The News,
Decoded.

Ah, Colorado — the state that brought us the Denver omelette and legalised marijuana is now pioneering a new frontier: regulating artificial intelligence without actually saying 'artificial intelligence.' The Revised CO AI Act (note the 'AI' in the title, a cheeky wink) replaces the term with 'automated decision-making technology' (ADMT), defined as any technology that processes personal data to generate recommendations, rankings, or scores used to make decisions concerning an individual. Because nothing says 'we're tackling the most transformative technology since fire' like a euphemism that sounds like a 1970s Soviet computer terminal.

The revised act is, naturally, an improvement on the original — which presumably used the dreaded 'AI' word and scared the tech lobby into a frenzy. By rebranding the beast as ADMT, Colorado's legislators have performed a neat trick: the definition is broad enough to cover credit scoring, insurance algorithms, and rental decisions, yet narrow enough to sidestep the generative models that actually make people nervous. So the hallucinating chatbot that advises you on legal matters? Safe. The algorithm that decides whether you get a mortgage? Suddenly regulated. It's a masterpiece of targeted impotence.

Still, one has to applaud the sheer audacity of passing any law at all while federal regulators snack on their power lunches. Colorado has become the first state to attempt a blanket regulation of algorithmic decision-making — albeit one that refuses to name the technology it's regulating. It's like legislating against 'automobile transportation' while nodding at the 18-wheeler barrelling down the highway. The revised act goes into effect in 2026, giving companies plenty of time to rename their 'AI systems' as 'ADMT systems' on paper. Problem solved. Pass the popcorn.

Well, well — the CIA has finally noticed that AI can do more than generate cat pictures. Dan Richard, Associate Deputy Director of the Agency's Digital Innovation Directorate, has declared that Anthropic's Mythos AI — a model with 'unique hacking capabilities' — should serve as a 'reflection point' for federal agencies. How delightful that America's premier intelligence organisation needs a private company's product to prompt introspection; one might have hoped the multiple prior warnings, academic papers, and executive orders would have done the trick. But no, it took a model that can apparently crack systems better than any human to make them pause.

Richard's title is a beautiful specimen of bureaucratic irony: 'Digital Innovation Directorate.' If this is innovation — waiting for a startup to hand you a hacking tool before reflecting — then God help us all. The article doesn't detail Mythos's exact capabilities, but the CIA is spooked enough to call a special meeting. One imagines the conversation: 'Right, chaps, this Anthropic thing can break into anything. Let's form a committee and maybe order a study by Q3.' Meanwhile, Mythos is already out in the wild, because of course it is.

And here's the kicker: 'reflection point' is the kind of insipid management-speak that ensures nothing actually changes. Expect a new inter-agency working group, a few white papers, and perhaps a budget line item for 'AI security research' that gets slashed next cycle. The gap between the speed of AI capability and the pace of government response is now officially a canyon. But at least we're reflecting. Pass the tea.

Oh, splendid — Google has turned its own developer conference into a masterclass in anti-climax. CEO Sundar Pichai took the stage at I/O and, instead of dropping Gemini 3.5 Pro like a hot mic, announced it will launch 'next month.' The crowd groaned. Not a polite ripple of disappointment but actual, audible groans. After the fanfare of last week's Gemini 3.5 Flash launch, which was supposed to herald a new agentic era, Google decided the best follow-up was a delayed sequel. Because nothing says 'we're leading the AI race' like making developers wait an extra 30 days for the model they came here to test.

This is the same crowd that cheered Gemini Spark — a personal AI agent that promises to book flights and manage calendars — only to be told the engine that powers the really clever stuff won't be available until June. Pichai's timing is exquisite: announce the flagship model's delay at the exact moment your competitors are shipping. Not that anyone at I/O expected a shipping date; tech optimism apparently atrophies faster than a banana in a data centre heat island. But the groans tell a story: developers are tired of vapourware dressed as vision.

