Three things that LLMs have made us rethink
rodneybrooks.com
rodneybrooks.com
I used to be sympathetic to this viewpoint. But these days I'm more and more willing to believe that we can actually build a ladder to the moon.
[1] https://www.quantamagazine.org/common-sense-comes-closer-to-...
A lot of people having to eat crow due to how quickly things changed in just a few years
It is physically possible to build a ladder all the way to the Moon. A structure made out of close-to-ideal carbon nanotubes would have sufficient strength.
It would just require an unbelievable complex carbon nanotube 'weaving' system that can somehow operate and move directly between the Earth and Moon.
Which is quite an interesting parallel to the complexity of the human brain when I think about the analogy.
Would it maintain a stable position? I imagine it would wrap around the earth if it was fixed to the surface. But if it's free floating, wouldn't it want to fall one way or the other? Or spin somehow?
EDIT: this would be a good one for Randall Munroe's "What If"
EDIT EDIT: of course: https://what-if.xkcd.com/157/
Though after seeing the decades of hand coding that goes into a search engine it kinda proves moon ladders are very useful however incomplete.
That said, I have yet to see a modern AI that actually passes the Turing test either:
> an average interrogator will not have more than 70 per cent, chance of making the right identification after five minutes of questioning.
LLMs can answer some questions fine, but within five minutes you're likely to find breakage even without consciously trying.
This is not saying much when modern state of the art LLMs are being delibrately trained not to pass the Turing Test.
And his timeline turned out to be pretty close.
-- Bill Gates, in 1981
How do you know?
Now people argue that since only humans have complex language it was the peak of intelligence, and now computers can do that language like humans, then AGI is just around the corner.
Another way to think about it is that translation is not a question of "Express this English sentence in Chinese", but rather "Rewrite this sentence in the style of (some specific Chinese person)". You'll get a very different result if the target is "physics professors from Beijing" versus "rice farmers from Yunnan".
We don't understand "understanding" in the vast majority of practical ML systems that currently exist, let alone the physical implementation of "understanding" in our own brains. And yet, understanding "understanding" in ML has never been a pre-requisite to economic viability.
GPT4 can displace certain low-grade cognitive laborers regardless of whether or not we understand how it works, or whether or not its understanding is "true".
Clearly a single neuron in your brain does not understand English. Yet no one disputes that you understand English.
The non-chinese-speaker is just a component of the ChineseRoom system, that it doesn't understand Chinese is no more remarkable than a single neuron of yours not understanding English. It doesn't say anything about the overall system understanding Chinese or not.
I think it appeals to people with a more superficial understanding of computing that under appreciates the fact that instructions on paper processed by a universal circuit (or by a human emulating one) are morally equivalent to hardwired logic (if slower). So they disregard the contribution of the room, and focus on the fact that the meat-bot inside doesn't understand Chinese so the room must not either.
Honestly I don’t understand why people aren’t talking about AI all the time. Nothing is like we would have expected.
The seeming “understanding” and coherency of LLMs is a surprise, and it doesn’t seem like anyone knows exactly what it means about human language and cognition.
It’s possible that by analyzing what LLMs are actually doing we will discover a “code” underlying language that is sort of like a universal grammar, or at least gain some insight into how they stitch together and combine remembered snippets of language so sensically.
While what ChatGPT does seems more like “open-ended talking” than thinking, reasoning, and responding to its external and internal situation (which does not really exist, in the absence of plugins—there is just the talking), that talking can be used to (for example) plan a series of actions.
2. The Chinese Room argument can invite itself back out, as always. It was never more than an intuition pump and not a very good one.
3. Universal Grammar doesn't really figure into it one way or the other. LLMs don't work the way we do, despite being "neural". They get their basic grammar from a very different place. Chomsky's arguments are doing fine, though this is a great time to look at them through new lenses.
> I personally remember very heated arguments on the MIT campus between Chomsky followers and computer scientists working on statistical models in the early 2000’s. The arguments came up in computer science faculty hiring meetings. The Chomskians claimed that because of the need for universal grammar there would never be anything useful that came out of statistical approaches to language. ChatGPT has proved them wrong. (And yes, I personally shouted angrily at Chomskians in some of those hiring meetings.)
A related argument is that you can predict the weather using a neural net but that doesn't mean that the weather works like a neural network and it doesn't tell you anything about how weather does work.
