Build · founder · 8 min read

State of AI — Week of May 25, 2026

Anthropic predicts AI trains its own successor by 2028. OpenAI solves an 80-year math problem. Trump kills the AI safety EO. What founders need to understand.

Published May 25, 2026 ·

weekly-roundup state-of-ai ai-tools news anthropic openai security

This was not a quiet week in AI. Between Anthropic’s co-founder predicting recursive self-improvement by 2028, OpenAI autonomously disproving an 80-year-old geometry conjecture, the White House killing its own AI safety executive order after direct calls from Musk and Zuckerberg, and a supply chain attack ripping through GitHub, OpenAI, and Mistral — it was a week that deserves more than a feed scroll. Here’s what matters and what doesn’t.

Anthropic’s official view: AI trains its own successor by end of 2028

On May 20, Anthropic co-founder Jack Clark delivered the Cosmos Lecture at Oxford. The headline prediction: a 60%+ chance that an AI model will fully train a successor version of itself — “recursive self-improvement” — by the end of 2028. Clark called it an “intelligence explosion” and said the term is now in official Anthropic research documents, not just theoretical speculation.

He also predicted: AI making a Nobel Prize-winning discovery within 12 months, companies run entirely by AI generating millions in revenue within 18 months, and bipedal robots assisting tradespeople within two years.

These are not the words of a fringe AI doomer. Clark is Anthropic’s co-founder. Anthropic is Anthropic. And Anthropic is simultaneously running Claude Code, building commercial APIs, targeting a $900 billion valuation, and stating internally that there’s a 60% chance AI is training its own successor within 30 months.

What this means for founders: Nothing actionable on a Tuesday morning. But if you’ve been treating “AI will keep getting incrementally better” as your planning assumption, that assumption may need to widen its range of scenarios. The tools you’re using today will be substantially different in 18 months. The category you’re building in may look unrecognizable in 3 years. That’s not a reason to stop building — it’s a reason to think carefully about durable advantages that don’t rely on specific model capabilities staying static.

OpenAI solved a problem mathematicians couldn’t crack in 80 years

On May 20, OpenAI announced that an internal reasoning model autonomously disproved the Erdős unit distance conjecture — a problem in discrete geometry that Paul Erdős posed in 1946. The model produced a 125-page proof connecting elementary geometry to deep algebraic number theory that mathematicians in the field had never thought to apply here. Fields medalist Tim Gowers called it “a milestone in AI mathematics.”

The relevant fact is not the specific math. It’s the methodology: the model was given the problem statement and worked independently. It was not trained on this problem. It did not retrieve an existing solution. It found a novel proof in a domain where the current experts hadn’t been looking.

What this means for founders: If you’re building a product that involves research synthesis, pattern recognition across large knowledge domains, or novel problem-solving, the capability gap between what AI can do today and what it could do six months from now is worth taking seriously. “AI is good at summarizing but not reasoning” is no longer a safe framing.

Trump killed the AI safety executive order — who made that call

The White House AI executive order — a voluntary 90-day pre-launch review framework for frontier AI models, with NSA involvement in classified testing — was cancelled on May 21, hours before the signing ceremony. Invitations had already gone out.

The mechanism: David Sacks, Elon Musk, and Mark Zuckerberg all called Trump directly between Wednesday night and Thursday morning. Musk and Zuckerberg warned the order could slow AI development. Sacks called the president Thursday morning. The order was dead by Thursday afternoon.

This EO was partly drafted in response to Claude Mythos — Anthropic’s most powerful model, which autonomously discovered zero-day vulnerabilities in every major operating system during internal testing and which Anthropic chose not to release publicly. The national security argument for the review framework was real. The CEO argument for killing it was louder.

What this means for founders: The current US regulatory posture is “don’t regulate AI development.” For founders building AI-powered products, this is mostly good news in the short term — less compliance overhead, faster iteration. The risk is downstream: no review framework means no standard liability framework either, which matters if your product causes harm and you need to prove you built responsibly. Documenting your own responsible AI practices is worth doing regardless of what the White House does.

