Anthropic visitor: models building models is near as Claude now leads 26% of internal model R&D
robseamans · x · 2026-10-15
After spending days at Anthropic, Arpitrage distills two takeaways backed by internal charts: per-employee code output hit 8x the pre-2025 baseline in Q2 2026, and Claude now leads 26% of Anthropic's model R&D tasks (up from 1% in March), evidence RSI and rapid takeoff may be near. He argues lab arms races crowd out safety investment and proposes frontier labs open-source models to blunt reliance on Chinese open weights. The post also flags Senate news: with Trump opposing regulation, labs including Google are seeking antitrust safe harbor to legally coordinate on pacing frontier AI (Schiff-Banks bill).
More from AGI Musings
- Redwood's Ryan Greenblatt pulls timelines forward: automated coder by Feb 2028, full AI R&D automation by late 2028 — ben_j_todd · 2026-10-15
- Ben Todd: dramatic AI recursive self-improvement hinges on how fast research returns diminish — ben_j_todd · 2026-10-15
- Token prices are subsidized like early Uber, and GPT-6 Astra insisted it was right on physics evals — UltraRareAF · 2026-10-15
- A decade of degrees rendered useless: one person's take on skills devaluing — Paimaamu · 2026-10-15
- New AI labs will be hit by the Bitter Lesson harder than anyone, dev says — eigenron · 2026-10-15
- levelsio: Normal people already use AI in advanced ways while tech lives in a bubble — joonasvirtanen · 2026-10-15