The AI-powered Era of Scientific Discovery Is Here - Ep. 25 with Dr. Bradley Love

[00:00] Imagine explaining quantum mechanics to a dog. Quantum mechanics is a very well-specialized theory, but it's just beyond a dog. And like, you know, maybe at some point we're the dogs, you know, like we're not... [00:10] can understand what's going on. And we build tools, we can already see beyond our eyes with telescopes, and we can add up numbers we can add, you know, with calculators and computers, do all these amazing things. So maybe like, it's just going to be beyond us. And we'll have to accept it and take different forms of explanation that are much more general. Or maybe we're just going to be in this predictive world. And it's not even dystopian, like everything could be better. We'll just have stories about how this stuff works. So like, maybe scientists will be like [00:40] things and like making it explaining it to how it how it works and providing meaning to the systems. [01:00] Bradley, welcome to the show. Yeah, thanks, dude. Thanks for having me here. [01:03] Yeah, I'm super excited to have you. So for people who don't know, you are the Professor of Cognitive and Decision Sciences in Experimental Psychology at UCL. And you are also building one of the main builders of a large language model that is focused on helping people do better neuroscience research called Brain GBT. And I'm super psyched to have you on. This is going to be like a slightly

[01:33] deep into science and using AI for science and sort of like how science might change as a result of AI. But yeah, super psyched to have you. [01:41] No, I'm excited too. And like you hinted at, I think there's larger ramifications that go beyond neuroscience where we developed this model for that should affect all your listeners. [01:52] Cool. So like, tell us, like, let's get started with BrainGPT. Like, tell us about what it is. [01:56] Sure. I mean, first maybe I'll give some of the motivation. So unlike probably a lot of your listeners, I was never really a big tool development person. [02:06] But I just saw how science was going and it's exponentially increasing literature, literally increasing and you just can't keep up. It's just not really a human... [02:18] readable literature. And so just kind of getting a grip on it, it seems like we need tools to do that. And [02:26] People are making all kinds of tools, particularly using large language models, but we kind of had a different take on it. So there's a lot of great work that we call like backward looking, which is not to be pejorative. It's work involving summarization of the scientific literature, kind of writing instant reviews or almost meta-analyses in cases. And that's great. And that's valuable. [02:48] But we wanted to focus more on like what I think is really important in science, namely prediction. So what we call forward looking and [02:56] uh can we actually predict the outcome of studies before they happen [02:59] And so we just really have this project, and I'm happy to dive in as far as you like, but we wanted to see can large language models

[03:09] both just off-the-shelf models, but also models that we fine-tune on 20 years of the neuroscience, scientific literature, can they actually predict the results of neuroscience studies, experiments, better than human experts like professors of neuroscience? I mean, in short, there's a lot to say, but in short, they can. They're a lot better. [03:28] That's fascinating. Wait. So just to go back to the original motivation, it sounds like the first thing was like, okay, there's like way too much information. There's way too much science being done for anyone to keep up. But it sounds like the thing that you're building, BrainGPT, isn't necessarily about summarizing the research. It's about predicting what future research might hold. So like, what's the connection between not being able to keep up with the literature and kind of like future research predictions? [03:58] understanding of the past. So not necessarily like a nice clean text that's summarized, but, you know, a model that could draw on thousands of findings through different literatures at different levels, because, you know, like neuroscience is multilevel. It has everything from psychology and behavior all the way down to really low level cellular molecular. [04:17] findings, things involving DNA and so forth. And so, I mean, no one could really draw on all that information, but it might be that biology is really messy. [04:26] It's not like how computer scientists create these abstraction layers. We have the hardware, the software and so forth. So to make a prediction, you might very well need to draw on all that information. And why do you want to make a prediction? Because I mean, well, first, there's so many uses in science for this. So you maybe want to run a more informative study. So if, you know,

