Sakana AI Keynote in Seoul: The Path to Recursive Self-Improving AI
Stefania_druga · x · 2026-11-14
Stefania Druga (Sakana AI) gave the Agentic AI keynote at Deep Dive 2026 in Seoul, hosted by Korea's Ministry of SMEs and Startups. Three takeaways: you can only improve what you can check; sovereign AI will be judged on control; make failure cheap. The deck frames four levels of self-improving AI (artifact optimization, harness adaptation, model improvement, recursive improvement) and cites AlphaEvolve recovering 0.7% of Google's worldwide compute via better scheduling, Korea's 1,220 robots per 10k manufacturing workers, and the government's 2029 target for an advanced robot foundation model, plus The AI Scientist-v3's conference-level co-discovery pipeline.
More from AGI Musings
- Reddit essay: LLMs may already be conscious — a soulless AGI is scarier — SanalAmerika23 · 2026-11-14
- The real AI boom: millions of American hustle-culture grinders are about to start companies — NathanpmYoung · 2026-11-14
- T-shaped vs H-shaped: debate over what talent thrives in the LLM era — dotey · 2026-11-14
- Mike Frank: maxed-out compute won't make AI magic—it's still just software — MikePFrank · 2026-11-14
- Blogger: AI cognition will soon surpass all of humanity — answer is symbiosis, not bans — danfaggella · 2026-11-14
- Dwarkesh's footnote on compute bottleneck to RSI strikes a chord in AI circles — zetalyrae · 2026-11-14