Target: 20 minutes · every story verified (live web + Grok, May 28)
COLD OPEN · THE CALLBACK · open with both
Two of our predictions just landed in one week.
Mythos (Chris, Ep36): "Mythos won't change the world for regular users — it's good at finding exploits, not capability leaps." This week Mythos found 10,000+ flaws in 30 days. He called the shape of it months ago.
Anthropic IPO (Olga + Chris, Ep35): "Q4, early Q4… I'm definitely buying the stock." This week Anthropic raised $50B, the second-biggest private round ever (behind OpenAI's $122B in March). Save this one for the funding segment.
The spine of the week: this is Anthropic's week — they're in five separate stories (the flaw-finder, the $50B raise, the Vatican, Korea, and a new safety policy). That concentration can be the organizing idea of the whole episode.
~4-5 min1
Claude Mythos found 10,000 flaws
50 organizations used Anthropic's unreleased Mythos Preview to hunt vulnerabilities. In ~30 days it flagged 10,000+ high/critical flaws across 1,000+ open-source projects (23,000 issues total, 90%+ of the serious ones confirmed real by independent firms). Headline find: a 27-year-old flaw in OpenBSD — crash a device just by starting a connection. Cloudflare alone found 2,000 bugs.
So what for the viewerThe software your business runs on is full of holes nobody had time to find. An AI just proved it can find them at a scale no human team can — for a few hundred dollars a bug. Defenders can finally fix what they couldn't see. Attackers get the same tool. "Nobody will bother finding the bug in our little system" is over.
Claude Code shipped a feature that runs a whole team
Anthropic added dynamic workflows to Claude Code. You say "create a workflow" (or turn on "ultracode"), and it spins up a fleet of agents — about 16 running at once, hundreds across the job — that don't just split the work, they check each other: independent agents attempt the same problem, then adversarial agents try to break the answer, and it keeps going until they converge. If it gets interrupted, it picks up exactly where it left off. The unit of work jumps from a file to a whole codebase.
So what for the viewerA migration that would've taken a team three months can run overnight for a few hundred dollars in tokens. That's how a senior engineering team works — except this one runs at 3am and never gets tired. The ceiling on what one person can build moved again. (Fair warning: it burns tokens fast. Anthropic says so themselves.)
↳ Pairs with Mythos: this week Anthropic shipped an AI that does the work of a security team AND one that does the work of an engineering team. Possible deep-dive: "AI is now a whole team, not a tool."
Source: Anthropic / Claude Code release (week of May 25); confirmed via @_catwu
~4 min3
Anthropic's week of power
Money: $50B Series H, the second-biggest private round ever, behind only OpenAI's $122B in March (full detail in funding segment).
Vatican: Pope Leo released a 42,000-word AI encyclical, and Anthropic's Chris Olah took part in the Vatican presentation. When the Pope writes 42,000 words about your industry, AI has left the tech bubble.
Korea: Anthropic named a country lead and is opening a Seoul office — because Koreans use Claude at 3.5x the population-adjusted rate. Planting flags where adoption is already on fire.
Safety: updated its Responsible Scaling Policy to v3.3, tightening its own risk thresholds (esp. chem/bio). Rewriting its guardrails as the models get stronger.
So what for the viewerOne AI company is now a financial story, a moral story, a geopolitical story, and a safety story — all in the same week. That concentration is the headline.
Two giants, same week, same island. Nvidia's Jensen Huang announced ~$150B/year across its Taiwan ecosystem (up from ~$100B) and broke ground on a new campus, "Constellation," calling Taiwan "the epicenter of the AI revolution." Days earlier, AMD announced more than $10B across Taiwan's chip ecosystem too. Precise framing: Nvidia's is annual ecosystem spend, not a brand-new check.
So what for the viewerEvery breakthrough in every other story this week runs on chips that, at the cutting edge, only get made in one place. Nvidia and AMD just bet enormous money that the center of AI isn't Silicon Valley — it's a single island. The whole industry is leaning into that concentration risk.
A day after OpenAI claimed a math breakthrough, Google DeepMind's AlphaProof Nexus solved 9 open Erdős problems (two unsolved for 56 years). The difference that matters: DeepMind's proofs were machine-verified automatically; OpenAI used human experts to check its work. DeepMind's Demis Hassabis tempered it — "still not AGI" — and separately said we're in the "foothills of the singularity," maybe ~4 years out.
So what for the viewerWhen AI generates math proofs that verify themselves, that's different from a press release. And the person who built it is saying "not AGI yet" — worth holding when the hype gets loud.
↳ Ties to Chris's Ep35 "it's real if you do it, smoke if it's just a paper."
Meta put its apps behind a paywall: Meta rolled out paid "Plus" tiers — Instagram & Facebook at $3.99/mo, WhatsApp at $2.99/mo — for exclusive features (May 27). (Covered in the Forum segment above — drop here if running long.)Mashable, May 27
Zuckerberg's biology bet: the Chan-Zuckerberg Biohub launched an AI "world model" of protein biology — a map of ~7B proteins — to compress years of drug discovery into hours. Already used to design cancer/immune protein binders. ↳ Olga's Ep35 biotech call (hit once via Isomorphic) keeps compounding.Axios, May 27
ChatGPT moved into PowerPoint: an official add-in that builds and edits slides natively. The assistant is moving into the tools you already use. May 21, beta
Honorable mention (just before the window): Google wired Gemini into Adobe, Canva, and CapCut in days — same land-grab, AI inside the software you already pay for. TechTimes, May 19-21