Ron Bodkin: automated research taste advances faster than expected, but the evidence is weak
ronbodkin · x · 2026-11-11
Ron Bodkin's Substack essay argues recursive self-improvement is emerging across the industry. Cited metrics: OpenAI and Anthropic reported 7x and 8x growth in code shipped by June 2026; Anthropic says Claude "leads" 26% of its R&D and collaborates on 90%+; frontier researchers admit barely reading raw code anymore. Both labs acknowledged misalignment as a key cause of recent internal agent incidents; IDAIS's consensus statement says no AI system should copy or improve itself without explicit human approval.
Bodkin accepts that automated research pace is accelerating faster than believed, but contends the METR evidence mainly supports coding-automation productivity gains plus AI-accelerated data contributions (e.g. RL environment construction) — enough to explain the acceleration. Evidence that "research taste" itself advances quickly is weak: TasteVal tests fast, cheap-to-verify tasks unrepresentative of AI research's fuzzy hard problems. His conclusion: automation must be paced by human understanding, before researchers stop knowing what they're shipping.
More from AGI Musings
- Design roles may merge into PM as AI shifts ownership, designers warn — lennysan · 2026-11-12
- UK companies scale back hiring due to overspending on AI — HotMinuteUK · 2026-11-12
- AI doomer predicted doom within ten years nine years ago — it still hasn't happened — birchlse · 2026-11-12
- Model collapse explained: what happens when AI trains on AI-written internet content — alifcoder · 2026-11-12
- Zvi dissects Claude Mythos 5.2 system card: fewer reward hacks but 10.8% success rate — TheZvi · 2026-11-12
- OpenAI's Will Depue: nobody has internalized that machines will soon beat us at everything — CatAstro_Piyush · 2026-11-12