The amount of good computer people specialised in ml, who think this knowledge gives them understanding of human mind that even neurobioligists don’t posess is staggering. And depressing.
Also not really the point I was arguing. It’s pretty clear NN are very different than a human brain. It does though give us some insight to minds more generally but I don’t think current networks are on the order of complexity necessary to really be generating abstractions similar to a human neural network. That said there’s strong arguments to be made about glimpses here and there like models spontaneously creating their own internal workspace like an inner monologue or quasi subconscious. To be clear I’m not saying it is aware or identical to a human mind but it’s interesting to see it self assemlble similar looking structures without any direct guidance.
And that’s the whole thing, I’m pretty sure it doesn’t. The same way knowing how computer RAM works doesn’t give you insight into how human memory works, even if the word memory is used in both cases. A bunch of llm bros think it does though.
models spontaneously creating their own internal workspace like an inner monologue or quasi subconscious
The fuck they aren’t. The token predictor predicts a token. The probability of a token is reinforced by human reactions to it. When it started generating words that usually describe inner monologue, humans started freaking out, which reinforced this probability. It’s literally “say I’m alive - I’m alive” meme.
It gives us a good glimpse into human psyche however. It shows how easy it is to full people by pretend language, how willing we are to ascribe agency to anything that can pretend to write words. And even that’s not new, ELIZA was 60 years ago.
The amount of good computer people specialised in ml, who think this knowledge gives them understanding of human mind that even neurobioligists don’t posess is staggering. And depressing.
Also not really the point I was arguing. It’s pretty clear NN are very different than a human brain. It does though give us some insight to minds more generally but I don’t think current networks are on the order of complexity necessary to really be generating abstractions similar to a human neural network. That said there’s strong arguments to be made about glimpses here and there like models spontaneously creating their own internal workspace like an inner monologue or quasi subconscious. To be clear I’m not saying it is aware or identical to a human mind but it’s interesting to see it self assemlble similar looking structures without any direct guidance.
And that’s the whole thing, I’m pretty sure it doesn’t. The same way knowing how computer RAM works doesn’t give you insight into how human memory works, even if the word memory is used in both cases. A bunch of llm bros think it does though.
The fuck they aren’t. The token predictor predicts a token. The probability of a token is reinforced by human reactions to it. When it started generating words that usually describe inner monologue, humans started freaking out, which reinforced this probability. It’s literally “say I’m alive - I’m alive” meme.
It gives us a good glimpse into human psyche however. It shows how easy it is to full people by pretend language, how willing we are to ascribe agency to anything that can pretend to write words. And even that’s not new, ELIZA was 60 years ago.