The general message here seems to be that inference-time brute-forcing works as long as you have a good search and evaluation strategy. We’ve seemingly hit a ceiling on the base LLM forward-pass capability so any further wins are going to be in how we juggle multiple inferences to solve the problem space. It feels like a scripting problem now. Which is cool! A fun space for hacker-engineers. Also:
> My mental model for LLMs is that they work as a repository of vector programs. When prompted, they will fetch the program that your prompt maps to and "execute" it on the input at hand. LLMs are a way to store and operationalize millions of useful mini-programs via passive exposure to human-generated content.
I found this such an intriguing way of thinking about it.
> We’ve seemingly hit a ceiling on the base LLM forward-pass capability so any further wins are going to be in how we juggle multiple inferences to solve the problem space
Not so sure - but we might need to figure out the inference/search/evaluation strategy in order to provide the data we need to distill to the single forward-pass data fitting.
> My mental model for LLMs is that they work as a repository of vector programs. When prompted, they will fetch the program that your prompt maps to and "execute" it on the input at hand. LLMs are a way to store and operationalize millions of useful mini-programs via passive exposure to human-generated content.
I found this such an intriguing way of thinking about it.