3 ms·
> Given a prompt, they use machine learning to try and come up with the most likely set of characters that follow the prompt. Funnily enough this is almost exa
by shawabawa3 4y ago
> Given a prompt, they use machine learning to try and come up with the most likely set of characters that follow the prompt.
Funnily enough this is almost exactly how human children work when they're learning to talk
- raducu 4y agoIt doesn't really matter how we learn the features that we learn; I've heard numerous times that ML models tend to learn architectures/neurons that biological neurons do; I've seen convincing evidence about image recognition back in the day that this is the case. I read that animals dream in the woumb partially to pre-train their nervous system, including vision stuff. But I'm convinced GPT-3/4 did learn a lot of features/generalization/reasoning, even if part of a "text prediction". Sure, it's predicting text... but to do so it learnt some calculus and quantum mechanics -- it's just an example, I don't know the cognitive building blocks to explain how humans/future AI do their reasoning.