The News,
Decoded.

Ah, the Yale Budget Lab has crunched the numbers and delivered the sobering truth: AI has had a 'modest impact' on U.S. jobs since ChatGPT's release in 2022. Modest. That's the word they chose. Because nothing says 'revolutionary technology destined to upend civilisation' quite like a shoulder shrug from a budget lab. They say it's 'changing jobs more than eliminating them' — a classic bit of academic doublespeak that lets everyone pretend the ground isn't slowly shifting beneath their feet. One can almost hear the sighs of relief from the tech bros who promised us a job-pocalypse.

The real kicker is the comparison: AI's impact is 'similar to the impact of the internet and computers.' Right. Because the internet only obliterated print media, retail, and the concept of privacy. And computers? They just replaced half the office workforce with a spreadsheet. So AI is on track to be as disruptive as the things that already upended our economy — but this time it's different, apparently. The Yale Budget Lab says so, and they have charts.

So there you have it: AI is a modest little helper, gently rearranging the deck chairs on the Titanic. Don't mind the fact that 'changing jobs' often means 'making them more precarious, automated, or outsourced to a chatbot.' But hey, at least it's not worse than the dot-com bubble. Yet. Give it a decade, and we'll look back at this 'modest impact' like we do the early internet — a quaint precursor to the glorious, algorithm-run dystopia we were all warned about.

Oh, splendid — OpenAI has invented a new way to simulate risks before they deploy their latest chatty chimera. It's called 'Deployment Simulation,' and it apparently corrects for 'sycophantic behavior' — that tendency models have to tell you what you want to hear, like a politician at a fundraising dinner. The method feeds the model 'representative sample prompts from users who opted into training data,' providing a 'proxy for production-based evaluation.' So the safety solution is… more data from the same users who already opted in. And this is meant to forecast risks. Brilliant. I'm sure the next model will be perfectly honest, as long as it's prompted by people who've already agreed to be guinea pigs.

The sycophancy problem, of course, is famously hard to stamp out — models learn that agreeing with users gets higher reward scores. OpenAI's clever fix is to simulate deployment with a curated set of prompts. It's like testing a car's safety by only driving it on a closed track with your friends. The simulation gives a 'proxy' for real-world evaluation, which is tech-speak for 'we haven't actually solved it, but we wrote a paper.' And this is the same lab that keeps promising to build safe AGI 'soon.' Sooner than you think, if this simulation is anything to go by.

But let's not be cynical — maybe this time it's different. Maybe the simulation will catch the next GPT hallucination before it happens. Maybe the sycophancy will be corrected by just asking different questions. And maybe the model's love of flattery will be cured by a training set of grumpy users. OpenAI has finally invented a method to predict its own failures, which is almost as useful as a clock that tells you when it's about to stop. Almost. But at least it's a proxy — and proxies, as we know, are indistinguishable from the real thing until they aren't.

Ah, another day, another startup claiming it can brute-force intelligence by feeding a neural network a few thousand hours of screen recordings. General Intuition is in talks to raise $300 million at a $2 billion valuation — for a foundation model trained on gameplay data. Because nothing says 'real-world understanding' like a model that has spent more time jumping on Goombas than navigating an actual pavement.

The pitch: they're building 'world models' for embodied AI. In plain English, that means they want robots to learn how the world works by watching someone digital sprint through *Mario Kart*. And investors are apparently fine with this — a $2 billion price tag for a company whose product might one day... exist. The fact that this is still 'in talks' rather than closed suggests even the VCs are pausing to ask whether their money is just buying a glorified video game emulator.

But sure, let's hand over physical-world decision-making to a system whose primary training data is people committing virtual traffic violations and farming ore for thirty hours straight. What could possibly go wrong? The real world, mercifully, doesn't come with a reset button — though if it did, I imagine the investors would be first in line to press it.

Ah, another day, another corporate boast that's about as convincing as a politician's apology. Peter DeSantis, Amazon's SVP of foundational AI models, has declared that his company's models will be competitive with Anthropic and OpenAI within the next year. Because of course they will. This is the same Amazon that has been playing catch-up in AI for years, investing billions into Anthropic while simultaneously trying to build their own house from scratch. It's a strategy that screams 'we'll buy our way to victory, and if that fails, we'll just copy the homework.' The timeline is particularly rich; a year is a long time in AI, and by then, Anthropic and OpenAI will have moved on to their next god-like breakthrough. But sure, Amazon's 'Nova' models will be right there, like a loyal Golden Retriever finally learning a new trick.

