Hacker News Reader

Comment by cal85 | Hacker News Reader

Can someone help me understand: is ChatGPT essentially just using the GPT-3 text-davinci-003 under the hood, with a bunch of prompt-prefixing (similar to what is going on in this guy's Siri Shortcut)? Or is it using a significantly more powerful model?

View on HN

ChatGPT is GPT-3.5 plus some human feedback reinforcement learning[1] to steer it away from the things that tanked Tay. Meaning they had a bunch humans test the thing out, rate responses, and incorporate that into future training.

    [1]: https://en.wikipedia.org/wiki/ChatGPT#Training

From what I understand, they refer to it as GPT 3.5. So, that plus the prompt prefixing as you mentioned.

It was either text-davinci-002-render or text-davinci-003-render

AFAIK that's exactly what it is.

If so, doesn’t that make it trivial for anyone to make a ChatGPT-like app using the GPT-3 API? If so, I would’ve expected several ChatGPT-like tools coming out in the days following GPT-3’s release. I don’t understand why it took so many months - and then, why it was OpenAI themselves that finally released the first interactive thing on top of it. It’s puzzling.

Reply on news.ycombinator.com