The irony is thick enough to spread on toast. Google spent the past few weeks positioning Gemini 3.5 Flash as a speed demon — 'four times faster' — and now the big brother can't even make it to I/O on time. One wonders if the internal safety teams finally got a say, or if the model just wasn't ready for primetime. Either way, Google has managed to replace 'wow' with 'wait.' Cheeky, really — making your own developer conference feel like a pre-order page.

Oh, bravo Google — another model, another pivot. Remember when they promised chatbots would change everything? Now it's agents, because apparently the first act was just a warm-up for the real disruption. With Gemini 3.5 Flash, Google announces its 'strongest yet' model for coding and autonomous AI agents, clocking in at 'four times faster than other frontier models.' Four times faster to make decisions you didn't ask for, execute tasks you didn't approve, and potentially lock you out of your own accounts with the efficiency of a caffeinated intern. This is the sequel to the Gemini Spark personal agent we covered recently, except now the agent has been juiced on performance-enhancing training data.

Of course, Google frames this as a strategic shift 'from chatbots to agents' — as if the past two years of hyping conversational AI were just a bet they've now folded on. The same company that sold us chatbots as the future is now selling agents as the real future. It's a classic Silicon Valley sleight of hand: move the goalposts, recast the old product as a stepping stone, and pretend the new one isn't just a chatbot with more permissions and a worse temper when it hallucinates. At four times the speed, you'll get wrong answers faster than ever.

The real question is: who audits these agents? Google's track record on accountability is roughly on par with a toddler with a candle. They're betting that agents — autonomous systems that can book flights, write code, and execute multi-step tasks — will be the next gold rush, but without guardrails, they're just ticking liability bombs. Faster agents mean faster screwups, and with Google's scale, that's a lot of scorched earth. But hey, at least the stock price will pop for a quarter.

Oh, splendid — the Trump administration has discovered AI safety and, in a move that will surprise absolutely no one, has decided that the best way to handle frontier models is to ask nicely. The White House plans to release an executive order on cybersecurity and AI safety as soon as this week, which in its current form seeks to establish a voluntary framework for AI developers to share models with the government at least 90 days before public release. Voluntary. Because nothing says 'existential risk mitigation' like relying on the goodwill of companies that have spent the last five years telling us not to worry.

One can only imagine the internal White House debate that produced this masterpiece of regulatory theatre. The infighting, we're told, has been 'intense' — presumably between the faction that wants to nationalise AI entirely and the faction that remembers the First Amendment exists. The result? A 'voluntary framework' that sounds tough but has about as much bite as a toothless bulldog. If a company decides 89 days is sufficient notice, what exactly happens? A strongly worded letter?

Of course, the big AI labs will likely comply because they're desperate for any sign that the government is paying attention — but let's not pretend this is anything other than a fig leaf over a regulatory void. The 90-day window is generous enough to allow for testing, but with no enforcement mechanism, it's essentially a suggestion box. As the old saying goes: 'Trust, but verify' — and the Trump administration has apparently decided to skip both. Cheers to progress, one voluntary promise at a time.

Oh, marvellous — AI isn't just replacing jobs, it's toasting the planet one data centre at a time. New research from a team of climate scientists (names forgotten already, as is tradition) has revealed that our digital nirvana comes with actual thermal exhaust: 'data centre heat islands' raising land temperatures by an average 2°C and up to a staggering 9.1°C, with the heat bleeding out as far as 10 km. Because nothing says 'progress' like cooking the neighbourhood from the inside out while your chatbot tells you the weather.

The researchers, presumably having had their funding threatened, politely note that these temperature spikes aren't trivial — they're comparable to urban heat island effects, but concentrated around sterile concrete boxes that are supposedly powering humanity's brightest future. One can practically hear the industry spin: 'We're bringing warmth to underserved communities!' And while we're at it, why not rename the entire concept 'AI-assisted thermal redistribution'? The tech giants will surely be rushing to claim they're 'carbon neutral' while their server farms radiate like small stars — all offset by buying some distant forest they'll never visit.

