[alt0_]

I think one of the techniques underexplored in all the hype is guiding the evaluation process depending on the context. I.e. if you're generating code, it has to satisfy the parser for the given language. If the token is unsatisfactory, throw it out and try another one. Thought chains could be generated in a similar way (you can do so with special tokens, see "Recursion of Thought").

But yeah overall GenAI tends to remain hype-over-substance.

[iLoveOncall]

The main blocker for this is that LLMs are slow. Imagine waiting 3 seconds for your output in a pretty happy case, only for it to be invalid and have to wait an extra 3 seconds, with again non negligible chances of being wrong.

We envisioned doing this for an SQL query generator at work but with our constraints a single query already takes 15 seconds.