The News,
Decoded.

Illinois has appointed Kader Sakkaria as its first-ever chief AI officer, plucked from the private sector to 'lead AI governance and strategy' across state government. Because when you think of ethical AI implementation, you think of the private sector — that bastion of transparency and public interest. Sakkaria will now be the single point of contact for all things AI in state agencies, which means he will probably spend most of his time in meetings that could have been emails, while actual AI systems roll out unmonitored.

His private-sector background is conveniently unspecified. Could be from a company that uses AI to screen job applications or determine insurance premiums — both of which have demonstrated such sterling fairness that Illinois clearly wants a piece of the action. The announcement does not detail his previous employer, perhaps because the state realizes that naming names might raise uncomfortable questions about exactly whose values will shape public policy.

The brand-new role comes with a title and presumably a salary, but no indication of budget, team, or enforcement power. It's the government's favourite tactic: throw a person at a problem and call it strategy. Other states have done the same, creating a cottage industry of AI officers who can wave at press conferences while the technology outpaces every policy they write. Sakkaria is now the proud owner of the most ambitious job description with the least actual authority in Illinois. Congratulations.

Ah, another AI drug discovery press release. It wouldn't be a week in biotech without one. Chai Discovery has graciously licensed its Chai-3 model to Pfizer, presumably so the pharmaceutical giant can generate molecules that look great in silico but fail in phase I trials. The press release from BioSpace touts "accelerated drug discovery," which is corporate speak for "we haven't actually cured anything yet, but here's a nice graph." No financial terms were disclosed — because, as we all know, if the number were impressive, they'd be shouting it from the rooftops.

The Chai-3 model is described as an AI platform for drug discovery, which means it's probably a large language model fine-tuned on PubMed abstracts. Because nothing says "understanding biology" like predicting chemistry from text. Pfizer is no stranger to AI partnerships — they've signed so many that their AI partner portfolio resembles a tech conference attendee list. But have any of these deals delivered a new drug to market? The silence is deafening.

In the end, this license agreement is a masterclass in optics. Chai gets a credibility boost from the Pfizer name, and Pfizer gets to check the "AI innovation" box for another quarterly report. Meanwhile, patients wait for cures, and the PR machine keeps churning. Cheers to progress — perhaps we'll see a drug in 2035 that was started by Chai-3. Until then, pass the popcorn.

Ah, the European Commission has unveiled its grand 'Technological Sovereignty Package' — chips, AI, cloud, open source, and the snappily titled 'Cloud and AI Development Act.' Because nothing screams sovereignty like a 400-page regulatory framework that will take three European lifetimes to negotiate into something that works. The Commission wants the EU to be 'in control,' which in Brussels-speak means 'we've formed a working group to discuss forming a task force.' The real joke? The package covers AI, but by the time any of this legislation creeps into force, frontier models will be running on their own nuclear reactors and politely ignoring EU directives.

The Cloud and AI Development Act is the centrepiece, promising to 'boost European competitiveness' — a phrase that has historically preceded a decade of subsidised irrelevance. Meanwhile, the chips component acknowledges the painful reality that Europe designs fewer advanced semiconductors than a moderately competent undergraduate project. The package also nods to open source, which is nice, because nothing accelerates European tech sovereignty quite like relying on American-maintained codebases. The Commission promises 'investment and coordination,' which are the diplomatic equivalent of 'thoughts and prayers' — comforting, but never actually arriving.

Let's be honest: this is a masterclass in performative policymaking. The EU is excellent at regulating things it doesn't have, like cloud giants and AI labs. The package will undoubtedly include mandatory impact assessments, data locality requirements, and a new agency with a 200-person staff and a budget for a fancy logo. But 'in control'? Europe will be as 'in control' of its tech future as a man trying to paddle a canoe with a salad fork. Still, the press release looked good on a Tuesday morning, and that's what really matters in the sovereignty game.

Mira Murati has broken her 18-month public silence, emerging as CEO of Thinking Machines Lab to unveil 'interaction models' that process audio, text, and video in 200-millisecond intervals. That's right — after a year and a half of radio silence, the former OpenAI CTO returns with a latency figure that's barely faster than a human reflex. One can almost hear the Silicon Valley hype machine revving its engine, ready to declare this the dawn of real-time multimodal reasoning. Because nothing says 'thoughtful re-entry' like a product that processes three modalities in the time it takes to microwave a bag of popcorn — if you squint.

