It's not that I agree with everything Chomsky says or writes but I can't shake the feeling that he's much more intelligent, eloquent and gentlemanly than any of his critics.
Chomsky used some examples in his article of what LLM's would get wrong. It is clear that he hasn't even tried, because chatGPT in fact gets those examples right. Why does he get a free pass for such errors?
Sounds like Aristotle when he claimed that women do not have wisdom teeth for reasons (he made up), instead of bothering to check.
You might want to check your wife's teeth :)
I have my own reservations about his opinion but would never state them in the combative and sneering tone often employed by his critics.
Can't help but feel like they personally have a problem with him, and they're using this piece as a cudgel to "take a swing".
If we are evaluating LLM's purely on outcome alone, it seems only fair to compare them to humans that don't have fully functioning cognitive capabilities. We wouldn't discredit a schizophrenic person that hallucinates, we know that this person can also see reality for what it is, most of the time. I just don't think that finding absurd examples, usually after careful prompting, is a slam dunk against LLM's at all.
They can, it's just that Chomsky's dismissal of "weight adaptations" as insufficient is not actually justified by any evidence. The Transformer model augmented with memory is Turing complete, so it can in principle learn to compute anything using only weight adaptations.
* On the Computational Power of Transformers and its Implications in Sequence Modeling, https://arxiv.org/abs/2006.09286
Our brains are much more complex than an artificial NN, though. So the mere thought that the two somehow share functionality after a wildly different learning process is somewhat far-fetched.
Well sure, because CFGs can have unbounded depth, and all real brains are finite. But they can learn bounded CFGs:
https://arxiv.org/abs/2112.09174
> Our brains are much more complex than an artificial NN, though.
Yes, brains have a few orders of magnitude more parameters, and lots of evolutionary baggage. It's not clear how many parameters would be needed without that evolutionary baggage. Time will tell.
> So the mere thought that the two somehow share functionality after a wildly different learning process is somewhat far-fetched.
I don't know what precisely you mean by "share functionality", but it doesn't seem like a stretch to suggest they are in some sense isomorphic given Turing equivalence. Rule 110 doesn't look like it should share much with Turing machines or the lambda calculus or SKI combinators, but these are all equivalent in a formal sense.
> in some sense isomorphic given Turing equivalence
Then you're brain is just like any CPU, which is very far from what such a statement tries to express in the context of NNs.
Yes, the brain is like a CPU + a specific program running on that CPU. The properties of that program are where the isomorphism comes in. So to come full circle back to your original point, we don't yet know that "weight adaptations" are not a sufficient operational model for that program.
In contrast, the world is too complicated to learn through exposure. Something has to structure our learning process. Language is (in principle) not learnable, only when you make enough assumptions. Those assumptions must be coded somewhere in your brain and probably genes.
Tabula rasa sounds nice, but it is untenable and utterly lacks evidence.
What did Scott Aaronson contribute of comparable impact?
People who rely on LLMs to do their work in domains they understand find them useful in the sense of hefting another useful tool as humans have done throughout their existence. That's it. Among serious users of LLMs as a productive tool there's no wistful wonder that they may be speaking to a truly understanding intelligence. So Chomsky and the rest of the critics don't really have a foot to stand on when they say things like
> It is at once comic and tragic, as Borges might have noted, that so much money and attention should be concentrated on so little a thing — something so trivial when contrasted with the human mind, which by dint of language, in the words of Wilhelm von Humboldt, can make “infinite use of finite means”, creating ideas and theories with universal reach.
Why is it comic and tragic? LLMs are genuinely useful tools and one of my biggest productivity boosters at work since I learned how to make a computer do my bidding. Of course huge amounts of money are being poured in when it's so helpful across a number of domains for getting work done.
Really, folks, there's nothing to see here. Make use of these marvelous tools and continue doing what humans have done best, a truly innate human inheritance, to use tools to support our goals.
> However useful these programs may be in some narrow domains (they can be helpful in computer programming, for example, or in suggesting rhymes for light verse), we know from the science of linguistics and the philosophy of knowledge that they differ profoundly from how humans reason and use language.
I think you and he agree on your main point, but his main point is about something else.
Not that this is really relevant to this topic, but that's simply not true. He doesn't support Russia's interests, he's simply acknowledging that when dealing with powerful nations, we have to consider their interests and reach compromises where their interests conflict with ours in order to avoid the horrors of war. You know, diplomacy 101.
Has been doing that for decades.
He is intellectually dishonest and morally bankrupt. This makes his thoughts on other topics suspect as well.
Moreover how do you do diplomacy with Hitler2.0?
Thinking that giving up yet another piece of (for you) unknown land to one of those will bring peace makes you sound really naive.
Speaking of intellectual dishonesty making other thoughts suspect...