Live data from Hacker News

Show HN: Tiny Code Improver

github.com

11–12 of 12 posts

Re: Show HN: Tiny Code Improver

#11
post #4

Claude has 100k context for around 2$ per million tokens. With GPT4's 4-8k token limit, anything but very small projects in their early phase can benefit from this. Also GPT4 would be far too cost prohibitive

Good idea about Claude, never tried it myself. But I've heard that GPT-4 is superior at the moment. Is it true? Also thought about GPT-4 prices, but from my experience: If I use it intensively for a 3-4 hours it costs about $5. This is much cheaper than anyones hour rate.

Re: Show HN: Tiny Code Improver

#12
post #5

What is the perf of GPT4 vs GPT3.5 in the 'reasoning', 'reflection', 'critisism' and 'resolver' tasks like in your project? I see that you have commented out gpt3.5 and replaced with GPT4 in config yaml. Was GPT3.5 perf too bad? I don't think many people have GPT4 API access. If this requires atleast GPT4 to be effective, it might take a while before anyone else in the community can take it up.

GPT3.5 works not bad for generating, but it is not very good at self-reflection and self-critisism. Or maybe my prompts are not so good.
Post reply on HN