Comment by NaiveEmpiricist
1 day ago
Norvig (and Chomsky) are talking at cross purposes.
Norvig confuses engineering for science. If an engineer wants something that works (and guarantees that it will work) then you use the most effective tool available to achieve that goal. No problem there. But there's absolutely no reason to think that tells you anything scientific about the phenomenon you're modeling. For example, aeronautical engineering says very little about how birds fly. And - currently relevant - LLMs learn to use and understand language, but they use millions of times more training data, and millions of times more computing power, than a child does. So an LLM may be a great technology, but it's probably not a good model for how humans learn and use language.
Chomsky OTOH doesn't see any use for statistics in science, at least not in linguistics. Fortunately for the field of linguistics, Chomsky's influence in linguistics is waning and linguists are beginning to use statistical tools to do science.
> Norvig confuses engineering for science
I don’t think Norvig is confused. He explicitly addresses this point right at the top of the article.
His argument is that successfully engineering a system using a statistical method provides some bits of evidence for the hypothesis that the system itself is statistical.
Which is obviously true, so long as we take care to weigh evidence properly.
> there's absolutely no reason to think that tells you anything scientific about the phenomenon you're modeling
No, this is not how evidence works. If you can accurately model a system with method X and you cannot come up with a model that works with method Y, you should think it’s more likely (but not certain) that X is the actual physical model.
And indeed modern cognitive science (see Predictive Processing) also has a bunch of evidence for the brain as a hierarchical predictor, rather than a rule engine.
To your points - planes are not model organisms of birds, but they do shed insight on the bounds that birds must obey, eg thrust-to-weight etc. And the sample efficiency point really needs to die. The human brain is born with an extremely complex prior baked in. An LLM begins pretraining with almost nothing (just a very simple architecture and a few simple algorithms like attention), and has to learn the basic architecture of cognition from scratch. A better baseline would be how many samples were processed by every animal in our genetic ancestry that had a brain, since brains existed.
> No, this is not how evidence works. If you can accurately model a system with method X and you cannot come up with a model that works with method Y, you should think it’s more likely (but not certain) that X is the actual physical model.
Could you expand on this point?
Try to model radioactive decay via a process other than statistical and you will almost certainly fail, since that is our best understanding of it.
I was also going to suggest coin flips, but it’s not quite the same… I think any chaotic system is generally best modeled that way, and if you cared about the effect the system has on other things, it’s unlikely you can do better since measuring the initial state will not be feasible.
1 reply →
i agree. but it's uncanny that stochastic generation of language yields a form of recognizable intelligence.
in college, i was somewhat involved in cogsci wrt semantic mapping of linguistic concepts based on human surveys (this was well before LLM/AI-word2vec), and it surprises me today how the mathematical properties of relatedness/distance seem now to drive machine intelligence, something which we did not consider at the time at all.
maybe i'm getting in my metaphysical sentimental stage, a turn a lot of adults make after middle age, but i wouldn't fully dismiss the fact that there may be an equivalence.
Indeed. Who'd have thought that mere statistics could do that? Hat's off to them - they made the darn thing work.
There's pretty interesting (I think) research in how effectively it's possible to scale down LLMs. If you're interested, take a look at the BabyLM conferences:
https://babylm.github.io/
(I have no connection to them.) The work there demonstrates that gap between what we can do now with human-relevant amounts of training text, and what remains to be discovered. Quite a bit, it looks like.
"Mere statistics" is how physical reality works; I don't mean the language, the syntax
I mean shifting influence of gravity relative to an object with enough mass type ratios are everywhere
The language of math started as visual art to depict such ratios; in modern times this ground truth associations is frequently overlooked or never understood to exist in the first place without engagement
Differential diagnosis in medicine is exactly the same process; such processes and their successes are everywhere.
The issue we have in software is the outcome isn't as obvious as jabbing a blood draw needle in a patients eye; we spent a lot of time wasting time on illusory needs
“Mere statistics” is heavily underselling what the transformer architecture does. Unless you think “mere ion channels” is a sensible way to think about the human brain.
> stochastic generation of language yields a form of recognizable intelligence.
Can you use a human language well enough and it not sound intelligent? And no, I don't mean just syntactically correct because semantics is a part of human languages too. I think the astounding thing is that they managed to model a language this well. It's like trying to learn English without a functional understanding of the word "Please".
My take is that a lot of the intelligence we natively ascribe to humans exist in human languages and just like the nutrients in your food, by harnessing a language that possesses latent intelligence, you too can behave like an intelligent person. And how did the nutrients of intellect got in there in the first place? Just like languages evolve nowadays: via contributions from speakers over time. Not unlike soil fertilisation.
And language itself is so very 'stochastic' too, due to complex subjectivity. Every word floats on a variable sea of individual choice.
