Hacker News Reader

I feel like any deviation from the syntax LLMs are trained on is not productive.

Sure you can represent the same code in fewer tokens but I doubt it'll get those tokens correct as often.

Yeah, big plus one from me. I recently tried to investigate some sort of alternative encoding to/from “the prompt,” and was swiftly told that was both not possible and would work against me. As you pointed out, the LLMs are trained on language and language itself is often not terse. Trying to skirt that will cause the LLM to calculate the vectors poorly because the relation between the input tokens and its training data doesn’t really exist.

Reply on news.ycombinator.com