Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

That's just a matter of fine tuning


That "just" is doing some heavy lifting! GPT-4 is just a few matrix multiplications, how bad can their moat really be?


Not sure what the snark here is for: It would be trivial to produce a dataset where the model asked you questions then fine-tune on that.

People already do it with chain-of-thought and you could get away with a few dozen examples if you wanted to try this.


Out of boredom I decided to prove this too: I asked ChatGPT and Claude for ~200 samples in total.

Just uploaded the examples as-is to OpenAI, selected 3.5 as the model to fine-tune and about 20 minutes later I had my model.

Works fine, asks good questions, can ask more than 1 follow up question if needed, and actually changes its answers based on the clarifying questions.

https://imgur.com/a/SsXunVN


I'd bet a synthetic data set could do the job effectively.


Do you have an example model I could try that does this?


Try Pi by inflection. It asks a lot of questions.


I tried it, it just asked me how my day was going. I don't think this is doing exactly what I have in mind. But its a step in that direction




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: