search

StampyAI / stampy-chat

Conversational chatbot to answer questions about AI Safety & Alignment based on information retrieved from the Alignment Research Dataset
https://chat.stampy.ai
MIT License
13 stars 7 forks source link

Inject glossary into generated answers #73

Closed FraserLee closed 1 year ago

FraserLee commented 1 year ago

Closes #69

FraserLee commented 1 year ago

One thing of note: right now the glossary is tiny (only 23 terms, many duplicates). I'm not certain where existing efforts are at to grow it, but it feels like we could add a lot of value for relatively little effort by expanding this. We could even try things like auto-generating glossary entries for every existing stampy article where it could make sense using gpt-4, then manually reviewing and correcting them. Minuscule effort, potentially a pretty good end result.

Current keywords: agi, ai alignment, artificial general intelligence, chain of thought prompting, chain-of-thought, embedded agency, embedded agent, existential risk, goal misgeneralization, goodhart's law, instrumental convergence, instrumentally convergent goals, interpretability, large language model, llm, ood, optimizer, oracle, orthogonality thesis, out of distribution, terminal goal, terminal goals, "the big g,"