The GitHub supply chain attack that hit 3,800 repos in 18 minutes

TeamPCP — a financially motivated cybercrime group — compromised a trojanized version of the Nx Console VS Code extension (2.2 million installs, verified publisher status) and pushed it to the Visual Studio Marketplace. It was live for 18 minutes on May 18. That was enough.

The payload: a silent credential harvester that targeted GitHub tokens, AWS keys, npm tokens, 1Password vaults, and Anthropic Claude Code configuration files (specifically ~/.claude/settings.json). Using harvested credentials, TeamPCP exfiltrated approximately 3,800 GitHub internal repositories. Confirmed victims: GitHub, OpenAI (two devices, certificates being revoked June 12), Mistral AI (extortion demand), the European Commission.

If you had the Nx Console VS Code extension installed, rotate your credentials now: GitHub personal access tokens, AWS keys, npm tokens, and anything in 1Password.

The broader pattern is the real story. Traditional security watches for known malware signatures. VS Code extensions, npm packages, and PyPI distributions are plain-text scripts that operate at a different layer entirely. The time from publication to confirmed credential harvest was 18 minutes. Standard tooling is not designed for that timeline.

What this means for founders: If you’re building with Claude Code, Cursor, or any other AI coding tool in a professional environment, your developer credential surface area is wider than it was two years ago. Short-lived tokens, hardware-backed credential storage, and routine rotation aren’t optional anymore. The vibe coding security guide covers the basics; add “extension publisher verification” to your practice.

Meta, Intuit, and the week that explained what AI costs

On May 19, leaked audio of a Meta all-hands surfaced. Mark Zuckerberg explained that Meta had been monitoring employee Gmail, coding sessions, and internal tool usage to train AI models — then 8,000 employees received layoff notices the same day. Separately, Intuit announced 3,000 job cuts explicitly tied to AI-driven efficiency.

This is the clearest articulation yet of the economic model: companies train AI on expert human behavior, then restructure headcount. The AI infrastructure investment (Meta committed $125 billion in capex for 2026) gets justified by the labor cost reduction. The math works for shareholders. It’s brutal for workers.

What this means for founders: Two things. First, if you’re building a product in a space where enterprise AI efficiency is replacing human roles, the tailwind is real — but the sales conversation is increasingly political. Procurement processes are adding scrutiny to “replaces X jobs” products. Second, if you’re a one-person or small-team founder using AI to do the work of a larger team, you’re living proof of the other side of this argument: AI lets small teams compete with large ones. That’s still the best framing for what you’re doing.

The Gemini creative hub: Adobe, Canva, CapCut coming to one chat window

Less ominous news from the week. Google announced that three major creative platforms are integrating directly into Gemini: Adobe (via Firefly AI Assistant, giving Gemini access to Photoshop, Premiere, Illustrator, and Lightroom workflows), Canva (Magic Layers — generate in Gemini, edit individual layers in Canva), and CapCut (mobile video editing inside Gemini, announced May 21, rolling out soon).

The Canva integration is the most immediately useful. You generate an image in Gemini, and Canva unlocks it with every element on a separate editable layer. For a founder building marketing assets, that’s the workflow where you currently switch between four tools. Gemini AI Ultra subscribers can try Magic Layers now; the Adobe and CapCut integrations are rolling out over the next few weeks.

This continues the pattern from Google I/O: Gemini is positioning as an orchestration layer, not a single tool. Whether that’s a better experience than using the individual tools directly will depend on how smoothly the handoffs work in practice. The early Canva integration reports are promising.

What to actually do this week

If you’re building with AI tools: rotate your developer credentials if you had Nx Console installed, watch for the macOS ChatGPT app forced update before June 12 (OpenAI is revoking its signing certificate), and note that Gemini AI Ultra gets the Canva Magic Layers integration now.

If you’re watching the bigger picture: Jack Clark’s predictions are Anthropic’s institutional view, not speculation. The pace of capability improvement is not slowing. The right response isn’t anxiety — it’s building with genuine durability in mind, keeping your stack modular enough to swap components as the tools change, and not over-investing in capabilities that AI will likely commoditize in the next 18 months.

The week’s most important sentence may actually be from Clark: “What do you do with a tremendous amount of abundance in many, many different fields of science all at once? Today’s institutions have very, very narrow pipes through which you push new drug candidates.” That’s the question for the decade, not just the week.