How likely is it that free will emerges from the very randomness which you cite and discard as “The output of an AI is completely determined by its inputs, model weights, and any random numbers that are used in the computation”? IIRC the LLM chooses the next token randomly after it calculates the energies and probabilities as , then the token is selected with the probability . If, say, the two biggest probabilities are close to each other, then the choice is dependent on the random number. Otherwise, the choice depends almost entirely on the LLM’s preprocessing which makes other options practically impossible. However, the preprocessing is itself impacted by context and the LLM’s knowledge.
I suspect that the closest equivalent of such a situation which the humans can encounter is facing a choice between two alternatives of similar expected distributions of utility-like functions, or between a good and bad alternative, while the distributions themselves are governed by human knowledge, intuitions or even emotions experienced at the time via processes whose description cannot be contained in this margin depends on mechinterp applied to the human brain (or, if the dilemma is posed before an LLM, to the LLM).
I think LLMs typically use pseudo-random numbers for these calculations, and so are still ultimately deterministic given the seed of the pseudo-random number generator, such that if you re-ran the AI on the same inputs, with the same weights and PRNG seed it would give the same output, with no room for a metaphysical agent to be exerting free will.
We might be talking past each other, possibly my fault if I’m using a non-standard definition of “free will.” But what I’m trying to get at isn’t about decision theory and the way individuals choose between different options to maximize a utility function, it’s instead about the relationship between physical matter and consciousness, and the strong intuitive feeling that my subjective consciousness is an agent that is causing physical events in a way that is not reducible to deterministic physical causation or randomness. And wondering if this is real or an illusion, and noting that AIs certainly do not have some unexplained metaphysical property like this, because their processing of inputs into outputs can be fully traced from end to end. Possible also that your point was about this and I just didn’t correctly understand it. Anyway thanks for the comment!
How likely is it that free will emerges from the very randomness which you cite and discard as “The output of an AI is completely determined by its inputs, model weights, and any random numbers that are used in the computation”? IIRC the LLM chooses the next token randomly after it calculates the energies and probabilities as
I suspect that the closest equivalent of such a situation which the humans can encounter is facing a choice between two alternatives of similar expected distributions of utility-like functions, or between a good and bad alternative, while the distributions themselves are governed by human knowledge, intuitions or even emotions experienced at the time via processes whose description
cannot be contained in this margindepends on mechinterp applied to the human brain (or, if the dilemma is posed before an LLM, to the LLM).I think LLMs typically use pseudo-random numbers for these calculations, and so are still ultimately deterministic given the seed of the pseudo-random number generator, such that if you re-ran the AI on the same inputs, with the same weights and PRNG seed it would give the same output, with no room for a metaphysical agent to be exerting free will.
We might be talking past each other, possibly my fault if I’m using a non-standard definition of “free will.” But what I’m trying to get at isn’t about decision theory and the way individuals choose between different options to maximize a utility function, it’s instead about the relationship between physical matter and consciousness, and the strong intuitive feeling that my subjective consciousness is an agent that is causing physical events in a way that is not reducible to deterministic physical causation or randomness. And wondering if this is real or an illusion, and noting that AIs certainly do not have some unexplained metaphysical property like this, because their processing of inputs into outputs can be fully traced from end to end. Possible also that your point was about this and I just didn’t correctly understand it. Anyway thanks for the comment!