jsm on Nostr: Some solid LLM predictions. Not completely convinced though on how bearish the ...
Some solid LLM predictions. Not completely convinced though on how bearish the analysis of reasoning models is for two reasons.
1. There are a lot of fields where problem solving verification in training is an option.
2. I think it probably is possible to generalize the abstract text patterns of reasoning and analysis so the model can use them in domains it hasn’t been specifically trained in to improve performance.
https://www.lesswrong.com/posts/oKAFFvaouKKEhbBPm/a-bear-case-my-predictions-regarding-ai-progress
1. There are a lot of fields where problem solving verification in training is an option.
2. I think it probably is possible to generalize the abstract text patterns of reasoning and analysis so the model can use them in domains it hasn’t been specifically trained in to improve performance.
https://www.lesswrong.com/posts/oKAFFvaouKKEhbBPm/a-bear-case-my-predictions-regarding-ai-progress