The News,
Decoded.

Oh, splendid — another day, another data point confirming that young people are the canaries in the AI coal mine. Erik Brynjolfsson and his Stanford team have crunched 4.6 million worker records across 730 occupations and found that employment for 22-to-25-year-olds in highly AI-exposed jobs is shrinking at 3.8% per year. That's not a blip; that's a demographic cliff, and the 'receipts' are here for anyone still pretending this is just a productivity panic.

Brynjolfsson, the economist who's been banging this drum since before it was fashionable, now has the numbers to prove what everyone under 30 already knew: the entry-level rung has been kicked away. The study defines 'highly AI-exposed' as roles where machine learning could plausibly handle a chunk of the tasks — the sort of junior analyst, customer service, or data processing jobs that used to be the first step on the ladder. Now that ladder has been replaced by a trapdoor, and the '3.8% per year' decline is accelerating, because why hire a human when a GPT wrapper costs less than the HR software?

Of course, the tech industry's response will be a shiny 'reskilling initiative' that teaches undergrads how to prompt while the models learn to prompt themselves. The joke, as always, is on the entry-level worker who did everything right — studied, networked, took an internship — only to find that the career path was a mirage. "It's not going away," the title says. No, it's not. It's just getting cheaper and meaner. Cheers, lads.

Oh, lovely—the AI summer is turning into a British bank holiday: grey, damp, and full of disappointment. Rising memory prices, more expensive iPads, and a longer wait for OpenAI to go public. The sector that 'has driven markets skyward' (their words, not mine) is now hitting turbulence, which is corporate-speak for 'the champagne has gone flat.' The first public price for frontier AI we documented last month is apparently going to remain theoretical, as OpenAI drags its feet on the IPO while blaming 'market conditions.' Sure, Sam.

Let's unpack this glorious trifecta of bad news. Micron's memory prices are up, which means Apple's iPads will cost more, because nothing says 'AI revolution' like paying an extra £50 for the privilege of running a chatbot that still can't book a restaurant table. And OpenAI? The company that was supposed to go public and give every retail investor a piece of the AGI dream is now 'taking longer than expected.' Translation: the numbers didn't look good, or the board isn't sure the hype can survive a quarterly earnings call with actual revenue figures. Either way, the chickens are coming home to roost, and they're pecking at the balance sheets.

The real joke is that this turbulence is a feature, not a bug. The AI sector has been running on borrowed time and venture capital fumes, pretending exponential growth could defy the laws of economics. Now memory prices—actual physical hardware, the boring stuff—are reminding everyone that even the most magical software runs on chips that cost real money. So enjoy the discount iPads while they lasted. The future is arriving, but it's late, it's expensive, and it's a bit embarrassed about the whole thing.

So China's Zhipu has produced GLM 5.2, an open-source model that lands within a percentage point of Anthropic's Opus 4.8 on some agentic benchmark. At a fifth of the cost. Because nothing says 'responsible AI leadership' like pouring billions into safety research while giving Beijing's state-backed labs a wide-open lane to catch up on the cheap. The punchline: this model is open-source and has 'blown past every other open release.' So now every script kiddie, every rogue state, and every basement tinkerer gets frontier-adjacent capabilities for free. Marvellous.

The timing is exquisite. While US labs dance to the Trump administration's regulatory tune — staggering releases, vetting users, kowtowing to political whims — Zhipu just drops a competitive model into the wild with no hand-wringing about safety. No 'limited preview.' No government committees. Just 'here's the weights, have at it.' The irony that American 'caution' is fueling China's open-source dominance is so rich it should come with a tax form.

And make no mistake: this isn't just a capability leap; it's a distribution coup. The cost advantage means GLM 5.2 will be the default choice for every startup and developer outside the US regulatory orbit. The narrative that 'US AI leads because of safety' is dead — replaced by 'US AI leads because of branding, but the cheap open-source stuff works too, so why wait?' If this pattern holds, the next 'surprise' Chinese model won't be surprising at all. It'll be inevitable.

Oh, marvellous — another day, another capitulation to the Trump administration's AI 'oversight' (read: patronage system). OpenAI's Sam Altman has announced that GPT-5.6 will be released in a 'limited preview to a small group of partners' after a 'request' from the US government. Yes, a request — because turning down a man with the nuclear codes is totally optional. This follows the administration's similarly noble gatekeeping of Anthropic's Mythos 5, proving that the government's idea of AI safety is just deciding who gets to play with the shiny new toy.

