3 ms·
It seems like right now the testable hypothesis "LLMs generate text that is fundamentally different from human beings." That's effectively the Turing test. Curr
by jfengel 3y ago
It seems like right now the testable hypothesis "LLMs generate text that is fundamentally different from human beings." That's effectively the Turing test. Current LLMs do a better job of passing the Turing test than things in the past, but it doesn't take a lot of effort to distinguish.
It's difficult to generalize because it can be tuned to do any one thing. It's the whole process of doing anything that requires the full apparatus of a human being, and there is no sign that LLMs are approaching that any time soon.
Which is the other problem. We're talking about what LLM's might one day do, rather than what they currently do. It's entirely possible that one day LLMs will be as flexible as human beings, training themselves for every new scenario. I have reason to doubt it, but the basis of that doubt is only noticing the mechanical difference between brains and LLMs. I cannot prove that the limit cases will remain different.