3 ms·
Yes, DaVinci is a base model that hasn't been put through RLHF which is where that sort of bias gets introduced. Unfortunately there's no base model for GPT-4.
by mike_hearn 3y ago
Yes, DaVinci is a base model that hasn't been put through RLHF which is where that sort of bias gets introduced. Unfortunately there's no base model for GPT-4.
If you force things by demanding a list of 10 books then I can imagine that would hack around it, but then if there aren't actually ten such books you're going to get hallucinations again. For example, the first item on the GPT-4 list doesn't exist (the author does).
- Spivak 3y agoFair, let me try gpt4 with your original prompt and see what I get. The book you're thinking of is "Dissolving Illusions: Disease, Vaccines, and The Forgotten History" by Dr. Suzanne Humphries and Roman Bystrianyk. The book discusses the history of vaccines, including smallpox, and challenges the popular belief that vaccines were solely responsible for their eradication. The gpt4 api gives zero fucks. It really does seem like it's just whatever logic they layered on to ChatGPT.
- mike_hearn 3y agoYou're right, I can also reproduce that behavior with the API. Fascinating. I really thought they were the same model but ChatGPT4 is clearly a lot more RLHFd compared to the API. Not only is the API honest but it's also way less wordy. Practically every answer from ChatGPT in the sample conversation had to be truncated because it includes an essay about the scientific consensus.
- lucubratory 3y agoThe API model is just as RLHF'd as they disclosed in their 'paper'. The difference you're seeing is very likely to just be the ChatGPT pre-prompt telling it not to engage in harmful activities.