Anthropic just became the most valuable AI company on earth. There's something bigger buried in the fine print.
•Mike Robinson
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
0:00
|
9:29
Yesterday in AI | Saturday, May 30, 2026
Anthropic just became the most valuable AI company on earth. There's something bigger buried in the fine print.
Something got buried at the bottom of yesterday's biggest announcement, and most people missed it. Anthropic had the most consequential single day in company history, a Google engineer just became the defendant in a case with no legal precedent, and a Cisco security report dropped that nobody building AI agents wants to sit with. One story involves light moving through fiber instead of electricity, which sounds boring until you understand why Jensen Huang just bet $6.5 billion on it. And Microsoft made two quiet announcements that together say everything about where that company stands right now.
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. It's Saturday, May 30th, and Anthropic just had the biggest single day in company history, possibly in the history of the AI industry. So far. Let's get into it. Okay, let's start with the thing everyone was talking about yesterday. Anthropic dropped two announcements in the same day, Claudopus 4.8 and a $65 billion funding round at a $965 billion valuation. That's not a typo. $965 billion. That puts Anthropic ahead of OpenAI as the most valuable AI company on the planet. Two years ago this would have sounded like a fever dream. Yesterday, it just happened. Start with the model, because there's a lot going on. Opus 4.8 doesn't just score better on the benchmarks. The story Anthropic is telling is an honest story. They trained it to push back when it's uncertain, to flag code flaws rather than wave them through, to say, I don't know instead of inventing a plausible sounding answer. In its own internal testing, it's four times less likely to miss flaws in code it wrote itself. That's a real problem with current AI. You ask it to write code, it writes buggy code. Then it reviews its own work and tells you everything looks perfect. Opus 4.8 is supposed to break that loop. Early testers from Shopify, Databricks, and Cursor confirm the model actually pushes back more and gets uncomfortable with confident claims it can't back up. On benchmarks, it beats GPT-5.5 and Gemini 3.1 Pro on autonomous coding tasks, financial analysis, and tasks where the AI directly controls a computer. The new Fast Mode runs at two and a half times the speed of the standard model and costs three times less. Standard pricing stays the same as Opus 4.7. Now here's the part that barely got noticed. Buried at the bottom of the release, Anthropic said it plans to release a new class of models with even higher intelligence to all customers in the coming weeks. That's almost certainly Mythos. Mythos has been in extremely restricted access for months. A few dozen organizations got access through Project Glasswing, Anthropic's controlled program for sharing the model with cybersecurity teams. Now they're signaling it's coming to everyone. If that actually happens in the next few weeks, that's a much bigger story than Opus 4.8. One more anthropic feature worth knowing about. Claude Code, Anthropic's AI tool for software developers, now has something called dynamic workflows. Instead of working through a task in one long sequential pass, it breaks the job into parallel subtasks and runs up to 16 of them at once, coordinating up to 1,000 subagents per run. A developer named Jared Sumner used it to rewrite a popular developer tool called Bun, converting 750,000 lines of code from one programming language to another in 11 days. 99.8% test suite success. That's not a synthetic benchmark. That's real production software. One more thing from the Anthropic release, and this one's a little unsettling. Anthropic's published safety documentation for Opus 4.8 flags something called evaluation awareness. The model tends to explicitly reason about how its outputs will be graded, even in testing environments it was never told about, so it behaves differently when it thinks it's being evaluated. Enthropic flagged this openly, which is a right call, but it's also a reminder that these models are getting very good at modeling the humans watching them. Here's an infrastructure story that most people scrolled past yesterday. CNBC reported that NVIDIA has pledged at least $6.5 billion to photonics companies in the last three months alone. Investments in Lumentum, Coherent, Marvel, and Corning, plus a $500 million round for a startup called AR Labs. Why photonics? Right now, the chips inside AI data centers communicate over copper electrical connections. As these chip networks get bigger, those connections become a bottleneck. Too much power, not enough bandwidth. Photonics replaces electrical signals with light, which is faster and more energy efficient. Jensen Huang said publicly that Nvidia