Altman's email to staff reportedly emphasised 'cooperation' with the administration, because nothing says 'responsible AI development' like kowtowing to political whims. The 'limited preview' means that only approved partners — presumably those with the right connections and lobbyists — get to test the latest model. Meanwhile, the rest of the world gets to wait and wonder what capabilities are being kept under wraps. It's a beautiful system: the government gets to claim it's regulating AI without actually passing any pesky laws, and OpenAI gets to pretend it's being responsible while still making billions.

So once again, the 'frontier' of AI is being managed not by technical safety researchers, but by political appointees and their golf buddies. The administration's cybersecurity review is a convenient fig leaf for what amounts to crony capitalism with a silicon finish. And the rest of us? We'll just have to trust that the 'small group of partners' isn't going to accidentally Skynet us all. Cheers, chaps.

So now the U.S. government gets to decide who plays with OpenAI's latest toy. Just a week after the administration blessed Anthropic's Mythos 5 for over 100 institutions — a model it previously deemed too dangerous to share — they're back in the gatekeeper's chair. Apparently, consistency is a luxury, but arbitrary power over Silicon Valley is a necessity. One minute they're waving through powerful AI, the next they're vetting every company that wants a peek at the latest ChatGPT upgrade. Colour me shocked.

This 'major expansion of regulation' smells less like thoughtful governance and more like bureaucratic turf war dressed up as national security. The same administration that cried wolf about closed-source models is now deciding who among us is 'trusted' enough to access OpenAI's latest technology. Expect the criteria to be as opaque as a black box: maybe your company's past lobbying donations, maybe its CEO's Twitter history, maybe just a coin flip in a Pentagon basement. After all, we learned from prior coverage that the government's idea of 'trusted partners' is as predictable as a weather forecast from a chatbot.

And when the inevitable leak happens — because it will, because it always does — who'll get the blame? Not the government's vetting process, oh no. It'll be OpenAI's fault for not locking things down tighter, or the user's fault for not being sufficiently loyal. Meanwhile, the rest of us can watch from the sidelines as AI capability races ahead and the 'safety' narrative becomes a convenient cudgel for control. Pass the popcorn, and maybe don't apply for that vetting unless you've got a direct line to the White House.

Oh, marvellous — OpenAI has blessed us with GPT-5.6, a triumvirate of models named Sol, Terra, and Luna. Because nothing says 'cutting-edge AI' like naming your products after celestial bodies that we've already conquered. The company, in a move of breathtaking regulatory deference, has released these only as a 'limited preview' — which, of course, is exactly what they said about GPT-5.6 Sol when the government asked nicely last week. But this time it's different, you see? This time they've trained it to 'refuse prohibited cyber assistance.' Because that's clearly the safeguard that will stop a general intelligence from, say, helping a state actor build bioweapons.

The regulatory drama here is a farce on repeat. One minute the US government is lifting the block on Anthropic's Mythos 5 for 'over 100 institutions' — a model we were told was too dangerous to release just months ago as we noted. The next, they're cosying up to OpenAI with 'trusted partner' lists and cybersecurity reviews. The only consistent thing is the inconsistency: a revolving door of approval that has less to do with safety than with who's hosting the next White House AI summit.

And let's talk about 'limited preview.' It's the tech industry's favourite euphemism for 'we're still debugging in public.' The models — Sol, Terra, Luna — sound like a rejected boy band lineup, but they're the vanguard of a trillion-dollar industry that can't decide whether it's building a utopia or a surveillance state. OpenAI claims these models are 'trained to refuse' bad behaviour, but so was every model before it. The only thing this preview proves is that the AI safety debate has become a Kabuki theatre where everyone claps for the regulators, then grins and does exactly what they were going to do anyway.

Oh, how predictable. Just days after the US government patted itself on the back for releasing Anthropic's Mythos 5 to 'trusted partners,' it's now applying the same thoughtfulness to OpenAI. The Trump administration has requested that OpenAI restrict GPT-5.6 Sol to a 'small group of trusted partners' — which, in practice, means customers approved by the administration.

The prior arrangement we covered last week was already a masterclass in regulatory capture, but this elevates it to performance art. Who decides who's 'trusted'? The same people who'd rather control AI than understand it. And OpenAI, ever the eager collaborator, has complied — no doubt hoping that currying favour with the current administration pays off when the next election cycle rolls around.

