Mindreading with Jean-Rémi King

26:54We don't actually try to inject this knowledge. We just have these AI algorithms that we can use off the shelf, open source models, either produced by our colleagues or by the rest of the scientific community. And these algorithms, they're not trained to mimic the brain. They're trained for whatever other purposes, to be chatbots and to recognize cats from dogs in images. But what we observe empirically is that training these algorithms tend to make them generate representations which are comparable to those of the brain. Okay, first of all, that is scary AF. I mean, it's fascinating and it's really cool, but it's also kind of scary.

27:37Tell them what AF means. It's scary as... But the reason why it's a little scary is because, on the one hand, it kind of diminishes us as this crowning jewel in all of creation with the zenith of intellect that we believe that we hold. We ain't the zenith of anything. Right. Yeah. Couldn't we do with a little bit of humbling every now and again? I don't know about you. No, no. Here's how you get out of that. Here's how you get out of that. Go ahead. We are so brilliant. Right. We created something more brilliant than ourselves. I wouldn't say this quite yet because AI is really limited in many ways today in spite of the hype.

28:18I understand the emotional reaction, but frankly, I also think that there is a source of marvel here. For the first time, we have AI systems or systems that we trained for a task. The task is surprisingly arbitrary or even mundane. For instance, trying to predict the next word given the preceding words. I mean, that is what all LLMs do. Exactly, yeah. Large language models. Thank you, yes, large language models. And this simple task pushes the algorithm to generate hidden latent representations which resembles those that we have in our own heads. And that suggests something to me which is very profound, right?

28:57There exists general principle that push these systems, biological artificial systems, to generate a similar computational path, a similar set of representations. So is there a similarity between the brain and how it processes data in its architecture and that of a large language model that it's learning in a very similar way to the human brain? Because as I understand it, the original idea of neural nets as invoked in computers was an attempt to mimic what we thought our brain wiring was doing. and we learned that that's not really how our brains work. So it's just dangling there now as its own thing with its own utility, but it's no longer a biological analog.

29:43Yeah, so the history of AI and neuroscience intertwined quite a lot, but for a long time it was, these links were metaphorical. The idea of a neural network was, it wasn't, I think, a useful concept, but the goal was not to be as close to the brain as possible. In fact, it's really a huge simplification this idea of artificial neurons as compared to what was already known at the time. And you've had these bridges between the two disciplines for many decades. What is different now is that this comparison is not just like conceptual, like kind of loose. It's very precise. We can quantify the extent to which the activation patterns in the brain and the activation patterns in these AI systems do look alike or not.

30:29and even though these systems are not built for that purpose. Now, having said that, I also feel the need to mitigate these results because this is a tendency that we have, but we also see a lot of edge cases where this does not work. So typically if you take the very best model, the largest model, this similarity tends to break down. So we do have cases where what we call the convergence of representations between AI systems and the brain is not monotonous, is not systematically the same. All right. One thing I haven't sort of got to grips with, the speed at which image, back a brain, and how quickly that is and how quickly you're able to then process that data back through an algorithm.

31:16In the head? Yeah. In the head, it's quite slow, actually. So when you look at reading, for instance, you flash a word onto your retina. This takes for about 70, 100 milliseconds to really blow up the visual cortex in the occipital lobe. And from there, you'll get another 50 milliseconds for this visual information to be processed. A millisecond is a thousandth of a second. A thousandth of a second. So 50 thousandth of a second would be five hundredths of a second. Yeah, that's correct. I'm not good with math, especially not in my native language. We're messing in my head if you're not good with math.

31:52But so yeah, around 100 milliseconds, this is really when the activity peaks in the visual cortex. for the sensory processing, let's say. And then this information is being analyzed into edges that will eventually construct the representations of letters and of morphemes of words. And this is around 200 milliseconds, so one-fifth of a second. And then it takes another 200 milliseconds, so around 400 milliseconds, does the semantics part of words really rise in the brain and is broadcast to a wide variety of brain regions. And so this process is relatively slow. It takes about half a second for you to analyze the sensory.

