It's not about mysticism, the fact that it's "basically a spreadsheet" isn't referring to the fact that it's running in silicone. It's referring to the fact that you can see the whole NN as a big association table, where for each possible word in the input, it knows the next "most likely" token, where likelihood is calculated based on how often such a token would have appeared after similar inputs in the training corpus. This is how it works - it doesn't perform meaningful computations, it doesn't "consider the situation" or anything like that, it's not even really interactive (in the sense that it doesn't update anything based on the information you give it, it simply searches a different part of the spreadsheet based on the conversation so far) - it's just a gigantic spreadsheet that you are probing.
I, and probably many of the people commenting on this, including the author of this submission, agree that silicon computers can someday become AGIs, at least in principle; or, equivalently, that humans are just biological computers. But a language model (or an NN in general) is not a general-purpose computer.
What is your evidence that neural networks do not perform "meaningful" computation?
Still, it's important to note that "ANNs are Turing Complete" actually means "particular classes of ANNs are Turing Complete" - that is, for any Turing Machine, you can construct an equivalent ANN from that class. Here is a paper [0] showing this for the Transformer architecture that is quite relevant to GPT-3 (not sure if LaMDA is also based on this). Note that even here, there are some caveats - you have to use a positional encoding, and the Transformer must have some particular layers.
Still, this doesn't mean that a particular ANN (by which I mean, the result of training, like the copy of GPT-3 that is accessible via APIs) is Turing Complete, or even that the training mechanism used is Turing Complete. The proof in that paper only shows that for every TM there exists a corresponding Transformer ANN that produces the same output for the same input. That doesn't mean you can train this ANN based on backprop/gradient descent. Even more, it definitely doesn't mean GPT-3 itself can perform any computation, in the sense that my CPU can be programmed to, or produce any program in the sense that a C compiler can.
Note that even the universal function approximator result works the same way - it shows that for any (continuous and differentiable in at least one point) function there exists an ANN with certain properties that approximates it to arbitrary precision [1]. This doesn't mean that such an ANN can actually be constructed using backpropagation or other known training methods, from any input.
It also means that there isn't necessarily any ANN that can arbitrarily approximate a non-continuous function, such as tan(x) or f(x) = {1 for x >= 0, 0 for x < 0}, or continuous everywhere but differentiable nowhere functions like the Weierstrass function.
[0] https://arxiv.org/pdf/1901.03429.pdf
[1] https://en.wikipedia.org/wiki/Universal_approximation_theore...
A lot of our mental lives is the regurgitation of past tropes. Does failing to have a novel thought disqualify someone as having sentience?
That is easily proven. Ask the dumbest person you know to find their way through a simple labyrinth they've never seen before, and you will see that they eventually succeed - which they can't have done without performing computation.
That if you feed all the variables of the current state of the universe to an algorithm that knows all the laws of the universe, we can replicate life 1:1 to the last atom, from the beginning of the universe to any arbitrary point in time, the same way, every single time?