[04:49] Brain GPT predicts that your study is going to work out 99.9% certain as you expect. There's really no information gain in that. There's no reason to run the study. [04:58] By the other, on the other hand, you know, if you, the system says, oh no, this is unlikely to give the pattern results you expect, but you have an intuition that the literature has gone off course and there's a systematic bias, you know, that same bias is affecting what brain GPT is trained on. Then like, in some sense... [05:16] you're like trailblazing you're making a really impactful discovery [05:20] So really everything, getting into reproducibility is a huge, and replication is a huge issue in science. Like most findings just don't actually replicate. And so I think we could in the future, very near future, [05:30] use systems like this to kind of get a handle on what's true, what we could count on, [05:36] what next step we should take and scientific discovery yeah is it like when you're training it how are you taking into account the fact that like you know p hacking occurs and like maybe the literature is like somewhat damaged like if are you training it on literature how do you filter that out i guess yeah i mean right now not so much but that's something that's uh definitely of of interest so i mean the way i see it i mean i don't think scientists see it this way i mean [06:06] - Yeah. [06:06] most people think of papers as like individual contributions or discoveries, but I think of [06:12] each contribution is as even if it's not p hacked is really flawed noisy incomplete but you have this like tapestry of thousands of papers and so hopefully if you just aggregate it's almost like in machine learning if you do an ensemble so it's like an ensemble of everybody

[06:29] hopefully has the signal in the correct direction. I mean, that's why I raised the possibility before that there could be systematic biases or issues, but I think a lot of [06:39] the problems just really come down to underpowered studies, like statistically. So that opens the door to p-hacking or just, you know, a careless mistake. And so hopefully there's not that many systematic flaws. And if we could just not start reasoning about individual papers, but about thousands of papers at once, [06:57] I think we'll get the signal. [06:59] Thank you. [07:00] That's really interesting. And you said it predicts better than like neuroscience experts, like which studies are going to work and which won't. Like, what are the boundaries around those kinds of predictions? Like, what is it good at predicting? [07:12] Yeah, yeah. So the way we tested it, I mean, kind of taking [07:17] our cue from what people do in computer science and machine learning, where they make benchmarks, you know, like the ImageNet benchmark was... [07:25] really critical to developing computer vision models. So we made our own [07:29] Brain Bench model and. [07:32] What we did is we looked to the Journal of Neuroscience, which is – [07:36] kind of a standard, well-respected journal in neuroscience. And the reason we chose it is because it gets to your question that there's [07:43] It really covers all of neuroscience. There's five subsections to it. It goes from everything from cognitive behavior, like pretty high level stuff, to like cellular molecular systems, to developmental psychiatry type stuff. [07:57] And so our benchmark would have these five subscores. And what we did is we took recent publications from this journal that are unlikely to have leaked into the training sets of models. And we trained some models from scratch where we know what they're trained on. But what we did is we just subtly altered these abstracts. So scientific abstract, just to back up for...

[08:21] readers that haven't read a scientific paper, they tend to be structured that there's a bit of background at first, a couple sentences, [08:27] Then, um, you know, like the method, what kind of. [08:30] experiment was run and then the result. And so we just altered the result in very subtle ways, like kept the kind of linguistic flow. So if it was like, blah, blah, blah, interior hippocampus, we changed it to blah, blah, blah, posterior hippocampus. And so basically just, or something was this brain activity increases, but it decreases. [08:49] And so we just, so it's a multiple choice basically with two options. [08:53] And we tested basically neuroscientists and a whole slew of large language models, including some that we fine-tuned on neuroscience literature and just compared their accuracy. [09:07] That's really interesting. And I guess like for you, like what is the what do you think is the current sort of frontier for neuroscience? Like what is the what are the interesting problems right now and how does how does brain GPT fit into that? [09:23] Yeah, I mean, there's so many things. I mean, in some ways, neuroscience has been around... [09:29] for a while, but in other ways it's still like a really young discipline, say compared to physics. [09:34] And, you know, some people could even say it's almost like pre sort of there's no standard model or anything like that in physics. So pre paradigm. Yeah, exactly. Yeah, exactly. That's what I was that's what I was grasping for. Thank you. So there's still like I mean.

