A leaked tape. A fired workforce. And a city that just put a 1-in-20 odds on an AI economic shockwave.
•Mike Robinson
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
9:23
Yesterday in AI | Friday, May 22, 2026
A leaked tape. A fired workforce. And a city that just put a 1-in-20 odds on an AI economic shockwave.
A secret recording surfaced this week that reframes everything companies say about "productivity monitoring," and the person talking is one of the most powerful tech executives alive. Meanwhile, a general-purpose AI disproved something mathematicians have believed since 1946, verified by three of the world's top math minds. Then there's the AI company that just hit a financial milestone nobody expected this soon, a government report that assigned specific probability weights to the scenario where AI destroys millions of jobs, and a White House executive order that was almost signed, then wasn't. Six stories that all land differently than the headlines suggest.
Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, Bluesky, or Substack. If you like the show, please take a minute to rate and review it so others can find it!
Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. Today is a hybrid episode because you guessed it, I ran out of time recording this. So thanks for your patience while I figure this out. I'll start us out and come back around at the end to finish us out. It's Friday, May 22nd, and this week the AI industry apparently decided Thursdays are for dropping bombs. We've got a leaked Zuckerberg tape, a new math breakthrough that's been 80 years in the making, a government report that just put a dollar figure and a probability on AI's worst-case economic scenario, and a White House executive order that almost happened. Let's get into it. Let's start with the story that had every tech worker's jaw on the floor. A leaked audio recording from a Meta All Hands meeting surfaced this week, and it's honestly a lot. In the recording, Mark Zuckerberg explains that Meta has been monitoring employee activity across Gmail, GChat, their internal tool called MetaMate, and VS Code, which is the coding software most engineers live inside to train its AI models. His reasoning, the AI learns from watching really smart people do things, and elite engineers make better training data than outside contractors, which is, if you set aside everything else, a genuinely smart training strategy. It's the then fire 8,000 of those really smart people part that makes it something else entirely. Here's how the timeline played out. Back in April, Meta quietly installed keystroke and mousetracking software on employee computers. Their public explanation was mild. The models just needed to learn how people click drop-down menus. Routine stuff. Then at an April 30th All Hands, Zuckerberg gave a rather different explanation on tape. He acknowledged the leak risk, telling staff it was not strategically in your interest to share details. Then on Monday this week, 7,000 workers got reassigned to AI-focused teams, framed as a productivity upgrade. Then on Wednesday, 8,000 employees received layoff notices starting at 4 a.m. Singapore time. The leaked audio hit the same day. Here's the thing. Meta will survive this. Zuckerberg is building toward AI native product development, and his numbers back it up. $60 billion in profit last year, $115 to $145 billion in planned AI capital spending this year. He's playing a longer game. But this story is a warning shot for every company with an employee productivity monitoring program. Meta is just the one that got caught explaining the real logic on tape, right before firing some of the people it was watching. The line between helping you work better and training your replacement just got a lot blurrier. Now it's time for a mathematically different kind of story. OpenAI announced this week that one of its internal AI models has disproved an 80-year-old belief in discrete geometry, specifically related to Paul Erdos' 1946 unit distance problem. Before your eyes glaze over, here's the plain English version. For eight decades, mathematicians operated on a shared assumption about how many same-length links you can draw between a set of points. OpenAI's model found a completely different answer using algebraic number theory, a different branch of math than anyone had tried before. The solution was verified by Tim Gowers, Noga Alon, and Thomas Bloom. Those are three of the most respected mathematicians alive. The proof checks out. Sam Altman called it kinda a big milestone. That might be the single rarest case of an AI CEO underselling something. OpenAI's researcher Alex Wei put it better. Math is a leading indicator of what is to come. And that's exactly the point. DeepMind's Alpha Proof was built specifically to work on proofs. OpenAI's model is a general-purpose AI. It walked into an 80-year-old geometry problem and found something no one had found before. If an AI can do that in pure mathematics, the implications for drug discovery and material science are hard to overstate. OpenAI is planning to release this model publicly