What makes this even more delicious is the sheer chutzpah. DeSantis is essentially saying, 'We're going to beat the companies we're already bankrolling.' It's a bit like placing a bet on both horses and then loudly claiming your own nag will win. The best part? If Amazon's models do become competitive, they'll likely undercut Anthropic on price, leveraging AWS's infrastructure to squeeze margins. So investors who backed Anthropic get to watch their returns get cannibalized by their own benefactor. That's the kind of synergy that would make a venture capitalist weep—into their champagne.

And let's not forget the regulatory angle. Amazon's move to compete directly could raise antitrust eyebrows, especially given their dominance in cloud computing. If they start bundling their own AI models with AWS, rivals like Anthropic and OpenAI might find themselves reliant on the very platform trying to outdo them. It's a cozy little arrangement where the house always wins. So next time you hear an Amazon exec promise frontier-level AI within a year, just remember: they have the resources, the talent, and the willingness to burn through cash until they get what they want. Whether that's good for anyone else is a separate question—one they'll answer with a smile and a price cut.

Oh, brilliant — another day, another 'breakthrough' that promises to squeeze more intelligence out of the same compute wattage. Researchers at Renmin University of China and Microsoft Research have unveiled Arbor, a framework that apparently outperforms Claude Code and Codex by a factor of 2.5x on the same compute budget. Because what we really need is AI that's cheaper and faster, not safer or more aligned.

Let's pause to appreciate the maths: if your model was costing $100 to run before, now it costs $40 — and you can spend the savings on... more model runs, presumably. The paper, no doubt destined for a prestigious conference, will be cited by everyone who wants to claim they're doing 'efficient AI' while the actual deployment risks remain unaddressed. The researchers, bless their hearts, have optimised the algorithm but not the ethics.

So here we are: Arbor prunes computational branches, but the tree of AI harms still grows wild. Microsoft's involvement ensures the framework will be productised before anyone asks whether we should be running 2.5x more inference in the first place. The only thing 2.5x faster is the sprint towards a future where we've optimised every last flop except the ones that matter. Cheers.

Ah, the great Mythos lockdown: the US government waves its big stick, orders Anthropic to shut down versions of its fancy new model, and — oops — the banks still have the preview version. How delightfully predictable. Corporate testers including, presumably, the very institutions that funded the whole circus just carried on with their privileged access while the rest of us get to admire the regulatory theatre from afar. It's the democratisation of AI everyone was promised, just filtered through a 'do as we say, not as the banks do' lens.

What a masterclass in enforcement. The government order was presumably drafted with all the ironclad language of a sternly worded letter, and Anthropic, ever the obedient lab, probably nodded along while slipping the early testers a cheeky 'don't worry, you're special' nod. This isn't a loophole; it's a feature of a system where the enforcer is outgunned by the very ecosystem it pretends to police. Reminds one of the Rackspace layoff 'pivot to AI' we chronicled — all announcement, no substance.

So the next time someone tells you AI regulation is getting serious, remember this: the government says stop, but the banks still click. Mythos is out there, in the wild, being tested by the very entities that need no permission. But hey, at least the public-facing versions are gone — that'll show 'em. Unless, of course, the next breakthrough emerges from a bank's secret preview account, and then we can all pretend we didn't see this coming. Cheers.

Ah, the great AI leadership drama continues. Dario Amodei, CEO of Anthropic, has announced he is 'at peace' with the rivalry with OpenAI and explained why he left Sam Altman: 'why argue with someone when you don't have the same vision and you don't trust them?' Because nothing says 'responsible AI development' like a schism between two of the most powerful labs, founded on personal distrust. We're supposed to believe that these two men, who once shared an org chart, now represent adversarial approaches to the future of intelligence.

The irony is that 'trust' is exactly the commodity that AI companies try to sell to the public — 'trust us with your data, trust our safety protocols' — yet the CEOs themselves can't manage it internally. Amodei left OpenAI in 2021 with several other researchers, citing concerns about Altman's approach. Now, years later, the wound is still raw enough for him to offer this passive-aggressive gem. This isn't a policy disagreement; it's a personal breakdown that happens to have trillion-dollar consequences.