[09:52] A lot of what I do in research, I do like empirical research modeling, but I feel like a lot of time I'm just trying to figure out like what the question is or how to. [09:59] frame things. But, you know, people are, I mean, there's so many questions and some of them are relevant to [10:06] I mean, it goes from really high level stuff to really low level stuff. Like how, I mean, we all think like, you know, [10:11] We have synaptic change, and that's the basis of learning, but even that's up for grabs. [10:16] how that works even at the very low level, like how is a memory encoded in the brain? Is it in sort of those synapses, the weights as we do in neural networks? Is it something more internal to the cell? How does error propagate from learning? So that'd be relevant to like, you know, deep learning. Does the brain do gradient descent? Does it do something else? And that's high level stuff like, [10:36] is like when we understand space, is that the basis for higher level mental concepts like freedom, justice, or just, you know, chair? Or is it like just a general learning mechanism? So there's all these issues that go from very low level to high level. It's such a huge field, like I can't really say there's [10:52] There's one issue and I kind of gravitate towards the ones that might have a bleed over or [10:58] transfer into AI machine learning personally. [11:00] And what are the ones that you think are like the most kind of bleeding over into AI machine learning? [11:06] Yeah, well, I mean, one that I don't work on that seems like an obvious candidate is just like transferring power consumption ideas, you know, so like, I mean, modern GPUs are... [11:19] amazing and they power transformers which power ai revolution and other technologies

[11:25] But of course, like, you know, these data centers are going up every day and it's stressing the grid. There's carbon impacts and so forth. Whereas like our brains are doing a lot of computation, but I guess you just have to, you know, eat a sandwich or something and it's a lot less power consumption. So that's sort of neuromorphic computing applications. [11:47] I'm just excited for like a world where like Microsoft data centers, like they just like order a big, big load of subway, you know, and it drops. Yeah, exactly. It's like a reverse matrix or something. Just give it a salad. It doesn't use us as batteries. But yeah. [12:04] Yeah, that's really funny. Yeah, but possible. I mean, there's so many things. I mean, I'm really interested in the higher level stuff. So I mean... [12:13] I still think where people are better is like somehow we have some tricks for how we represent the world and tasks and. [12:20] promote generalization. I mean, in some sense, why we're so excited, everyone's so excited about large language models is [12:25] That they're like, you know, base or foundational in some sense that. [12:29] you could apply them to other tasks, not just the tasks they're trained on, whereas previous generations of machine learning models, even the great convolutional models that somewhat cracked object recognition or like AlphaGo doing... [12:44] you know, doing its games and so forth. Like, um, [12:47] Those are, you know, specialized models. And I think people are still the kings of that flexibility. So there's probably some secret sauce to gain still from humans and how we represent situations and generalize and link things up.

[13:03] That's really interesting. Wait, can we see a demo of it? Like, do you have a way to like use BrainGPT that we can look at? Oh, you know, it's... [13:11] Actually, so your show, there's a lot of like, good guests that do interesting things with prompting. So this is actually a failure of prompting, which might be even more interesting your guests and give them a new way. [13:23] to interact with large language models. So it's not going to be visually exciting because it's just going to be me talking, but maybe it'll be like informative. So we tried. [13:34] doing kind of what you would expect we would do, like, you know, firing up GPT-4, which we did back in the day, like the previous version and saying, give it a few prompts like, hey, we're going to give you two versions of you're a nurse, pretend you're a neuroscientist. We're going to give you two versions. [13:49] of a scientific abstract in neuroscience one of them is true one of them is altered which one is the original so it was above chance at that but it's like my memory's bad that maybe it's like 61 so that's actually like pretty much human level but if instead instead of [14:05] interacting with the model, you actually just have access to the weights, which you don't for GP24, but you do for like Vama family of models, [14:14] for a lot of other models, for Mistral, Free Models and so forth, the Balkan models. There's a lot of models out there where there's access to weight and Microsoft's Fee models, a smaller accessible model. With those models, you could actually just compute what language researchers called perplexity, which is how surprising the text is in the model. So it's like given what the model's been trained on. So it's like basically, you know, eating all the scientific literature and it's set all the weights.

