Artificial-intelligence systems such as large language models (LLMs) can exhibit behaviour that seems strikingly intelligent. This has led many people, including some experts, to wonder whether these systems could have some form of consciousness — the capacity for subjective experience. However, we think that existing AI systems are not conscious, and that, along current trajectories of development, they are not likely to become conscious. Greater clarity about whether AI could become consciousness will be crucial as society grapples with the challenges posed by this increasingly powerful technology.
Linear algebra + sigmoids are enough to approximate any function from R to R; in particular, a small set of a few layers of LLM neurons can connect and interact in pretty novel ways. There are some things they can’t do, and those things aren’t computable. If our brains can do non-computable things, then our brains are super-turing (and this, given the evidence, seems pretty unlikely right?).
I’m not a neurologist or a philosopher so I think this may not be something I can speak intelligently about, but it seems to me that the question doesn’t really live in the function approximation, but in the gaps. Doing a single train of thought is the easy part. I think what you need is an “always online” feature to keep thinking and some way to be stateful (make memories) without constantly retraining the whole network or whole sets of layers.
This is, I think, the benefit of the parallelism and asynchrony of the “individual neuron” discussed in the book.
(book? The linked article is like 4 pages. References are somewhat larger, but nothing over 60 afaict?)
They claim that silicon is essentially incapable of doing what is needed; if you disagree with this, then computationally everything you can build with a circuit board + SSD + RAM you can build with an LLM with a deep layer or 7. In particular, you can do online learning of a sort.
I note that wasps and some beetles don’t seem to have memory whatsoever, and plasticity seems awfully low. The authors, and I think most people, would say they are probably a little conscious.
Sounds to me like you just need a context window with memories that self prune based on the dynamic context window of the model. That’s what we do. We forget things, and when they’re remembered we can access the “biological database” of associated memories on the fly. Because the associated information isn’t deleted, only the short term memory from being included in the context window of each update of our own neural networks. Some people have greater context windows, and see therefore seen as smarter, because they can access more information and include it in the processing of each moment in time.
There’s zero reason a digital representation of this system could not do the same thing. No matter how much I’m ridiculed for daring to suggest it 😅