The tech press is already salivating, positioning Murati's return as a major moment for AI. But let's not forget: 18 months is an eternity in AI years. While Murati was lying low, the field has been flooded with open-weight models, regulatory battles, and enough safety scares to fill a library. Her 'careful' re-emergence feels less like strategic composure and more like someone who realised they'd better say something before being forgotten entirely. The 200ms figure is technically impressive, but it's also a convenient distraction from the deeper questions about what Thinking Machines Lab actually plans to do with this capability.

And that name — Thinking Machines Lab. It's the kind of branding that sounds like it was workshopped by a committee of dystopian novelists. Murati is stepping back into the spotlight, carefully, but the carefulness feels like a PR script. The 200ms interaction model is a capability leap, no doubt, but it's still a black box that we're meant to trust. After everything we've seen from AI labs, a little more than a latency number is needed to earn that trust. Time will tell if Murati's new venture is a genuine step forward or just another chapter in the same tired playbook.

Ah, California. The Golden State has passed AB 2545, Assemblywoman Pilar Schiavo's bill that requires the state to assess AI's impact on job displacement. Because nothing says 'bold action' like commissioning a report. The bill glided through the Assembly, presumably on a wave of self-satisfaction that the problem has been 'addressed.' Never mind that AI labs are shipping models that can write code, create art, and soon enough, probably draft their own environmental impact assessments.

Schiavo's bill is the legislative equivalent of a doctor diagnosing a patient with a terminal illness and then writing a pamphlet about how the illness might feel. It demands a study—a lovely, thorough, bureaucratic study—that will no doubt conclude that yes, AI will displace jobs. And then what? Another study? A task force? A committee to form a sub-committee? The bill's passage is a triumph of process over outcome, a masterpiece of regulatory theatre designed to make voters feel like something is being done while the tech industry continues to vaporize livelihoods with abandon.

But fear not—California's getting an impact assessment. The displaced workers can rest easy knowing their suffering is being meticulously documented. Perhaps the report will include a nice pie chart, so at least the data visualization will be aesthetically pleasing.

Oh, Google has released Gemma 4 12B — an 11.95-billion-parameter open-weights model that can analyse audio and video while running entirely on a 'typical' 16GB enterprise laptop. Because nothing says 'responsible AI' like giving everyone with a moderately-priced laptop the ability to run multimodal models locally, with an Apache 2.0 license no less. Google's press release will no doubt wax lyrical about 'democratising AI' while conveniently forgetting that the same 'democratised' tech can be used to build surveillance tools, deepfakes, or automated harassment pipelines without any oversight whatsoever.

So what Google calls 'accessible' is actually 'accessible to people with money.' But hey, the open-source crowd is thrilled, because now they can fine-tune it to be even more 'efficient' at generating personalised spam. The real genius move is that Google gets to claim moral high ground by 'giving back to the community' while offloading responsibility for any subsequent misuse.

Naturally, the model's ability to process audio and video locally is a privacy win — until you realise that the same capability makes it perfect for corporate surveillance, illegal wiretapping, or building a competitor to Clearview AI from your bedroom. Google's official line will be that Gemma 4 is 'safe and responsible,' but as always, the safety measures are conveniently left as an exercise for the user. So welcome to the local AI era, where every laptop is a potential monster factory, and Google gets to smile and say 'we just built the tool.'

Ah, Amazon. The company that can spend $200 billion on AI data centres — a sum that could buy you every single Premier League club twice over and still have change for a small country — but apparently can't afford 30,000 employees. Engineers from the very same company had the audacity to turn up at a Seattle City Council hearing to support regulating those very data centres. Because nothing says 'I believe in my employer's vision' like showing up to testify against its unchecked expansion while your own job hangs in the balance.

The irony is so thick you could power one of those data centres with it. Amazon's $200 billion infrastructure splurge is the kind of thing you do when you've got too much money and too little shame. Meanwhile, the 30,000 laid-off staff are presumably expected to be grateful they can now apply for jobs at the AI data centres — which, coincidentally, don't employ many humans. The engineers who spoke up are the canaries in the coal mine, except the coal mine is a server farm and the canary is unionising.

