Anthropic's 207k-conversation study: AI could speed up math/CS research 15x, wet-lab fields only 3-4x
AI寒武纪 · wechat · 2026-11-06
Anthropic analyzed 207,049 research-related Claude conversations (Aug–Sep 2026, spanning Claude chat, Claude Code, and Claude Science) and surveyed 528 active scientists. Conclusion: full AI automation could boost output in pure cognitive fields (CS, math, statistics) 15–16x, while experiment-heavy fields stay capped at 3–4x by physical bottlenecks.
Key findings: computer scientists are 4.3% of researchers but account for 32.7% of Claude Code sessions; scientists mainly use AI for data analysis (33.7% of chat, 48.1% of Code) and paper writing (32.0%), while hypothesis generation (4.2%) and data collection (11.5%) — the steps scientists deem most important — see minimal AI use. RL-trainability explains the capability spread (correlation 0.78): verifiable, simulable tasks score high; physical or regulated tasks near zero.
Production-function scenarios: full adoption of current AI takes over 27–52% of scientists' time (CS output 2x); 75% automation of non-physical steps → CS 8x, stats 7x; 100% → 15–16x in cognitive fields, with compute cost the new ceiling. Wet-lab work consumes 40% of labor/budget in experimental fields. 40% of surveyed scientists expect fully autonomous research projects by 2030, 64% by 2035; half think AI's first recognized major discovery comes after 2035. Risks flagged: scarce good hypotheses, paper inflation, AI-assisted peer review, bio-safety concerns.
More from AGI Musings
- Gary Marcus: AI isn't over — what may end is insane valuations and data center spending — GaryMarcus · 2026-11-06
- IIT Goa professor redesigns exam: students write their own questions, then answer them — kingslyj · 2026-11-06
- "We had the Transformer paper before LLMs — nothing even close since," dev argues — dosco · 2026-11-06
- Toby Ord: jump-plus-steeper-slope mixed hypothesis beats a steeper slope alone — tobyordoxford · 2026-11-06
- Readback Labs proposes mandatory audits requiring AI execs to prove they understand AI R&D — ronbodkin · 2026-11-06
- Oxford's Toby Ord: Capability curves look like a step-change, not a slope shift — tobyordoxford · 2026-11-06