It's a pleasure to have Adam, my colleague and friend, and Yan who's been with us before. Yan, you really are all over the news right now. I've gotten so many people forwarding articles about you this week. It all kicked off on Wednesday. Do you want to discuss it? I can just say the headline. The headline was the equivalent of Yann LeCun, Chief Scientist, leaves Meta. Do you care to comment?
I can neither confirm nor deny.
Okay. So all of the press that's here to get the scoop cannot get the scoop tonight. You can come afterwards and buy a drink and see how far you get.
Oh, really? I had one, but that wasn't the Frenchman had some wine upstairs.
So we have this era where every time any of us turn on the news, look at the computer, read the paper, we're confronted with conversations about the societal implications of AI. Whether it's about economic upheaval, the potential for political manipulation, or AI psychosis, there are a lot of pundits discussing this. It is a very important issue. I want to push that towards the end of our conversation because what a lot of people who are discussing this don't have is the technical expertise that's on this stage. I want to begin by grounding this in that technical scientific conversation. I want to begin with you, Yan, about neural nets. Here's this instance of biomimicry where you have these computational neural networks that are emulating human networks. Can you describe what that means, that a machine is emulating human neural networks?
Well, it's not really mimicry. It's more inspiration. The same way airplanes are inspired by birds. The underlying... I thought that didn't work. Copying birds with airplanes? In the sense that airplanes have wings like birds and they generate lift by propelling themselves through the air, but the analogy stops there. The wing of an airplane is much simpler than the wing of a bird, yet the underlying principle is the same. Neural networks are a bit like that: to real brains as airplanes are to birds. They're much simplified, but perhaps some underlying principles are the same. We don't actually know because we don't really know the underlying algorithm of the cortex or the method by which the brain organizes itself and learns. So we invented substitutes. Birds flap their wings, airplanes have propellers or turbo jets. In neural nets, we have learning algorithms that allow artificial neural nets to learn in a way we think is similar to how brains learn. The brain is a network of neurons interconnected with each other, and the way the brain learns is by modifying the efficacy of the connections between neurons. The way a neural net is trained is by modifying the efficacy of the connections between simulated neurons. Each of those is like a parameter; you see the number of parameters of a neural net. The biggest neural nets at the moment have hundreds of billions of parameters if not more, and those are the individual coefficients modified by training.
And how did deep learning emerge in this discussion? Deep learning came along the path after thinking about neural nets, and this has been since the 80s or earlier.
Yes, roughly the 80s. Early neural nets that were capable of learning something useful, at least in the 50s, were shallow. You could basically train a single layer of neurons: feed the input and train the system to produce a particular output. You could use those to recognize or classify relatively simple patterns, but not really complex things. People in the 60s realized that progress required training neural nets with multiple layers. They built multi-layer nets but couldn't train all the layers, only the last layer. Until the 1980s, nobody found a good way to train those multi-layer systems, mostly because the neurons were the wrong type: they were binary. Neurons in the brain are binary; they either fire or don't. People wanted to reproduce that with simulated neurons that would either be active or inactive. For modern learning algorithms to work, called backpropagation, you need neurons with graded responses. That only became practical in the 1980s. People had the idea before but couldn't make it work. That caused a renewal of interest in neural nets in the 1980s. They had been largely abandoned in the late 60s and then came to the fore again in the mid to late 80s. That's when I started graduate school in 1983. There was a wave of interest that lasted about 10 years, then interest waned again in the mid-90s until the late 2000s when we rebranded it into deep learning. Neural nets had a bad reputation in computer science and engineering, so we rebranded it into deep learning and brought it back. The results in computer vision, natural language understanding, and speech recognition convinced people it was a good thing.
Now Adam, you at a very young age were interested in theoretical physics, not specifically computer science, and you were watching some of this unfold from afar. What's the catalyst that sweeps up so many people decades later? There's this time where there's great interest, great success in handwriting recognition or visual recognition, but it's not sweeping up the world. What happens that brings us to this point where we're all talking about large language models?