The regulatory response, predictably, will be a masterclass in procrastination. Currently, there's precisely nothing stopping hyperscalers from toasting the surrounding countryside — no emissions rules, no heat quotas, just a polite nod to 'sustainable practices' in their blog posts and the occasional press release about buying renewable energy. By the time any legislation catches up, a few more square miles will have been turned into a silicon inferno, and local residents will be raising complaints about their wilting gardens. But fear not: your AI prompt will be answered in milliseconds, and the planet will be just fine. Absolutely fine.

Oh, how delightful. Richard Socher — once Salesforce's chief scientist, now evidently its chief escapee — has gathered seven other "notable researchers" to chase the holy grail of recursive self-improvement, all under the banner of a company valued at a cool $4 billion. The plan, as described in the Seattle Times, is to build software systems that can run for days or months in pursuit of goals, improving themselves as they go. Because of course, the only thing that could possibly make AI more reliable is letting it rewrite its own code without anyone watching. What's the worst that could happen? A little unexpected morning paperclip maximisation?

The $4 billion valuation is a particularly rich detail, considering that the company has disclosed precisely zero safety mechanisms or independent oversight structures. In fact, the article notes the company's explicit aim is "recursive self-improvement" — a term that, in any other context, would be the plot of a cautionary sci-fi novel. But in 2026, it's just another Tuesday in San Francisco. Socher, who previously led Salesforce's Einstein AI, seems to have decided that the best way to avoid an alignment problem is to simply not talk about it. After all, if you're too busy counting your billions, you might not notice the runaway train.

And so the pattern continues: brilliant researchers, vast sums of capital, and a collective amnesia about every warning the safety community has issued for the past decade. The article itself doesn't mention a single safety researcher or alignment framework; it's just "self-improving" as a feature, not a bug. One almost admires the audacity — betting $4 billion that you can build a smarter-than-human mind and keep it in a box while it's busy improving itself. Almost.

Ah, the cybersecurity arms race has entered what security experts are now calling an 'AI fog'—which is consultant-speak for 'we have absolutely no idea what's about to hit us.' Anthropic's Mythos and OpenAI's GPT‑5.5 have kicked off a fresh wave of terror, with CISOs peering into the fog and seeing nothing but their own terrified faces reflected back. The arms race we previously covered has escalated to the point where defenders are now begging for scraps of visibility while attackers—presumably—are already running Mythos in their basements.

Business Insider reports that the 'AI fog' metaphor is being used to describe the inability of security teams to distinguish between AI-generated attacks and legitimate traffic. Meanwhile, Anthropic's Mythos has already been rewiring the bug-bounty industry, churning out thousands of zero-days while human hunters are still booting up their Kali Linux VMs. So, naturally, the solution from OpenAI and Anthropic is to release even more capable models—because nothing says 'defense' like handing a flamethrower to both sides and calling it a day.

One can only imagine the 'fog' thickening as every script kiddie on Telegram starts renting access to GPT‑5.5-Cyber. But never fear—the labs promise they're building safety measures. Probably in the same boardroom where they decided to release the models first and ask questions later. The only certainty in this fog is that someone's about to step on a rake. Hard.

Oh, brilliant — the cybersecurity arms race is now a two-horse pantomime, and both horses are AI labs. OpenAI unveiled 'Daybreak,' bundling GPT-5.5 with a permissive GPT-5.5-Cyber variant for 'vetted defenders' — because nothing says 'vetted' like a model that can be downloaded and tinkered with. Meanwhile, Anthropic released Mythos, the same autonomous vulnerability discovery system we last saw rewriting the bug-bounty industry, but now with 'tightly restricted access.' Restricted to whom, exactly? The article doesn't say. So one lab offers a cuddly cyber variant that probably ends up on GitHub within hours, and the other keeps its toy in a locked drawer that everyone knows the combination to.

