claudio

• claudioMar 23, 2025, 10:21 AM

  1 point

  0

  in reply to: Yair Halberstadt’s comment on: How I force LLMs to gen­er­ate cor­rect code

  At this stage, the tool doesn’t make assumptions about the context size, also because you can use any LLM. I rarely pass the whole project to Unvibe: I just make it iterate on a module/​folder of the project: basically anything that is touched by the unit-tests. This is also what you’d do as a developer.
  But an idea for larger “folders” could be to put in the prompt just the definitions of functions/​classes, without the actual code. In general, I try to work at the margin, assuming LLMs will get more accurate and the context will keep growing. “Write code as if the LLM that comes out in 2 years exists already and ignore the details” has been a good heuristic.

• claudioMar 22, 2025, 11:23 AM

  2 points

  0

  in reply to: Rasool’s comment on: How I force LLMs to gen­er­ate cor­rect code

  Yes! The reason being that when you do your first post it takes ~2 days to get it actually published: by virtue of the LW sorting algorithm, nobody saw it because it’s “old” by the time it’s out. I’m happy I got some feedback now

• claudioMar 22, 2025, 8:06 AM

  3 points

  0

  in reply to: Gordon Seidoh Worley’s comment on: How I force LLMs to gen­er­ate cor­rect code

  That is very important, in fact Unvibe passes the source code of tests to the LLM. But in the latent space of the LLM there are many possibile implementations that the LLM doesn’t know how to distinguish. Having the unit tests score as ground truth helps a lot