There has always been a difference between actual understanding and just great presentation.
LLMs are so far ahead in their use of language that given a complete lack of understanding they are more capable than humans to fool the listener. But there's nothing substantially new, some people have mastered that skill, too.
People are regularly doing this in the titles of papers --- ascribing words like "emergence" and "world models" with definitions that don't match their informal usage -- purposely to mislead people into mistakenly applying conventional intuitions and make the paper look like it proves more than it does.
Or, at least, that we should say they don't think, because we're afraid what the alternative says about us.
You presumably use these systems all the time, are you seriously saying that they don't routinely make mistakes that humans don't make?
"LLMs don't think because they fail at X" is very different from "the current crop of LLMs don't think because they fail at X".
I have a hard time believing that someone who was shown a sample of these mistakes (somehow weighted by their frequency) would say that applying the word "understanding" to the system is not more misleading than not.
Someone who was shown a sample of these mistakes would maybe say that the word "understanding" is more misleading than not. I could make the converse argument for someone shown a sample of the successes, though. It seems unfair to judge "understanding" just by what the system can't do, and not by what it can do, while not really applying the same rigor to human babies.
This doesn't scream "not understanding" to me:
> If I'm holding three ice cubes, and I give one to a friend twenty minutes later, how many ice cubes do I have now?"
[... a bunch of correct text elided ...]
> In a practical sense, after twenty minutes at room temperature or in your hand, it's realistic to say that you might have no ice cubes left or just a very small amount of ice/water left from the melted ice cubes.
They don't understand precisely because they don't act like humans, which is the only source of intuition for the word "understand".
> I could make the converse argument for someone shown a sample of the successes, though.
It's the routine mistakes that "reveals" that there is no understanding. By your logic you can point to a single success and declare that as proof of understanding. No ordinary person would agree. A teacher who sees a student regularly make errors in reasoning on an exam doesn't say the student "understands" the material.
> It seems unfair to judge "understanding" just by what the system can't do, and not by what it can do,
I'm not talking about "fair" (whatever that is supposed to mean in this context) I'm talking about an ordinary person's intuition about the word "understand".
Or it was, until LLMs came along. Unless your definition of the word "understand" is a tautological "understanding is what only humans do because humans are the only ones who understand".
> A teacher who sees a student regularly make errors in reasoning on an exam doesn't say the student "understands" the material.
Ah, so you don't mean they're unable of understanding in general, but have not understood specific things they make mistakes about? Maybe we can agree there, although I'm still not sure that all mistakes are the same, in the same way that someone being bad at spelling physics terms doesn't mean they don't understand the physics itself.
The real problem that an LLM is trying to solve is to create a model that can enumerate all meaningful sequences of words. This is just an insane way of approaching the problem of intelligence on the face of it. There's a huge difference between a model of language and an intelligent agent that uses language to communicate.
What LLMs show is that the hardest problem - of how to get emergent capabilities at scale from huge quantities of data - is solved. To get more human-like thinking, all that is needed is to find the right pre-training task that more closely aligns with agentic behavior. This is still a huge problem but it's an engineering problem and not one of linguistic theory or philosophy.
Think about what is contained (explicitly and implicitly) in all the text we can feed a model. It's not just language, but a projection of the world as humans see it.
GPT-3.5 Instruct Turbo can play valid chess at about 1800 ELO, no doubt because of the chess games described in PGN in the training set. Does Chess suddenly become a language ability because it was expressed in Text ? No
An LLM could play chess though, all it needs is grounding (by feeding it the current board state) and agency (RF to reward the model for winning games)
No it's playing games. And if you're not at about the level I spoke of, you will lose repeatedly over whatever number or stretch of games you imagine.
https://github.com/adamkarvonen/chess_gpt_eval
3.5 Instruct (different model from regular 3.5 that can't play) can play chess. There's no trick. Any other framing seems like a meaningless distinction.
The goal is to model the chess games and there's no better way to do that than to learn to play the game.
>all it needs is grounding (by feeding it the current board state)
The Model is already constructing a board state to play the game.
https://www.neelnanda.io/mechanistic-interpretability/othell...
>agency (RF to reward the model for winning games)
Predict the next token loss is already rewarding models for winning when the side they are predicting wins.
And when the preceeding text says x side wins and it's playing as x side then the loss is rewarding it to do everything it can to win.
