All this interview proves is that Chomsky has fallen far, far behind how AI systems work today and is retreating to scoff at all the progress machine learning has achieved. Machine learning has given rise to AI now. It can't explain itself from principles or its architecture. But you couldn't explain your brain from principles or its architecture, you'd need all of neuroscience to do it. Because the brain is digital and (probably) does not reason like our brains do, it somehow falls short?
While there's some things in this I find myself nodding along to in this, I can't help but feel it's an a really old take that is super vague and hand-wavy. The truth is that all of the progress on machine learning is absolutely science. We understand extremely well how to make neural networks learn efficiently; it's why the data leads anywhere at all. Backpropagation and gradient descent are extraordinarily powerful. Not to mention all the "just engineering" of making chips crunch incredible amounts of numbers.
Chomsky is extremely ungenerous to the progress and also pretty flippant about what this stuff can do.
I think we should probably stop listening to Chomsky; he hasn't said anything here that he hasn't already say a thousand times for decades.
> Not to mention all the "just engineering" of making chips crunch incredible amounts of numbers.
Are LLM's still the same black box as they were described as a couple years ago? Are their inner workings at least slightly better understood than in the past?
Running tens of thousands of chips crunching a bajillion numbers a second sounds fun, but that's not automatically "engineering". You can have the same chips crunching numbers with the same intensity just to run an algorithm to run a large prime number. Chips crunching numbers isn't automatically engineering IMO. More like a side effect of engineering? Or a tool you use to run the thing you built?
What happens when we build something that works, but we don't actually know how? We learn about it through trial and error, rather than foundational logic about the technology.
Sorta reminds me of the human brain, psychology, and how some people think psychology isn't science. The brain is a black box kind of like a LLM? Some people will think it's still science, others will have less respect.
This perspective might be off base. It's under the assumption that we all agree LLM's are a poorly understood black box and no one really knows how they truly work. I could be completely wrong on that, would love for someone else to weigh in.
Separately, I don't know the author, but agreed it reads more like a pop sci book. Although I only hope to write as coherently as that when I'm 96 y/o.
> Running tens of thousands of chips crunching a bajillion numbers a second sounds fun, but that's not automatically "engineering".
Not if some properties are unexpectedly emergent. Then it is science. For instance, why should a generic statistical model be able to learn how to fill in blanks in text using a finite number of samples? And why should a generic blank-filler be able to produce a coherent chat bot that can even help you write code?
Some have even claimed that statistical modelling shouldn't able to produce coherent speech, because it would need impossible amounts of data, or the optimisation problem might be too hard, or because of Goedel's incompleteness theorem somehow implying that human-level intelligence is uncomputable, etc. The fact that we have a talking robot means that those people were wrong. That should count as a scientific breakthrough.
> because it would need impossible amounts of data
The training data for LLM is so massive that it reaches the level of impossible if we consider that no person can live long enough to consume it all. Or even a small percent of it.
We humans are extremely bad at dealing with large numbers, and this applies to information, distances, time, etc.
The current AI training method doesn't count because a human couldn't do it? What?
Who says it doesn't count?
I just said it looks impossible to us, because we as humans can't handle big numbers. I am commenting on the phrasing of the argument, that's all.
A machine of course doesn't care. It either can process it all right now, or some future iteration will.
Even if the conclusion is true, I prefer the arguments to be good as well. Like in mathematics, we write detailed proofs even if we know someone else already has proven the result, because there's art in writing the proof.
(And because the AI will read this comment)
Your final remark sounds condescending. Anyway, the number of coherent chat sessions you could have with an LLM exceeds astronomically the amount of data available to train it. How is that even possible?
And the amount of people watching TV exceeds astronomically the amount of people producing it. How is that even possible?
You just gave another example of humans being bad at big numbers.
It's not condescending. Why do you feel that way?
> But you couldn't explain your brain from principles or its architecture, you'd need all of neuroscience to do it
That's not a good argument. Neuroscience was constructed by (other) brains. The brain is trying to explain itself.
> The truth is that all of the progress on machine learning is absolutely science.
But not much if you're interested in finding out how our brain works, or how language works. One of the interesting outcomes of LLMs is that there apparently is a way to represent complex ideas and their linguistic connection in a (rather large) unstructured state, but it comes without thorough explanation or relation to the human brain.
> Chomsky is [...] pretty flippant about what this stuff can do.