[14:41] How aberrant is this text passage? And so if you do that, [14:46] it has two real advantages. The first one is it's just way more accurate so you could, it'll have a lower perplexity for the correct version than the incorrect version. [14:53] But what's also really powerful is that you could take the difference in perplexity, you know, how surprising these two passages are, and you could use that as a measure of confidence in the model. And it turns out that confidence is... [15:04] calibrated to its accuracy, what some of the models more confident when the difference in the perplexity of the correct and incorrect passage is larger, the models more likely to be correct. So that's really important for human machine teaming because you could put a human [15:16] and the machine together, and you could get a better result than either one alone. So we have some analyses like that we're writing up now for a follow up publication. But basically, it's interesting why the prompting doesn't work. And I think the issue is that [15:29] Usually in modelable choice, the difference between the options is pretty large. Whereas here, it's really subtle. Like that example I gave, anterior, posterior, increases, decreases. It's pretty like... [15:38] subtle and I think [15:40] Scientific literature, again, is noisy and imperfect. [15:43] And I also think this task is not like a lot of its... [15:47] tuning that it was given to be conversational, like it's reinforcement learning through human feedback or other supervised fine tuning. I know these models generalize really well the task, but I think this is still like kind of a weird task. [15:58] for them to do even with a few examples of prompting. So it just turns out you could really [16:03] To me, it's like a pattern recognition problem. If you could just like look at all the models weights to get its perplexity out. They do just they just do much better. As a matter of fact, they're like in the 83%.

[16:13] range. Um, so you get like a 20% bump and like humans are only at 63%. So they get sort of superhuman. [16:20] good by just being able to calculate the perplexities directly. [16:24] That's really interesting. So like, um, just to kind of play back what, what you just said is, um, yeah. [16:31] Rather than like prompt the model and be like, I'm going to run this experiment. What do you think? What you do is you like write a future neuroscience paper and then you run that through the model and calculate its perplexity. Yeah. I mean, you got the total idea. What we do is like so much simpler. Like we have these. [16:51] to abstracts like sort of the original version and the altered with the one with the altered result and we just get a number for each one like the perplexity how surprising is the original one how surprising is the altered one and the model just chooses whichever one's lower but you could do what you were doing like this is what i kind of see the [17:08] I think you're already seeing where this is going, because I think for scientific discovery, to me, this is the [17:13] first step, you have to be able to predict, like, is it, did this experiment go this way or that way? And if you did that, you could start [17:20] doing more interesting things like you're saying and build tools on top to say like, you know, what experiment should I run next and so forth? Or is this plausible? Like, would this work out? [17:29] how I think. And you could even pair that with generation because you could have the model [17:34] in prompt mode generate different patterns results then you could use it to actually to evaluate through perplexity which one [17:40] is most likely to be reasonable. [17:43] That's really interesting. Yeah, this is just sort of like, I think an interesting dovetail into like how does

[17:49] how might science change with this kind of with these kinds of tools and um for people who have been listening to this podcast i i've gotten on my soapbox like once or twice before about about what i think um and i'm sort of curious to like uh to do to to like explain it a little bit to you and then kind of like get your take as because i'm just like some guy with a newsletter and a podcast uh no no i think in this day and age like people that are placed like you are actually the [18:19] and linking things together. Like, I think it's really valuable. And I mean, you're being humble, but you should give yourself more credit. So I want to hear what you have to say. Thank you. I appreciate that. Well, I guess like the place I usually start is... [18:32] science seems to have been like really obsessed with explanations for a really long time, like basically since science was invented, which makes a lot of sense because explanations are extremely powerful for making predictions. And they're also like kind of beautiful to understand how the world works. And it has like finding scientific explanations has worked like really well in lots of areas of science, like physics or like chemistry, stuff like that. [19:02] the quote unquote soft sciences where like actual causal, um, [19:07] parsimonious scientific explanations are like really, really, really hard to find. Um, [19:12] And that's what makes studies hard to replicate and all that kind of stuff. And if you look at the history of psychology, we've been trying to figure out what depression is or anxiety is or whatever for like 150 years. And we have some different lenses on what it is, but we don't have the germ theory of mental illness, for example. We just don't actually know what it is and there's lots of different competing ideas.