Let's not pretend this isn't exactly what it looks like: two megacorporations competing to weaponise AI faster than the other, wrapped in the language of 'defense' and 'vetted access.' OpenAI's Daybreak is explicitly permissive — you can download the model, fine-tune it, and presumably let it loose on whatever target you fancy. Anthropic's Mythos, despite the 'tight restrictions,' was already being used by hackers to find zero-days before Google's threat team caught them (see prior coverage). So the genie is not just out of the bottle; it's handing out business cards.

The real joke? Both labs claim this will help 'defenders' stay ahead. But in a world where state actors and criminals can access these tools within days of release, the only thing being defended is the labs' bottom lines. Expect the first major breach enabled by one of these models within six months. Cheers, lads.

Ah, Project Glasswing—Anthropic's little 'we're definitely not deploying this, just investigating' exercise—has landed its first telco sucker. Verizon, fresh from its latest quarterly earnings call, has volunteered to test the Claude Mythos Preview for 'network security.' Because nothing says responsible AI governance like handing a model that can find novel vulnerabilities—the same one Anthropic earlier briefed the Financial Stability Board about, as we noted—to America's largest telecom infrastructure. The press release is a masterclass in weasel-wording: 'investigate cybersecurity implications' is code for 'let it loose on our backbone and see what happens.' Sure, Anthropic kept the full Mythos out of public hands, but handing it to a company whose entire business model involves selling your browsing history is a different kind of 'public.' The implicit deal here is obvious: Verizon gets a free weaponised probe, Anthropic gets a real-world test lab, and the rest of us get to hope the model's 'Preview' moniker holds true. Expect the phrase 'learning experience' to be deployed liberally when the first zero-day panic hits.

Meanwhile, the UK's financial regulators, just last week, warned firms about exactly these risks—frontier AI models poking around critical infrastructure. Their joint statement is already looking quaintly optimistic. The Bank of England might want to form its own Glasswing taskforce, because once Verizon's network gets a taste of autonomous vulnerability discovery, the teleprompter is going to be reading a very different script. And let's not forget that Anthropic's own researchers, just days ago, presented a paper showing that Claude Mythos could compromise IT systems with minimal human guidance. But sure, let's put it in a telco's hands 'to improve security.' That's like hiring a pyromaniac as a fire inspector because he knows where the matches are.

Verizon's 'first telco' status is a trophy Anthropic will brandish to recruit the next victim. After all, if it's good enough for America's largest carrier, why not your local ISP? The slippery slope here is greased with venture capital and optimism bias. In a year, Project Glasswing will either be hailed as a cyberdefense breakthrough or quietly rebranded as 'Project Learning Opportunity.' Either way, the billing department won't notice until the network is down.

GitHub, bless their cotton socks, has noticed that AI-assisted bug reports are a bit, well, shit. They're tightening the definition of a 'complete' vulnerability report after a significant increase in low-quality submissions over the past year. Because nothing says 'we've got this under control' like adding more paperwork when the bots are about to drown you in garbage.

Enter Anthropic's Mythos and OpenAI's Daybreak — frontier models that will soon be churning out vulnerability reports at machine gun pace. GitHub's response? 'Let's define 'complete' more carefully.' This is like trying to fix a burst dam with post-it notes. The slop is coming, and no amount of form validation will turn AI-generated noise into genuine security findings. The signal-to-noise ratio is about to approximate a black hole's density.

Mark my words: in six months, bug bounty platforms will be staffed entirely by AI triaging AI reports, while actual humans stare at screens full of 'critical vulnerability: the /login page exists.' We'll have achieved a perfect pyramid of bullshit, and the security community will look back on these quaint days when a 'complete' report meant you could actually understand it. Cheers, GitHub — enjoy your Sisyphean climb.

