01
SELF-CRITIQUE CAN IMPROVE OUTPUTS WITHOUT RETRAINING
The Self-Refine paper had a model critique and revise its own answers across seven tasks. Same model, no extra training. Scores went up anyway. Inference-time trick, not a successor model.
FIELD GUIDE · AI TEACHING ITSELF TRICKS
CAN AI MAKE ITSELF SMARTER?
AI already lends a hand building AI. Things get spicy if each upgrade makes the next one arrive quicker — and nobody can hit the brakes.
UPDATED 2026-10-12
TL;DR
THE STORY SO FAR
02
AI CAN DISCOVER ALGORITHMS THAT IMPROVE THE AI COMPUTING STACK
DeepMind's AlphaEvolve mixes language models with auto-graders and evolutionary search. Its discoveries improved data-center scheduling, hardware, software, and parts of AI training. A genuine feedback path from AI research tooling into AI infrastructure — but humans still wrote the tests and approved the deployments.
03
FRONTIER LABS NOW TREAT AI SELF-IMPROVEMENT AS A SAFETY-RELEVANT CAPABILITY
OpenAI's 2025 Preparedness Framework tracks AI self-improvement as a capability category, alongside long-range autonomy, autonomous replication, and safeguard-undermining. They're measuring it. They haven't declared it loose.
04
AUTOMATED IMPROVEMENT IS EASIEST WHERE SUCCESS CAN BE SCORED
AlphaEvolve shines when candidate programs can be run and auto-graded against hard numbers. Messy open-ended science is way harder than optimizing with a clean score.
LATEST FROM AI NEWS PAGE
FROM THE FRONT PAGE + ZIMBO
KEEP READING
NEXT FIELD GUIDE →WHEN IS AI ACTUALLY AGI?There is no AGI referee. Labs disagree about whether the finish line is breadth, autonomy, human-level performance, or replacing valuable work.MORE QUESTIONS
- WHY DOES PETER THIEL FEAR THE ANTICHRIST?
A billionaire who builds surveillance software fears a world government promising protection. Douthat pressed him in 2025. Then came the secret lectures, the pope, the AI safetyists — and an Iran-war break with Trump.
- DID THE FUTURE STOP IN 1970?
Peter Thiel made his fortune on the internet and keeps insisting the internet was not enough. Flying cars, Mars, AI, longevity and the meaning of progress — the other half of his worldview.
- IS THE U.S.-CHINA AI RACE MAKING ANYONE SAFER?
Washington is trying to choke off China's AI — chip bans, blacklists, and now Taiwan making smuggling a crime. Beijing just returned the favor, locking down its own AI exports. Meanwhile the models keep getting cheaper, better, and closer together. A field guide to the argument over whether the race protects anyone, or just speeds everything up.
SHOW YOUR WORK
SOURCES
One phrase. A pile of very different loops.
- SELF-CRITIQUE CAN IMPROVE OUTPUTS WITHOUT RETRAININGSelf-Refine: Iterative Refinement with Self-FeedbackarXiv
- AI CAN DISCOVER ALGORITHMS THAT IMPROVE THE AI COMPUTING STACKAlphaEvolve: A Gemini-powered coding agent for designing advanced algorithmsGoogle DeepMind
- FRONTIER LABS NOW TREAT AI SELF-IMPROVEMENT AS A SAFETY-RELEVANT CAPABILITYOur updated Preparedness FrameworkOpenAI