+It is from this understanding that some of the various flaws of LLMs begin to +make sense. For example, the reason why they cannot accurately cite sources is +because, firstly, they do not know what a source even is, but secondly, because +it is simply looking back into its data and checking what usually follows +within that context with the parameters of “citing the source,” which is why it +so often simply makes them up. The truth is that its source is all its +data mashed together probabilistically based on the input prompt and the +context of the overall “conversation.” +
+ ++Getting back, however, to a question raised earlier about the level of +consciousness of these LLMs, although in the analogy given above the man surely +has a rational soul and a human intellect, it is also evident how it is not +necessary to make use of these higher faculties in order to do what these +machines can do: it is merely pattern recognition and probabilistic +computation. This is something that even the beasts could do if sufficiently +trained (think of the example of a parrot). The machine has no concept of the +meaning symbolized by these words; in fact, it does not see them as +symbols at all, but tokens with numerical values. And if it cannot comprehend +meaning, then it certainly cannot reason on the basis of meaning. What is more, +this is not simply a question of needing more training or more data, it is a +matter of the process itself; for no matter how much you train the man in the +analogy with more data and better pattern recognition techniques, he never will +have actually learned Chinese until he starts to associate meaning those +symbols and is thus able to reason a response instead of merely guessing what +tokens go next. +
+ +