[19:42] And my kind of contention is that machine learning AI approaches allow us to kind of like unbundle explanations from predictions so we can get good predictions without having to explain anything, having any good theories about what's underneath. [19:58] And so we sort of turn this like science problem into an engineering problem, which is like instead of explaining depression, we just predict it. And if you can predict it, one is you can start to help people. So you can predict when it's going to happen to someone. You can predict what interventions are going to work. And then two, if you've developed a good enough predictor for that kind of phenomena in the world, like theoretically, the theory exists in the neural network somewhere. [20:28] They're more interpretable than human brains are in general. And so you may be able to actually find a good causal theory of depression inside of a neural network if it's sufficiently predicting, which I think is really cool. Or what you may find, which I kind of feel like is probably the case, but I'm curious what you think. What you may find is that... [20:49] the theories for higher level phenomena like depression are so big that they don't fit in our rational brains. And so that's why like a good clinician has a good intuitive sense of what's wrong with someone and can kind of like help them, you know, [21:05] help them overcome their depression. And they can give some sort of explanation for what's going on, but really a lot of stuff is just happening subconsciously. And so I think the larger impact of all of this is we may start to realize that

[21:22] Um, we put too much emphasis on our like rational problem solving logical brains and not enough emphasis on our intuitive brains. Um, and, uh, and AI may help us like kind of re re realize the like power of intuition because AI like makes intuition sort of like, um, usable, um, in the same way that like, like logic, you can write down logical arguments. You can also build an AI predictor that has human intuition in it, which I think is like really, really cool. [21:52] change how we do science. So instead of doing small scale studies, we will just try to do much larger, larger scale studies where we're getting lots of data together. And that's the project of science. And then we build predictors on top of that. So that's a very, very long spiel. There's a lot in there, but I'm sort of curious what you think. No, I mean, I largely agree and could even [22:15] say further, I mean, certainly, like, I mean, I've been trained as, um, like a cognitive psychologist and kind of more migrated to machine learning. Yeah. So like a lot of things we do are exactly like you described. So [22:27] it's not if you ask a great baseball player how they hit the ball they'll tell you something but it's pretty much useless and doesn't describe what they do at all which is [22:36] really one probably a reason why a lot of great athletes end up being like terrible coaches and managers. There's not really that strong of a correlation and [22:45] Um, [22:46] But yeah, I mean, I might take it further. So I don't think it's a hard versus soft science thing. I think this...

[22:54] really nice coupling of explanation prediction like i fear it might be a historical uh phenomenon and it really only applies to some aspects [23:05] of physics because, you know, you could do things that [23:08] I don't know if, I mean, not even psychology, but just in biological sciences, like everything is so messy. It just might be that there's, [23:17] 10,000 variables and they're all interacting and these variables are at different levels, like we're saying from DNA to behavior and there's all these feedback cycles and delays. [23:26] And that just might be what it's like. And if that's what it's like, [23:29] That's just not going to have a clean explanation. [23:33] that will be human, understandable. And so at that point, it almost becomes like storytelling to have [23:39] An explanation, the explanation, the prediction, much like hitting the baseball or, you know, it's going to diverge more and more with time. Unfortunately, again, this is not how I want the world to be because I was always somebody. [23:51] that would try to come up with the crisp explanation even when i do modeling i don't want to just say the model worked i want to understand the principles behind it and how they interacted to give rise [24:00] to the behavior of the model. [24:03] you know, have understanding, intuitive sense of understanding and a solid scientific explanation. But I don't know if that's our [24:10] our future um [24:12] You know, I use this joke sort of like, you know, I mean, [24:16] imagine like explaining quantum mechanics to a dog you know it's like quantum mechanics is a very well specified theory but it's just beyond a dog and like you know maybe at some point we're the dogs you know like we're not gonna understand what's going on and we build tools we can already see beyond our eyes with telescopes and we could add up numbers we can add you know with calculators and before computers do all these amazing things so maybe like it's just going to be

