5 ms·
When I try to get my gamedev code written by gpt it seems like its more like 50% right 50% wrong but revising the wrong takes a long time.
by Madmallard 4y ago
When I try to get my gamedev code written by gpt it seems like its more like 50% right 50% wrong but revising the wrong takes a long time.
- EGreg 4y agoOften it is far longer than writing it all yourself. What I see as the advantage is taking doc comments or documentation and fine-tuning the engine on it. Not embedding more words as numbers. Not adding it to a prompt. Literally fine tuning the model. Does Alpaca or any LLaMA version let you do it? Cause Chat-GPT doesn’t. And anyway I wouldn’t want to give our whole code over to some third company to re-use the way an artist’s work got jacked because it was publicly posted. I want to self-host the fine-tuned model. THAT would be the killer feature, for businesses. The Web attracted businesses, not people. People moved over later. Does Alpaca or any LLaMA version let you do it? Giving it (with its existing weights) a gigabyte of text files and letting it fine tune itself on that?
- code_runner 4y agoGiven a large enough context window the need for fine tuning diminishes a little bit.... The cost of fine tuning these massive models is crazy right now. Fine tuning a smaller model can get you better performance for specific tasks in many cases. A lot of models are out there and totally free to adapt as needed via hugging face etc.
- EGreg 4y agoHugging face?
- dragonwriter 4y agoHugging Face is kind of the github of AI/ML: https://huggingface.co/ https://huggingface.co/