Right, I’m not saying they are equivalent or that LLMs work exactly like the brain. We completely agree that LLMs on their own do not do what brains do. I’m talking about the general principles here. My key point is that a lot of simple behaviors can combine together to produce complex emergent behaviors.
The real question is whether high level patterns human brains produce, which we agree are substrate agnostic, can be expressed within an artificial system. The underlying mechanics of the two systems don’t have to be the same. Similarly how a combustion engine and an electric one both achieve the same end result in a very different way, it seems very likely to me that the patterns the brain produces can also be supported by a completely different substrate. I’ve seen no evidence to suggest that there is something special about biological substrate that makes it a unique platform in this regard.
The reason why saying LLM is a next token predictor is reductive is because this view ignores all the complexity associated with the token being produced. It’s similar to the way we use language a single word at a time, and you could say my brain is the next word predictor. Yet, each word exists within a context, and it’s a token associated with a web of ideas in my head at any one time. Ignoring all that and focusing solely on the word I produced is reductive and completely misses how my brain actually arrived at that word.
This is not like how human reasoning works in the vast majority of cases, and calling it “exactly like” human reasoning is begging the question.
What evidence is there that this is not how human reasoning works, pretty much all the evidence that I’ve seen strongly suggests that human reasoning is very much a stochastic process. Again, the algorithm is almost certainly different, but the principles appears to be very much the same.
As for human analog, the “number of letters” test that most models have gotten better at (but still hallucinate on sometimes), also succeed on nonsense words that are not in their training set. And they work because they are engineered, hard-coded, to break down words into individual characters. That is not an emergent property of the NN like it is in humans.
Humans hallucinate all the time and produce nonsense statements regularly. And just because something is engineered doesn’t mean it works on a different principle. It’s entirely possible for evolution to come up with a similar solution to one that’s consciously designed. The selection pressure is thermodynamics in both cases. Either a system evolves towards a certain state being moulded by energy gradients, or it is designed towards that state because the designer is trying to minimize energy usage and maximize performance. So, whether something is emergent or not is not really a relevant point here.
And you’re right, transformer is just a piece of a much bigger puzzle. In my opinion, active memory formation and dynamic learning are likely the next steps. The real elephant in the room is that we need to delineate training and inference phases. That’s the clearest sign of the limitations of the LLM approach. A human brain can learn efficiently on the fly from few shot examples, and to integrate new information continuously. The current approach is not efficient enough to allow for that, and the fact that the model needs a huge numbers of samples to identify invariant patterns within them itself suggests that there is something wrong with the approach.
There is some research being done in this direction already such as this paper, there are ideas for creating an equivalent of short term memory that’s used in the active context, and then these memories get consolidated into longer term memory layers. So, you could have a base model that is pretrained, and then have active learning on top of that which might get a lot closer to having the plasticity our brains exhibit. Another really interesting approach is MemOS which would allow packaging up skills into composable building blocks that can be shared. So, a model could learn a task, package it up and then another model can reuse it.
It’s also possible that the whole transformer approach is a dead end entirely. For example, SpkingBrain looks like a promising alternative.
So, to sum up, I think LLMs exhibit a lot of similar behaviors to biological brains, which suggests they likely operate on similar underlying principles. But I don’t think they’re equivalent to brains, and at best they mimic what specialized brain structures are doing and you would need to build out other structures to get closer to mimicking what the brain does as a whole. But it’s also possible that they’re just a dead end entirely, and that this approach is not efficient enough to do on the fly learning which is a key aspect of general intelligence.
Right, I’m not saying they are equivalent or that LLMs work exactly like the brain. We completely agree that LLMs on their own do not do what brains do. I’m talking about the general principles here. My key point is that a lot of simple behaviors can combine together to produce complex emergent behaviors.
The real question is whether high level patterns human brains produce, which we agree are substrate agnostic, can be expressed within an artificial system. The underlying mechanics of the two systems don’t have to be the same. Similarly how a combustion engine and an electric one both achieve the same end result in a very different way, it seems very likely to me that the patterns the brain produces can also be supported by a completely different substrate. I’ve seen no evidence to suggest that there is something special about biological substrate that makes it a unique platform in this regard.
The reason why saying LLM is a next token predictor is reductive is because this view ignores all the complexity associated with the token being produced. It’s similar to the way we use language a single word at a time, and you could say my brain is the next word predictor. Yet, each word exists within a context, and it’s a token associated with a web of ideas in my head at any one time. Ignoring all that and focusing solely on the word I produced is reductive and completely misses how my brain actually arrived at that word.
What evidence is there that this is not how human reasoning works, pretty much all the evidence that I’ve seen strongly suggests that human reasoning is very much a stochastic process. Again, the algorithm is almost certainly different, but the principles appears to be very much the same.
Humans hallucinate all the time and produce nonsense statements regularly. And just because something is engineered doesn’t mean it works on a different principle. It’s entirely possible for evolution to come up with a similar solution to one that’s consciously designed. The selection pressure is thermodynamics in both cases. Either a system evolves towards a certain state being moulded by energy gradients, or it is designed towards that state because the designer is trying to minimize energy usage and maximize performance. So, whether something is emergent or not is not really a relevant point here.
And you’re right, transformer is just a piece of a much bigger puzzle. In my opinion, active memory formation and dynamic learning are likely the next steps. The real elephant in the room is that we need to delineate training and inference phases. That’s the clearest sign of the limitations of the LLM approach. A human brain can learn efficiently on the fly from few shot examples, and to integrate new information continuously. The current approach is not efficient enough to allow for that, and the fact that the model needs a huge numbers of samples to identify invariant patterns within them itself suggests that there is something wrong with the approach.
There is some research being done in this direction already such as this paper, there are ideas for creating an equivalent of short term memory that’s used in the active context, and then these memories get consolidated into longer term memory layers. So, you could have a base model that is pretrained, and then have active learning on top of that which might get a lot closer to having the plasticity our brains exhibit. Another really interesting approach is MemOS which would allow packaging up skills into composable building blocks that can be shared. So, a model could learn a task, package it up and then another model can reuse it.
It’s also possible that the whole transformer approach is a dead end entirely. For example, SpkingBrain looks like a promising alternative.
So, to sum up, I think LLMs exhibit a lot of similar behaviors to biological brains, which suggests they likely operate on similar underlying principles. But I don’t think they’re equivalent to brains, and at best they mimic what specialized brain structures are doing and you would need to build out other structures to get closer to mimicking what the brain does as a whole. But it’s also possible that they’re just a dead end entirely, and that this approach is not efficient enough to do on the fly learning which is a key aspect of general intelligence.