They’re useful for tasks that are easy to verify but hard to generate, and code is one meaningful example. It seems trite to say after so many others, but I bu

They’re useful for tasks that are easy to verify but hard to generate, and code is one meaningful example. It seems trite to say after so many others, but I built a web app in one day by pair coding with GPT-4, which I am sure would have taken weeks, mostly learning the quirks of the various frameworks involved by sifting through the noise of the web. The LLM wasn’t correct 100% of the time, but I could iterate with it when there was a problem and debug. It probably helps that I was using hugely popular libraries, but I think this means there’s a ton of promise in fine-tuning these models on internal codebases/docs.