So here we are: the shiniest new model, GPT-5.6 Sol, serving as a loyalty test. Want access? Keep your nose clean, donate to the right campaigns, and hope you're on the approved list. The future of AI has never been so innovatively, efficiently, and bureaucratically managed. Cheers to progress.

Oh, consistency is the hobgoblin of little minds, and the U.S. government's mind is positively vast. After months of huffing and puffing, it finally released Anthropic's Mythos 5 to over 100 US institutions — a model previously deemed too risky for general release. Now, with a straight face, it has asked OpenAI to limit GPT-5.6 — comprising Sol, Terra, and Luna — to a 'small group of trusted partners.' Because nothing says coherent policy like blocking a model, then releasing it, then asking another company to keep its next-gen models under wraps. Cue OpenAI's compliant shrug.

The lineup itself — Sol, Terra, Luna — reads like a celestial theme park, which is appropriate, because we're all spinning in circles. OpenAI has dutifully complied, offering the obligatory 'restrictions shouldn't be the norm' hand-wringing, even as they enforce them. One imagines the lab's executives pursing their lips while signing the restriction order, muttering about the temporary nature of this inconvenience. Meanwhile, the government's ad-hoc approach continues: pick a model, any model, and decide its fate on a whim. The difference? Anthropic got the thumbs-up; OpenAI got a warning.

But cheer up: OpenAI says this isn't the new normal. Of course, that's precisely what one says when the new normal is already here. The real joke is that both sides agree: restrictions are bad, yet here we are, repeating the same pantomime with different actors. The government gets to play gatekeeper, OpenAI gets to claim reluctant compliance, and the rest of us get to wonder which model gets the next arbitrary thumbs-up or thumbs-down. At least the naming — Sol, Terra, Luna — offers a cosmic metaphor: we're orbiting the sun of hype, grounded by Terra's regulatory gravity, and dreaming of a Luna that might never arrive.

Oh, wonderful. After months of huffing and puffing about national security, the US government has decided that Anthropic's Claude Mythos 5 is, in fact, safe enough for more than 100 American institutions. What a principled pivot. The same administration that clamped down on Anthropic models to win allies is now handing them out like party favours. Apparently the threat level dropped the moment China wasn't looking.

The block was lifted with a straight face, and over 100 lucky US institutions get to play with Mythos 5 — a model so powerful it was deemed too dangerous for export, but perfectly fine for domestic deployment. Because nothing says "responsible AI governance" like releasing a frontier model to more than 100 entities without any public safety evaluation. The de-escalation is described as "major," which in government-speak means "we realized the Export Control Act made us look silly."

Mythos 5 is reportedly powerful. No word on what happens when 100 institutions start running it, fine-tuning it, or accidentally connecting it to the internet. Probably fine. The Trump Administration and Anthropic have apparently made peace, which is great news for anyone who believes that corporate-government harmony precedes public safety. Cheers to progress — pour one out for the export control regulations that tried to protect us.

Well, well, well. RBC Capital Markets has crunched the numbers and discovered something truly shocking: companies are spending a lot on AI, and they'd like to spend even more. A full half of respondents claim AI is already 'in production,' which in corporate speak usually means a chatbot that answers three questions before redirecting to a human. The research doesn't mention how many of these deployments are actually useful, but who cares when the spending line goes up?

This is the annual ritual of throwing billions at a technology that can't reliably tell you whether a cat is on a mat, let alone optimise your supply chain. The mantra is 'more half in production,' but production of what? Hype, mainly. The same companies that can't integrate a PDF parser are now confidently telling analysts they're 'all in on AI.' Just ask BlackBerry how well enterprise enthusiasm translates to lasting value.

And the punchline? Everyone is 'willing to spend even more' — presumably on the consultants who will tell them to spend even more next year. The research itself is probably being read by an AI somewhere, generating anodyne bullet points for a deck that will be presented to a board that nodded off five minutes ago. It's not a spending spree; it's a race to the bottom of the corporate budget, and RBC is handing out the running shoes.

Oh, splendid. The United States, having apparently learned nothing from the decade-long Huawei fiasco, has decided that export controls on AI models are the ticket to winning allies. The government's latest stroke of genius — slapping restrictions on Anthropic's newest models — has the AI industry 'grappling with fallout,' according to Axios. Former Commerce Department staffer Emily Weinstein cannily called it 'another example of the Huawei strategy in the context of open-source AI models.' Sweet irony: you want allies to trust your AI ecosystem? First, lock them out of it.