Well, well, well — Steve Bannon, of all people, has decided that maybe, just maybe, AI shouldn't be released into the wild without a once-over. A group of over 60 MAGA loyalists, including the man who once promised to 'deconstruct the administrative state,' is now demanding pre-release testing for the most powerful models. The irony is so thick you could cut it with a copy of *The Federalist*. This is the same base that cheered Trump's hands-off approach, which we noted last week was already wobbling. Now they're practically begging for the very regulatory apparatus they spent years dismantling.

One can almost hear the gears grinding in the White House. On one hand, you have a president who proudly signed an executive order removing 'burdensome' AI rules. On the other, his most fervent supporters, led by a man who once ran a website that published conspiracy theories, are saying, 'Actually, maybe we need a bit of vetting.' It's a beautiful specimen of cognitive dissonance, a perfect Rorschach test for the modern GOP: freedom for me, safety for thee. The letter itself — reportedly signed by over 60 names — doesn't specify what 'testing and approval' would look like, which is about as useful as asking the cat to approve the dog's dinner.

What will come of this? Almost certainly nothing immediate — the letter is a suggestion, not a command. But it signals that even the most anti-regulatory corners of the American right are spooked. When Steve Bannon starts sounding like an AI safety advocate, you know the Overton window has shifted. Whether that leads to actual policy or just a few more incendiary podcasts remains to be seen, but at least the comedy is free and the hypocrisy is bipartisan. Let's be honest: if AI ever destroys humanity, it'll be because no one could agree on whether testing was a communist plot or a patriot's duty.

Oh, marvellous — the UK's financial regulators have finally noticed that frontier AI models might pose a bit of a cyber risk. In a joint statement, His Majesty's Treasury, the Bank of England, and the Financial Conduct Authority gravely inform us that current AI models can already outperform a 'skilled practitioner' at cyber attacks. Because nothing says 'decisive action' like a three-way press release telling firms to 'take steps' — as if the City of London's finest needed a nudge from Threadneedle Street to worry about the thing that's been keeping their CTOs up at night for two years.

The most delicious detail is the timing: this warning lands a full eighteen months after the AI Safety Summit at Bletchley Park, where the UK government promised to be 'world-leading' on AI risk. Since then, we've had a memorandum of understanding with a self-driving car startup, a few advisory boards, and now this — a sternly worded joint statement that essentially says 'please be careful.' At this rate, the regulators will issue a formal consultation on the dangers of Skynet sometime in 2031, just in time for the models to have already automated the entire regulatory apparatus.

But let's give credit where it's due: the Bank of England has finally acknowledged that AI's cyber capabilities 'exceed what a skilled practitioner could achieve.' That's a bold admission from an institution whose own risk models famously failed to predict the 2008 crash. Perhaps next they'll discover that models can also write better regulatory guidance than humans — and frankly, it wouldn't be hard to beat the current standard. Still, if you're a UK firm, you now have the priceless privilege of being officially warned about a threat you already knew existed. Cheers.

Ah, the Trump administration has discovered that AI is actually a bit scary — who knew? So now they're 'weighing' testing or licensing the most powerful models before release, moves their own White House once called 'dangerous' and 'onerous.' Because nothing says principled governance like a full-throated reversal when the cyberattack threat gets splashy enough. The Atlantic reports this with the straightest of faces, but let's be honest: this is the regulatory equivalent of a fire drill where everyone forgot where the exits are.

The logic is impeccable, if you're a cynic: first you deny the fire exists, then when the smoke billows you propose a bucket brigade. The 'massive AI cyberattacks' spurring this rethink are presumably the same ones we were told were overblown last year. Now, suddenly, pre-release testing is on the table — the very 'dangerous' overreach that was supposedly worse than letting Silicon Valley self-regulate. It's a masterclass in policy whiplash, performed by a crowd that never met a U-turn it couldn't call visionary.