[24:42] beyond us and uh we'll have to accept it and take different forms of explanation that are much more general and much more about like well we built the system when we put these constraints on or for like taking it back to brain gp it could be like [24:55] It's like, it might be really general stuff, like a psychology and neuroscience related fields. Does knowing about behavior help prediction about neuroscience? Well, I don't know. Let's train a model that has both. [25:05] psychology and neuroscience and it does it predict neuroscience better okay then those fields are related that's the kind of explanation but it's not like [25:12] You know, like, um, equals MC squared or something. It's not very clean or. [25:17] So maybe we're going to have stuff like that in the future as explanation. Or maybe we're just going to be in this predictive world. And it's not even dystopian. Like everything could be better. But like we just really will just have stories about how this stuff works. So like maybe scientists will be like the new priest just interpreting things and like making it, explaining it to how it works and providing meaning to people. [25:40] the systems, but yeah, I don't know. I don't think [25:43] I don't know if we're quite there yet, and many of my colleagues would disagree, but I could see a world in which, like, um, explanation and prediction. [25:51] unfortunately diverge. And it's not because anybody did something wrong. It's just because the world's really complex. [25:57] systems are complex and we're not just, our brains aren't built to make sense of this kind of stuff. What is, what do you think is the, like, if one of your colleagues that would disagree, like, what would be like a reasonable counter to this, like, this viewpoint?

[26:13] Yeah, I mean, maybe I'm so on this train, I have trouble simulating it. But I think a lot of things would just be, you know, appeal to the past and really elegant theories that unlocked understanding of, you know, the past. [26:31] an explanation. And I mean, I'm, I'm, I couldn't even point to things I've done myself that have that flavor, but a lot of the arguments that I come across, they're not really forward looking. They're more like in the past, it was this way, but you know, in the past, we didn't [26:47] have, I mean, we're having increasingly more large data sets available and there's more and more, you know, different types of data to link together, not just in the commercial world, but in the scientific world, like massive DNA databases and disease databases and, [27:04] brain recordings from the same people and um and the scientific literature just gets bigger and bigger and bigger so it just seems like you know it's sort of like a sensor fusion problem at some [27:15] connecting these dots. So I, [27:17] Again, I don't think that was what it was like [27:20] In the in the in the past, but I mean, I guess first. Yeah, so I think also scientists tend to like look up to physics too much and it hasn't worked that way there and not being an expert in physics. I'd be surprised if. [27:34] Many parts of physics aren't already like what I'm saying, because it's probably at some point things are going to get many, many variables complicated there, you know, so that not everybody's trying to do the grand unified theory, you know, most people are like.

[27:47] I don't know. I'm guessing there's something sort of [27:50] veers into the real world like trying to build a a containment field for for fusion or something i i bet there's tons of machine learning going on there and again i know nothing about that it just seems like the kind of thing where it's going to get [28:04] complicated there's going to be many variables to balance and time series and it's not going to come down to like [28:10] some simple equation with a square in it or something. Yeah. That, that, yeah, I think, I think that makes sense. I guess, I guess like what I'm sort of pondering right now is like, I obviously like feel like there's like, we're, we're sort of approaching the world where like, maybe we don't jettison explanations entirely, but like our emphasis shifts a bit, particularly in science from like X, like clean explanations to like the stuff we're talking about, like predictions and stuff like that. And using different models to like draw different conclusions, which means we're not necessarily always getting down to the brass tacks of like, [28:40] we know how every causal interaction happens, but we can generally know what is going to happen or all that kind of stuff. [28:50] But I guess one of the things that your dog analogy reminded me of is just like, how do we make a world where that is true? That's also not a huge bummer. Like, I feel like there's a lot of interesting things about it. But like it also, I can imagine like thinking like thinking towards a world like that and being like, I guess we're like just fundamentally limited. Like that sucks. You know, what do you think? Yeah, I mean, I'm more positive.