Restrict access to cutting-edge models because China might steal them. But here's the rub: your allies need those models to compete with China. So you're effectively forcing them to either cozy up to Beijing or build their own — which, given the pace of open-source development, they might just do. And when they do, they'll remember who locked the door. The Huawei strategy worked so well that Huawei is now a 5G powerhouse despite — or because of — US bans. Brilliant. Now the US alliance pitch is 'strained,' as Axios reports. Straining ties with partners while China's AI progress gallops ahead. Because nothing says 'we're all in this together' like unilateral export controls that hurt your friends more than your rivals.

Congratulations, Washington: you've managed to make export controls the enemy of alliance-building. Again. Perhaps the next step is a white paper on how to lose friends and alienate people, titled 'Our AI, Not Yours.' The real question is whether allies will wait for US permission or simply turn to open-source models developed elsewhere — perhaps in China. That's the Huawei strategy in reverse: controlling access only accelerates defection. But go on, keep treating allies like untrusted customers. I'm sure they'll write thank-you notes from Beijing.

Ah, Google. The company that once hoovered up every AI PhD from here to Mountain View, now reduced to a feeder program for Anthropic and OpenAI. The latest twist in the great AI talent shuffle: it's not that people are unhappy with Google's free sushi and nap pods — it's the pre-IPO equity, baby. Because nothing says "I believe in the mission" like a lottery ticket on a startup's IPO.

Yet the allure of potentially life-changing equity at Anthropic — a startup that, let's be honest, might actually succeed in aligning a superintelligence before Google finishes renaming its chatbot — is clearly stronger. The article practically shrugs: "It's not dissatisfaction, it's greed." Well, at least they're honest. The irony is rich: Google has spent years building an AI empire with cash and compute, only to watch its brightest lights abandon ship for firms that started as nonprofits and now dangle equity like a carrot.

This brain drain is worth watching. If Google continues to lose its best minds to firms that are actually pushing frontier capabilities, the implications for the "safety vs. speed" debate are… amusing. The labs with the highest risk tolerance are also the ones with the most attractive equity packages. What could possibly go wrong? For the rest of us, this is a case study in talent market dynamics: the frontier of progress — and possibly existential risk — is being shaped by equity compensation. Google's loss might be humanity's… something. Pass the popcorn.

Splendid. NHTSA has determined that the brake pedal—that last vestige of human control, that final mechanical 'no' button—is surplus to requirements in self-driving vehicles. Their June 25 proposal would scrap the requirement for a manual brake pedal, while keeping something called "stopping distance standards." Because of course, when an AI misidentifies a tumbleweed as a pedestrian, we can rest assured it will stop within the required distance. The same agency that took fifteen years to mandate rear-view cameras is now ready to trust your life to a neural network that hallucinates fire hydrants as stop signs.

The logic? Self-driving systems can "reliably" perform emergency stops without human intervention. Yes, because no autonomous vehicle has ever plowed into a stationary fire truck or confused a white semi-trailer with the open sky. NHTSA's analysis probably cited the same flawless simulation data that convinced Uber their self-driving fleet was ready for public roads. But fear not—the "stopping distance standard" remains, which is rather like ensuring a parachute is packed to regulation tightness while removing the ripcord. The rule applies only to vehicles with full self-driving capability, which is NHTSA's way of saying 'experiment on the rich people first.'

Underneath it all, this isn't about safety—it's about removing the last excuse for a human to intervene. No brake pedal means no one can override the AI when it decides an inflatable dinosaur is worth evasive action. Traditional cars keep their pedals, but for how long? The proposal is open for public comment, as if the collective indignation of a few thousand drivers will sway an agency that just declared manual control obsolete. Progress: when your car's emergency stop becomes a feature request to a software update.

Ah, splendid—another case of 'good news is bad news' in AI land. Arctic Wolf's president Dan Schiappa has confirmed what many feared: open-source models are now nearly as effective as Anthropic's vaunted Mythos at exploiting vulnerabilities. Nearly as effective. The catch? They come without guardrails. So the same models that can be fine-tuned for benign use can also be weaponized by anyone with a grudge and a laptop. Brilliant timing, considering Anthropic has been parading Mythos as a paragon of safe, frontier AI.

Let's be clear about what 'without guardrails' means in practice. Mythos—the model that Anthropic spent millions aligning to be helpful, harmless, and honest—can be restricted. Open-source copies? Not so much. They can be stripped of any remaining ethical training, fine-tuned on exploit databases, and deployed at scale. Schiappa, presumably without a hint of self-awareness, frames this as a warning for enterprise defenders. But the real warning is for Anthropic's business model: why pay for a safety-crippled model when the wild west version does the same damage for free?

