awesome-ai-constitutions repo collects AI constitutions & model specs with generalization research
dhadfieldmenell · x · 2026-11-01
A new GitHub repo, awesome-ai-constitutions, compiles AI constitutions and model spec documents in chronological order, with quotes from studies on which principles generalize.
- An AI constitution is a human-readable document that shapes a model's values during training via reflection/judging, tracing back to Anthropic's Bai et al. 2022 paper; OpenAI calls the same kind of document a model spec.
- Unlike runtime system prompts, constitutions act at training time; character docs are often written in the model's own voice ('I am...') while constitutions use directives ('Claude should...').
- The repo is early-stage (2 stars, 13 commits) but serves as a useful index for alignment and model behavior spec research.
More from Safety
- Model-refusal failover chains quietly lower your real safety floor — petrusenko_max · 2026-11-01
- Building an agent-sec library: provenance tags vs. indirect prompt injection and the Lethal Trifecta — sassyalfred · 2026-11-01
- Researcher says Anthropic banned him for "training a competitor" despite never distilling Claude — stanfordnlp · 2026-11-01
- Proton engineer confirms quotes in critical story were real but stripped of context — uwukko · 2026-11-01
- Scoop: Moonshot and DeepSeek reportedly distilled OpenAI's Astra after Anthropic crackdown — peterwildeford · 2026-11-01
- Grok API to enable text watermarking by default across all models starting Nov 7, 2026 — XFreeze · 2026-11-01