For example, if I compose your post as :
--
"Good prose embodies intellect gradually evolved by myriad humans over time, so it wows me to see LLMs create it, though I soon realised that the intelligence is not being created by the LLM but rather is already present in the human-evolved input datasets - the prose itself! - which they statistically arrange."
--
..the particulars of the language are entirely different (and omit some of your chosen nuance) but most readers would receive the same "core truth" from it, the same measure of beneficial information.
Which highlights the great fluidity of language; that is another of the key underlying mechanisms allowing LLMs to flourish.
1 reply →
> i agree. but it's uncanny that stochastic generation of language yields a form of recognizable intelligence.
It's funny, I used to think it was complete and utter horseshit when literature professors would say that the meaning of a text is created by the reader, not the author.
Since LLMs have become commonplace, I have completely reversed my opinion on that.
That you see recognizable intelligence in the output of stochastic generation is because you are imposing meaning on it. It is why "claudish" is so hard to read--you are trying to impose meaning on a sequence of words that was stochastically generated, and it takes mental effort to do that.
I'm not so sure about that.
If Claude can write formal mathematical proofs (that a machine can check), or it can write software that works, where's the head of the reader that imposes the meaning?
i am not an expert in any of this, but intuitively i largely feel the same... the character we see as "claude" is mostly in our heads; its almost like a "super rubber duck" phenomenon or something
Reading what you wrote, I must interpretively attempt to construct a model of what you may have meant and what you may believe, because on the surface it contains multiple claims that are frankly false in my experience (itself a probabilistic world model).
> a sequence of words that was stochastically generated
As was your comment (and mine, and the others on this page and elsewhere).
With Claude I know that there is no source meaning or belief, but it helps to impose Dennett's "intentional stance" and treat it as if there were, up to a point. That there is such a "point" allows be to manipulate it in my attempts to extract factual information implicit in its training data, expressed probabilistically via its stochastic processes.
I fail to see what exactly is explained, or clarified by the theory here. Postmodernist philosophers are notorious for writing texts which make readers struggle. Surely, it doesn't mean "Foucault-ese" is somehow more stochastically generated than "Claudese".
The existence of LLMs is telling us something, surely it means something. I'm not sure we've quite figured it out yet but its probably something along the lines of 'intelligence might not be as complicated as we thought' ... a fourth narcissistic blow to humanity after Copernicus, Darwin and Freud.
Freud doesn't belong in that group. Freud was a fraud.
1 reply →
We train them on the sum of human knowledge available in digital format, so I'm not sure that's an indictment, and the large models are incredibly complicated to the point it's hard to reverse engineer how they work so well and what they're actually doing to build internal models to predict the next token. Several billion to several trillion parameters is anything but simple.
2 replies →
The narcissism might not be in thinking 'intelligence might not be as complicated as we thought’ - the narcissism might be in thinking ‘LLM intelligence is the way we thought’.
5 replies →
> it's uncanny that stochastic generation of language yields a form of recognizable intelligence.
It most certainly does not. It's obvious to anyone who critically examines LLMs, without letting the hype override sense, that there's zero intelligence there.
To the cynic, Chomskyism is exactly what Thomas Kuhn warned you about "normal science".
That is, you can apply for grants and write papers in the transformational grammar paradigm endlessly. You can pose and answer questions that make sense within in the system.
Chomskyism was disappointing in two respects: (1) was that it was an inadequate basis for machine understanding of text, (2) was that structuralists were hoping everything in culture could be understood in terms of "... having the structure of language" but syntactic structures weren't all adequate for that task.
(1) and (2) both pointed to Chomskyism being hermetic and not something really useful or inspiring to other fields and maybe just a little less "true".
Actually, there have been a number of rounds of paradigm shift (or attempted paradigm shift) in linguistics over the past 70 years. The unique thing about linguistics is that they've been compressed into the lifetime of the same central figure.
Also, I'll mention that there is a lot of empirical knowledge about natural language syntax in generative grammar, even if the details aren't quite right about their theoretical underpinnings. And a lot of them were "low-hanging fruit" waiting for someone to point out how to find them and characterize them. That Chomsky did that is the best argument for his intellectual value - permitting all that "normal science" to happen.
Science doesn't guarantee technological applications, whether or not that science is correct.
Finally, just saying: use of a word like "Chomskyism" makes you sound like one of those creationists who call evolution "Darwinism".
So we can't use the word "creationism" without validating their use of "Darwinism"?
This is wholly fallacious ... you may disagree with what you interpret as the implications of attaching "ism" to Chomsky, but in no way does that equate it to attaching "ism" to Darwin -- each of these attachments and their validity are independent.
BTW, even what you see as pejorative in such an attachment is a cultural interpretation ... e.g., it is far more common among British evolutionary biologists than American to view "Darwinism" neutrally -- Dawkins for one uses the term and describes himself as a Darwinist. See, e.g., https://www.edge.org/conversation/richard_dawkins-chapter-3-...
