4 ms·
Very much yes to your first sentence, but I don’t think “just plug something into ChatGPT” is a fair summary of your (very interesting) post, so you’re fine wit
by codeflo 3y ago
Very much yes to your first sentence, but I don’t think “just plug something into ChatGPT” is a fair summary of your (very interesting) post, so you’re fine with me.
I’ve seen several statements recently, including one in a talk by a Microsoft researcher working with OpenAI, that claim that all the RLHF safety training introduces human biases and makes the models worse at abstract reasoning. I really wonder why a relatively small amount of human feedback has that effect, while training on this gigantic human-written corpus doesn’t seem to introduce the same biases. That hints at something interesting going on.
- mlajtos 3y agoWe are gonna need to prefix all prompts with "From now on, act without human biases."
- panarky 3y agoTo human ears, the problem of 3 hens, 3 eggs, 3 days is a simplified and more obvious problem than the original: "If a hen and a half can lay an egg and a half in a day and a half ..." But it's the same problem in the same form, and GPT4 nails it without extra goofy prompting like "don't replicate human cognitive biases". A more interesting question would cross logic with problem solving. "If a hen and a half can lay an egg and a half in a day and a half, how many hens are needed to fill Yankee Stadium with eggs in a year?" But GPT4 doesn't have very good intuition on how to estimate the volume of Yankee Stadium: Yankee Stadium has a seating capacity of approximately 50,000 spectators. Let's assume each spectator occupies a space of 2 cubic meters (which is a generous estimate, including seating and legroom). Total volume (approx.) = 50,000 seats × 2 m³/seat = 100,000 m³