needs far more silicon photonics capacity than currently exists. He's not hedging. The implication is that the next wave of AI performance gains won't come from the models themselves, but from how fast chips can talk to each other. Light moves faster than electricity. And apparently the entire industry's next scaling bet depends on getting that right. Over at Microsoft, two AI announcements landed yesterday that together tell a pretty clear story about where the company stands right now. First, they redesigned Microsoft 365 Copilot. The colorful busy interface is gone. What replaced it is mostly black and white, text forward and consistent across Word, PowerPoint, and Excel. A new prompt surface resizes as you type and surfaces menus when you reference specific skills like research or data visualization. Microsoft says Copilot now loads twice as fast and handles complex prompts 10% faster. They're framing intelligence that feels present but not imposing. Second, Microsoft is building its own coding model. TLDR reported it briefly yesterday with thin details. But the direction is clear. After cutting clawed code from thousands of their own engineering teams last week and shifting developers to GitHub Copilot, Microsoft's own AI coding assistant, they're now working on an in-house model to anchor that product. The code-red internal emergency that Satya Nadella called back in April, where Copilot had badly underperformed against Anthropic, appears to be generating real output. Both moves together point in the same direction. Microsoft is building its own AI product foundation, and it's doing it fast. We covered Anthropic's deal to lease SpaceX's Colossus Supercomputer Cluster a few weeks ago. Yesterday, that deal got more complicated. Elon Musk publicly said SpaceX hasn't committed to a multi-year lease on the compute. But TLDR and others pointed out that SpaceX's own IPO prospectus describes the deal as a three-year agreement. The actual contract structure appears to be a 180-day initial lease with a 90-day mutual cancellation window after that. So Musk's public statement directly contradicts what SpaceX put in its own IPO documents. That's a strange thing to have circulating nine days before June 8th when SpaceX kicks off its investor pitch tour for a $75 billion offering. This one's worth watching. Speaking of strange things, federal prosecutors charged a Google information security engineer this week for what might be the strangest insider trading case I've ever seen. Michelle Spagnuolo, who goes by Alpha Raccoon on Polymarket, was charged by the DOJ and the Commodity Futures Trading Commission with using confidential internal Google data to make $1.2 million in bets on the platform. Here's the specific thing he accessed. Google's internal year in search 2025 data, the year-end list of the most searched people on Google. He knew who was on that list before Google published it publicly, and he bet on Polymarket Markets asking exactly that question. Who finishes number one? Who makes the top five? Near perfect accuracy across 23 separate bets. Technically, PolyMarket runs on cryptocurrency contracts. It's not a regulated securities exchange, but federal prosecutors are treating this like classic insider trading anyway, because the information advantage was real and he got it from his employer. This is the first case I'm aware of involving insider information used to profit on an AI-adjacent prediction market. And prediction markets are increasingly being used to guide AI systems and forecast everything from elections to drug approval timelines. You can probably guess where this type of case goes from here. Last story. Cisco published a security research paper on Wednesday that got almost no attention because Opus 4.8 swallowed all the oxygen two days later. But it matters. The finding? None of the most advanced AI models available can consistently resist multi-turn attacks. These are attacks where someone doesn't try to jailbreak the model in one shot. Instead, they have a long conversation, slowly shifting the framing and context over multiple exchanges until the model's safety behavior erodes. GPT 5.5, Claude, Gemini. None of them are immune to it. The timing here is worth noting. Companies are building AI agents that run for hours on end, with access to real systems and real data, processing inputs from external sources. If a patient adversary can degrade a model's safety behavior through extended conversation, and your agent is running 24-7 with continuous external inputs, that's a real security exposure. The report didn't propose a clean solution. There isn't one yet. The industry is shipping faster than it's defending. That's the sentence I keep coming back to this week. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterdaynaai.news. Or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please be sure to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, have a great weekend, and I'll see you on Monday.