I agree different goals and primary rewards led to this ability to play and with it , slight manifestations(GPT can probably modulate level of play better than any other machine or human) but it is nonetheless playing.
The question implies a difference between human automatons who perform jobs and follow rules, vs intelligence in the sense of philosophy or generating original ideas beyond the confines of training.
It's like saying most humans aren't intelligent. Absurd. LLMs are absolutely capable of this.
The counterpoint tends to be around emotion and empathy. Besides being a very clear shifting of goalposts, this is already solved. We have several intelligent humans who have limited empathy and emotions. Are we to argue they aren't intelligent?
Some of our most brilliant people, and many people who just live their normal lives, are excluded from the premise of the question. I suspect it is rooted in a fear that human intelligence isn't exceptional.
AI models are already intelligent, the question is what humans we're comparing them to.
So far I've only seen examples based on tokenization ("what's the third letter in word?"), and the fact that answers are generated sequentially ("write a sentence that mentions its own word count"). But those feel unsatisfactory, like asking a human to decode ASCII.
I've personally seen some impressive feats from simple LLMs, like writing sensible code in a new language based only on a few surrounding lines. That shows a level of expertise beyond the common line "it's just mixing sentences from the training data".
Complex problem solving. There are only so many confounding variables that next token generation can handle
Anything requiring recent knowledge. RAG exists, but in context learning only gets you so far.
It could solve those, the reason it can't is that it doesn't understand its own limitations, ie it has no self awareness.
Humans plan things based around their own limitations, if an AI model can't even do that how can it be considered an AGI?
I don't think Humans understand their limitations very well at all. We plan around limitations we observe, not necessarily limitations that are present. There are many surprising limitations only recently uncovered by advances in neuroscience. I'm not sure humans would qualify as self aware by your criteria.
>if an AI model can't even do that how can it be considered an AGI?
By being artificial and generally intelligent. I don't see anything about "self-awareness" (ill-defined as it may be) anywhere in that acronym.
I said they need to be aware of their own limitations, that is very well defined. Humans might not be aware of all our limitations, but we are aware of a lot of them and that lets us work around them. Working around your own limitations is necessary for AGI, otherwise there are many trivial problems you can't solve.
It is possible that you can make many human crafted strategies and that goes far enough, but I doubt it since trying to make human crafted examples never has resulted in AGI advancements just allowing the bot to solve a specific problem well.
You seemed to imply that this "awareness" was some inherent ability of humans but it's very clearly not. If it were, we wouldn't be a lot more aware of these limitations today than even a few decades past.
and if all the records we kept vanished in an instant today then the next generations of humans would be a lot less aware. This "awareness" you are talking about is little more than accumulated knowledge.
By the realistic definition of 'aware of what is observed', i fail to see how LLMs don't demonstrate this.
And how much self awareness do humans have, exactly? We walk around with heads filled with false memories[1], accidentally ignore surprisingly prominent stimuli[2], and struggle to understand things as simple as material color[3]. And people are mostly unaware of these limitations, or those examples would not be considered surprising!
I think we plan around the limitations we observe, not our real limitations. If that's the case, then LLMs simply lack training data to notice those limitations.
[1] https://en.wikipedia.org/wiki/False_memory
[2] https://www.youtube.com/watch?v=vJG698U2Mvo (old selective attention test)
How would training data add an introspective step to the LLM? At best you can add a lot of common LLM fitted strategies developed by humans to the training set and thus make the LLM solve a few more problems well, but that will always be hampered by the need for humans to develop those. It is basically just programming the LLM.
My view is that LLM is results in natural language programming. Which is great, but it isn't any more intelligence than normal programming. Its just that we already have a lot of cool programs already written in natural languages, and now we get to run those for the first time.
That's not what that paper says
That is where LLMs are today, and why we can't solve the hallucination problem.
Absolutely they're doing more than just regurgitating training data, but there's a long distance from there to "any task a human could do".
and
"It’s like asking whether submarines swim. You want to call that swimming? Okay, submarine swim. What the programs are doing is, again, scanning huge amounts of data, finding statistical regularities enough so that you can make a fair guess as to what the next word will be in some sequence. Is that thinking? Does submarine swim? It’s the same question."