32:34How fast would machine learning do it? Or let's say an OCR, how fast would it know? In terms of inference, the machine would be much faster. It would be just a few milliseconds. A few milliseconds to do the whole process, and we take like a full half a second. Absolutely. So we're basically like, duh. Well, at the inference stage, so what we... Stop giving AI ideas about what to do with us when it becomes our overlord. What is fast is what we call the inference stage, right? So once you already train the algorithm, using it is actually very fast. What is typically slow is loading the information onto the graphic card, but once it's there, it's actually very fast.

33:20However, training these algorithms is ridiculously slow, right? If you want to train an LLM today, a large language model today, you need trillions of words, which represents many, many, many lifetimes of just reading all of the text that we've created in humanity. That takes us back to what we were talking about earlier. So in order for it to know, it has to see all the words. In order for us to know, we just have to see a word and then something similar, and we're like, oh yeah, it's that. So that's what we're, for instance, a ball. If you show us a ball, you can show us one ball. We've never seen a ball before in our life and you show us a ball and then you show us a basketball and we'll say, that's a ball.

34:04And then you show us a baseball and we'll say, that's a ball. But the machine is like, well, I have never seen that before. So that's the difference. Yeah, this is one of the many differences. I mean, when we emphasize similarities, we emphasize the similarities is because we are in a field of differences. Everything is different. The architecture is different. The type of data that they receive is different. The training, of course, the physical insensation is different. This is also highlighting why we are all the more surprised and interested in the fact that in spite of all of these differences, we can still find similarities in the way they process information.

34:41Wow. And you've got, when you've got these sort of caps that you're putting on, they're sensitive enough to be able to operate at that sort of speed? I know you say it's slow, but that for me is really quite fast. So it depends on the device. So with functional magnetic resonance imaging, fMRI, you get a snapshot of brain activity approximately every two seconds. So a lot can go on within two seconds. However, if you take magnetor encephalography, you can get a snapshot every millisecond. So you'll get a much more well-resolved signal in time. but the spatial resolution now is much lower. So you tend to have a blurry image, let's say, of brain activity.

35:19So you have a trade-off between these different technologies. Wow. So it's kind of like cosmology. The better your looking tools get, the easier it's going to be. That's the general truth of science. It's just going to be so much easier for you to figure out anything you need to figure out. It's just a matter of we got to be able to see it faster and then know more clarity. I got a question. When I think of the brain, I think of it as organ. And there are these parts of the brain that are similar from one person to another, even if there's differences in detail. Are we to believe that the brain knows that in advance how it would divide up its territory?

36:02Or are we all just socialized the same way? We all grow up in a civilization, and so we all have the same influence on our developing brain for it to take the shape the way it does. So this is a very profound question. There is a tension. I'm not a historian of science, but there is a tension in the field that dates back to philosophy between empiricists, people who think that the structure of representations come from the data to which you are exposed, and rationalist, the people who would rather emphasize the importance of innate representations and innate structures. So you have, for instance, Plato on the one hand, if we take this back to ancient Greece, that would really be in the rationalist point of view with this idea that there exist innate representations in ourselves and ultimately we can approximate them with reasoning, whereas other people, and I think that the whole study of AI is really on the extreme empiricist side is just let's take a blank system and just press a lot of data onto it and ultimately this system will manage to perform a task.

37:11And what is interesting nowadays is to see that irrespective of whether the representations are innate or acquired through exposition, not necessarily culture, but even just sensory data, they seem to at least have some similarities. This is what I think is interesting in the case of this comparison between AI and the brain for language. The brain is obviously structured very differently to these AI algorithms. And obviously there must be some innate structuring in our brain. This is why only humans have language in a sense of being able to combine words together in order to reason and to communicate.

37:51This is not an ability. It reminds me of the New Yorker comic, two dolphins swimming together. Right. They're in a water show. Oh, okay, like the sea world type. The sea world, they're two and one says to the other, of the humans. They face each other and make sounds, but we're not sure they're actually communicating. Right, that's pretty funny. But there are a lot of experiments. Our brain's bigger than our brain. For some of them, not all of them. But the reason why there is, I mean, there's been a lot of experiments on behavior with dolphins, but also with apes, to try to see whether they would be able to combine concepts.

