It sounds like you're asking whether the output of a neural network is a deterministic function of its input. For many LLMs, you can make that answer yes with the right combination of parameters (temperature = 0) and underlying compute (variance in floating point calculations can still introduce randomness in model outputs even when the model should theoretically return the same answer every time).
There are some ways to introduce stochasticity:
1. Add randomness. The temperature or "creativity" hyperparameter in most LLMs does this, as do some decoders. The hardware these models run can also add randomness.
2. Add some concept of state. RNNs do this, some of the approaches which give the LLM a scratch pad or external memory do this, and continuous pre-training sort of does this.
How this affects people's perception of LLMs as thinking machines, I don't know. What if someone took every response I ever gave to every question that was ever asked of me in my life and made a Chinese Room[1] version of me? A lookup table that is functionally identical to my entire existence. In what contexts is the difference meaningful?
A LUT version of you is inductive. Every observed input/output pair does not uniquely identify your current state. Much like a puddle left by a melted ice cube indicates its volume, but little to nothing of its shape.
Post LUT-you genesis, applying property based fuzz testing would quickly reveal that the LUT-you is one of an infinite number of LUT-yous that melts into the puddle of historical data, but not the LUT-you that is the original ice cube.
People can not be reduced to lookup tables even in theory. No one even knows how a single cell does what it does let alone an entire organism like a person.
I'm not making an abstract claim about neural networks because all numerical algorithms like neural networks can be reduced to a lookup table given a large enough hard drive. This is not practical because the space required would exceed the number of atoms in the known universe but the argument is sound. The same isn't true for people unless a person is idealized and abstracted into a sequence of numbers. I'm not saying no one is allowed to think of people as some sequence of numbers but this is clearly an abstraction of what it means to be a person and in the case of the neural network there is no abstraction, it really is a numerical function which can be expanded into a large table which represents its graph.
>People can not be reduced to lookup tables even in theory
Sure you can. Simply enumerate all of the physical states that the atoms in your body could be in. Any finite-sized object has a finite number of possible states, and so can be represented by a finite lookup table.
Then give some concrete numbers for the states of the atoms. My argument is not abstract, it is very concrete. Give me a neural network and I can generate the graph and prove the equivalence between the network and its graph representation as a table of tuples.
Where are you going to get all that time and space to build a lookup table? Are you sure you're able to measure all state at enough precision to make an accurate table?
Any theory that asserts exhaustive coverage of people would need to take all relevant aspects of reality into consideration, so I suggested some of the trickiest things that are relevant.
Unfortunately for me, they are so tricky that they "don't count" (try, genuinely, to model the reality bending capability of people in a theory, I would love to see that!).
Much to the disappointment of my teenage self who would really have liked the shape-shifting spell to work, I don't see any evidence we can bend reality.
--
> consciousness?
I think this is a red herring. We can talk about P-Zombies, but we lack the means to determine if some random human (let alone AI) is one.
Isn't one of the big criticisms of AI at the moment the fact that they do this slight more than humans, and we can point and laugh at them?
(While conveniently forgetting that half of us were Yani and the other half Laurel, that half of us were blue and black while the other half were white and gold, etc.)
> The Science of the Gaps will do I suppose?
A reference to the ever diminishing role for God in the late 19th century onwards, but I'm not sure how you're using it here?
All this considered, and during this process, did you happen to form any conclusions (or ~"update weights"), consciously or unconsciously (in reality)?
I think it's interesting how the human mind can "know" whether things that are unknown can be modeled, or not, and I happen to believe that this phenomenon occurs within reality (where I believe the comments within this thread are), bending that portion of it. I also believe that this phenomenon is fundamental.
But then, there "is" "no evidence" for any of this...and we all know what that means!
Then I would say your theoretical model is wrong or incomplete or makes for a circular argument (it's an assumption and not proven that finite matter evolving through time reduces to a lookup table).
Simply not true. Of course it is comforting for computer people to believe the world they live in is a giant computer, but that is not our real reality.
"""the number of bits required to perfectly recreate the natural matter of the average-sized U.S. adult male human brain down to the quantum level on a computer is about 2.6×10^42 bits of information (see Bekenstein bound for the basis for this calculation).""" - https://en.wikipedia.org/wiki/Orders_of_magnitude_(data)
(That said, I think quantum physics makes it "all a Markov chain" rather than "all a lookup table").
You are making the assumption that your body consists of a static set of atoms, but your body is a living thing. Your lookup table would end up containing the entire universe to account for extremely remote possibilities.
My "what if someone made a lookup table of everything I ever said in response to something else" hypothetical is pretty flimsy - I realized that right after writing it.
The point I wanted to make is that concepts of sentience, consciousness, reasoning, intelligence, etc. are very philosophically loaded ideas.
Responding to your comment, I don't think anyone credible is arguing that a human being is somehow the same as a neural network. I think the question at play here is "what constitutes reasoning?" - and more specifically "can a deterministic process reason?"
This is not a new debate at all - an abacus can tell us truths about the world, but we don't consider the abacus intelligent. Is GPT-4 somehow different, or is it a very large abacus?
As a numerical function it can be implemented on an abacus so I don't think it's any different from a large enough abacus. It's practically not feasible but theoretically there is no idealization or abstraction happening when numerical calculations on a computer are transferred to an abacus.
There are some ways to introduce stochasticity:
1. Add randomness. The temperature or "creativity" hyperparameter in most LLMs does this, as do some decoders. The hardware these models run can also add randomness.
2. Add some concept of state. RNNs do this, some of the approaches which give the LLM a scratch pad or external memory do this, and continuous pre-training sort of does this.
How this affects people's perception of LLMs as thinking machines, I don't know. What if someone took every response I ever gave to every question that was ever asked of me in my life and made a Chinese Room[1] version of me? A lookup table that is functionally identical to my entire existence. In what contexts is the difference meaningful?
[1] https://en.wikipedia.org/wiki/Chinese_room