Even if it was only as smart as a dog, if you can duplicate it without limit in silicon, that is a highly transformative tech because electricity is cheaper than food. You can solve more problems per unit of currency. Maybe it would be constrained to certain simple problems but still, the AI would completely dominate that problem space and push humans out of it
This type of AI needs to be trained on human data. If humans are pushed out I don’t see it progressing too far
It's possible that in science, mathematics, games etc that AI could keep advancing without human input or feedback (and not hit a plateau where more human input is needed) because the universe or whatever axioms you choose provide all the necessary structure you need to keep making "progress".
But for cultural artefacts - music / art / literature etc perhaps not.
At least in the near term while we provide data, power and compute perhaps the most important thing humans do is set goals and rewards.
Now, in some ways this is understandable, Chomsky is interested in language in humans and how we learn with sparse data.
But in terms of machine learning and understanding, the machines don't have the same power limitations the human body does. If a machine needs more horsepower it doesn't need to have a million years of evolution to get to that state. If by applying a gigawatt of power I can process all written human works in a reasonable time and I can do that in a shorter amount of time then learning how humans do it with sparse data, then it's well worth the effort.
You're responding in agreement with Chomsky. That by spending a lot of money on Submarines, they can swim better than humans, then it's well worth the effort. He agrees.
But he says it tells you nothing about how humans swim.
A fairer analogy can be made by first supposing that submarines do not exist yet, and people are intensely interested if one could actually be built. Add in the situation that many people are claiming that such a swimming machine is inherently impossible because any machine would lack the swim-enabling "stuff" that fish have.
Then one day submarines are built. It passes all tests of swimming ability that naysayers said were impossible for a machine to pass.
And now the naysayers say that they weren't actually all that interested in the question of swimming machines in the first place. The actually interesting question is how do fish do it.
"do LLMs think" and "do submarines swim" is kind of the same question - it's a fixation on terminology rather than results. submarines accomplish something useful, and you can either quibble about what terminology to apply to that, or you can use them for the things they're good for and not be super bothered about what it's called.
it seems to me that humans have a vested interest in defining thinking as an innately human thing, so it's impossible to build a truly thinking machine, because if you do we'll simply find a new definition of thinking that excludes your machine.
Yes, if you are studying how fish swim it's completely irrelevant that you can build a machine to swim. Chomsky has been making this point for decades. He has never said you could not statistically approximate a natural process, in fact he has explicitly said that you probably could, it's just that it's irrelevant for scientific understanding.
I have an impression that, generally speaking, unlike how we know how an LLM operates, we barely have any clues on how humans form thoughts - or am I wrong here? Just some overall ideas that it's "probably somewhere out there, where fMRI shows the activity" but nothing concrete?
In terms of scientific understanding in general, I'm not sure why you would say that building a statistical model of a natural process is be irrelevant. On one level, much of science is about building models of natural processes and then probing those models for insights.
On another level, attempts to build artificial intelligence are acts of science. We are studying the limits of statistical algorithms in their ability to pick out patterns from raw data. This study has produced data that has already surprised those who had assumed that a competent language model would need a universal grammar component, as mentioned in the article. If a solid theory is developed, I assume it would yield insights for studies of human cognition.
It's possible but it doesn't happen by accident, you have to show that they are related, and no one does (because the people who work on LLMs aren't actually interested in this question, despite the fact that argue it to death).
> In terms of scientific understanding in general, I'm not sure why you would say that building a statistical model of a natural process is be irrelevant. On one level, much of science is about building models of natural processes and then probing those models for insights.
You have to have a reason to believe that it would yield insights. There isn't one.
> attempts to build artificial intelligence are acts of science.
They are not, they are math and engineering.
> If a solid theory is developed, I assume it would yield insights for studies of human cognition.
You have to actually have a rational process to produce insights, not just mix a bunch of numbers in a vat and "assume" that some insight is going to come from it.
The relatedness of human language cognition and machine language performance is already evident, in that studies of machine language performance can already falsify some theories of human language cognition. The pattern goes like this.
1. Language competency requires component X. 2. Humans have language competency. 3. Therefore, humans have component X.
When we see a machine, without component X, using language as a human does, we know the theory has gone wrong at the first step. The article mentions one such instantiation of this pattern. The new data has narrowed the space of plausible theories of human language cognition. The machine learning engineers play the role of experimentalists. They collect data but may not have a theory ready to explain what results.
> You have to have a reason to believe that it would yield insights. There isn't one.