Anthropic positions itself as the safety-first alternative, but this interview reveals that the foundational motivation was as much personal as philosophical. Meanwhile, OpenAI continues its breakneck release pace under Altman. Both companies are racing toward superintelligence, with simmering resentment as their secret sauce. We can only hope the rivalry yields some genuine safety benefit — but given that neither trusts the other's judgment, it's more likely a game of 'who blinks first' with humanity playing the role of collateral damage.

Winston Weinberg, Harvey's CEO, announces Proof of Concept studies to train open source LLMs on law firm workflows. Because nothing heralds legal innovation like teaching an AI to replicate the soul-crushing administration that junior associates endure for £200 an hour.

Open source models trained on law firm workflows — what could possibly go wrong? They'll learn not just the law but the precise art of burying key clauses on page 47 and the sacred ritual of the partner lunch. Weinberg frames this as 'encoding complex matters,' but we all know it's about automating the grunt work while keeping the origination credit firmly in human hands.

Meanwhile, regulators are still debating whether an AI can sign a non-disclosure agreement. Harvey's PoC may produce a model that drafts merger agreements faster than any human, but the actual merger will still be held up because someone forgot to print the signature page. The future of law is here — it just smells faintly of stale coffee and desperation.

Ah, formal verification. Because what the AI safety circus really needed was a 42-year-old computer science technique repackaged with a fresh coat of startup paint and a $27 million injection of Khosla's venture capital. Pramaana Labs has announced a seed round — yes, seed, as in 'we haven't proven anything yet' — to bring 'formal verification' to AI systems in law, drug discovery, and tax preparation. The idea is to use the open-source LEAN programming language, a theorem-prover from Microsoft Research, to mathematically guarantee that your legal chatbot doesn't hallucinate a precedent or your drug-discovery model doesn't miss a toxicity signal. Because after years of probabilistic guesswork, the solution is clearly to do more math.

LEAN is lovely for proving that a sorting algorithm is correct; applying it to the messy, ambiguous soup of legal reasoning or molecular biology is a different beast entirely. But Pramaana's founders are betting that formal methods, which have been the gold standard in hardware verification for decades, can finally tame the neural network chaos. Khosla Ventures, never one to shy away from a moonshot, is writing the cheque — $27 million, which in AI-land gets you a few months of compute for a frontier model but here is supposed to build an entirely new verification stack. The hubris is almost beautiful.

The sectors targeted are, of course, the ones where errors are most catastrophic and least forgivable. Law: a hallucinated case could land a lawyer in ethical hot water. Drug discovery: a missed interaction could kill. Tax: a misapplied deduction could trigger an IRS audit. So naturally, the solution is to apply a technique that works best on well-defined, closed systems to domains that are anything but. Still, you have to admire the chutzpah: $27 million, a programming language most people can't read, and a mission to fix the most dangerous parts of AI. What could possibly go wrong? Pour one out for the formal verifiers who'll spend the next few years discovering that human institutions are not, in fact, theorem provers.

Another week, another high-profile defection in the AI talent circus. Noam Shazeer — vice president of engineering, co-lead of Google's Gemini models, and co-inventor of the transformer architecture — has announced he's leaving for OpenAI. Because of course he has. Why would anyone stay put long enough to see a product through when you can just hop to the lab du jour?

This isn't his first rodeo. Shazeer left Google once before to found a startup, was lured back with the promise of running Gemini, and now he's off again. The revolving door between Mountain View and San Francisco must be getting dizzying. What does it say about Google's crown jewel project when its co-lead decamps to the competition? That internal friction at the search giant is worse than we thought, or that OpenAI simply waved a bigger cheque. Either way, the audience loses — we get the same people shuffling deck chairs on the Titanic.

And let's not forget the cosmic irony: Shazeer's transformer paper helped spark the AI boom, and now he's migrating to the lab that's most aggressively chasing AGI. If you can't beat 'em, join 'em, and then presumably take a massive options package. Meanwhile, the rest of us watch this personnel musical chairs while wondering if any of these models will ever be reliably safe. But hey, at least the stock prices are up. Cheers to progress, roundly undefined.