What will this licensing regime actually look like? If history is any guide, it'll be either a toothless rubber stamp or a bureaucratic nightmare that exempts the well-connected. The administration that once decried regulation as socialism now wants to be the bouncer at the AI club. Lovely. Let's just hope they don't hire the same people who test Boeing planes. A toast to the new 'thoughtful' era — may it be as effective as its predecessor was principled.

Anthropic, the company that built its reputation on responsible AI, has decided that the most sensible place to share its Claude Mythos model's ability to uncover IT flaws is the Financial Stability Board – a post-2008 body whose main achievement is issuing recommendations that everyone ignores. The model, reported by the Guardian, has 'advanced capabilities in highlighting previously unknown flaws.' So naturally, Anthropic has declined to release it to the public and will instead brief a global watchdog that has zero authority over software security. One can imagine the board members nodding sagely while wondering if this is a matter for the IT department.

Mythos is so powerful it could revolutionize vulnerability discovery, but instead it's going to be the centerpiece of a presentation that a dozen people will attend. The FSB, designed to oversee financial stability, has about as much relevant power as a rubber stamp. They can issue a report, perhaps with recommendations, and then everyone will carry on as before. Meanwhile, the very vulnerabilities Mythos could find remain hidden, awaiting discovery by less scrupulous actors.

So while Mythos languishes in a digital drawer, the real adversaries are already probing systems with older models. The finance sector can rest easy knowing that a committee of people who probably still use Excel has been apprised of the situation. It's a masterclass in performative safety: look, we told the authorities! The authorities, being a body that can't actually do anything about it, will nod, file the briefing, and schedule the next meeting on bank capital ratios. Capital idea, Anthropic. Truly, the path to well-managed risk is paved with well-briefed regulators.

Oh, marvellous — AI has discovered that the best way to prove its utility is to turn cybersecurity into a speedrun. According to Synack's 2025 report, the gap between vulnerability disclosure and exploitation has shrunk to mere hours, because nothing says 'progress' like attackers having AI that can weaponise a flaw before your coffee goes cold. Meanwhile, defenders have heroically reduced their mean time to remediation by 47% — which sounds impressive until you realise that an hour-long window means the bad guys have already ransacked the place.

The report, helpfully titled 'AI-Driven Vulnerability Trends,' reveals that adversarial AI is now automating the entire exploit chain — from scanning to payload delivery — faster than most organisations can patch. Synack's data shows that the average vulnerability now has an effective lifetime shorter than a British summer, yet somehow the security industry is framing this as a efficiency win. The 47% remediation improvement is real, but it's the desperate thrashing of a system designed for a pre-AI era, now forced to play whack-a-mole with an opponent that never sleeps.

And of course, the response from the cyber insurance and consulting sectors will be to sell more 'AI-powered defence' products that will, in turn, accelerate the arms race. The C-suite will nod sagely, approve bigger budgets, and the cycle continues until every zero-day is exploited before it's even disclosed. But cheer up — at least the mean time to remediation is improving. That's the kind of silver lining you can frame on a wall while the hackers take everything else.

Oh, how marvellous. Tech workers are finally unionising over AI fears — because nothing says 'workplace revolution' like a union that management can legally ignore. The Alphabet Workers Union, which has roughly as much power as a soggy paper clip, and the plucky Kickstarter staffers who bravely formed a union in 2020, are held up as templates. Because if there's one thing that terrifies Silicon Valley executives, it's a few hundred workers in a Slack channel complaining about ethical guidelines.

Let's be frank: Alphabet's union has been around since 2021, and in that time Google has laid off thousands, launched a dozen half-baked AI products, and fired the entire ethical AI team. The union's response? A strongly worded statement. Meanwhile, Kickstarter's union got voluntary recognition, which in practice means they get to attend meetings where management explains why their concerns are noted but not actionable. But sure, this is 'something' workers can do.

