Google's recursive self-improvement push: training kernel 23% faster, TPU gains
imjustnewatai · x · 2026-09-12
The author traces Google's public steps toward recursive self-improvement: the AI co-scientist (Feb 2025) generates, critiques and refines research proposals, described by Google as "recursive self-improvement with increased compute." AlphaEvolve (May 2025) used Gemini to optimize algorithms used to train Gemini — a key training kernel got 23% faster, cutting total training time by 1%, and a chip-design improvement landed in an upcoming TPU. The author's guessed route: automate more AI research, improve training and hardware, build a stronger model, then feed it back. The metric to watch: whether each generation gets better at building the next.
More from AGI Musings
- Biologist: Blocking AI to protect research jobs would be morally indefensible — QuintinPope5 · 2026-09-12
- From capable AI to economically trustworthy AI: the next frontier for agents — fnlog0 · 2026-09-12
- Ex-Anthropic/OpenAI researcher: AI may replace humans in much research within a year — rohanpaul_ai · 2026-09-12
- AGI will arrive like an economic depression, not an overnight singularity — paigeinsf · 2026-09-12
- X debate: could near-future AI models invent the math to prove P vs NP? — airkatakana · 2026-09-12
- Kenneth Stanley: The Fields Medal Letter Shows 'The Path Matters More Than the Destination' — typewriters · 2026-09-12