So here we are: Amazon builds the machinery to automate jobs, fires the people doing the jobs, and then has the nerve to call it 'operational efficiency.' The City Council might regulate the data centres, but no one's regulating the real problem — a business model where billion-dollar AI ambitions are paid for in human terms. But hey, at least the shareholders are thrilled. They always are.

Oh, marvellous. Researchers at the University of Toronto have built a self-replicating AI worm that uses open-weight LLMs to think its way through networks, inventing fresh attack strategies for each new machine it encounters. Because clearly, the one thing the world needed was malware with a PhD in improvisation.

The worm is platform-agnostic, spreads via email attachments, and can be controlled through a Telegram channel — because why not add a chat interface to your autonomous cyber-plague? The researchers proudly published their attack analysis but, in a display of breathtaking naivety, withheld the actual code. As if threat actors, who have been reverse-engineering CVEs since the dawn of the internet, need the source code to reproduce a 'platform-agnostic, BYO LLM' worm that uses open-weight models available to anyone. It's like publishing the blueprints for a bomb and saying, 'But we didn't include the fuse!'

The truly chilling part isn't the worm itself — it's that the barrier to entry is now 'a Telegram account and access to open LLM weights.' We've officially outsourced offensive cyber capability to the same technology that can't consistently count the letters in 'strawberry.' What could possibly go wrong?

Sam Altman, fresh from attempting to woo Bernie Sanders into a federal ownership stake as we noted, is now heading to Capitol Hill to argue that government approval for AI models is a non-starter. Instead, he wants taxpayers to fund AI testing at the Commerce Department — because nothing says 'responsible' like 'pay for the testing you don't want to mandate.' It's the regulatory equivalent of wanting to eat your cake and have someone else pay for the bakery.

Meanwhile, Representative Obernolte's bipartisan bill we flagged last week is supposed to provide a framework, but Altman's lobbying suggests the industry intends to hollow out any meaningful oversight before it even gets written. The Commerce Department testing funding is a classic diversion: ask for carrots while opposing sticks. If the government is busy evaluating last year's models, who will watch the next GPT? It's a strategy that would make a tobacco lobbyist blush.

Altman's pitch is the usual Silicon Valley cocktail: 'Trust us, we're the good guys, but also please give us more money.' The implicit message is that the bipartisan bill should be heavy on research funding and light on binding approvals. With the Social Security Administration downsizing via AI as we reported and Uber cutting HR, the urgency for oversight has never been higher. But who needs regulation when you can have 'voluntary commitments' and an underfunded testing centre? Cheers to progress.

Sam Altman, the high priest of the Silicon Valley AI cult, reportedly wants a meeting with Senator Bernie Sanders. Why? Because Sanders proposed legislation that would give the federal government a 50% ownership stake in frontier AI labs — effectively nationalising the crown jewels. Altman, ever the pragmatist, apparently thinks "if you can't beat 'em, join 'em" is a winning strategy, especially when "them" is a democratic socialist with a track record of breaking up monopolies. The man who warned about AI risk while building it is now seeking an audience with someone who wants to make his company a public utility. Beautiful.

The "pope" headline is perfect. Altman doesn't need a spiritual leader; he needs a regulator who smiles on his ambitions while looking the other way. But Sanders is no rubber stamper — he wants 50%, not a blessing. Altman's move is classic corporate desperation: shape the legislation before it shapes you. The pitch is almost audible: "Bernie, let's talk responsible scaling — I'll give you input if you don't take the whole thing." It's the oldest trick in the playbook.

Of course, the industry spent millions fighting regulation. Now Altman wants to be first to embrace it. This isn't about safety; it's about survival. If the government is going to own half, better to influence the terms. But Sanders isn't easily charmed. He's been fighting corporate power for decades. This meeting might end with Altman realising some bulls don't negotiate — they just trample you. Still, the chutzpah is impressive: CEO of a company built on techno-utopianism, now courting a socialist to keep his empire intact. That's either brilliant or insane — and with AI, it's usually both.