38:27And there are some experiments that show that in some edge cases, they are able to do this. But for now, we don't have any evidence suggesting that you have any other species that can learn this vast amount of concepts and be able to combine these concepts together in order to produce a sentence or to understand a sentence, a new meaning that they've never heard before. So this ability must be, to some extent, input in our genome and be an innate structure. Well, it has to be. I mean, we're the only—and it's so funny because it's disassociated from everything else that we are and have, language.

39:06For instance, I can be deaf, dumb, and blind, and you can teach me any language. I don't have to have an actual reference like everybody else does. So, I mean, we are truly unique in the way that we do communicate. I don't know if, I mean, I'm sure other animals communicate too. I'm not sure. Yeah, all animals communicate. But we're very unique in that if I don't know how to communicate with somebody I meet from halfway across the world, we will find mediums that allow us to know each other's language. Right. And this is coming back to the whole empiricist versus rationalist tension. This is why there is something very interesting here.

39:49So we established that the human brain must have some genetic or innate properties for it to acquire language. And this is why it differentiates itself in part from other animals. And we also know that these deep learning algorithms, they have very little what we call inductive biases. The architecture that we use in deep learning, they are remarkably blank and versatile. And so it's really the data with which they are trained that pushes them to build the representation that they have. And nevertheless, it seems to be comparable, at least to some extent, to those of the brain. Not in every way, but in some ways.

40:24And so that suggests that no matter where you come from, whether you come from this really rationalist type of approach to cognitive science or much more from an empiricist approach, there seems to be some sort of convergence between these two approaches. What I want to get into now is the application of your research, where it could go as we progress with this. Now, I said in the opening about people who can think but can't speak. Is there an opportunity with this research to give them a voice, to have their understandings made public, made aware? Right. So you have indeed a lot of patients who suffer from an inability to communicate typically because of a brain lesion, so either a traumatic brain injury or an anoxia, that will lesion the part of the brain which is responsible for, for instance, motor control.

41:20So they will be paralyzed or lose an ability to move their facial movements. And there are now a few teams that have shown that it is possible to put a set of electrodes in the motor cortex and to use these neural signals to feed in algorithms that can then be used to do a brain-to-text translation and allow the individuals to regain communication abilities. So this is already something which is happening with invasive approaches, so with electrodes which are implanted with neurosurgery. One of the goals, of course, is to try to see whether it would be possible to push this approach with non-invasive devices, which do not require brain surgery, in order to rehabilitate communications in patients, but also perhaps to diagnose.

42:09So sometimes you have patients who do not respond, but they are awake. It's a paradoxical state which occurs sometimes after a coma. And in these patients, you want to know whether they don't communicate because they're not conscious of the environment or whether they don't communicate because they are fully paralyzed, for instance. Or they just don't like you. And perhaps they just don't want to, which is actually an issue, right? If you have lesions to the part of the brain that are intrinsically linked to motivation, that could also be a cause of a lack of action. I'm tired of talking to you.

42:43I'm just done. And so for these patients, having devices that would allow us to, well, allow them to communicate, but even to allow us to know whether indeed they are conscious or not of the environment is certainly of a prime use. Are we likely to find that sort of ability in the near future or are we having to wait? For invasive electrodes, this is already happening. You know, our boy Elon. That's what he wants to do. Neuralink. Yeah, I want to put a chip in everybody. You can do that through the vaccines. Well, no, that's Bill Gates. Let's get our billionaire streak. Okay? Elon wants to put a chip in your head.

43:25I mean, an electrode in your head. Bill Gates already did it.

43:51I think the limits of the technology, as Neil has pointed out, will be reduced because of AI, and they'll find solutions sooner. But it's the ethics of being able to potentially sort of decode the brain's messages and then reverse engineer it so as you can read someone's mind. It's the ethics of that being possible because I think that's going to freak not just Chuck out. You know, I'm going to be honest though. Isn't that already happening when you look at metadata that's taken from our phones and our location and other phones that are around us? couldn't you pretty much tell what I'm thinking?