[29:20] For people, I mean, first for most people, and like, again, not denigrating most people, [29:25] if they're healthier uh the economy grows the climate doesn't collapse and so forth [29:31] They're probably going to be pretty happy with the future. If we have a future where we could predict disease and... [29:38] had a [29:39] you know, have sustainable energy, everything, you know, if, if, if these predictive tools unlock those kinds of discoveries, people, most people be happier, but people, I guess you and me, like, we want to actually have some insight into, you know, [29:53] what's going on and [29:54] I guess there, it's just going to maybe be different forms of explanation. I remember as an undergraduate, when neural networks are really out of style and learning about them and having fun, you know, programming up from scratch and see my own little back prop nets and stuff like that, and didn't really understand what was going on in the weights. But I almost felt there was a different kind of understanding or something deeper by this kind of system could do this problem. So there might be that kind of [30:20] explanation exactly like what you're saying where you can't like trace it all out you might even be able to do what the judea pearl type people that really emphasize interventions and systems like tweaking i don't like zapping this neuron or or changing something or doing an intervention and who knows in our climate and then being able to predict how that [30:38] Cascades forward. I mean, [30:40] I don't know, maybe the kind of explanations would be the kinds of variables that are involved, some gross characterization of the system and the dynamics, and it'll be just like you're saying, not a... [30:50] charting out the whole causal network, because that just might not really be something that's possible. And it's probably not even what the machine learning systems are doing. So I don't even know if you could distill it. Yeah.

[31:05] Yeah, so... [31:07] I just think maybe what we accept as an explanation could change over time too. I mean, people change and [31:13] our expectations of [31:15] I mean, how we understand the world must be so different now than 500 years ago. [31:19] What we would find satisfying [31:23] So maybe it's just going to change again. And we're worrying about something that people in 100 years from now just won't even think twice about. [31:31] Yeah, I kind of think that that's true. Like I do like you mentioned the idea of like, okay, how do we tell if two fields are related? Well, we run it through a model. And if it says it's related, then it is. And that's like, maybe that's not satisfying to us today, but it will be satisfying to like a future generation of scientists. And I think the conditions under which that would be satisfying is like, [31:51] we have a model that is like so standardized and like well used that everyone has an intuition for like [31:57] what it is and trusts it to some degree. And in that case, getting new results from that model, I think will be fascinating and really interesting in the same way as finding some underlying causal network or something like that. It's pretty clear we're moving to a world where using models to judge things is a really important way to make progress, [32:21] in science, but also just like in AI in general, like we're kind of like reaching that point too, or in like mechanistic, mechanistic interpretability, like it's so complicated that you have to use models to like find features of models and all that kind of stuff. And I just think that that feels like a place that's pretty, we're, we're, we're moving towards.

[32:39] Yeah, yeah, it could be. I mean, obviously, there's tons of really smart people working on interpretable AI, both building it in from the start, but also [32:50] kind of in a post-examination of these models. So even if we're moving towards that world, maybe along the way that will help us [32:59] transition or you know maybe there'll be a breakthrough but i'm just skeptical because if the underlying reality [33:04] is so complicated and we have a model that appreciates that underlying reality [33:10] then the explanation might not might might basically be the weights of the model in the training set you know so at that point i don't know what you do but maybe we'll have [33:19] some purchase, some [33:23] some transition and obviously that works important and when we make these models, we're never going to blindly trust them. So we're going to need all kinds of checks along the way and [33:35] mitigate bias and so forth. But yeah, I think we're going to that world just because that's the world that things are going to get done in. Yeah. Yeah. I guess I'm curious, like you referenced, I don't know, like a couple of minutes ago, like one of the reasons this is so hard is maybe the way reality works is it's like 10,000 variables weekly interacting. And that's really hard to create a prediction around, or sorry, an explanation around. Can you explain what that, like why that [34:05] for example, like genomics or something like that? - Yeah, I mean, I think everything, or like why we're doing what we're doing right now in this conversation, it's just everything.

[34:14] is is that way i mean even in in physics i mean you think of like everything's like this ideal idealized spherical you know particle in some vacuum or something and none of that stuff would even work in the real world when you have more variables like the wind resistance the friction this and um the material stressing and but um [34:37] Yeah, so I just... [34:39] Yeah, sorry, could you, maybe, I think I got lost too many variables. Could you even like repeat the question? I mean, I just think everything is, is, has this flavor to it, you know, when it gets closer to the real world. But yeah, so yeah. [34:55] I mean, a lot of the issue, too, is like... [34:58] Anything biological, it's probably beyond biology. So like, you know, even in social sciences and economics, you have all these interacting elements and like in an economy, and that's just incredibly complicated. [35:11] But in biology, of course, you have the cell itself is really complicated. Then you have all these interactions there and in history. And you mentioned DNA and [35:19] how things, you know, how pro how the proteins are expressed, but also then with [35:25] You could think of it also, I mean, I think every, I'm like a physicalist from a philosophy standpoint, so I think things... [35:31] even if you have to study higher level things like psychology, neuroscience, and ultimately, there is like this lower level explanation. It's just again, too many variables, humans can't make sense of it. But you know, you could think of like all our goals and what we do as also affects