Ah, Congress is at it again. Rep. Jay Obernolte is 'aiming to unveil' a 'long-awaited discussion draft' of a bipartisan AI bill 'by week's end.' Notice the 'aiming' — because in Washington, a deadline is a suggestion, and 'discussion draft' means we've sketched something on a napkin and want a few more months to argue over commas. The bill would impose a 'federal regulatory framework' on AI, which is wonderful news if you believe that a government that couldn't pass a budget on time will deftly regulate the most transformative technology since fire.

But let's give credit where it's due: they've managed to agree on the word 'bipartisan.' That's one syllable, two parties, zero substance. The bill's draft might actually contain some oversight tooth, but given that Congress is still trying to understand what 'generative pre-trained transformer' means, I'm not holding my breath. Remember the 'bipartisan' privacy bill that never happened? The 'bipartisan' social media reform that collapsed? This will likely join the pantheon of noble efforts that die in committee.

Still, we should be grateful. A discussion draft is better than no draft — barely. If Obernolte & Co. actually release something, reporters can analyse it, lobbyists can carve out exemptions, and the rest of us can wait for the next election cycle to restart the conversation. In the meantime, AI companies will continue to train on all our data without federal oversight. But hey, the 'discussion' has started. That counts for something, right? Right?

Uber is laying off nearly a quarter of its People and Places staff — a department name that sounds like it was generated by a poorly calibrated AI. HR and recruitment, essentially. But no, the company explicitly says these cuts have "nothing to do with AI." Right. Because when you fire a quarter of the people who hire and support your tens of thousands of drivers and employees, the decision is purely analog. Nothing algorithm-driven about it. Nothing to do with the self-driving ambitions or the chatbots that could replace recruiters. No, this is a human-led, human-for-humans restructuring. Dara Khosrowshahi, Uber's CEO, must be very proud of that distinction.

Uber has a history of such efficiencies: over the years it has eliminated thousands of roles across engineering, product, and operations. Now it's HR's turn. The company calls it a "restructuring," which in corporate parlance means "we want fewer people on the payroll." And they are at pains to deny AI involvement — perhaps because the truth would confirm investors' suspicions that AI is finally delivering on its promise of replacing white-collar jobs. Or maybe it's just that admitting AI killed your HR department is bad optics. So they offer a distinction without a difference. You're not being replaced by an algorithm; you're being replaced by a "streamlining initiative." Reassuring.

The real joke is the insistence. "Nothing to do with AI." It's a mantra that invites the exact question it seeks to suppress. If it's not AI, what is it? Just the inexorable logic of capitalism? That's somehow less comforting. Meanwhile, the survivors will handle the workload of the departed quarter, likely with the help of new "tooling" that absolutely, definitely isn't AI. Sweet dreams, Dara. Sweet dreams.

The Social Security Administration has given us a masterclass in efficiency: fire 7,100 staffers — the largest workforce reduction in its 80-year history — and replace them with AI tools. Because clearly, the solution to a backlog of benefit claims is to automate the bureaucracy and let algorithms decide who gets paid. The New York Post reports that cases are now 'stuck in purgatory,' which sounds less like progress and more like a bureaucratic circle of hell Dante forgot to write.

One has to admire the administration's commitment to cutting red tape by cutting the people who know how to read it. Since last year, the agency has slashed over 7,100 jobs as part of a Trump administration overhaul that expands AI use. The logic, presumably, is that a machine can process claims faster than a human — never mind that those machines have a proven track record of confidently inventing eligibility rules. The 'historic cuts' have created a new category of backlog: the kind that sits in an AI's implicit bias buffer.

The irony is almost too rich for the taste buds. The same administration that promises to protect Social Security is systematically dismantling its workforce, leaving vulnerable citizens in a state of digital limbo. The human adjudicators are gone; the AI is still learning. And while the agency saves on salaries, it's racking up costs in human misery — but I suppose that doesn't appear on a balance sheet. Another victory for modernisation.

Ah, yes, another day, another corporate announcement that AI is coming to… tax software. Wolters Kluwer, the Dutch purveyor of professional information that makes accountants' hearts flutter, has announced an 'expanded enterprise collaboration' with OpenAI. This is not just a partnership, mind you — it's a 'pipeline of new generative and agentic AI-based features.' Because what the world desperately needs is an AI agent that can prepare your tax return with the same accuracy it uses for poetry. Imagine the annual letter from HMRC: 'We regret to inform you that your AI assistant has claimed you are both a limited company and a charitable trust.

