I don't think that's really a more accurate description at all; but even if it is, it sounds like you're implying that it's merely the scale of the mechanism that qualifies what it does as genuine "understanding". If none of the few hundred assistants knows what the Chinese characters they're receiving, converting, and outputting mean, the guts of the thought experiment remain the same as if it was just one person.
Okay, let's consider the more accurate (I would rather say my description of the "room" was more realistic) question first. Large language models are based on neural networks with, I'm reading from multiple sources, many millions of nodes and up to tens of billions of connections. Let's say there's a mere one million nodes and ten billion connections.
If you're going to execute the LLM algorithm with a person and paper, which is theoretically quite possible, at the very least you need to have the weight of each interconnection written down on a list. At a hundred entries per page (hope you brought your reading glasses) and a thousand pages per book, that requires a hundred thousand books. Sounds like a library to me. By nature the calculations for a neural network involve multiple parallel computations. You'll want to keep track of the cumulative sums of the inputs of each neuron as you go; at a thousand per white board, you need a thousand white boards. (You can reduce the number of white boards you'd need, at the cost of more work rewriting them more often.)
Now to do the calculations, at ten seconds per interconnection (hopefully the books are well-organized to minimize how often you have to go back to the stacks to fetch new ones, and you're a prodigy at addition and multiplication of high precision real numbers in your head, and of course you never need to eat or sleep or take any breaks at all), will take you 3,169 years. The final steps of sorting the weights of the final layer of nodes and then looking up the matching Chinese character(s) are trivial by comparison. But, I'm pretty sure that's per word of the output, so you'd better clear your schedule for a few hundred millennia if you're planning a conversation or to answer some involved question. And I sure hope it's an important question being answered!
Okay, so you don't absolutely need those hundred assistants to finish before the sun burns out. But I'd bet you'd be glad to have them.
Now, you might be thinking that such a computation-intensive neural network is the wrong way to design the Chinese room. It's certainly not what Searle had in mind. At the time, it was already known that just looking up a list of possible inputs in a book and following some simple if-then and/or random-choice instructions to choose a response is nowhere near adequate to create the semblance of intelligent discourse. (Computer game NPCs work that way.) And that making the book bigger wouldn't work either, even if it filled the room.
What was considered the most likely method for passing the Turing test was to compile a database of facts about the world and then use algorithms of logical deduction to figure out facts or hypotheses about the intent of the input questions and what output would correctly resolve it. That sounds more like something that might fit in a room.
But, it isn't. To be useful for general AI (or the illusion thereof), the database of facts has to include "common-sense knowledge" that requires millions of entries. At a hundred cryptic facts per page, it's still thousands of books. (Why cryptic? Because of course, if the knowledge base was in English you'd learn Chinese in the process.) Which is fine, but it turns out there's no way to tell in advance which facts will be relevant. That kind of logical deduction, the kind used for proving mathematical theorems for instance, works by trial and error, exploring an ever-increasing space of possibilities, and practically every step along the way requires searching through the database and trying each entry. The amount of computation becomes intractable even for machines that can do the LLM neural network calculations routinely. It might seem there should be some clever way around this, but the researchers who worked on it for decades consider the problem so intractable that there have been arguments that it proves AI that can do, well, what it turns out LLMs can do, to be completely impossible. The human brain just "somehow knows" which information is relevant to a problem at hand in a way that a computer can't imitate, according to that argument.
The approach does work where only a smaller database of specialized knowledge is needed, and it can also work when specialized hardware like Watson can muster up enough cycles to manage it. But chat with Watson still quickly reveals more severe limitations than we see with LLMs. And if you were to encode Watson in a room full of books, the necessary steps to run the algorithm by hand would far exceed the neural net version. Then you really would need sun-burns-out amounts of time in the Chinese room.
Why do I harp on that when you've already acknowledged the possibility that maybe you really do need the huge library of books and the hundred assistants, and claimed that doesn't change anything important? Because it's not the scale of the
mechanism that matters (that just determines how long it takes, whether it's one person or a hundred assistants, or even ten billion people each doing one calculation nearly simultaneously so the answer would come out lickety-split), but the amount of information the mechanism takes into account. Whether it's hundreds of thousands of books of neural net weights, or tens of thousands of books of facts, or heck, in principle you could do it with one long fixed sequence of printed if-then choose-your-own-adventure instructions in which case the number of books would more than fill the known universe but you'd only have to consult a tiny fraction of them to produce any single response. In all those cases the room doesn't fit in the parameters of Searle's description of it.
And of course, all those conceptual Chinese Rooms (and LLMs) are toys, in terms of data content and parallel computing power, compared to a mature human brain. 80+ billion neurons, 100 trillion connections...
There's nothing strange about a whole system being able to do things that its component parts cannot do individually. What individual part of an airplane can fly? What individual organism can evolve? Which soldier was the one who surrounded the enemy formation?
Darat seems to imply that as long as the machine - the combined whole - is outputting a coherent message then the internalities of the process don't matter, the machine "understands Chinese" regardless of what's happening inside it. That's the concept I'm challenging.
Since this is the science subforum rather than the philosophy one, what test for "understanding" other than the ability to respond coherently to probing questions about the concept(s) in question are you proposing?
Consider the question I suggested for ChatGPT (which Pulvinar kindly tested out for real) about opposites. Or simpler questions like "which two of these words are opposites: cold, fast, dark, hot, night." Or "Write a sentence that uses opposites." Questions/tasks like these are how we test students' understanding of concepts, and we interpret correct answers as demonstrating understanding. How else?
