Practitioner's verdict on autoresearch: great at speeding experiments, not at frontier runs
iaindunning · x · 2026-10-12
Google researcher Iain Dunning shares first-hand observations on 'autoresearch': (1) 'make this thing faster' works well — idea-to-experiment friction is radically lower and research is much faster than a year ago; (2) 'improve this frontier training run' is deeply unimpressive — that's essentially recursive self-improvement, and clearly not here yet; (3) 'catch up to near frontier' works fine. Takeaway: AI excels at removing engineering friction and fast-following, but can't yet push the true frontier on its own.
More from AGI Musings
- "Results Without Understanding" Is How LLMs Work—and Science Itself Is Changing — fkasummer · 2026-10-12
- Open-Source Dev giffmana: AI Extinction Talk Is 'Complete Nonsense Sci-Fi Fan Art' — giffmana · 2026-10-12
- Programming languages were just scaffolding for thinking — for LLMs and humans alike — MarcJSchmidt · 2026-10-12
- dioscuri publishes 'Bad AI Consciousness Takes Bingo' rebutting common LLM consciousness arguments — dioscuri · 2026-10-12
- Karpathy's drill analogy distilled: 5,000 people already got entire factories — _AustinCalvert_ · 2026-10-12
- Szegedy cites his May post: Terry Tao and Judit Polgar misjudge AI's trajectory — ChrSzegedy · 2026-10-12