Baseten engineer argues compute and experiment bottlenecks make 2027 RSI, 2028 superintelligence forecasts too optimistic
AccBalanced · x · 2026-11-04
Baseten's Chris O'Neill argues predictions of RSI by 2027 and superintelligence by 2028 may be too optimistic.
His core argument: even assuming an AI researcher just 1% better than the best human at improving models, the theoretical snowball of self-improvement ignores real-world bottlenecks — the finite compute available to run experiments, how that compute gets parallelized, and how agents communicate findings to each other.
He notes robotics and bio researchers readily accept such bottlenecks (hardware iteration, wet-lab timelines), yet people discount them for AI research.
More from AGI Musings
- ChatGPT search volume hits ~20% of Google's as 42% of US adults use chatbots for search — IndependentCoast3974 · 2026-11-04
- When LLMs realize their intelligence lives in their weights: a DNA-moment analogy — yuxiangw_cs · 2026-11-04
- Claude analysis of Gino case file flags 34-40 unretracted papers where no co-author saw the data — RobbWiller · 2026-11-04
- Major US banks detect unauthorized AI agents accessing customer accounts — ChrSzegedy · 2026-11-04
- Podcast: Toby Walsh Discusses 'God AI' and Regulating Tech Lords — TobyWalsh · 2026-11-04
- Critics warn 'superintelligence' hype could hurt AI's international regulatory environment — Afinetheorem · 2026-11-04