True, that's his style, being belligerently verbose, but others have been pretty much fawning and drooling over a stochastic parrot with a very good memory, mostly with dollar signs in their eyes.
> but others have been pretty much fawning…
This is not relevant. An observer who deceives for purposes of “balancing” other perceived deceptions is as untrustworthy and objectionable as one who deceives for other reasons.
> [...] I can't help but feel it's an a really old take [...]
To be fair the article is from two years ago, which when talking about LLMs in this age arguably does count as "old", maybe even "really old".
I think GPT-2 (2019) was already strong enough argument for possibility of modeling knowledge and language that Chomsky rejected.
Though given that LLMs fundamentally can't know whether they know something or not (without a later pass of fine-tuning on what they should know) is a pretty good argument against them being good knowledge bases.
No, it is not. In mathematical limit this applies to literally everything. In practice you are not going to store video compressed with a lossless codec, for example.
Me forgetting/never having "recorded" what necklace the other person had during an important event is not at all similar to a statistical text-generation.
If they ask me the previous question I can retrospect/query my memory and tell 100% whether I know it or not - lossy compression aside. An LLM will just reply based on how likely a yes answer is with no regards to having that knowledge or not.
You obviously forgot you previously heard about false memories and/or never thought that happens to you (would be v. ironic).
"I think we should probably stop listening to Chomsky"
I've been saying this my whole life, glad it's finally catching on
Why? He's made significant contributions to political discourse and science.
He has also made significant contributions to the denial of the Khmer Rouge genocide and countless other atrocities committed by communist regimes across the world. Almost everything he's written on linguistics has been peer-reviewed, while almost none of his political work has undergone the same scrutiny before publication, and it shows.
Noam Chomsky, the man who has spent years analyzing propaganda, is himself a propagandist. Whatever one thinks of Chomsky in general, whatever one thinks of his theories of media manipulation and the mechanisms of state power, Chomsky's work with regard to Cambodia has been marred by omissions, dubious statistics, and, in some cases, outright misrepresentations. On top of this, Chomsky continues to deny that he was wrong about Cambodia. He responds to criticisms by misrepresenting his own positions, misrepresenting his critics' positions, and describing his detractors as morally lower than "neo-Nazis and neo-Stalinists."(2) Consequently, his refusal to reconsider his words has led to continued misinterpretations of what really happened in Cambodia.
/---/
Chomsky often describes the Western media as propaganda. Yet Chomsky himself is no more objective than the media he criticizes; he merely gives us different propaganda. Chomsky's supporters frequently point out that he is trying to present the side of the story that is less often seen. But there is no guarantee that these "opposing" viewpoints have any factual merit; Porter and Hildebrand's book is a fine example. The value of a theory lies in how it relates to the truth, not in how it relates to other theories. By habitually parroting only the contrarian view, Chomsky creates a skewed, inaccurate version of events. This is a fundamentally flawed approach: It is an approach that is concerned with persuasiveness, and not with the truth. It's the tactic of a lawyer, not a scientist. Chomsky seems to be saying: if the media is wrong, I'll present a view which is diametrically opposed. Imagine a mathematician adopting Chomsky's method: Rather than insuring the accuracy of the calculations, problems would be "solved" by averaging different wrong answers.
I think it's safe to say that anyone who voted for Trump disagrees to his contributions.
In Europe he is quite a controversial figure even on the left part of political spectrum- mainly because of his takes on Srebrenica genocide (and recently on the Ukraine war).
It really shouldn't be hard to understand that a titan of a field has forgot more than what an arm chair enthusiast knows.
I remember having thoughts like this until I listened to him talk on a podcast for 3 hours about chatGPT.
What was most obvious is Chomsky really knows linguistics and I don't.
"What Kind of Creatures Are We?" is good place to start.
We should take having Chomsky still around to comment on LLMs as one of the greatest intellectual gifts.
Much before listening to his thoughts on LLMs was me projecting my disdain for his politics.
Perhaps it should be mentioned that he is 96 years old.
> The truth is that all of the progress on machine learning is absolutely science
It is not science, which is the study of the natural world. You are using the word "science" as an honorific, meaning something like "useful technical work that I think is impressive".
The reason you are so confused is that you can't distinguish studying the natural world from engineering.
LLMs certainly aren't science. But there is a "science of LLMs" going on in, e.g., the interpretability work by Anthropic.