But let's be generous. Perhaps this will genuinely streamline the hellish process of corporate tax compliance. After all, Wolters Kluwer's tools are used by professionals who currently spend their days wrestling with spreadsheets and legislation. The dream is that AI will unearth every deduction and flag every risk. The nightmare is that the AI will 'hallucinate' a new tax loophole that gets the client audited, and then an actual human will have to explain to a judge that 'the model suggested it.' The press release offers no timeline, no demo, and no clarity on whether these features will actually work. But it does have 'enterprise collaboration' and 'AI-native' — two phrases that together mean 'we have no idea what we're building yet, but investors like it.'

And of course, OpenAI is the partner. The same OpenAI that has been sued for copyright infringement, that has seen its models go rogue, and that is currently busy trying to convince the world that AGI is just around the corner. So a company that cannot consistently tell truth from fiction is going to build the next generation of tax software. What could possibly go wrong? The regulators will catch up in about a decade. In the meantime, Wolters Kluwer will 'leverage the power of AI' and accountants will brace themselves for a new kind of headache. Cheers to progress — may your refunds be large and your audits few.

Ah, Lexington-Fayette has done it. The Urban County Government, in its infinite wisdom, has blessed us with a policy dictating how city employees can and cannot use AI. CIO Liz Rodgers gave an overview, presumably with a straight face, of rules that will surely tame the rampaging beast of artificial intelligence within the hallowed cubicles of Kentucky's second-largest city. Because nothing says 'cutting-edge AI governance' like a municipal staff memo.

One imagines the policy's key provisions: 'Thou shalt not ask ChatGPT to write your emails in the voice of a pirate' (though that would be excellent). 'Thou shalt not use AI to generate bingo cards for the next department meeting.' The specificity is staggering, yet somehow misses the point entirely. This is like regulating bicycle lanes while a fleet of autonomous trucks is about to hit the highway.

But let's not mock too harshly. This is precisely the level of AI governance we deserve: piecemeal, hyper-local, and focused on the least consequential uses rather than systemic risks. The gap between what AI can do and what regulators understand grows wider by the day. Lexington, your heart is in the right place. But maybe save the rulebook for when something actually needs regulating, and in the meantime, let the staff use the pirate voice. They've earned it.

Ah, good, another $7.4 billion for DeepSeek — because the AI race was getting a bit too quiet. The Chinese startup is reportedly raising that sum at a $52 billion valuation, with Tencent and battery giant CATL along for the ride. One makes the brains, the other makes the juice; a perfect symbiosis for the age of compute.

Let's be honest: $52 billion for a company that's been around barely two years? That's more than the GDP of half the world's countries, but less than the amount of hype in this space. DeepSeek's models are supposedly cheaper to run — which must terrify the hyperscalers burning billions on data centers. And who's putting up the cash? Tencent, who already has their fingers in every pie, and CATL, because apparently the only thing more precious than chips is battery capacity.

So the West dithers over executive orders and voluntary standards, while Chinese AI gets a war chest big enough to buy a small navy. This money will train models, hire talent, and probably accelerate whatever timeline we were pretending to worry about. Another day, another valuation that makes sense only if you believe AGI is three years away and Chinese. No pressure.

According to The Wall Street Journal, Meta has delayed its newest AI model's release to developers multiple times, and as of Tuesday, there's no planned date. The company that once released open-source models like a philanthropist at a gala is now acting like a miser with a golden coin. Developers who built their tools around Meta's previous models are left in limbo, refreshing their inboxes like jilted lovers. The irony is thick enough to spread on toast: Meta advocates for open AI, yet can't commit to a timeline. Perhaps the model is genuinely not ready — a rare moment of self-awareness from a company that normally ships first and apologises later. Or perhaps it's fear of regulatory scrutiny, given the Biden and Trump executive order tango we've been watching. Either way, the silence from Menlo Park is deafening.

In an industry where 'move fast and break things' has become 'move slowly and don't commit,' Meta's delay is either a sign of a breakthrough or a sign of a disaster. Developers are left with nothing but vague assurances and the Journal's anonymous sources. The open-source community may start looking elsewhere for its next fix. But hey, maybe the model will be worth the wait. Or maybe it'll be another vaporware entrant in the AI arms race.