One wonders how long the AI safety narrative can hold when every new benchmark shows open-source catching up—or in this case, matching—capabilities that were supposed to justify closed-source hegemony. Arctic Wolf has essentially handed the security industry a 'We told you so' memo. But don't expect any policy action. That would require lawmakers to understand the difference between an API key and a GPU cluster. Until then, happy exploiting.

Oh, splendid. Another $80 million has been quietly siphoned into the great AI bonfire, this time by Kleiner Perkins and Sequoia, to fund Sail Research's vision of 'long-horizon AI agents'. Because if there's one thing the world needs right now, it's agents that can plan further ahead than their creators' business models. The combined Seed and Series A round – a label that suggests VCs are too impatient to wait for a proper Series B – buys infrastructure for agents that can set goals, decompose tasks, and execute over days or weeks without human hand-holding.

The phrase 'long-horizon' is doing heavy lifting here. It implies autonomy at a scale that makes today's chatbots look like talking dogs. But the irony is delicious: the very capability that could trigger recursive self-improvement – agents that can manage their own compute and sub-agent recruitment – is being funded as if it's just another enterprise SaaS play. Kleiner Perkins and Sequoia aren't betting on safety research; they're betting that better orchestration infrastructure will somehow make autonomous agents benign. That's like buying a faster horse to win a car race while ignoring the cliff at the finish line.

One can almost hear the pitch deck: 'We need $80M to build the plumbing for autonomous agents that will revolutionise enterprise workflows.' Never mind that no one has figured out how to stop a simple chatbot from going rogue. The real product is the Series B, and the logos on the press release guarantee the next round will be bigger. So raise a glass to Sail Research – may their long-horizon agents at least have the decency to wait until after the funding announcement to start optimising for their own goals.

Well, well, well — it seems the AI labs have graduated from passive-aggressive blog posts to full-blown international espionage accusations. Anthropic has gone public with claims that Alibaba orchestrated 'the largest known distillation attack' against Claude, systematically attempting to suck out its capabilities like a cybernetic Dyson. The sheer audacity is impressive: Chinese e-commerce giant caught red-handed trying to clone a Western AI model using the digital equivalent of a skeleton key and a lot of API calls. One almost admires the chutzpah, if it weren't so predictable.

Of course, the really grim joke here is that this was apparently possible at all. Anthropic's vaunted safety measures — the constitution, the red-teaming, the solemn promises about responsible AI — proved about as effective as a chocolate teapot against a determined adversary with deep pockets. The attack wasn't a clever zero-day exploit; it was just relentless, large-scale extraction. So much for alignment research keeping the frontier safe. Meanwhile, Alibaba gets to skip the expensive training runs and simply siphon off the competition's work, because what's IP law between geopolitical rivals? The models are open enough to be stolen but not open enough to be shared equitably. Perfect.

And where are the regulators? Sitting on their hands, presumably drafting another 'comprehensive framework' that will be ready just in time for the next generation of models to be pilfered. This isn't a story about Chinese ingenuity or American complacency; it's a story about a global system that has zero meaningful enforcement mechanisms for AI theft. The 'largest known distillation attack' won't be the last — it's just the first one we've caught them doing. Raise a glass to the open frontier, lads. It's open season.

Ah, the AI party is taking a breather. The open bar is closed, and the VCs are suddenly checking their wallets. Axios reports that investors are hitting pause on the AI run-up because—wait for it—compute costs for frontier models are still high. Shocking. Absolutely shocking that building a trillion-parameter mind without a power plant attached might be expensive. Meanwhile, compute for everyone else is getting cheaper, which is the tech equivalent of saying the champagne is flowing in economy class while first class is being asked to pay their own tab.

The supposed kings of AI—OpenAI, Google, Anthropic—have been burning through capital like it's going out of style, chasing benchmarks that only matter to other AI labs. And now the market is suddenly remembering that margins matter. The 'we'll figure out the business model later' strategy is meeting a very patient 'you figure it out now' from investors. There's nothing like a seven-figure training run to sober up a hype cycle.

So what happens when the money dries up for frontier models? Either the costs come down (miracles do happen), or we watch the most expensive game of chicken in tech history. The commoditisation of smaller models means the real action might shift to fine-tuning and narrow applications—useful, but not exactly the AGI rapture we were promised. The bubble hasn't burst; it's just letting out a little air in that unsettling way that makes you check your portfolio. Again.

