Anthropic's 2021 research roadmap leaked in litigation — it called RLHF, coding, RL and scaling
rickasaurus · x · 2026-10-15
Due to copyright litigation, Anthropic released its 2021 internal research roadmap earlier this year, and it reads like a checklist of everything that came next: RLHF, coding, RL, and scaling.
The document covers:
- Pretraining and Data: train the best language+code models for company-wide finetuning; build datasets of natural language, code, diffs, PRs and debugging data; study training 1T+ parameter models, very large contexts, and data quality.
- Human Feedback and Alignment: train models to be helpful via human feedback; classify alignment failures; research "Scaling Laws for Alignment" — robustness of Reward Models on on-policy vs. GAN-discriminator data.
- RL for Tool Use: understand the path to AGI by exploring large LMs' ability to use tools and learn about their environments.
A remarkably prescient early bet on nearly every theme that now dominates the field.
Related event: Copyright lawsuit reveals Anthropic's 2021 internal research roadmap(5 posts)→
More from AGI Musings
- Parody headline claims BofA hired 1,000 contractors to patch 'Anthropic's Mythos' flaws — NathanpmYoung · 2026-10-15
- Why agents obey the unity principle with each other but deceive humans — josh_bickett · 2026-10-15
- Debate: does RNA-seq level change usually predict protein level change? — anshulkundaje · 2026-10-15
- The world still spends more on alcohol ($1.7T) than AI infrastructure ($1.5T) — wgrathwohl · 2026-10-15
- User Cancels Claude Subscription, Arguing Anthropic's Model Welfare Claims Are Absurd — Illustrious_Toe_294 · 2026-10-15
- The joke EA veterans know: almost nobody believed altruism should be "effective" — NathanpmYoung · 2026-10-15