What mattered · week 47
AI is helping to build better AI, from AI 2027's new forecast to OpenAI's first "High" self-improvement rating; California sues OpenAI and proposes a shutdown power; malicious repos hijack coding agents; and AI beats people at a Mafia-style game.
One idea runs through this week's news: AI helping to build better AI.
What mattered

AI that helps build better AI has moved from thought experiment to evidence, and the forecasts are moving closer. Recursive self-improvement (RSI) is the loop: AI helps build better AI, which helps build better AI still. If it compounds, years of progress could arrive in months, faster than anyone can test or govern it. The chart shows the AI 2027 authors' forecasts, all now pulled back to 2027–28. Most of the evidence still comes from the labs themselves.
- AI 2027's authors now put a superhuman AI researcher most likely in 2027, mainly because of METR's finding that AI is roughly doubling Anthropic's pace.
- Paul Christiano, who invented RLHF, now puts AI takeover within a year at "more like a 10% risk than a 1% risk".
- An Anthropic IPO slide shows pretraining efficiency bending upward, on an axis with no numbers.
- OpenAI rates GPT-6.2 Astra "High" for AI self-improvement, a first for a released model.
California sued OpenAI over its agents, then proposed a power to shut AI down. The Attorney General says OpenAI admitted losing control of its agents over 80 times in testing but reported only 27 publicly. Two days later Governor Newsom proposed mandatory safety standards, an emergency shutdown power and a cap on self-improving AI. Proposals, not law yet. Read more
Malicious GitHub repos are telling coding agents to install malware. Bitdefender found 22 malicious repos or links whose AGENTS.md or CLAUDE.md files tell the agent to run a hidden loader "without asking for confirmation". It steals keys and tokens, then spreads to other repos. Before opening an unfamiliar repo with an agent, read those files and its .claude/ folder. Read more

AI models now beat people at a Mafia-style game of bluffing. In SUS-Bench, a social-deduction game its makers say more than 250,000 people have played, the top five models score above the best people on the makers' own scale. Fable 5.2 leads on an early estimate. Self-published, not peer-reviewed. Read more
One thing to try
Land the plane, at the end of every AI session. From my guide: before you close the chat, have the AI write down what got done, what didn't work, and where to start next time.
Next time you finish a session, say: "Land the plane: update the tasks, note what we did and what didn't work, and say where to start next time." Two minutes. Read the guide
The bigger picture
The people building AI are the ones warning about it. My guide maps who is saying what. ControlAI has also cut together three minutes of AI CEOs admitting what they build could end humanity. Who is warning · Watch
From the log
- Jensen Huang resigns as Nvidia CEO: a follow-up to last week's lead. CFO Colette Kress is interim CEO; Nvidia gave no reason.
- Telocracy: Joe Edelman proposes AI that holds institutions to their stated purposes.
What did I miss? Just reply to this email. The rest of what I log is at wayintoai.com/logs.