soon. It's worth watching very closely when it drops. While we're talking about math, Anthropic told investors this week that it expects Q2, 2026 to be its first ever profitable quarter. Revenue is projected at $10.9 billion, more than double what they did in Q1. The growth is being driven almost entirely by surging professional demand for Claude, and particularly Claude Code, Anthropic's AI coding tool. But there's a catch. This profitable quarter might not hold. Anthropic has massive compute costs coming later in 2026, and those could push them back into the red. One profitable quarter doesn't mean the company turned a corner. But the timing is still interesting. Anthropic crossed $1 trillion in valuation on secondary markets back in April and is currently raising $40 to $50 billion more. Going profitable, even temporarily, reframes that fundraising story in a real way. The AI race suddenly has a new scorecard, actual revenue. How about some more math and money with a small side of Doom? New York City's Comptroller Office, the city's independent financial watchdog, dropped a report this week that does something I haven't seen a government actually do, put actual odds on how AI could reshape the economy. The report lays out five scenarios. The most likely at 35% probability is called AI empowered economy, your basic soft landing version where AI helps more than it hurts. Close behind at 25% is AI falls flat, where the technology disappoints and adoption stalls, then job replacement at 20%, productivity boon at 15%, and at 5%, something called AI shockwave. That last one is worth pausing on. In the AI shockwave scenario, there is 5.4 million private sector job losses nationally, 2 million of those specifically in office industries, lower GDP growth, a larger drop in the share of income that goes to workers. For New York City, the report models sizable hits to employment and tax revenues straight through 2030. Now, 5% sounds small, but government budget offices don't model 5% scenarios unless they take them seriously. What this report is really saying is, we've assigned a 1 in 20 chance to an AI-driven economic crisis, and we need an early warning system before it hits, that's a meaningful shift. Researchers and bloggers have worried about AI displacing jobs for years. A city government assigning it a 5% probability and building a live dashboard to track it puts the concern in a different category. Let's talk policy for a minute. The Trump administration drafted an executive order this week that would have been a real shift in how Washington thinks about AI. The plan, require leading AI companies to voluntarily submit their most powerful new models to the government for review 14 to 90 days before public launch. The Office of the National Cyber Director would coordinate the reviews. Companies would hand over models, and in return the government would surface security gaps before adversaries could exploit them. There was also a proposal for a shared database where researchers and companies could report security flaws they find while using AI systems. Sounds reasonable, right? Then at the last minute, the signing was postponed. The reason given was concern that the order could hurt the AI industry. So, the government drafted a plan to require companies to do something most of them already do voluntarily, then decided not to require it. But here's what I'd note: even a delayed order is a signal. Washington is increasingly treating the most powerful AI systems as critical infrastructure. The hands-off era is probably ending, and companies that already test their models for security before release will be ahead of everyone else when rules eventually land. Here's my last story. Google published its AI coscientist system in Nature this week, one of the most prestigious scientific journals on the planet. The new feature they're calling hypothesis generation works like this. Multiple research agents are pitted against each other in what Google describes as idea tournaments. Each agent proposes scientific hypotheses. They argue with each other. The best ideas survive. The system then surfaces novel hypotheses that human researchers in biology and chemistry labs can actually test. If that playbook sounds familiar, it should. It's the same logic as AlphaGo. Competitive self-play to get smarter. AlphaGo beat the world champion at Go by playing against itself millions of times. These research agents get sharper by competing against each other. It pairs with the OpenAI math story, though in a different way. OpenAI's model walked into an 80-year-old math problem and solved it on its own. Google's co-scientist is built to hand the most promising ideas to human researchers and let them do the testing. Two different bets on where AI fits in scientific discovery. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterday inai.news. Or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast, please be sure to rate and review it so others can find it. Thanks. That's all for this Attack of the Clones edition of Yesterday in AI. Stay curious, and I'll see you tomorrow.
