I believe LLMs are closer to biological life than to minds.
My completely made-up crude model of a cell is a huge list of short programs (DNA) + the goo everything floats in. Whenever some molecule binds to a section of DNA the "program" is run (which might produce a protein, interact in complex ways with other proteins, trigger more "programs", etc.). The DNA access is inherently parallel, with limits having to be implemented explicitly. In a system like that, all high-level behavior is emergent. There is no part of DNA that corresponds to any abstract drive, only evolution lining up conditionals in clever ways.
This makes small improvements easy. You can add something to improve this or that a tiny amount, and in a big enough environment there will always be things that correlate with success. Just like you can always add patches to a legacy codebase. And, like in programming, if you have to deal with spaghetti code you don't understand, it is much better if it consists of lots of tiny functions rather than a few big ones.
I think LLMs are sort of like this. More importantly, though - I think minds are not. Minds build structure. They are fragile and sensitive to small changes. While a brain can recover from a lot of damage (e.g., a stroke), it can only do so over time. Maybe a better analogy here is a market. Neurons allocate metabolic "reward" to dendrites (partly) based on the value of the information. A slightly late copy of the same signal will get pruned, just like a 20% worse offering will get no buyers on the free market, not 20% fewer buyers. This should lead to a constant increase in understanding over time, even absent any new data to learn from. I don't think anything like this is true of LLMs?
I believe LLMs are closer to biological life than to minds.
My
completely made-upcrude model of a cell is a huge list of short programs (DNA) + the goo everything floats in. Whenever some molecule binds to a section of DNA the "program" is run (which might produce a protein, interact in complex ways with other proteins, trigger more "programs", etc.). The DNA access is inherently parallel, with limits having to be implemented explicitly. In a system like that, all high-level behavior is emergent. There is no part of DNA that corresponds to any abstract drive, only evolution lining up conditionals in clever ways.This makes small improvements easy. You can add something to improve this or that a tiny amount, and in a big enough environment there will always be things that correlate with success. Just like you can always add patches to a legacy codebase. And, like in programming, if you have to deal with spaghetti code you don't understand, it is much better if it consists of lots of tiny functions rather than a few big ones.
I think LLMs are sort of like this. More importantly, though - I think minds are not. Minds build structure. They are fragile and sensitive to small changes. While a brain can recover from a lot of damage (e.g., a stroke), it can only do so over time. Maybe a better analogy here is a market. Neurons allocate metabolic "reward" to dendrites (partly) based on the value of the information. A slightly late copy of the same signal will get pruned, just like a 20% worse offering will get no buyers on the free market, not 20% fewer buyers. This should lead to a constant increase in understanding over time, even absent any new data to learn from. I don't think anything like this is true of LLMs?