> Studying the task of how to fake it certainly leads to insightful subproblems galore.
...
> I myself shall certainly continue to leave such research to others, and to devote my time to developing concepts that are authentic and trustworthy. And I hope you do the same.
...
> Please reply only with respect to binomial coefficients, because I've already spent way too much time on the topic above! The topic is timely, and important enough not to ignore completely, but it's emphatically not for me.
Knuth is a legend and a genius. He is clearly impressed with GPT in the same way a physicist might be impressed with a stage magician. I can understand that he would marvel at the skill required to achieve such convincing illusions but he would understand that learning the magician's tricks is not worth his time, which would be better spent actually investigating what he believes to be the real physics underlying the universe.
However, I feel his shots at GPT here are a bit cheap. We don't know if GPT is an illusion or if it is a leap in the right direction. Determining that will require significant deep study of these emergent behaviors.
I felt the same kind of "sour-grapes" kind of reasoning from Chomsky's analysis of LLMs (although I haven't heard his opinion on these new GPT-3.5/GPT-4 models). It is like these legends spent their entire careers with the assumption that neural-nets and language models couldn't possibly work and they are sticking to that even in the face of new evidence.
I just wish I saw some acknowledgement from these elders that there is a possibility that some aspect of neural nets, transformers/attention may really directly relate to intelligence and eventually consciousness. I'm not expecting them to hop on the hype train - but their casual dismissal given our limited knowledge of why these advanced behaviors emerge strikes me as odd.
Knuth's response here reminds me a bit of Einstein's rather dogged commitment to the "god does not play dice with the universe" philosophy. Just like non-determinism of Quantum Mechanics was a bit of a thorn in Einstein's side, the non-determinism and probabilistic nature of AI seems to put off Knuth from recognizing the long term value.
This isn't about being a "magician" - it's more about that probabilistic non-deterministic computation can provide immense value and can be the building block for a whole new class of approaches to solve problems.
It is very interesting to compare Knuth's position on LLMs to Einstein's position on quantum physics and I think it is apt.
At least Einstein was explicit in his distaste for non-determinism. Knuth does not specify in this exchange why he believes these LLM approaches are inauthentic. He does demonstrate the untrustworthy-ness of the current models but he doesn't provide any evidence that shows the approach is incapable of creating trustworthy models in principle.
Even on the topic of trustworthiness, it is an interesting kind of criticism in that we are holding AIs based on LLMs to a higher standard than we would hold any human. Could you imagine a vox-pop style on-the-street interview where an average passer-by was asked the same questions that Donald Knuth posed to the LLM? How many people would even be able to formulate a coherent answer to the questions about Beethoven, Rogers and Hammerstein, or The Haj? Yet somehow the imperfection of these answers from an early-generation LLM is enough to completely dismiss the entire approach.
If you give the person internet access and some time to answer the question, then most people will do better... or at least they'll say they don't know.
...
> I myself shall certainly continue to leave such research to others, and to devote my time to developing concepts that are authentic and trustworthy. And I hope you do the same.
...
> Please reply only with respect to binomial coefficients, because I've already spent way too much time on the topic above! The topic is timely, and important enough not to ignore completely, but it's emphatically not for me.
Knuth is a legend and a genius. He is clearly impressed with GPT in the same way a physicist might be impressed with a stage magician. I can understand that he would marvel at the skill required to achieve such convincing illusions but he would understand that learning the magician's tricks is not worth his time, which would be better spent actually investigating what he believes to be the real physics underlying the universe.
However, I feel his shots at GPT here are a bit cheap. We don't know if GPT is an illusion or if it is a leap in the right direction. Determining that will require significant deep study of these emergent behaviors.
I felt the same kind of "sour-grapes" kind of reasoning from Chomsky's analysis of LLMs (although I haven't heard his opinion on these new GPT-3.5/GPT-4 models). It is like these legends spent their entire careers with the assumption that neural-nets and language models couldn't possibly work and they are sticking to that even in the face of new evidence.
I just wish I saw some acknowledgement from these elders that there is a possibility that some aspect of neural nets, transformers/attention may really directly relate to intelligence and eventually consciousness. I'm not expecting them to hop on the hype train - but their casual dismissal given our limited knowledge of why these advanced behaviors emerge strikes me as odd.