Leaked Anthropic doc from July 2021 reportedly foresaw 1T models, test-time compute and RL for tool use
SonglinYang4 · x · 2026-10-14
A post circulating on X shares a purported Anthropic internal document from July 2021 (authenticity unverified), whose foresight is drawing wide attention:
- Discussion of training a 1T-parameter model with RL
- Hardware risks including TPU and Trainium
- Constructing datasets from language, code, diffs and PRs
- Various coding tools
- Scaling laws, reward models, test-time compute, long context and memory
- Even RL for tool use
The poster marvels that none of these ideas would surprise anyone today, yet all appearing in one 2021 document shows a striking cognitive gap of several years. Anthropic has not confirmed the document's authenticity.
Related event: Copyright Lawsuit Exposes Anthropic's 2021 Research Roadmap(4 posts)→
More from Companies & People
- Tübingen relaunches Virtual Humans course with open-source SMPL viewer toolkit — CSProfKGD · 2026-10-14
- Dan Jeffries: AI feed is all secretly paid influencer campaigns, nobody actually uses the products — Dan_Jeffries1 · 2026-10-14
- Missing Sci-Twitter: Scientist launches QED's Insight Feed to rebuild it — OdedRechavi · 2026-10-14
- Jensen Huang pushes back on job fears: 'AI is actually creating jobs' — TinfoilTricorn · 2026-10-14
- A 12-month roadmap to becoming an AI engineer in 2026 without a CS degree — mdancho84 · 2026-10-14
- Mathematician questions why elementary schools test kids on property names — AlexKontorovich · 2026-10-14