44:32Well, I don't know. But what I can say is these are certainly topics that come up very often. And there are several things to say. The first thing is that what is possible today in terms of decoding brain activity is really limited to specific cases like perception and motor control. And the reason for this is because we know what the person sees, so we can attach the image to the brain patterns. However, as soon as we try to do this in imagination, for instance, as we mentioned before, then things become drastically more difficult. Not just because of an inability of the algorithm to work, but really because a signal is just not there.

45:14That means it's not likely that you will anytime soon read someone's dreams. Until you get a signal booster. from a statistical point of view for fundamental research there is research on the science of dreams however there is all of the evidence point out to the fact that it will be very difficult with the state of knowledge that we have to have a device which can read your mind in the way that people think with your train of thought and all this and the reason for this is because even with the largest multi-million dollar type of devices that are being used the signals remain extremely noisy and it's very difficult to go beyond this.

45:58So the physics of the signal that we pick up is really the main constraining factor, not the AI algorithm part. So the AI algorithms can be used as a useful tool, but in terms of the signal that you can pick up, You can't generate the input necessary for them to do a good job. Yeah, the data that can be collected with these devices remains extremely, extremely noisy. And so from that point of view, the risks seem limited. Now, this is the current state of affair, but our role here as scientists is also to say what is possible, what is the state of the art, and to share this through the research, through open sourcing and all this.

46:38That's the reason why we do this work openly. In science, you're always limited by your signal to noise. Absolutely. So if the signals, you'd have to add up days, weeks, months of measurements to pull a signal out of that noise. But you want to have then, this only works if you have the same signal that comes up again and again. Whereas when people think of mind reading, they think of reading the mind at a given instance. You don't think of the same thing again and again and just repeat this until your noise averages out. So this is why currently all of the evidence suggests that there is not a systemic risk.

47:14However, technology continues to evolve and we want to make sure that the risks are limited and this is also why we engage in these kind of discussions, of course, to ensure that the discussion does not just happen within the scientific community but with the rest of the... So when is the time to make that determination? Is it now before you actually have the equipment to measure this? What determination? The determination as to the ethics, like codifying the ethics themselves. Guardrails going forward. When do you come up with those guardrails? because if you come up with them after you're able to do it, the horse is out of the barn, as they say.

47:51Yeah, absolutely. So this has already started, right? There is already a lot of regulations on what you can and cannot do. For instance, I work in France, and so we have the GDPR in Europe. It's constrained the way the data that is being collected from brain imaging can be used. In France, for instance, you're not even allowed to do neuromarketing. You're not allowed to use brain data for marketing purposes. So this discussion is obviously already engaged. And along the way, we need to continue and update these decisions with the state of knowledge that continues to evolve, yes. What was the movie Minority Report?

48:25Yeah, yeah. Where they had this sort of scenario. The precogs. Yeah. Precogs. I mean, that's... It's a great movie. That's everyone's default thought as regards this research and where it leads to. And I think that's what scares them. And I think they'd be grabbing not just for the guardrail. I mean, do you look to... Well, precogs, you were not digging out of their head what they saw from the past. You were digging out of their head what they foresaw in the future. In the future. Right, so that was different. So they would see you committing a crime that you haven't committed yet. That you haven't committed yet, but you were definitely going to commit.

48:55And then they'd just go arrest you. Right. They started doing that. The crime rate went to near zero. It's kind of like immigration in America right now. Oh, how interesting. You pre-arrest people. You just pre-arrest people. So what is the end game for Metta in this? I actually don't know why Meta hired me in the first place. I can only tell you what we're trying to do within our team. So the goal here is, well, the goal is well posed, right? We have now some preliminary evidence suggesting that you have similarities between AI systems and the brain. And that suggests something which to me is very intriguing, that there exists these general principles that shape the information processing in AI system and the brain.

49:41So discovering what those laws are and also trying to understand what is missing in AI systems for them to be as intelligent, as efficient as us remains a major topic of research. So this is why we're pushing on this frontier to better understand the brain and make better AI algorithms. So if you're able to achieve that, people are going to feel an invasion of privacy. They're going to feel thought security becomes potentially compromised. I mean, you said there's discussion over the ethical point of view. Are we looking at those sort of features as well? So, yeah, it's the same topic that we briefly discussed before.