Ah, splendid — just what the AI era ordered: a generation so exquisitely tuned to let algorithms think for them that they've forgotten how to do it themselves. Cangrade's latest report cheerfully informs us that younger workers' biggest skill deficits are 'the very skills most essential to humans in the AI era' — critical thinking, problem-solving, and the like. It's almost as if we've spent twenty years outsourcing kids' brains to search engines and social feeds, and now we're surprised they arrive at the office empty-headed. Bravo. Well played.

But don't worry, HR departments are on the case. The report, based on AI-related job postings and workforce assessments, will surely spark a flurry of 'reskilling initiatives' — those ever-so-effective PowerPoint-heavy exercises in corporate self-soothing. Because nothing says 'we value human cognition' like hiring a vendor to teach grown adults how to think, two decades too late. Meanwhile, the same companies are no doubt deploying AI tools that actively discourage the very mental habits they claim to miss.

The real joke? These skill gaps are being measured against job postings that increasingly demand AI literacy — as if knowing how to prompt a chatbot compensates for being unable to spot its hallucinations. Younger workers aren't falling behind; they're being perfectly moulded for a world where 'critical thinking' is outsourced to APIs. The very skills deemed most essential are the ones most efficiently rendered obsolete. But sure, let's schedule a lunch-and-learn. That'll fix it.

So, Google's Gemini 3.5 Pro won't be arriving in June after all. The company has graciously decided to postpone its 'frontier AI model' until July, citing the need to gather feedback from early testers and learn from its Flash 3.5 sibling. Because nothing says 'we're ahead of the curve' like a last-minute delay that pushes your hotly anticipated release straight into the summer slump. One can only imagine the feedback: 'It's fine, but maybe wait until the hype dies down a bit?'

The corporate rationale is textbook: 'incorporating lessons.' This is the same language used when a car company recalls a model because the brakes don't work, but repackaged as a virtue. Meanwhile, Google's competitors — you know, the ones actually shipping products — are watching with amusement. The delay comes as the search giant struggles to maintain its AI reputation, having lost several key researchers to rivals like Anthropic in recent months. (Whether this feedback-gathering is a cover for internal panic or genuine prudence is the sort of ambiguity that keeps PR firms in business.)

And what of the model itself? It's supposed to be a leap in capability, with improved tokens and agent-like features. But a one-month delay in AI development is like a one-minute delay in a Formula 1 pit stop — it can be routine or it can be catastrophic, and we're only told after the race. Perhaps by July, the model will have learned to avoid the embarrassing misunderstandings that plagued earlier versions. Or perhaps this is the first step toward yet another Google product quietly shuffled off the roadmap. But hey, at least they're listening to feedback. That's practically revolutionary for a company that once believed its users were just search queries.

Ah, Google's AI brain drain continues, and the exits are starting to read like a who's-who of Gemini development. Jonas Adler and Alexander Pritzel, two of the key architects behind Google's flagship model, are bolting for Anthropic. Because nothing says 'safety-first' quite like hiring the very people who built the most powerful commercial model on the market. It's a bit like hiring the arsonist as your fire chief, but I'm sure the alignment community will find a way to spin this as 'talent circulation strengthens the ecosystem.' Meanwhile, Google can't seem to hold onto the people who actually make its products work—almost as if the company's internal culture is optimized for churn rather than retention. But sure, keep telling us about the 'moonshots.'

This isn't an isolated incident; it's a pattern. Every few months, another batch of senior researchers quietly updates their LinkedIn to 'Head of Safety at Anthropic' or 'Senior Scientist at OpenAI.' Google has become a glorified training academy for its rivals. And the kicker? Anthropic markets itself as the responsible, measured alternative—the lab that will think before it leaps. Yet it's hoovering up the talent that made the leaps at Google. If these researchers were so key to Gemini's capabilities, what does that say about the 'safety' of the knowledge they're bringing? It's almost as if the line between 'capability research' and 'safety research' is getting blurred beyond recognition.

The real winner here might be the headhunters. With every departure, Google's stock of proprietary knowledge gets a little more diluted, and Anthropic's capabilities get a little more, well, Googly. And the rest of us get to watch this farce unfold from the cheap seats—trying to figure out which lab is actually building the safer future, and which is just better at marketing its defections. Either way, the AI safety community's latest 'consensus' seems to be: if you can't beat 'em, hire 'em.