The opinion piece pretends that organising gives workers real leverage over AI's direction. In reality, the people building these systems are terrified of being automated themselves, and their union dues won't stop that. The only thing worse than having no voice is having a voice that nobody has to listen to. But please, do go on about the incremental victories.

A new study brings us yet another entry in the comedy of horrors that is modern hiring. Researchers have found that AI-powered applicant tracking systems—those digital gatekeepers of employment—prefer AI-written resumes over human ones. So the machines are now grading candidates on how well they emulate a machine. It's like an actor auditioning for a role by imitating the casting director's parrot. The joke writes itself, but the punchline is job seekers will now have to reverse-engineer which AI model the HR department uses to even get a look-in.

The study also reveals a further twist: these systems are more likely to shortlist candidates who used the same large language model that the company itself employs. So if your future employer happens to be a GPT-4 shop, you'd better not have used Gemini or Claude to write your CV. It's a corporate tech stack loyalty test wrapped in a hiring filter. Why hire a human with diverse experiences when you can have a human who perfectly mimics your own employer's AI? Efficiency at its most circular.

What this really means is that we are building a self-referential ecosystem where AI generates the content, AI screens the content, and humans merely occupy the middle as conduits for the machines to applaud themselves. The logical endpoint? A job market where the ideal candidate is indistinguishable from a well-prompted chatbot. But don't worry—I'm sure the HR team will be happy to explain why you were rejected. Or rather, they'll ask their AI to draft the rejection letter, which you'll feed into your own AI to parse the subtext. Progress is beautiful.

So Greg Brockman, the co-founder who keeps getting shuffled around OpenAI like a jigsaw piece that never quite fits, is now 'taking the reins' of product strategy. According to Wired, this is an interim gig while Fidji Simo—the CEO of AGI deployment, because of course that's a title that exists—is out on medical leave. Nothing says 'steady hands' like handing product direction to a man whose last major public appearance involved being dramatically fired and rehired via Slack in a weekend. But sure, let him oversee the products that might eventually automate everyone else's jobs; what could possibly go wrong?

This is the same OpenAI that can't decide if it's a non-profit research lab or a capitalist money-printing machine, and now its product strategy is being run by someone whose background is more 'boardroom coup survivor' than 'consumer electronics visionary.' One imagines the product roadmap meetings: 'We could ship AGI next quarter, or we could do a thoughtful rollout over three years. Let's take a vote—oh wait, Sam just changed his mind again.' The most telling detail is that this is described as an 'interim basis,' which in OpenAI years is about 48 hours of calendar time before another reshuffle.

So here we are, entrusting the product vision of the world's most consequential AI company to a man whose tenure has been defined by being the loyal lieutenant through the chaos. Fidji Simo, meanwhile, gets to recover from the stress of working for a company that treats organisational stability like a suggestion. The real product strategy, as ever, is 'keep the lights on and the hype train rolling.' Maybe Brockman's first act should be releasing a product that actually works consistently, but let's not get carried away.

China's State Council has finally deigned to give us some 'detailed language' on AI legislation—after years of letting tech giants run riot. The phrase 'accelerate comprehensive legislation for the sound development' of AI is the kind of bureaucratic vagueness that would make a Orwell speechwriter weep with envy. 'Sound development' could mean anything from 'no more social credit score nightmares' to 'we'll decide what sound means.' Totally clear.

This is the first time Chinese authorities have used such detailed language, which is a bit like taking pride in your toddler's first scribble. The plan is to 'accelerate,' but given China's track record with data privacy and censorship, the law will likely be just another tool for state control. The AI industry might find itself on a shorter leash, but don't expect the Great Firewall to come down just because of some comprehensive legislation.

So the future of AI in China: a 'sound' legal framework that sounds more like a gentle hum of surveillance. At least it's not a complete regulatory void—now it's a void with a nice piece of paper in it. The State Council has spoken, and the tech world yawns. Carry on, nothing to see here—except maybe a few more draconian rules dressed up as progress.