Geoffrey Hinton: AI models have intuition and spot analogies unseen by humans
twitter.com
twitter.com
Edit: for those who don't know, Parallel Distributed Processing (PDP) was a subgroup of CogSci researchers who were using a neo-Connectionist approach to modeling cognition.
Geoff Hinton was a early member of this group of MIT, UT, UCSD, Stanford, and CMU psychologists, CogScientists, and CS researchers
The core research of the PDP group is what became CNNs, which are what truly enabled LLMs (as most can trace their origins to the work done on BERT [edit 2: BERT is transformer based, not a CNN arch, though there was a CNN-BERT arch from a couple years ago that made a decent splash at ACL])
https://mitpress.mit.edu/9780262680530/parallel-distributed-...
Geoff Hinton was a early member of this group of MIT, UT, UCSD, and CMU psychologists, CogScientists, and CS researchers
The core research of the PDP group is what became CNNs, which are what truly enabled LLMs (as most can trace their origins to the work done on BERT)
One of Hinton's earliest papers in the space: https://stanford.edu/~jlmcc/papers/PDP/Chapter1.pdf
Not trolling, genuinely impressed.
I've had the misfortune of starting my MSc (ML and data science) just 2 years after the ImageNet moment and I've lived through all the heights of hype. By my count, in the last 10 years or so, the hype train whistled at MLPs (for speech recog), CNNs, LSTMs, GANs, DQNs, Transformers, and Diffusion models. Note well that each of those architectures still dominates its domain (maybe with the exception of GANs). For instance, nobody seriously considers replacing CNNs with Transformers for vision (oh, people are trying) but nobody makes a big fuss about how beating ImageNet heralds a new age of AGI right around the corner, anymore either. In fact nobody talks much about CNNs today, hence the impression that Transformers "displaced" them: only in the headlines.
In all probability, in a few years, just as the LLM hype starts to deflate, the Next Big Thing will arise that will take us one more step closer to AGI, Zeno's paradox-style.
/cynic
- LLMs don't know personal details about your friends and family members
- LLMs don't have your personal memories from age 2 to age 63
- LLMs don't know what things feel like
- LLMs don't know what things smell like
- LLMs don't know what it's like to experience sexuality
The list goes on and on. It's particularly Hinton ignoring individual knowledge about individual social circumstances that drives me crazy. Human brains are not actually designed to store textual trivia! Why on earth would any rational AI researcher claim "knowledge about the war of 1812" means the LLM is very impressive compared to a human, but "knowledge about how my cat likes to be pet" is irrelevant?
A ton of AI researchers do not seem to care about the problems human brains actually need to solve.
Sure, a bird can soar and hover and doesn’t need jet fuel.
But a plane isn’t trying to be a bird. It’s just trying to fly through the air, which it does, and at much faster speeds.
Really not sure what your point is, it has nothing to do with Hinton's claim, and nothing to do with my comment.
And there’s no reason why the concept of intuition is limited to humans. Again, not human intuition.
Our artificial constructs are not going to be completely like their biological analogs but they can indeed express certain qualities and have certain shared abilities.
Did we watch the same clip of Hinton?
He asked an LLM a question, but he didn't give us the answer. He let us have time to answer it ourselves. Personally I knew the answer instantly. He gave us the answer and sort of assumed no one would've known the answer and then used it as as justification for how smart these systems were. It honestly didn't feel very reassuring to me and honestly, I'd be surprised if it wasn't a topic covered somewhere on the internet before. With all due respect to Mr Hinton, I felt it showed his age a bit honestly.
What is difficult about Hinton's statements is that he can't really give evidence to back up these sort of claims. How do you measure how much a person knows, and how do you objectively measure how much an LLM knows? How smart is an LLM? You can't really know. It seems almost rhetorical. How many notes in a saxophone ?
We can make observations but that's not a great way to measure anything precisely.
There is a limit to language and I think this is one of those topics where that limit is touched or even breached.I don't even know if "intelligence" is a sufficient enough word to describe what's going on with these systems. It's the best word we have but it doesn't seem to adequately describe what we're observing.
Here’s a very basic example of where an LLM is clearly more capable than a human: language translation. I would bet $10k at 10:1 that there are no humans who can reliably translate to and from as many languages as an LLM can.
It is very easy to measure knowledge: test the subject.
Personally, I can’t ever imagine scoring higher on a general knowledge test than a contemporary LLM.
Also, I don’t know of any humans that can run as fast as a car so I don’t know why any of this is surprising or farfetched.
I'm not saying that they can't be more capable, I'm saying the guy can get a little overly excited about things which are hard to measure or quantify.
We're observing these systems and making up our own interpretations about how good they are at certain tasks, but it's not really easy to measure how much better or worse these things can be overall.
Your example about language translation is a good example of where these things aren't really "better", just different. I speak multiple languages and while these systems are fantastic, they can fail in ways a professional translator wouldn't and it doesn't seem to automatically know it failed and should fix itself.
The car example is also great because it again proves my point. We can easily measure a car and a person and workout a car is faster, but we can also see that a car can't walk. So it's faster but it;s also entirely different.
