Hacker News Reader

Comment by davidhs | Hacker News Reader

Very impressive. Its programming skills are still kind of crappy and I seriously doubt its reasoning capacity. It feels like it can deep fake text prediction really well, but in essence there's still something wrong it it.

View on HN

Not sure I agree. The way you interact with LLMs in context of programming has to be tuned to the LLM. Information has to be cut down to show just what is important and context windows are a bit of a red herring right now, as LLMs tend to derail its solution from the target completely, the more information is at play. For some this is more trouble than it's worth.

In certain languages it's almost magical in terms of showing you possible solutions and being a rubber ducky to bounce your own logic off of. (Python, JavaScript, TypeScript) In certain languages it is hopelessly useless beyond commenting on basic syntax. (GLSL)

I tried GPT-4o earlier where I was iteratively asking it to write and in improve a simple JavaScript web app that renders graphs of equations and it had a lot of trouble with substituting slow and ineffecient code with faster code, and at some later point where I asked it to implement a new feature how the graph coloring is rendered it started derailing, introducing bugs and very convoluted code.

Yes, at some point ChatGPT "reaches the limit of new information", but is unable to tell you that it has reached the limit. Instead of saying "Hey, I can't give you any more relevant information", it simply continues to cycle to previously suggested things, starts suggesting unrelated solutions or details. Especially true with refactoring! When it has reached the limit of refactoring it starts cycling through suggestions that change code without making it better. Kinda like having a personal intern unable to say "no" or "I can't".

That is part of working with LLMs and what I meant before with "for some, more trouble than it's worth".

Reply on news.ycombinator.com