50:18We have an ongoing discussion to try to see whether AI and neuroscience developments are changing the risks associated with, for instance, mental privacy. As of now, the discussion is ongoing, but I don't think we have a change in the technology that changes the risk. What we observe is that it is possible to decode brain activity in certain cases, typically for motor control or for visual perception, but it is not possible to decode what you are thinking at a given moment or your train of thought or to extract your password from brain activity. The reason for this is because the signal that we have...

50:56It's not the thing that just takes the password out of your head. Right, like the readers that they have now that steal your credit card. That's radio frequency. All of the physics on which we base this analysis prevent us to work outside of the lab, right? So with an MRI, you need to plunge someone into a very high magnetic field. This is not something that can be translated for, I don't know, a consumer product. But what you could envision is a dystopic future where the state who has the power and the money to actually have a machine that could read your brain and during an interrogation extract information from you that you don't want to give up, basically like you violated my mental privates.

51:43So that's actually foreseeable based on just what we talked about today. Yeah, I mean, if we go down to the dystopic possibilities, is I suspect that the states will not need an MRI to force you to give away your password. But it is an important... That's a good point. I'm just like, I'm not going to your MRI machine. I refuse. Just like, yeah, okay, yeah. This baton says different. But still, if the risk does exist, we should try to characterize it to ensure what is the path to this risk. And this is part of the scientific enterprise too, yeah. Cool, man. You've got some new research that you're about to release into the public domain.

52:27Can you sort of expand upon that for us, please? Sure, yes. So far, we've done this comparison between AI systems and the brain with adult participants. And to some extent, this is frustrating because there is something which is missing here in the picture, which is the learning process, right? So in the case of language, we don't just want to understand how the brain process language, but we want to understand what makes it able to acquire it so efficiently. Like with just a few words, we acquire language. The average number of words that we hear is typically around a few thousands per day, a few dozens of thousands per day.

53:07And if you compare this amount of data to the data which is input for the training of AI models, this is really a droplet of information compared to the oceans of data that these algorithms use. And so what this means is that fundamentally, the architecture or the training principle that we use for AI, they are really, really mediocre, right? We need to understand much better how you can get to a system that learns much more efficiently. So if you train your AI on children, you may end up learning how we actually learn or acquire language, but then you're also going to have AI saying things like, I hate you so much, I hate you, you never let me do anything.

53:55but certainly it would be important to understand not necessarily to train AI models with this data but to understand the principles that allow young children to acquire language so efficiently this is one of the big marvels of our species and this is certainly what we try to understand so this is actually a work with a hospital the Rothschild hospital in Paris that has a unit for epilepsy in young patients down to two years old patients. So you have these patients who suffer from intractable epilepsy. Again, same patient as we mentioned before. We have electrodes that are implanted inside the brain in order to identify the location that is generating the seizure and who can stay for about a week in the hospital and listen in that context to audio books.

54:45And then we can time lock the brain responses to each individual world to try to understand how the representations of language are processed in these young patients and how this evolves with age. Let me see if I can offer a perspective here. I'm as big a champion of AI as the next person, but I still enjoy being human, and whatever I can do to distinguish being human from a machine, I will embrace, leaving me to wonder whether the true creativity of what it is to be human may actually lurk within the noise that can be never read by a machine. The first person to paint an impressionistic representation of reality.

55:37Could a machine have had that first thought? or is that a human being rummaging within the noisy confusion of our own brains, pulling out something that no one had done before, no one had imagined before, and in the end genuinely creating that which is human and can never be machine? I just wonder. That is a cosmic perspective. And that was beautifully said, except the first impressionist was just some dude who was nearsighted. It's all just fuzzy. It's all just fuzzy. It's just how he saw the world. People were like, what an incredible interpretation. He's like, what are you talking about?

56:26So, Zan Ami, thank you for visiting. Oh, yeah. All right, Chuck, always good to have you, man. Always a pleasure. Gary. Pleasure, Neil. Thank you. Thanks to you and Lane and others. for coming up with these topics. They keep coming up with them, so we're going to keep finding them. We chase them down. Yep. All right. This has been StarTalk Special Edition. Neil deGrasse Tyson, your personal astrophysicist. Do keep looking up. And in.