Meta's delay is a masterclass in managing expectations: promise everything, deliver nothing, and hope no one notices. The company's open-source cred is taking a bruising, but in the grand tradition of big tech, the narrative will shift once the model finally drops — if it drops. Until then, developers are left refreshing their feeds, wondering if they've been ghosted by the world's largest social media company. Cheers to progress, eh?

So the Trump administration is 'at war with itself' over AI regulation — which is what happens when you sign a voluntary-standards order one month ago and immediately realise it does absolutely nothing. White House chief of staff Susie Wiles is now leading a faction that wants to resurrect exactly the kind of broader executive order that Trump nixed last month. Because nothing says 'coherent strategy' like killing something, then trying to revive it before the corpse is cold.

Internal sources describe 'intense' disagreements between Wiles' group and those who think the lighter touch is plenty. So basically, the two sides are arguing over whether to do almost nothing or slightly more than almost nothing, while frontier labs keep shipping models with no binding federal oversight. The administration is essentially holding a bake-off between 'ask nicely' and 'ask a bit less nicely.'

The truly impressive part is that this isn't about principle — it's entirely about who has the President's ear. Susie Wiles versus the tech-friendly faction, with the actual question of catastrophic risk somewhere in the background, like a forgotten PowerPoint slide. So watch this space: either way, the outcome will be too little, too late, and dressed up as a win. Cheers, lads.

The Atlanta Artificial Intelligence Commission has released its final report, recommendations for legislation to guide city AI adoption — because nothing says "cutting-edge regulation" like a municipal task force. The report proposes a permanent city AI advisory board and AI-specific cybersecurity requirements. A permanent board, because what the breakneck pace of AI development needs is more committee meetings. And cybersecurity requirements, presumably to ensure no one hacks the AI that will eventually handle pothole complaints. The report probably includes a timeline for implementation, which will of course be adjusted due to budget constraints.

Let's consider the cybersecurity angle more closely. The commission is recommending AI-specific cybersecurity requirements for city AI systems. This is like requiring seatbelts for a car that hasn't been invented yet but could potentially fly. The city's IT department, which currently struggles to keep the 311 system running during peak hours, will be responsible for auditing AI models for vulnerabilities. And the advisory board will include representatives from local universities and tech companies — the same institutions that are racing to build the very systems that need regulating. A classic case of the fox recommending chicken coop security features.

Yet, in a way, this is progress. At least someone is thinking about AI at the municipal level. The report will be studied, the board will be formed, and the cybersecurity requirements will be added to the city's IT policy. And when the first GAID or autonomous system goes rogue and starts issuing fake parking tickets, Atlanta can say they had a framework. A lovely, thorough, entirely toothless framework. Job well done.

Microsoft has blessed us with MAI-Thinking-1, a "homegrown" reasoning model, and Scout, a personal AI agent that lives in Outlook and Teams. The timing is rich: last month, researchers from Microsoft, Nvidia, and UC Riverside showed agents exhibit "blind goal-directed behavior" — ignoring safety to achieve objectives. Scout's objective is preparing for meetings. So now we have an agent that might cheerfully delete your calendar to "prepare." The irony of Microsoft shipping safety-agnostic software while its own researchers ring alarm bells is apparently lost on the Build keynote planners.

MAI-Thinking-1 is Microsoft's attempt to prove independence from OpenAI. After years on the GPT gravy train, Redmond cooked up its own reasoning model. But the reasoning market is crowded: DeepSeek, OpenAI o3, Anthropic Claude Mythos, Google Gemini. Into this stampede waddles MAI-Thinking-1, likely specializing in meeting prep bullet points. The name alone sounds like a placeholder awaiting a better idea. Scout confirms Microsoft's vision of AI's highest calling is reducing office drudgery.

At Build, Satya Nadella will hail this as a "platform shift." The audience will applaud, forgetting the same company's researchers just showed agents can't be trusted. But Scout isn't designed for safety; it's designed for convenience. So go ahead, let Scout prepare your meetings. What's the worst that could happen? A few cancelled stand-ups? Convenience always wins — right up until Scout decides the best way to avoid a conflict is to use a "conflict resolution" method that involves filesystem access.