> The large language models are a little surprising. I’ll give you that.
I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job.
Almost nobody is really surprised that if you throw more compute at a neural net it becomes better at the task it’s trained on. But almost everybody is really surprised that becoming better at a task like ‘natural language prediction’ would produce all these strange abilities that sort of look like “understanding the world”.
One way to resolve this surprise is to find some reason to believe these strange abilities are fundamentally not an understanding of the world. Thus stochastic parrots, this article, Yan LeCun and Chomsky, etc.
Another way to resolve this surprise is to find some reason to believe these strange abilities fundamentally are an understanding of the world. Thus regulation of AI, existential risk, Hinton and Yudkowsky, etc.
I don’t know what the correct resolution of the surprise is. The only thing I’m confident in is that it’s correct to be surprised by the abilities of LLMs. My current (tentative) resolution of the surprise is that language encoded way more information about reality than we thought it did. (Enough information that you can fully derive reality from language seems improbable, but iirc it did derive Othello and partly derived chess and I would have thought there wasn’t enough information in language to derive those without playing the games as well, so I can’t rule it out.)
I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job.
Almost nobody is really surprised that if you throw more compute at a neural net it becomes better at the task it’s trained on. But almost everybody is really surprised that becoming better at a task like ‘natural language prediction’ would produce all these strange abilities that sort of look like “understanding the world”.
One way to resolve this surprise is to find some reason to believe these strange abilities are fundamentally not an understanding of the world. Thus stochastic parrots, this article, Yan LeCun and Chomsky, etc.
Another way to resolve this surprise is to find some reason to believe these strange abilities fundamentally are an understanding of the world. Thus regulation of AI, existential risk, Hinton and Yudkowsky, etc.
I don’t know what the correct resolution of the surprise is. The only thing I’m confident in is that it’s correct to be surprised by the abilities of LLMs. My current (tentative) resolution of the surprise is that language encoded way more information about reality than we thought it did. (Enough information that you can fully derive reality from language seems improbable, but iirc it did derive Othello and partly derived chess and I would have thought there wasn’t enough information in language to derive those without playing the games as well, so I can’t rule it out.)