Conflicting RSI evidence: NVIDIA says AI-assisted CUDA optimization went 5x, then 2x in a month
soumitrashukla9 · x · 2026-10-24
Stanford's Rishi Bommasani highlights conflicting evidence on recursive self-improvement (RSI) progress, most of it lab-sourced: OpenAI/Google evals show 15-minute time horizons for automated R&D tasks (slower than METR software horizons → slower RSI); Mirendil's results suggest pretraining at scale can be run by a single researcher (faster RSI); and per SemiAnalysis, NVIDIA engineers said CUDA optimization went 5x in a year with human experts (Oct 2025–Aug 2026), then 2x in one month with AI assistance, then another 1.5x with AI orchestration + MegaKernels. His team is building independent RSI evaluations and hiring.
More from AGI Musings
- 70k credentials leaked in Fortinet attacks; researcher says an LLM code review would cost less than one appliance — cyb3rops · 2026-10-24
- Companies say AI will handle junior work—then demand 8 years of experience — tech__unicorn · 2026-10-24
- 30-month study of 26,811 students: AI homework use raised scores 18% but exam scores fell 20% — lmoroney · 2026-10-24
- Rob Leclerc: AI safety overreactions undermine real concerns, let markets decide — robleclerc · 2026-10-24
- Gary Marcus echoes 'AI's Wuhan moment' alarm, calls for temporary ban on AI agents — GaryMarcus · 2026-10-24
- Santa Fe Institute's Eric Smith: why we shouldn't assume AGI is 'alive' — danfaggella · 2026-10-24