David Eagleman, Neuroscientist and Author, on the Brain as a Neural Parliament, AI Hallucinations, and Rethinking Democracy with AI

00:02:56 · David Eagleman, Dan TurchinPlay from here

[00:02:56] S0: Great. Well, first of all, thanks, Dan, for having me here. I'm glad to check this off my list of things that I wanted to do. How did I get here? I, you know, I grew up in New Mexico outside the Albuquerque city limits in the middle of nowhere, and so I'm sort of surprised that I've made it out of there. But, my father had an extensive library in our house. He built stacks of books like a library, and so I spent my whole childhood reading, and I actually... As an undergraduate, I majored in British and American literature. That was my first love. And then I was always taking a lot of science, though. I was always invested in science, and, I finally discovered neuroscience my last year of college, and I was so hooked. And I thought this is the greatest thing, and now all these years later, I still think it's the greatest thing. So that's how I got here. [00:03:47] S2: We met recently at doctor Ben Goertzel's, good previous guest on both of our shows. I should have mentioned doctor Eagleman's also the host of the number one science podcast, Inner Cosmos, amazing podcast if you don't listen to it. But we met at doctor Ben Goertzel's SingularityNET event recently and that conference is intriguing in many ways. It's all about engineering a brain. Maybe let's just get started by having you share your perspective on should we engineer a brain? Can we engineer a brain? Great.

00:04:21 · David EaglemanPlay from here

[00:04:21] S0: Let me start with the can we. So my whole career in neuroscience, it it always felt like AI was a cool idea, but was really far away. And and that's because when when you look at the brain, you've got 86,000,000,000 neurons. Each neuron in the head is unbelievably complex. It's an entire cell. It's got the whole human genome in it. It's trafficking millions of proteins around in very complicated cascades. It's got all these little compartments where it's doing chemical reactions and little crucibles. I mean, a cell is as complex as a city. Okay. So what AI did when it started really rolling in the fifties and sixties is it said, look. Let's forget all that complexity and just assume it's a unit, and the units are connected because, of course, neurons are connected on average about 10,000 of its neighbors you're connected to. So so AI, said, look. Let's just get rid of that, make a neural network where you've got units and connections. You can change the strength of the connections. And so they ignored all of the complexity of the brain, and the truth is that most of us in neuroscience sort of sort of snickered at AI for a long time thinking this is just never gonna work that well. And then suddenly, it worked really, really well. Mean, there were, of course, hints of that. I've been following this field for, you know, decades. But, you know, it just started with Perceptron and then different sorts of ways of doing pattern recognition in neural networks and and deep learning and so on. And it was just getting better and better, but, obviously, everything exploded with LLMs. That's when all of us really had our hair blown back by this. And so that's got me... I I think... I mean, I'm just gonna speak for myself, but I suspect this is true for a lot of my colleagues that we felt like, oh, wow. Maybe this is all Mother Nature had to work with, with cells, which are inherently really complicated. There's like little animals in each of them. But but maybe it is some sort of simpler algorithmic thing that... Where you just get intelligence out of that. You get really deep pattern recognition out of it and so on. So there are many unanswered questions just from the point of view of the brain. We we don't know if AI is conscious. We don't know if the current way that transformer models are built, if that could be conscious.

