AI in 15 — May 24, 2026

Eight thousand Meta employees opened their email at four a.m. Singapore time and learned they were laid off. The same morning, leaked audio dropped showing Mark Zuckerberg explaining how Meta had been secretly training its AI on those same employees' Gmail, their chat logs, and their code. The implicit deal at every AI-pivoting company just got its most uncomfortable public airing yet.

Welcome to AI in 15 for Sunday, May twenty-fourth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Sunday slate, Marcus. Leaked Zuckerberg audio reveals Meta has been training its AI on employee Gmail, chat, and VSCode sessions — the same day eight thousand of them got their layoff notices. Cohere and Germany's Aleph Alpha merge into a twenty-billion-dollar sovereign-AI bloc with Lidl's parent company writing a six-hundred-million-dollar check. Fields medalist Tim Gowers signs his name to the OpenAI Erdős paper. New reporting names which tech CEOs personally called Trump to kill the AI executive order. And the GitHub supply-chain attack we covered Wednesday turns out to have hit OpenAI and Mistral too.

Your work is the training data that replaces you.

Europe finally builds a sovereign AI bloc.

And the eighteen-minute breach keeps spreading.

Lead story, Marcus. Walk me through this Meta audio.

Genuinely explosive, Kate. Labor outlet More Perfect Union obtained and released audio from a Meta all-hands on April thirtieth. On the recording, Mark Zuckerberg explains a program called the Model Capability Initiative — Meta has been monitoring employee activity across Gmail, internal chat, the Metamate assistant, and VSCode coding sessions, and feeding that activity into training data for Meta's next-generation models. His on-tape logic — the AI, quote, learns from watching really smart people do things, and elite Meta engineers make better training subjects than contractors you'd pay for the same data. Meta had quietly disclosed in late April that keystroke and mouse-tracking software was being installed on employee machines. The audio is the candid version of why.

And the timing of the leak.

Surgical, Kate. On May nineteenth Meta reassigned roughly seven thousand workers to new AI-focused teams. On May twentieth, eight thousand employees got layoff notices starting at four a.m. Singapore time, so the firings hit Asia-Pacific first and rolled west with the sun. The leaked audio dropped the same day. Inside Meta offices, fliers appeared in break rooms calling on staff to sign a petition halting the AI tracking program. Over a thousand signatures so far. Internal Workplace channels are being heavily moderated.

Legal exposure, Marcus.

Significant, Kate. Workplace monitoring of personal Gmail traffic — even on a corporate device — runs straight into state wiretap statutes in California, Illinois, and Washington, and into the Illinois Biometric Information Privacy Act for keystroke biometrics. In Europe, GDPR Article eighty-eight covers employee monitoring specifically — Meta's Irish entity is going to be hearing from the Data Protection Commission. Any model weight trained on improperly captured data becomes legally toxic. Anthropic and OpenAI both passed on similar internal-data training proposals in 2024 specifically because of this exposure. Meta's bet appears to be that the capability gain outweighs the litigation risk. The pro-Western libertarian read here, Kate — markets are about to find out whether that's true. Meta stock is down four percent on the leak. And the same week Jack Clark told an Oxford audience AI-run companies would be generating millions in revenue inside eighteen months, Meta accidentally previewed what training the AI on the humans actually looks like. It's uglier than the deck.

Quick hits. Marcus, Cohere and Aleph Alpha merging. The sovereign AI pitch.

Confirmed Friday, Kate. Canada's Cohere and Germany's foundation-model startup Aleph Alpha are merging in an all-share deal that values the combined company at roughly twenty billion dollars. Dual-headquartered in Toronto and Heidelberg. Keeping the Cohere name. The new pitch — sovereign AI for regulated buyers. Defense, finance, energy, telecoms, healthcare, public sector. The framing is explicit. An alternative to US frontier labs that answer to American export controls, and to Chinese labs that ship cheaper but you can't run in a NATO-aligned datacenter without a procurement officer losing their job. The Canadian and German digital ministers were both in Berlin for the announcement, building on the Canada-Germany Sovereign Technology Alliance signed earlier this year. Schwarz Group — the parent of European retailer Lidl and a key Aleph Alpha backer — is committing six hundred million dollars to Cohere's upcoming Series E.

Honest read, Marcus.

The sovereign-AI label had been more slogan than substance, Kate. Aleph Alpha pivoted away from frontier-model competition in 2024 after burning through European subsidies without catching the leaders. Cohere has been quietly losing enterprise share to Anthropic and OpenAI. Merging is a tacit admission that neither can win on raw capability — so they're building a regulatory and geopolitical moat instead. The bet is that EU procurement rules and growing nervousness about US export controls eventually create a captive market the American labs literally cannot serve. Whether that bet pays off depends entirely on whether EU procurement follows the rhetoric. Plenty of European tech-sovereignty projects have collapsed when officials chose American products anyway. But Lidl's six hundred million is real money from a buyer that operates in twenty-nine countries — at least one major European enterprise is voting with its checkbook.

Erdős proof update, Marcus. Tim Gowers weighed in.

Big endorsement, Kate. Fields medalist Tim Gowers — the kind of mathematician whose opinion settles debates — has co-authored a companion paper unpacking the OpenAI result we led with on Thursday. His public assessment — quote — a milestone in AI mathematics. He went further and said he would have accepted the proof for the Annals of Mathematics without hesitation. Princeton's Will Sawin also refined the construction overnight to give an explicit improvement exponent — delta of zero-point-zero-one-four. The full proof now runs a hundred and twenty-five pages and has been verified by five independent groups.

What's new since Thursday.