So many physicists in the last years have pivoted from working on physics to working on AI. It really traces back to some of the work that Yan and others did to prove that it works. When it wasn't working, it was just this thing over there in computer science, not particularly interesting. Many physicists weren't paying attention. But after Yan and some of the other pioneers proved it would work, it became a totally fascinating subject for physics. You link up these neurons together in a certain way and suddenly you get emergent behavior that didn't exist at the individual neuron level. That seems like a subject that physicists, who spend their lives imagining how the rich pageantry of the world could emerge from simple laws, would find attractive. It immediately attracted the attention of many physicists. Nowadays it's a very common career path to do a PhD in physics and then apply it to an emergent system, but the emergent system is an emergent network of neurons that collectively give rise to intelligence.
Now let's do a lightning round because you raised the dreaded word intelligence. Everybody in this room very likely has interacted with something we're now calling an AI. These are all large language models. Before I ask you to define those, I want to do a lightning round of your yes or no response to certain things. Adam, yes or no? Are these AIs, these large language models, understanding the meaning of the conversations they are having with us?
Perfect. Yan's neurons are not stuck as binary values, isn't it? Exactly. It was my fault for giving you a binary choice. That allows me to ask the next question because it's not a foregone conclusion. Are these AIs conscious?
I think they'll one day be conscious if progress continues in the way that we're continuing. When is hard to say.
For appropriate definitions of consciousness.
Okay. We have some philosophers in the house, but we're not going to indulge in philosophical definitions of consciousness. Our hour would go and we'd still be here. I just heard that groan from our friends in the balcony. I have one other question. Actually, I have two in the lightning round. Are we on the precipice of doomsday or a renaissance in human creativity?
I have to throw this out to the audience, but I'll phrase it more colorfully. Will the robot overlords rise up against humanity? Yes? Hands up. Interesting. Okay. No? Hands up. Could you see that? It was a little more nose maybe, although the light is blinding. We'll come back and ask that again at the end. Here we are. These neural nets have been taught to execute a process we now call deep learning, and there are other kinds of learning that take off. What are large language models specifically, which have swept up the news and people's personal experience? We're mostly relating to large language models. What are large language models? Adam, maybe you could take that.
A large language model is something you've probably played with: ChatGPT, Gemini made by my company, various others made by other companies. It is a special kind of neural network trained on particular inputs and outputs in a particular way. At heart, it is mainly the kind of deep neural network pioneered by Yan and others, but with a particular architecture for the following task. It takes text in: the first few words of a sentence or the first few paragraphs of a book, and tries to predict what the next word will be. You take a deep neural network with a particular architecture and have it read, to first approximation, the entire internet. For every word that comes along on the entire internet, all the text data and now other kinds of data, you ask it what the next word will be. To the extent it gets it right, you give it a reward and strengthen those neural pathways. To the extent it gets it wrong, you diminish them. It starts off spewing random words. If you train it on a million words, it still spews random words. If you train it on a billion words, it might have started to learn subject-verb-object and sentence structure. If you train it on a trillion words or more, tens of trillions, it becomes the conversation partner you've hopefully played with today.
It strikes me as intriguing. Sometimes people get outraged at their chatbot when it leads them astray or lies. I've said it doesn't need to be words; it might as well be colors or symbols. It's just playing a mathematical game and therefore doesn't have a sense of meaning. I know Adam sort of objected to my summary. Do you think they are extracting meaning in the same sense we do when we compose sentences?
They're certainly extracting some meaning, but it's a lot more superficial than what most humans would extract from text. Most human intelligence is grounded in an underlying reality. Language expresses phenomena or concepts grounded in that reality. LLMs don't have any notion of the underlying reality, so their understanding is relatively superficial. They don't really have common sense the way we understand it. But if you train them long enough, they will answer correctly most questions people will ask. They collect all the questions everybody has ever asked and are trained to produce the correct answer. There will always be new questions or prompts for which the system hasn't been trained and might produce nonsense. So in that sense, they don't have real understanding of the underlying reality, or they have understanding that is superficial. The next question is how do we fix that?
I could play devil's advocate and say, how do I know what a human is doing is that much different? We're trained on lots of language. We get some dopamine hit or reward for saying the right word at the right time in the right grammatical structure. We back propagate and try to do better next time. In some sense, how is that different from what a human is doing? Maybe it's the sensory experience of being immersed in the world.
