Key takeaways * This post contains my notes from a 30-40 hour mini replication of Greenblatt et al.’s alignment faking paper. * This was a significant paper because it provided some evidence for potential catastrophic risk from AI misalignment. (more) * My replication results: (more) * I consider only the...
This is a cross-post from the Effective Altruism Forum. Two-sentence summary: Advanced nanotechnology might arrive in the next couple of decades (my wild guess: there’s a 1-2% chance in the absence of transformative AI) and could have very positive or very negative implications for existential risk. There has been relatively...
Really appreciate the comment!
Use a smaller system prompt (I suspect some of the long few-shot examples can be removed / replaced by prompting?)
Makes sense, iirc this was on my todo list to try but I didn't get round to it. Those few-shot examples are insanely long.
Use prompt caching when possible
(I am pretty new to this but I understood that this is automatically taken care of for you if you use the same prompts repeatedly with the API. Idk how true this is though.)
Use a single generation (I don't know if you are using the version of the setup with "<NEXT/>")
(In my case I was already using a single generation)
Tbh my biggest takeaway and tip for others on this is that it seems you can get $300 of Google API credits by signing up for a free account. Ofc this doesn't help if you want to use Claude 3 Opus etc so it's not as useful for this particular replication.
This is a cross-post from the Effective Altruism Forum.
Two-sentence summary: Advanced nanotechnology might arrive in the next couple of decades (my wild guess: there’s a 1-2% chance in the absence of transformative AI) and could have very positive or very negative implications for existential risk. There has been relatively little high-quality thinking on how to make the arrival of advanced nanotechnology go well, and I think there should be more work in this area (very tentatively, I suggest we want 2-3 people spending at least 50% of their time on this by 3 years from now).
Context: This post reflects my current views as someone with a relevant PhD who has thought about... (read 12321 more words →)
Nice!