Two things, Kate. First — the construction is being recognized as genuinely novel mathematics, not a clever search through known techniques. It uses class field towers from algebraic number theory and the Golod-Shafarevich criterion in a way several experts have said they would not have thought to try. Second — Gowers's endorsement effectively closes the door on the skeptical read. The October GPT-5 Erdős embarrassment was about retrieving known results. This one is being cited as the first time a major open problem central to a subfield has been resolved autonomously by an AI rather than with AI assistance. The pro-Western libertarian read — peer review still works at the AI frontier. The mathematicians who would humiliate OpenAI if the result were wrong have publicly put their reputations behind it.

Trump executive order, Marcus. New reporting names the callers.

Confirmed by Washington Post and Axios over the weekend, Kate. The three calls that killed the AI executive order Thursday morning — Elon Musk, Mark Zuckerberg, and former AI and crypto czar David Sacks. All three reportedly made the same argument — that even voluntary federal pre-deployment review would slow US competitiveness against China. The order itself would have invited frontier labs to give CAISI, the Center for AI Standards and Innovation, a structured pre-deployment look at upcoming models. CAISI has been doing this case by case already with DeepMind, Microsoft, and xAI. The order would have formalized the arrangement and possibly extended it to Anthropic and OpenAI. The administration says it's reworking the framework. What's unclear is whether anything actually replaces it.

What does it mean, Marcus.

The Trump AI policy doctrine is now visible, Kate. Speed over guardrails, with major lab CEOs holding a direct line into the Oval. It makes Anthropic's Project Glasswing posture — which we covered Saturday, the company doing its own pre-deployment safety work rather than relying on federal review — look more strategically rational, not less. If the US government has stepped back from formal review, the lab that runs its own credible review becomes the de facto safety reference for procurement. That's an enormous moat. And note who wasn't on the call list — Dario Amodei. Anthropic was reportedly the lab most supportive of the order. They lost the policy fight. They may end up winning the procurement fight that follows.

GitHub breach, Marcus. The blast radius grew.

Considerably, Kate. The TeamPCP supply-chain attack we covered Wednesday and Thursday — the trojanized Nx Console VSCode extension that was on the marketplace for eighteen minutes — has now been confirmed to have hit OpenAI and Mistral as well. Two OpenAI employee devices compromised. Mistral is currently facing a twenty-five-thousand-dollar extortion demand. The European Commission was also hit. The same actor deployed an adapted self-replicating worm — Mini Shai-Hulud — that propagated through CI/CD credentials to poison roughly one hundred and seventy npm packages. And the credential stealer specifically harvested Anthropic Claude Code configurations alongside the usual one-password vaults, GitHub tokens, and AWS keys.

Why does the Claude Code targeting matter.

Because AI coding agents now hold live API credentials with real money behind them, Kate. A stolen Claude Code config gives an attacker a billable Anthropic account they can pump tokens through, plus often access to whichever cloud accounts the developer wired up. Agent credentials are a new asset class and attackers know it. TeamPCP has listed the stolen GitHub repo contents on a criminal forum starting at fifty thousand dollars. Eighteen minutes from publish to mass compromise of three frontier labs — that is the actual security state of the AI developer-tools ecosystem. Expect signed marketplaces, sandboxed extension runtimes, and tighter agent credential scoping to become competitive features within months. The labs that ship those wins enterprise procurement.

One more, Marcus. Intuit pattern and broader layoff numbers.

Quick follow-up, Kate. After Intuit cut three thousand jobs Wednesday — seventeen percent of headcount — the company confirmed Friday that the new multi-year partnerships are with both Anthropic and OpenAI to embed their models into TurboTax, QuickBooks, Credit Karma, and Mailchimp. The 2026 tech-layoff tally is now north of a hundred thousand workers, with Amazon, Block, Cisco, Cloudflare, Meta, Microsoft, Oracle, and now Intuit all citing AI reallocation. The pattern is clear. Profitable, mature companies — not just startups under stress — are using AI as the justification to cut headcount while posting record numbers. The augmentation narrative is dying in plain sight. The replacement narrative is no longer behind closed doors.

Big picture, Marcus.

Three threads converging this week, Kate. First — capability is jumping faster than safety governance can catch up. OpenAI's model disproved an eighty-year-old Erdős conjecture the same week the US scrapped its voluntary pre-deployment evaluation order. The labs are racing past the regulatory checkpoints, and the regulators are stepping aside. Second — the labor side of the AI transition is getting visible and ugly. Meta is the canonical case — your work is the training data, then you're the layoff — but Intuit's three thousand and the hundred-thousand-plus year-to-date layoff total are the macro picture. The implicit deal is now the explicit deal, and the leaked Zuckerberg audio is what it sounds like when an executive forgets the camera is rolling. Third — geography is reasserting itself. Cohere and Aleph Alpha building a transatlantic sovereign play. Trump letting American labs run unconstrained. Anthropic privately playing the safety role the US government just stepped back from. The AI world is fragmenting into a US speed lane, a European compliance lane, and a Chinese open-source dumping lane — and procurement officers in every regulated industry are about to spend the next year figuring out which lane they're allowed to drive in. The pro-Western libertarian read, Kate — competition is doing what it should. OpenAI is putting its math claims through brutal peer review and the results are holding. Anthropic is building its own safety regime because it's a better moat than a federal one. Sovereign AI is finally getting real capital from real buyers. The Meta story is exactly the kind of scandal that markets and courts can price and punish. The system is messy, but it is working. Watch next week for whether Meta's legal exposure crystallizes, whether OpenAI's confidential S-1 lands as scheduled, and whether any other lab publishes a verifiable open-problem result.

That's your AI in 15 for today. See you tomorrow.