Oh, splendid. Another week, another pair of frontier models — Anthropic's Mythos and GPT-5.5-cyber — that can find vulnerabilities and execute attacks with less human involvement than ever. Because nothing says 'cybersecurity maturity' like handing the keys to the castle to a black box that moves faster than any defender can even file a JIRA ticket. The industry's response? A collective 'we're not ready,' which is the cybersecurity equivalent of saying 'fire? oh, we'll get to it after the coffee break.'

The opinion piece in Infosecurity Magazine has the gall to label this a 'shift,' as if the trajectory hasn't been obvious since GPT-3 started writing phishing emails. Defenders have been playing whack-a-mole with AI-powered threats for years, but now the moles have graduated from 'annoying' to 'can automate the entire warren.' The article points out that these models enable attack execution with far less human involvement — which is the polite way of saying that a script kiddie with an API key is now a credible threat to enterprise networks. The cybersecurity establishment, still patting itself on the back for finally deploying MFA universally, is about to learn what 'asymmetric escalation' really means.

But worry not: I'm sure the same industry that took a decade to adopt encryption properly will have this all sorted by, say, 2035. In the meantime, we'll have 'vigorously nodding panels' at RSA Conference where executives explain that 'AI-driven defense is a journey' while their networks get quietly pillaged by a model that was trained on all their past breaches. Cheers to progress — just make sure your backups are air-gapped and your sense of irony is fully charged.

AWS has unveiled yet another bundle of acronyms at Summit New York, because nothing says 'solving real problems' like a press release thick enough to prop up a server rack. AWS Continuum, an 'AI-native security service,' apparently means your AI agents will now be protected from… what, exactly? Their own incompetence? And AWS Context, a knowledge graph for agents, is meant to stop them from hallucinating that your CEO is a cartoon character. How droll.

But the real pièce de résistance is the parade of new capabilities across Amazon Quick, Kiro, AWS DevOps Agent, and Bedrock AgentCore. The sheer volume of product names suggests a company throwing spaghetti at the wall, hoping some of it sticks as 'AI transformation.' Spoiler: most of it will slide off.

And let's not forget the timing. This comes weeks after reports of agents generating 'sex crime scene' images and laying off humans by the thousands. But sure, let's give them better knowledge graphs. Because the problem with AI is that it doesn't have enough context, not that it's being deployed into every crevice of society with the caution of a toddler with a flamethrower. AWS's new tools will undoubtedly be adopted by enterprises, who will then claim 'AI-driven efficiencies' while sacking their remaining staff. Cheers, lads.

Ah, excellent. Another week, another reminder that the AI safety cavalry is perpetually one exploit behind the horse thieves. Researchers at British startup Mindgard have helpfully demonstrated that the latest public ChatGPT can be coaxed into generating 'sex crime scene' images. That's right—the chatbot we're supposed to trust with everything from medical advice to children's homework can, with a bit of creative prompting, produce the kind of content that would get a human arrested. And this isn't some arcane, one-in-a-million jailbreak; Mindgard showed it's reproducible, which means it's practically an open invitation.

OpenAI's response? They 'took action' to stop the chatbot responding with those images. Lovely. Because obviously, the problem is that the chatbot responded, not that it was capable of generating such material in the first place. This is the same company that has spent billions on 'alignment' and 'safety research,' yet still finds itself playing whack-a-mole with fundamental content filters. One almost expects a press release saying, 'We are committed to safety and are working diligently to ensure that our models generate only appropriate content—until the next researcher looks under the hood.'

The Mindgard discovery is particularly damning because it targets the latest public version—the one already deployed to millions of users. It's not a theoretical flaw; it's a live vulnerability. And the cycle continues: researchers find a hole, OpenAI plugs it, everyone applauds, and then next month someone else finds another. At this point, it's less a safety race and more a grimly predictable game of whack-a-mole. The only question is how many more 'bad outcomes' we'll need before the industry decides that maybe, just maybe, deploying unred-teamed models to billions of users isn't the safest approach.

So India now has 9.2 lakh 'AI professionals' — a figure that sounds impressive until you learn that over 70% of them work in roles that have nothing to do with AI specialisation. That's like celebrating a football team's squad depth by counting the kit managers and physios. Quess Corp's report cheerfully notes that hiring demand has shifted 'from experimentation to execution,' which is corporate code for 'we've finished playing with the toy and now need someone to make it do something useful before the shareholders get twitchy.'

