I was not thinking of an AI learning just like large language models "learn" today by adjusting their weights with training data. I was thinking of an AI that has the ability to learn build into it. One way I could imagine this could be done is as follows.
The AI has some memory, essentially just a big byte array. It will answer questions just like a current large language model, it will feed the input and the content of its memory [1] into a neural network and produce some response. In addition to this there would also be a neural network that generates memory update operations from the input and the current content of the memory in order to memorize information. And here I would imagine that this neural network will eventually become smart enough to decide what is worth memorizing and what should be discarded.
As far as I know we do not currently have such systems and it is not clear when we will have something like that. While what I described above seems more or less doable with current technology, it is not clear that it could actually work, that there is for example a realistic way to train something like this. Human brains, I would assume, neither do gradient decent nor explicitly update some memory cells, so maybe we are still lacking some key insights. But I am sure that large language models are not the final word on artificial intelligence.
[1] If the AI would have a gigabyte of memory, you could of course not easily feed the entire memory into a neural network at once. This would have to be done in chunks or the neural network itself would have to generate addresses of pieces of memory it wants fed into the neural network.
The AI has some memory, essentially just a big byte array. It will answer questions just like a current large language model, it will feed the input and the content of its memory [1] into a neural network and produce some response. In addition to this there would also be a neural network that generates memory update operations from the input and the current content of the memory in order to memorize information. And here I would imagine that this neural network will eventually become smart enough to decide what is worth memorizing and what should be discarded.
As far as I know we do not currently have such systems and it is not clear when we will have something like that. While what I described above seems more or less doable with current technology, it is not clear that it could actually work, that there is for example a realistic way to train something like this. Human brains, I would assume, neither do gradient decent nor explicitly update some memory cells, so maybe we are still lacking some key insights. But I am sure that large language models are not the final word on artificial intelligence.
[1] If the AI would have a gigabyte of memory, you could of course not easily feed the entire memory into a neural network at once. This would have to be done in chunks or the neural network itself would have to generate addresses of pieces of memory it wants fed into the neural network.