4 ms·

I think in some future we may be able to create something sentient, but it’s not this. AI is just a chain of matrix multiplications and some other tricks. Brai

I think in some future we may be able to create something sentient, but it’s not this. AI is just a chain of matrix multiplications and some other tricks. Brains are on a whole other level with much more complex processes, even for ants. It’s just that now we have a model that’s really good at mimicking human language, and we mistake language for consciousness, as if language should follow consciousness. I don’t think it does. These are my reasons: 1) Given that I know nothing about the mechanics behind sentience, I’d adscribe to the axiom that animals are definitely sentient, but most lack language. Then language isn’t a requirement for consciousness. 2) Since LLMs have been optimized for language generation and nothing else, they are missing the reproduction of whatever it is that is giving sentience to animals. There’s no way LLMs are sentient.

hackinthebochs 4y ago
>AI is just a chain of matrix multiplications and some other tricks. Brains are on a whole other level with much more complex processes This claim rests on the faulty assumptions that "just" matrix multiplications are simple and that the brain is necessarily more complex than matrix multiplications. A matrix multiplication is an abstraction of a collection of operations, the complexity of which scales with the number of parameters. Matrix multiplications can be arbitrarily complex depending on the number of parameters and their interrelationships. Even quantum mechanics, the most accurate description we have of the universe (including brains) has a formulation in terms of matrix operations (called matrix mechanics). So saying AI is "just matrix multiplication" is doing no work to demonstrate your claim. Conversely, the brain is complex, but it is made up of a complex arrangement of a handful of simple motifs. The cerebral cortex makes up 80% of the brains volume and the pyramidal neuron makes up the majority of that volume. And so complex does not mean heterogeneous composition. > Since LLMs have been optimized for language generation and nothing else This misses what is powerful about Transformers and LLMs. To model the output of some system in the limit of infinite samples, the model must perfectly capture the structure and information dynamics of the system. Any deficiency in the model will necessarily reveal itself in a difference in output at some point. Of course LLMs aren't trained in the limit of infinite samples, but the question is how much of the system (a human mind producing language) has it modeled? If sentience is a necessary feature of a mind producing the kinds of texts we produce, it is an open question whether LLMs have or will recover the structure that endows sentience.
- DanielVZ 4y ago
  > If sentience is a necessary feature of a mind producing the kinds of texts we produce, it is an open question whether LLMs have or will recover the structure that endows sentience. I kind of agree, but this is a big If. If it was necessary, then I’d expect some kind of breakthrough moment in which LLMs or some other model suddenly perfectly mimics consciousness and language instead of the gradual improvement based on the already massive training set that results in uncanny resemblance with infinite malleability.
  - hackinthebochs 4y ago
    Yeah, it would likely be a step change in quality at that point. I doubt the current architecture and training regime would take it in that direction. GPT essentially is a collection of personas to ideally mimic the widely different contexts in its training set. A singular sentient identity is counter-productive to modeling these varying contexts. Now if they could synthesize a large corpus of interactive chats and train on that, that may push it in the direction of a singular identity. Even then, I strongly suspect that some kind of recursive/recurrent architecture is necessary for sentience. Purely feedforward computation doesn't seem like it has the resources to result in the kind of tight unity of experience needed for sentience.

hackinthebochs 4y ago

>AI is just a chain of matrix multiplications and some other tricks. Brains are on a whole other level with much more complex processes This claim rests on the faulty assumptions that "just" matrix multiplications are simple and that the brain is necessarily more complex than matrix multiplications. A matrix multiplication is an abstraction of a collection of operations, the complexity of which scales with the number of parameters. Matrix multiplications can be arbitrarily complex depending on the number of parameters and their interrelationships. Even quantum mechanics, the most accurate description we have of the universe (including brains) has a formulation in terms of matrix operations (called matrix mechanics). So saying AI is "just matrix multiplication" is doing no work to demonstrate your claim. Conversely, the brain is complex, but it is made up of a complex arrangement of a handful of simple motifs. The cerebral cortex makes up 80% of the brains volume and the pyramidal neuron makes up the majority of that volume. And so complex does not mean heterogeneous composition. > Since LLMs have been optimized for language generation and nothing else This misses what is powerful about Transformers and LLMs. To model the output of some system in the limit of infinite samples, the model must perfectly capture the structure and information dynamics of the system. Any deficiency in the model will necessarily reveal itself in a difference in output at some point. Of course LLMs aren't trained in the limit of infinite samples, but the question is how much of the system (a human mind producing language) has it modeled? If sentience is a necessary feature of a mind producing the kinds of texts we produce, it is an open question whether LLMs have or will recover the structure that endows sentience.

DanielVZ 4y ago
> If sentience is a necessary feature of a mind producing the kinds of texts we produce, it is an open question whether LLMs have or will recover the structure that endows sentience. I kind of agree, but this is a big If. If it was necessary, then I’d expect some kind of breakthrough moment in which LLMs or some other model suddenly perfectly mimics consciousness and language instead of the gradual improvement based on the already massive training set that results in uncanny resemblance with infinite malleability.
- hackinthebochs 4y ago
  Yeah, it would likely be a step change in quality at that point. I doubt the current architecture and training regime would take it in that direction. GPT essentially is a collection of personas to ideally mimic the widely different contexts in its training set. A singular sentient identity is counter-productive to modeling these varying contexts. Now if they could synthesize a large corpus of interactive chats and train on that, that may push it in the direction of a singular identity. Even then, I strongly suspect that some kind of recursive/recurrent architecture is necessary for sentience. Purely feedforward computation doesn't seem like it has the resources to result in the kind of tight unity of experience needed for sentience.