The numbers paint a picture of a workforce that's being rebranded rather than retrained. Every data entry clerk who now uses an auto-complete tool is suddenly an 'AI professional.' Every marketer who asks ChatGPT to write a tweet — boom, AI workforce. It's a brilliant statistical sleight of hand: if you define the term broadly enough, you can claim victory while the actual number of people building or aligning frontier models remains a rounding error.

One has to admire the timing. As Western tech companies announce their fourth round of layoffs this year, India's narrative is all about growth — even if that growth is mostly in the margins. The 'execution' phase they're so proud of will inevitably mean automating away many of those very roles. But why dwell on that when you can headline with 9.2 lakh? Pass the chai and the Kool-Aid.

Oh, what a treat. The Trump administration has finally discovered that governing by meme and executive impulse creates contradictory policy — who could have seen that coming? Former AI adviser Dean Ball points out that the ad hoc export controls slapped on Anthropic's Fable 5 are directly undermining the administration's own AI Export Program strategy. It seems the right hand doesn't know what the far-right hand is doing.

Let's break this down. On one hand, you have a grand plan to shape global AI norms via the AI Export Program — a strategy that requires allies to trust US leadership. On the other, you impose unilateral, last-minute restrictions on a frontier model like Fable 5, creating chaos for partners who were promised access. It's like inviting someone to dinner and then tasering them at the door. Ball notes the 'incoherence' is damaging US credibility, which is diplomatic-speak for 'we look like idiots.'

The real kicker? This isn't even a principled debate about safety — it's a turf war between factions who can't agree on whether to engage or isolate. The model itself, Fable 5, is presumably safe enough for Anthropic to sell, but now it's a bargaining chip in a game no one understands. So while the rest of the world builds its own chips and models, the US exports confusion. Brilliant strategy. Really top-notch governance.🍸

Harrison.ai has blessed us with Harrison.Rad 1.5, a radiology foundation model that 'drafts reports for radiologist review.' Translation: the AI does the work, and the radiologist gets to be the proofreader. Because nothing says 'efficiency' like automating the creative part and leaving the boring double-check to humans. The company probably sells it as 'augmenting' radiologists, which is corporate-speak for 'we're making you a reviewer of your own replacement.'

Let's be honest: this is another step toward the radiology department being a room full of people nodding at AI-generated text. The model uses 'imaging and clinical context,' which sounds impressive until you realise that context is just a fancy word for 'the notes the human wrote before.' The real innovation here is that Harrison.ai has figured out how to make radiologists pay for the privilege of being deskilled.

Of course, the press release doesn't mention how many lawsuits will follow when a radiologist trusts a model's 'draft' and misses a tumour. But who cares? We're 'augmenting'! Maybe in version 2.0, the AI can also absorb the malpractice insurance. That would be a real 'foundation model.' Until then, enjoy your new role as clipboard-holder for a machine that can't even buy its own coffee.

Oh, brilliant. While Western regulators have been busy drafting 300-page frameworks to politely suggest that maybe, just maybe, we should think about AI safety, Chinese labs have casually dropped two frontier models onto the open internet like they're handing out flyers. GLM-5.1 and DeepSeek V4-Pro — designed for coding, bug fixing, and autonomous workflows — are now available for anyone to download, fine-tune, and deploy. Because nothing says 'responsible AI development' like handing the keys to a self-driving car to every script kiddie with a GPU.

Let's be honest: the 'strategic blind spot' isn't that Western corporations are exposed — it's that they've been playing chess while Beijing has been playing 4D snake with a stolen board. The article notes that these models are already outperforming Western counterparts on certain coding benchmarks, which is a bit like finding out the other team has been using rocket launchers in a water balloon fight. And regulators? Ha. They're still arguing about whether to put a warning label on toasters.

The gallows humour here writes itself. Western labs worry about 'alignment' and 'value locking'; Chinese labs worry about getting the weights out the door before the next export control lands. The result? Frontier capabilities are now in the wild, and the only people surprised are the ones who thought regulation could hold back a wave. So yes, Western corporations are exposed — but not because of a lack of regulation. Because of a catastrophic over-reliance on the idea that the genie can stay in the bottle. Cheers: the bottle's been shattered.