A typical LLM is trained on tens of trillions of words. It would take any of us something like half a million years to read through that material. Compare this with what a child perceives during the first few years of life. Psychologists tell us a four-year-old has been awake a total of 16,000 hours. There's about one byte per second going through the optic nerve. Every single fiber of the optic nerve, two million of them, is about two megabytes per second getting to the visual cortex. Over 16,000 hours, that's about 10^14 bytes. A four-year-old has seen as much visual data as the biggest LLM trained on all text ever produced. There is way more information in the real world, but it's also much more complicated: noisy, high-dimensional, continuous. The methods used to train LLMs do not work in the real world. That explains why we have LLMs that can pass the bar exam or solve equations, but we still don't have a domestic robot. We don't have level five self-driving cars. We certainly don't have self-driving cars that can learn to drive in 20 hours of practice like any teenager. Obviously, we're missing something very big to get machines to the level of human or even animal intelligence. Let's not talk about language; let's talk about how a cat is intelligent. We're not even at that level with AI systems.
Adam, you impart more comprehension on the part of the LLMs at this point already.
I think that's right. Yan is making excellent points that LLMs are much less sample efficient than humans. A human or a cat is able to learn from many fewer examples than a large language model. That is true, and it is a thing that is better about the architecture of animal minds compared to these artificial minds. But sample efficiency isn't everything. We see this frequently. Before large language models, when we built chess bots on top of large language models, they would play each other at chess a huge number of times. They played as many games as a human grandmaster has played and were still making essentially random moves. But silicon chips are so fast and parallel that they were able to play many more games than any human could ever play in their lifetime. They reached and surpassed the level of human chess players. They're less sample efficient, but that doesn't mean they're worse at chess; they're much better. So too with understanding. It is harder to get them to the same level with more samples, but once they've reached that, can we use the fact that they are more general and faster to push beyond? A cat is more sample efficient than a human; a cat learns to walk in a week, while a human takes a year. That does not mean a cat is smarter than a human or a large language model. The final question should be: what are the capabilities of these things, and how far can we push them? On almost every metric except sample efficiency, we've pushed large language models far beyond the frontier of cat intelligence.
I don't understand why we're not making cats.
The LLMs in question have much more accumulated knowledge than cats or even humans. We have many examples of computers being far superior to humans in tasks like playing chess. That's humbling. It just means humans suck at chess. We really suck at chess and Go even more. LLMs can accumulate huge amounts of knowledge and can be trained to translate languages from a thousand languages to another in any direction. No human can do that. They do have superhuman capabilities. But the ability to learn quickly and efficiently, to apprehend a new problem and come up with a solution, and to understand how the world behaves is still out of reach.
We've had recent successes where it's not the case that they're just looking up answers in a lookup table or pattern matching at a low level. They do pattern matching at a sufficiently elevated level of abstraction that they're able to do things they've never seen before that no human can do. There's a competition each year called the International Maths Olympiad, for the smartest finishing high school math teenagers. They're given six problems each year. The pinnacle of human intelligence. I have some mathematical ability; I look at these problems and don't even know where to start. This year we fed them into our machine, as did a number of other LLM companies. They took these problems they'd never seen before, completely fresh, combined different ideas, and got a score better than all except the top dozen humans on the planet. That's pretty impressive intelligence.
The question is, do they understand? We can look at the mathematics of the model, but it is a black box. It's not as though we can do that with the human mind either. To some extent it is a black box, but we presume it's doing calculations, moving matrices, working in some vector space. Are you describing understanding as a behavioral trait where it gives the right answers, or whether it deeply at the neural level understands?
I don't know if I understand that at the human level. I can't tell you what process I'm executing either, but I have some intuitive subjective experience that I understand the conversation. When I'm talking to you, I feel you are understanding. When I'm talking to ChatGPT, I do not. In my opinion, it is understanding. There are two pieces of evidence. One, if you talk to them about difficult concepts, I'm frequently surprised. With every passing month and new model, I am more surprised at the sophistication with which they discuss things. I encourage everyone to talk to these models. When science fiction writers imagined a Turing test-passing machine, they imagined it hidden in a basement. The way it worked out is we hooked it up to the internet and anyone can talk to it. Talk to these things and explore in areas you know to see both their limitations and their depth of understanding. The second piece of evidence is that they're not exactly a black box. We have access to their neurons, much better access than with a human. We can freeze them, replay them, write down everything, prod neurons in certain ways. In the field of interpretability, we see that when you feed them a math problem, there's a circuit that computes the answer that we didn't program; it learned to do that while predicting the next token. It learned that to most accurately predict the next word, it needed to figure out how to do math and build a proto circuit inside.