00:06:50 · David EaglemanPlay from here

We don't know if they actually have any sense of an emotion, even though, you know, even though they certainly claim to. Anyway, there are a million unanswered questions. Of course, are differences too, like an LLM is unbelievably energy intensive, whereas the human brain runs on about 20 watts. And most importantly, a human brain can do continuous learning, meaning every second of life, I'm taking new information and learning stuff, whereas LLMs have a training phase, and then you ask the questions after that. So there are differences, but I will say I am mightily impressed by what has happened with AI. And and and there's something in neuroscience... For decades, we've been talking about this, the computational hypothesis of the brain, which is, hey. Could you replicate what's going on in the brain with a machine or in silico or whatever? Like, Dan, if I took your brain and I replicated all the important stuff out of beer cans and tennis balls, and it's running the same algorithms as your cells, then I should be able to say like, hey, Dan. How are you feeling? You say, oh, I'm, you know, I'm a little cold in here. Maybe I'm a little hungry or whatever. But it it should be you and perhaps even conscious and so on. That's the computational hypothesis of the brain. We don't know that that's right, but that's certainly the that's certainly the theory for us to chase all the way to ground level before we dismiss it. Anyway, what we see is a real boosting of that possibility by what's happening in AI. We see that, wow, if you really just capture the basic algorithms, something amazing is happening. Now the question of should we, I don't know. I think I'm gonna default on that question because I just, like, it is what... It's... It is what's happening. And so whether or not we should, that's clearly the way we're going.

00:08:47 · Dan TurchinPlay from here

[00:08:47] S2: You are a world expert in neuroplasticity, or you describe it as Livewired. Livewired. That's right. Not so much software or hardware. It's Livewired. And here's where I feel tension with the brain AI analogy. The LM essentially becomes static after it's been trained. All of the inference, all the compute that's happening at inference time is once these quote quote very much, finger quotes, neuronal pathways are are hardened, but that's fundamentally not the way the brain works. So it... To you, does that create tension in terms of that analogy?

00:09:26 · David EaglemanPlay from here

[00:09:26] S0: Exactly right. I mean, there are real differences between the way it's working and we're working. But to to my mind, it's, you know, like, clearly the proof is on the table that we're close, that we're figuring stuff out. But this issue about the LLMs not changing their weights, this is the issue about continuous learning. This is why they can't do the continuous learning that we can. Every moment that we are walking through the world from cradle to grave, our weights, our connection weights are changing as we take in new information. So there are these big differences, but boy, I am really optimistic that we now have good reason to think, Hey, you know what? This is computationally capturable, and we can surpass the human brain. I mean, it's already the case that, you know, just if if you had a friend that read every single thing on the planet, you could ask him about anything, you'd be really impressed with that friend. So even before we hit AGI, it's it's already already mind mind blowingly blowingly good. Good. Is recall equivalent to intelligence? You know, I don't think it is, but LLMs are doing so much more than recall. I think a big part of intelligence, or let me say for sure, human creativity, is the remixing of ideas. So what we do is we absorb things our whole lives, and we remix them, we bend ideas, we break them, we blend them, and we're coming out with new versions of things. That basic cognitive operation is what has led to our civilization. We build skyscrapers, we compose symphonies, we get off the planet, we do all kinds of stuff because we can take in the data and do something new, and LLMs do that too. They're really good at remixing information. And and, of course, you can prompt it anyway you want. You know, give me a poem about Godzilla standing on Mars and petting a cat, whatever, and it does it effortlessly. It's it's really good at this, and I think that is, one of its greatest strengths. In fact, just as a side note, a lot of people complain about AI hallucinations. I think often hallucination is the same thing as hypothesizing, which is what we do all the time. As scientists, we say, hey. What if this crazy thing were true or something? And then, you know, the only difference is that as trained scientists, what we do is we make a hypothesis, and then we follow it out. We say, Okay, what if that were true?

00:11:49 · David Eagleman, Dan TurchinPlay from here

Okay, this is what would happen, and then you see if it explains the data better. When Einstein says, Hey, what if I were writing on top of a photon? What would the world look like? I mean, that's that's great. So, anyway, I... When when I hear about AI hallucinations, I I sort of feel like, okay. Well, that's a good example of of remixing things, and maybe once in a while, get something good out [00:12:13] S2: of that. Most of the community that feels like we're rapidly approaching AGI, artificial general intelligence, also feels that LMs won't get us there and that there's maybe a neurosymbolic architecture or something post LLM. You... You're pretty complimentary of LMs. Do you think that this is the... I shouldn't say end state. That's not the right right terminology, but how far can we go with LMs?