Quick update, this week OpenAI Announced the fine tuning API:
If you don’t know - fine tuning is about “training” an actual model like GPT-3 to respond in specific, more predictable ways. Right now, GPT-3 has a very well known controllability problem - it may answer questions randomly, off-brand, incorrectly, rant endlessly, repeat itself like a crazy person, or even say something grossly offensive. Which means you just can’t fully rely on it in many public-facing situations.
Fine tuning allows us to give GPT-3 many formal training examples which can lead to better outputs … this is a lot more than what we can currently do with a maximum of 2-3 examples in our prompt design. Fine tuning, in theory, should trigger GPT-3 to adjust the actual weights in its model to better fit the fine tuning training examples we give it and respond accordingly.
Advances in fine tuning and reinforcement are really important to make GPT-3 more reliable, especially in a commercial setting. With better controllability through fine tuning, there’s no reason why GPT-3 couldn’t be deployed in very serious ways in many public commercial organizational applications. With fine tuning, I see an opportunity where the GPT-3 developer community can start making more money, selling GPT-3 products, and taking on actual clients … now that the technology can act in the ways we want.
I’ll have more on this fine tuning stuff to come, either on the podcast or as 1-2 videos. To be honest, I’ve been very busy this week professionally but also with personal life stuff I have going on. So, please stay tuned and thank you sincerely for your continued support.
I also released a video yesterday on how you (yes you!) can already start playing around with free models which are similar, lighter versions of DALL-E. In the video, I provide some context, a brief guide, and also share cool art people I know are already making:
I hope you enjoy the video, and if you make anything cool feel free to @ me on Twitter and I’ll retweet it!