Do language models need sleep? Research from Carnegie Mellom #ai #LearnOnTikTok #chatgpt #news #EduTok

I'm gonna ask you a weird question, do language models need sleep? Now I know that sounds crazy, but Carnegie Mellon wrote a whole paper on this topic. See every AI model has a memory problem that I don't think no one really talks about as much. For the context window, you can think of it like reading a book, but every time you flip to a new chapter, your brain just forgets the last one. Which is what's happening with these AI models, they have a scratch pad of a memory and when that gets filled up, the old stuff gets wiped. Now before it gets wiped, the model gets one single shot to compress what it read into long-term memory. It's one pass, whatever it understood in that one look, that's what gets saved. And that's the problem because one pass isn't enough for hard problems. So Carnegie said what if instead of just one pass, we loop it. When the memory fills up, the model stops taking anything new and goes back over what it just read. Again and again, each loop, it's refining what it understood. Getting it better, this is what they mean by the model sleeping. Now this isn't a random analogy they just grabbed. This is maybe how our brains work at night. For example, our hippocampus replays our day and rewires it into long-term memory. And they tested this on real math problems. These are thousands of tokens long. And the result showed with no sleep, the model was basically lost, giving terrible hallucinated answers, but with four sleep loops, accuracy jumped 40% on the same problems. Now these are still small models, so I don't know if this holds up front to your scale. The paper doesn't really go into those details. But the fact that the fix for AI's memory problem looks this much like what evolution figured out for us, that's a little freaky. Follow and I'll keep you posted.