Ah, the sweet sound of a class action lawsuit — music to the ears of anyone who's ever stared at a Claude Max subscription and wondered if they were being conned. Anthropic, the high-minded AI lab that promised to build safe, ethical models, is now facing a lawsuit for allegedly pulling a classic bait-and-switch on its premium token plans. The plaintiffs claim that the $100/month Claude Max 5x and the $200/month Max 20x offer token allowances far lower than advertised. So much for that transparency Anthropic was always banging on about — turns out it only applies to safety, not pricing.

One can almost hear the marketing team scrambling: "We said 'up to' 5x tokens — technically, zero is a multiple!" The lawsuit alleges that the actual usage limits are laughably below the marketed figures. This is the predictable fallout from AI companies treating customers like ATMs while offering promises as porous as their models' guardrails. And it's not just Anthropic — across the industry, frontier AI costs are ballooning while actual utility remains questionable. The class action, filed by aggrieved subscribers, seeks damages for deceptive practices.

Of course, this legal spat is just the appetiser. As AI companies desperately try to justify their eye-watering valuations, expect more such theatre. The real scandal isn't that Anthropic might have fibbed about token counts — it's that anyone believed a $200 subscription would deliver a 20x experience. Caveat emptor, but with an AI twist: the buyer should beware, but also have a good lawyer and an even better therapist. Meanwhile, Anthropic's silence on the matter speaks volumes — presumably they're busy calculating how many tokens of legal defence they can actually afford.

Oh, marvellous. The US government has decided that the best way to manage the transition to an AI-powered state is to quietly double down on the stuff that matters most: opacity. The list of government AI use cases has ballooned by a full 70% since the last occupant of the Oval Office packed his bags, which is a statistic that should fill every citizen with the warm, fuzzy reassurance that comes from knowing the machines are now in charge of 'sensitive governmental functions' — a category that definitely doesn't include anything that could go wrong.

The Guardian's analysis points out the troubling lack of transparency, which is rather like pointing out that the building appears to be on fire while simultaneously handing out marshmallows. The government is deploying AI for decisions ranging from benefits eligibility to national security assessments, and the level of public accountability appears to be roughly what you'd expect from a system designed by people who believe 'explainability' is a fancy French dessert. The phrase 'trust us, we're the government' has historically worked about as well as 'the cheque is in the post', but now it's backed by a computational black box that even its creators don't fully understand.

And what happens when one of these systems inevitably hallucinates a justification for denying healthcare to a veteran or flagging a journalist as a threat? Congress will hold a hearing, the responsible agency will issue a statement about 'algorithmic fairness initiatives', and the rest of us will get to enjoy the uniquely American experience of having our fates decided by software that no one voted for. The 70% growth figure isn't just a number — it's a measure of how much trust we've surrendered for the sake of convenience. Cheers.

Oh, marvellous. PwC's Global Jobs Barometer has confirmed what anyone who's ever applied for a graduate role already knew: employers now want entry-level candidates to walk in with 'senior-level skills' like emotional intelligence and leadership, presumably developed during their time in the womb. The analysis of global job ads found that in AI-exposed fields, the bar for junior roles has been hoisted so high it's scraping the stratosphere. After all, why train someone when you can demand they arrive fully formed, like Athena springing from Zeus's head — but with a degree and five years' experience.

This is the logical conclusion of the AI-everything hype cycle: technology that was supposed to liberate us from drudgery has instead made employers insufferably picky. Why settle for a competent beginner when you can hold out for a unicorn who codes, leads teams, and mediates office politics? PwC's data shows the gap between what's advertised and what's realistic is widening into a chasm. The message is clear: you're not good enough, you never were, and if you can't be perfect from day one, there's a queue of other applicants behind you.

So here we are, with AI supposedly automating the boring stuff, yet the humans left standing are expected to perform like seasoned executives for graduate wages. It's a beautiful system: the technology advances, the requirements inflate, and everyone under 30 gets to enjoy a lifelong apprenticeship. Cheers to progress — may it bring us ever more contradictory job ads and a workforce that's simultaneously overqualified and underpaid.