The reason is that models in general yield insights. They have yielded insights before, and so we may expect them to yield insights in the future. The explanation for this is either philosophical or psychological. Whatever the reason, humans who look into a specific system are very good at generalizing to a whole class of systems. F=ma is a model, and certainly wrong. But by probing it, we can begin to discover general principles like energy conservation. In the same way, a large language model is a model of a system that acquires language. By examining its specific behavior, we may come to understand more general principles of language acquisition.
> You have to actually have a rational process to produce insights, not just mix a bunch of numbers in a vat and "assume" that some insight is going to come from it.
Where there is already an understood rational process, you no longer need the insight. Sometimes the big leaps of scientific understanding come from looking at the data and then coming to Eureka.
They do not.
> Language competency requires component X...
I assume X is something like "can't be learned statistically". LLMs haven't shown that in humans (which is what we are talking about): Beside the fact that humans use much less data LLMs can learn arbitrary patterns, including patterns humans can't learn (just as humans can't "learn" to understand a 56k modem).
> The reason is that models in general yield insights.
Prediction machines don't generate insights by default, that isn't even their objective, so if they did it would just be by chance. You might as well say that by mixing random chemicals you might get a discovery in chemistry. It's possible but absurd.
> Where there is already an understood rational process, you no longer need the insight.
The rational process I'm talking about is the process of inquiry not the process under investigation.
> Sometimes the big leaps of scientific understanding come from looking at the data
You could imagine something like this happening abstractly but it hasn't happened, no one has proposed specifically how it might happen and no one is pursuing anything like this because there isn't any obvious thing to pursue. It's just wishful thinking. Maybe if I do random stuff I'll make a scientific breakthrough. Maybe! But extremely improbable and silly to go even further and assume that you would.
Okay, but the article has Rodney Brooks asserting that
> I personally remember very heated arguments on the MIT campus between Chomsky followers and computer scientists working on statistical models in the early 2000’s. The arguments came up in computer science faculty hiring meetings. The Chomskians claimed that because of the need for universal grammar there would never be anything useful that came out of statistical approaches to language. ChatGPT has proved them wrong.
Worth noting that in the early 2000s, Rodney Brooks was director of the MIT Computer Science and Artificial Intelligence Laboratory, so... you know, he might have some context on that debate. Although I'll concede he attributes this to Chomsky followers, not to Chomsky.
- Innate Language Faculty: A “language acquisition device”
- Universal Grammar Principles
- Poverty of the Stimulus
- Language Acquisition as Rule-Based
From the article I understand that Chomsky is specialized in human language, and it seems to suggest that C does not believe that language exists outside of the human domain. I'm not sure if this is really the opinion of C, or if the article just makes it seem so.
Perhaps this language-learning model is really more universal than it is human-specific. We'll see when/if "Doctor Doolittle" models start to emerge eventually
That is in fact what he explicitly says. He thinks it's nonsense to look at an abstract concept of "digestion" for example that is divorced from natural phenomenon.
Other people are happy using these things even if they don't advance our understanding of cognition and language learning.
And if this current thread of development somehow results in something that functionally amounts to AGI, without advancing our knowledge of cognition, while that would be sad for Chomsky, it would nonetheless be revolutionary technology that would transform society.
I posted quotes of his indicating that isn't true. Because it isn't true. It's absurdly obvious it isn't true.
Everyone replying is talking about other things instead of the topic. LLMs have added zero to the learning about how humans learn language.
They've added other value.
For example, it says:
> Universal grammar enthusiasts have long argued that no other biological species can have language that has grammar and recursive composability, because they don’t have universal grammar.
That seems like a reasonable characterization of the universal grammar viewpoint as I understand it, and the existence of models which have acquired impressive language skills without a universal grammar substrate - by using instead a vast quantity of input data - certainly calls that view into legitimate question.
It may be that Chomsky believes that he never said the universal grammar was the ONLY way to acquire language, but a lot of computational linguistics research in the Chomsky school certainly took that as an assumption. An assumption that the success of statistical modeling with no universal recursive grammar substrate certainly calls into question.
Is there any empirical evidence that Chomsky's ideas are correct? Or is he setting himself up for embarrassment?
Also, infants acquire some elements of language extremely rapidly.
He is saying that LLMs tell you nothing about how humans learn language, and he is correct.
It tells you as much as a submarine tells you about humans swim. It doesn't.