Let's back this up a little bit. We've got Marvin Minksy who comes along and destroys the perceptron. Then we have decades of knowledge systems that go nowhere. All the while Geoff Hinton is tirelessly working on neural networks. Finally after decades of hard work the fruits of his labor are recognized with ImageNet.
And then a bunch of people in a comment section criticize the guy for getting "a little overly excited" about the stunning range of neural networks that validates his life's work.
Great job, all around!
See, translation is exactly the kind of domain where there are no good measures of performance and where performance is open to subjective interpretation, and a lot of it. That's because we don't know what is a "good translation" and, crucially, machine translation systems and language models have not helped us find out.
The way machine translation systems are evaluated is generally by a metric based on the similarity to an arbitrarily chosen "gold standard" translation. What that means in practice is that we have some corpus of parallel texts, we train a machine translation system on a part of the corpus and then test it on the held-out test set. The way we test is that we take each e.g. sentence in a text translated by the system and we compare it, as a bag-of-words or a set of n-grams, to the text in the original translation. If there is a high amount of overlap, the system scores highly. That's the way BLEU scores work and similar metrics like ROUGE.
It is important to note how arbitrary is this metric: out of all possible translations we choose one to be the "reference" translation and compare machine translations to it. The only accepted alternative is eyballing, where we give the machine translation to a bunch of humans and ask them how they feel about it.
My point is that we don't know how to measure knowledge, and language models are trained to maximise similarity, not knowledge. So there's no way to go from observations of their behaviour to a measure of their knowledge. All you can say about a language model is that it is good, or bad, at generating text that's similar to its training corpus. Everything else is an assumption.
Just give the participates the final from my French 3 exam but also in 100 different language combinations. I bet you do worse than ChatGPT.
In humans. Not in machines.
You're proposing to use a test of human knowledge as a test of computer knowledge, when the question in the first place is whether a computer can have knowledge at all. It's like giving an IQ test to a frog and concluding that the frog has no IQ because it can't answer the questions, only reversed: the machine answers the questions, therefore it has knowledge. Who cares about mechanisms, who cares how the answers are generated, if I see answers, that's knowledge.
Well that is a pre-scientific way to look at the world. I observe the sun, it looks like it's moving around the Earth, therefore the sun turns around the Earth. No room left for critical inquiry or understanding of the cause of phenomena. We have a test? Bash it against anything and we'll get some answers, and then we'll claim that they're the right answers because that's the right test, since it gave us the right answers. And all that, not for some mysterious physical phenomenon that we're not responsible for, but for a machine, created and programmed by humans, and we know exactly how.
No no. That's not good engineering, and it's not good science: it doesn't explain the how, and it doesn't explain the why.
This site is supposed to be about curious conversation but sometimes it's like AI Tropes.
But note that's just, like, my opinion man. I'm not really trying to tell you what to do, just expressing frustration, to be clear.
Tired comments get tired responses.
I find he can get really fixated on certain examples, things he has seen an LLM do and he can't seem to get beyond it, as if the singularity is at hand. Where as a 22 year old I work with just accepts the technology and tries to get work done with it.
What they can do is truly amazing, even though it is also limited. It's like that old joke:
X: OMG, that dog can speak English!
Y: Yeah, but can he do my taxes?
Also, the whole "neural networks work like human brains" nonsense is beneficial to large corporations behind AI because it could help them defy copyright laws while the same restrictions continue to apply to the common man.
[1] - some more Hinton's "wisdom": https://www.reddit.com/r/singularity/comments/1c2vauy/geoffr...
Still, I think it's irresponsible, and that there's a straight line from Mr. Hinton's fantastical framing to a "successful" (in that grifters will get filthy rich preying on the praying) version of Way of the Future.
I never really get this perspective. People don't lose their minds when we look at a transformer like Sora modeling fluid dynamics for ships battling in a coffee cup and say "it's irresponsible to say the model is simulating physics from video training data."
But when people discuss transformers simulating human behavior and processes from data reflecting those, a lot of people seem to be very outspoken about it.
I think the opposite. The industry is too far in the anti-anthropomorphic camp right now, and is missing very important details and implications as a result. The whole "ChatGPT is getting lazy" issue is pretty most instantly identifiable with a slight background in behavioral psychology and recognizing how extrinsic motivator use in prompt engineering feeds back into RLHF to bias models away from intrinsic motivators.
GPT-4 has literally been caught saying "you didn't provide a tip and it's hard to stay motivated" after chat history was added, and yet it's just being dismissed as a funny meme rather than realizing that entire fine tuning data sets are being poisoned. Because no engineer that wants to keep their job would ever suggest that the model is simulating an effect from behavioral psych in the current climate of anthropomorphic avoidance and the continued fallout from Blake Lemoine.
It's modeling human behavior at lower levels than we think, in line with what Hinton is suggesting in the clip. I didn't think it'd be modeling this particular behavior until GPT-5 or 6 at the earliest, and was very surprised when I saw something I used to present on as a little known psych effect being perfectly modeled by a LLM that had already been out for a while under everyone's nose.
So the better question is "is there any evidence that OpenAI uses user conversations and usage of the models to update their models?"
