AI Researchers Debate RSI: Compute Is the Binding Constraint on Recursive Self-Improvement
maksym_andr · x · 2026-11-15
Maksym Andr builds on Brian Huang's analysis of recursive self-improvement (RSI). The original thread argues critics miss two factors: how compute-constrained lab research really is—researcher compute pools already run experiments at 100% uptime, forcing hard prioritization among strong ideas—and how much headroom remains for "superhuman" research taste and execution.
Maksym agrees and generalizes: AI progress moves from one bottleneck (e.g., writing code) to the next (e.g., data), and RSI will certainly be compute-bottlenecked. He sees RSI as just one of many factors sustaining the historical exponential, but even extending the trend for years implies massive transformation. Includes an Our World in Data chart of world GDP over two millennia on a log scale.
Related event: AI debate: compute may be the bottleneck for recursive self-improvement(2 posts)→
More from AGI Musings
- 'Study dialectics' thread hints at AI synthesis wave coming in 2027 — ZeroStateReflex · 2026-11-15
- NVDA beat aggressive FY2026 revenue estimates by 4x, deflating S-curve takes — abhiadesai · 2026-11-15
- tszzl: Activism-shaped problems usually turn out to be technology problems — tszzl · 2026-11-15
- AI safety campaign calls for global ban on superintelligence and recursive self-improvement — marcvanderchijs · 2026-11-15
- Researcher argues boosting pretraining, not RL, is the safer path to AI alignment — maksym_andr · 2026-11-15
- Jacob Coxon, Anthropic's viral resignee, says he might return to an AI lab — rohanpaul_ai · 2026-11-15