The irony is that while (1) is currently dying a terrible death, (2) is on its peak influence of culture and politics
Officially structuralism had a crisis around the release of Syntactic Structures and the morphed into post-structuralism which has been through relentless post[n]-modernization and is whatever it is now.
> You can pose and answer questions that make sense within in the system.
Yes this is one of my critiques of the academic system (I have finished a PhD and will hopefully never work at a university ever again). Many papers in the introduction start with explaining how the study builds on plenty of previous research, but that alone doesn’t prove that something is a good direction to investigate. If academics were completely honest, many papers would write “We researched topic X because we got funding for it. But since we actually don’t care about X, we just wrote a proposal about our favorite niche topic that only one of the co-authors cares about, namely topic Y, and wrote it such that it sounds like X. Next we show that Y is actually very cool in the hope to convince others that it is cool plus to convince others of the conclusion that our most senior professor already found during the PhD which was 30 years ago. Apart from the professor, we don’t believe in the main argument of the paper but since he got the grant we’ll just go with it in the hope to one day be able to push our own ideas forward”.
> Norvig confuses engineering for science. If an engineer wants something that works (and guarantees that it will work) then you use the most effective tool available to achieve that goal. No problem there. But there's absolutely no reason to think that tells you anything scientific about the phenomenon you're modeling.
this distinction is superficial. Scientists use less precise models all the time. Like, you don't solve the schrodinger equation even when you're a scientist doing organic chemistry. You reach for lewis dot models. Science as a discipline is about building better models of the world, not uncovering fundamental truth.
> LLM may be a great technology, but it's probably not a good model for how humans learn and use language.
but from all models, it's the closest, "all models are wrong, but some are useful"
> but from all models, it's the closest
Care to back this up? What evidence do we have to suggest that back propagation, self attention, and byte pair encoding are mechanistically similar to human cognition? Similarity of the outputs says nothing about similarity of the machinery underneath. Turboprop engines tell us nothing about how birds fly - doesn't matter that they successfully transport millions of people around the world.
I would say that when it comes to higher level cognition, logic engines are more human like even if they only have a tiny fraction of the practical utility of LLMs.
> Turboprop engines tell us nothing about how birds fly
That is not true.
Turboprop engines and flying birds are different things, but there is a set of common properties.
A blade of a screw propeller functions exactly in the same way as a flapping wing of a bird, except that the wing cannot rotate without limits around the body of the bird, so it must stop, change angles and revert the motion, while the blade can rotate continuously around the axis, functioning on one side like a falling wing and on the other side like a raising wing. The similarities between propeller blades and insect wings are even greater, because insect wings cannot change their shapes so much as those of birds.
For now, our understanding of the networks composed of the neurons of animals is much less advanced than our understanding of wings and propellers, which is why there is a much greater uncertainty about whether there are any similar mechanisms in biological neural networks and in LLMs.
4 replies →
attention + residual seems kind of like Chomsky's merge operation.
LLMs behave like they're closest, but only during execution. During training, obviously, they don't resemble the human brain at all. No human reads a million books before they begin to speak.
And I'd be hard pressed to identify many cognitive characteristics of LLMs during execution that resemble human thinking, aside from those that are readily parrotable.
As a model of human cognition, LLMs deep suck.
Humans have billions of years of evolution going for them, that must count…
What reason to we have to believe that LLMs are the closest model we have for how humans learn language.
We considered the hypothesis that humans are a general intelligence that figures language out by exposure. The evidence has consistently come back that that is not how humans actually work. A model where humans are hardwired with a lot of knowledge about language, then fill in the gaps based on exposure still appears to be far more accurate.
The only thing LLMs show is that it is possible to learn language de-novo.
>closest model we have for how humans learn language
Is there some other competing model which can give results as good? If not, the statement seems true in the trivial sense that it's the _only_ model which we've observed to give good results.
1 reply →
> The only thing LLMs show is that it is possible to learn language de-novo.
I agree, but would add, "But not under conditions that humans learn language."
I think it's far too early to say LLMs are not how humans learn language.
Regarding "hard-wired"; it's not if you consider that "hard-wiring" of an individual is the result of millions of years of evolution.
Humans decidedly take much more data than LLMs. Like orders of magnitude. We happen to feed only distilled data to LLMs though. That's why they appear so efficient.
If you are counting by bytes and including all sensory input, sure. In terms of amount of text, it's not close in the other direction. It's very difficult to make an apples to apples comparison here.
You can obviously count by bytes. They are bytes for the model or for a person. Bytes to bytes is apples to apples.
You can't count by text, because text is not the only thing children learn on. So unlike the bytes comparison the text comparison is totally invalid.
3 replies →
The BabyLM challenge has shown that multimodal models don't seem to learn in a more data-efficient way with respect to tokens. Adding an unlimited amount of images doesn't help you learn syntactic islands faster.
Which entry are you referring to?
4 replies →