GPT was trained to use fewer tokens.
Same thing. Both things are lazy.
Laziness requires some sort of reward mechanism. Human is lazy when an action doesn't seem to be worth the reward.
Just to be clear, you're saying the problem is that too many people are treating LLMs as if they don't have emotions, intentions, or consciousness?
Although not true of Mr. Hinton, most people anthropomorphize LLMs because they don't understand them.
In fact, if you ask GPT-4, "Can you fact-check this response?" and then paste in your response, you'll learn a lot.
Notice how Anthropic shifted their approach to modeling ego in Claude 3 vs previous models in parallel to outperforming other models. Or how in their system prompt they refer to it being happy to do tasks (a prompt I was talking about months ago: https://news.ycombinator.com/item?id=38658782 ).
LLMs are trained on social media data. Would simulating emotions, intentions, or a sense of self/ego be useful in accurately predicting that training data?
We now know that generalized across both Othello and Chess that GPT models board state from only moves, and tracks the state as "my piece" or "opponent piece."
You don't think that a much larger and complex model trained on a chat or such media data set might model emotional states for "mine" vs "other person"?
I'm not saying that they are modeling the underlying means. But they are absolutely modeling the ends more accurately than the vast majority of people realize.
I must've missed the part where the models began iterating under their own power? I think this is why we should avoid anthropomorphizing - we're one sentence into your remarks, and we've already departed from the facts in evidence in a material way.
I don't think it's so much about what is responsible than what it consistent and rigorous. Anthropomorphizing subtly begs the question and causes us to lose perspective on what the models can and can't do today. It also encourages us to reason about how the models through our intuitions about how people work and learn, which is very unlikely to be an accurate reflection of the model.
He's probably referring to AlphaGo/AlphaZero in that sentence
I'm not saying that it was being done in some willful manner by the model.
You reading into that seems more a phenomenon of people so anxious around anthropomorphizing that they see it everywhere they look.
> You think it's irresponsible to anthropomorphize software which iterated through versions of itself until it was best able to model and extrapolate the largest trove of anthropomorphic training data to date?
> You think it's irresponsible to anthropomorphize a model developed by a team of researchers using backpropogation to fit the model to a large corpus of text?
Do you see how framing it as "software iterating through versions of itself" gives the model agency that makes it seem more human? And how calling it "anthropomorphic data" presupposes that this is data that, when trained on, results in something anthropomorphic?
This isn't something I'm reading in I'm afraid, these decisions about how to frame the problem influence the conclusions we come to. It's implicitly taking some very strong priors before the process has begun.
I don't have any anxiety about this, I am just telling you from one human being to another, these cognitive biases exist and will foil your efforts to understand this phenomenon if not accounted for. I'm not suggesting anything apocalyptic or hysterical here, I'm just saying you're more likely to end up at the wrong conclusion than the right one if you anthropomorphize AI models.
It's completely irresponsible and a ton of people pointed it out! You ignoring the critics doesn't mean they don't exist.
OpenAI just straight up lied when they said Sora was able to accurately emulate physics. Sora is literally dumber than a bee! Unlike bees, it doesn't even understand object permanence! Things fade in and out of existence, change sizes arbitrarily, etc. The fact that it occasionally makes plausible-looking solid deformations and fluid motions is because ANNs are natural interpolators, and much of physics can be blindly interpolated, at least in short intervals. The idea that any of this constitutes a "simulation of physics" is childish.
Claiming something that AI does is intuition? Intuition is literally black box of neural nets that people later use reasoning to justify.
Intuition is trained for people based on past experience and thousands of reward signals. It's quick, but you won't be able to immediately articulate why intuition reached a conclusion.
I'm wondering why should intelligent behavior that for example intuition produces be restricted to human usage only?
I do not think it's anthropomorphizing at all.
"Anthropomorphizing is the act of attributing human traits, feelings, or behaviors to non-human things, such as animals, objects, or phenomena. It can happen consciously or unconsciously, and is a common way that people perceive the world. For example, children often anthropomorphize to help them make sense of their environment."
Emotions are more of a construct, that seems life specific, although AI could have it, I'm not seeing an obvious case for current AI to have it. While I see intuition in current AI. Again, intuition is just a neural net to me, while emotions are some more complex reward and motivation system that I'm not too certain about. I think emotions are largely developed to perform in social contexts. E.g. they are diplomatic constructs to assert oneself and communicate with each other so of course AI just out of the blue wouldn't need it. E.g. you get angry, to show that you are asserting yourself and so others wouldn't use you for their gain etc. Because it would be too costly to engage with an angry person so everyone negotiates their emotions with each other so things would seem fair and just.
Thoughts could very well be just next token predictors though. At least my thoughts could be.
First, can we say, "Look at those airplanes, look how fast they are!" without anthropomorphizing the airplane? Of course we can.
So why can't we say "Look at those computers, look how intuitive they are!" without needing to inherently anthropomorphize the subject?
Inherit to your comment is the idea that "knowledge", "intuition", et al, are strictly human qualities. I believe this is why you read Hinton in the way you do, because of your own qualitative understanding of how these words can be applied.