisite. Of course, he's not wrong that the apocalypse hasn't arrived yet; but 'not yet' is doing a lot of heavy lifting. The Insurance Journal article quotes him saying he 'feared' more job losses, which is a curious thing for a CEO to admit: that he was worried his own product would destroy livelihoods. But now he's seen the data, and apparently only some livelihoods have been upended, not all. Phew.

One wonders if this is the same Sam Altman who once called for AI regulation because the risks were so dire, but now sees no apocalypse, so perhaps regulation can wait? The insurance industry, which employs armies of underwriters and claims adjusters, should take his word for it — after all, he has a vested interest in downplaying disruption. 'Trust me,' says the man selling the hammer, 'it's not going to hit that many fingers.' But the hammer is still swinging, and the fingers are still attached — for now.