[35:46] what goes on in our brain and how things wire up. So there's just so many interactions across the levels. And if you think of what engineers do, [35:55] They do go to great pains to reduce those interactions, but you know, evolution, physical reality doesn't care about this stuff. But you know, when an engineer makes abstraction layers, you know, you have the machine code and then maybe the assembly and then something like C or Fortran, then you have Python libraries on top. And I'm not doing a good job describing this and you have application and then like, [36:17] younger people today a lot of them don't even know there's a file system on the computer what a folder is because it's like so abstracted i mean which i guess is [36:24] Great, but you know, it's sort of like no one knows how to change the oil in their car or anything because we've extracted... [36:28] Away, but, but we build the things like that and a lot there's a lot of effort. [36:33] that goes into that, but I don't think nature or biology [36:38] respects that so it just makes it kind of a mess to unravel and of course even with our systems we have bugs and weird things happen like i remember decades ago the intel chip you know screwing up arithmetic or something from some hardware bug and it's just like really hard because it's spilled into other layers you know in the abstraction hierarchy [36:59] Yeah, and I think you've thought a lot about like... [37:03] The ways that trying to make natural interactions or natural processes human understandable actually might cause us to misinterpret them. You have one paper in particular called The Inevitability and Superfluousness of Cell Types and Spatial Cognition. Do you want to talk a little bit about that? Like introduce what the paper is, what the motivation behind it is, and what the results is that you found.

[37:26] yeah yeah it's really like a lot of our conversation is kind of leaning to this idea so for most of [37:32] neuroscience, a lot of the major discoveries, including multiple Nobel Prizes, have been somebody [37:39] recording from a cell, either in a human or a non-human animal, and basically having some intuition, some clear explanation. Like, for example, these are things [37:49] and the brain called play cells and they light up when a rodent's at a particular location and it's in its box they're like aha this is like the brain's gps system that's very intuitive and appealing but it's very also like simplistic because the situation the animal ones is very like contrived and we're kind of forcing this interpretation so what we found in this paper is many of these intuitive cell types that you could just interpret them and it seems like oh the brain [38:19] could make sense of this in a very, um, human understandable way. What we found is that when you take [38:24] larger um networks like deep networks and you put them in vr worlds a lot like the experiments neuroscientists run [38:31] on non-human animals, that you get these same cell types popping up, [38:37] even in random networks that don't serve the function of navigation or localization, like the place cells. So it's really like, [38:45] kind of a danger, I think, of trying to force [38:49] the intuitive understanding on the system, [38:51] Um, [38:52] It's just like you'll find it, but it's not actually how the system...

[38:56] works and even within [38:58] the field where people believe in these kind of easy to label intuitive cell types, what happens is there's this general pattern where someone will say, "Eureka, I found the place," and they'll literally get a Nobel Prize. [39:10] And then there'll be 30 years of research explaining how it's not really a place like, oh, well, if the room isn't perfectly round, [39:17] then you get like this distortion. Oh, if there's a reward like food, then it doesn't code for that. Oh, it depends on the history of the path. Oh, it depends on the viewpoint somewhat. And so it's like maybe we'd actually understand what was going on more if we would do kind of more of a [39:32] I don't know, respect the complexity of the system we're dealing with. In some ways, do something a little bit more bottom-up, still having mechanistic explanations and models. But, I mean, this stuff is just, like, so complicated, and it seems just ridiculous. It's almost like a... [39:47] Again, this is going to get me in trouble with people, and I think people hate this paper, even though it will be published and have impact. [39:56] It's almost like a kid's view of science, like, oh, I just mixed some chemicals and something happened and I tell people about it and it's a discovery. Or, oh, look, I found this butterfly. No one found this before. That's so important. But that's not really like what scientific explanations are like. And it's just the... [40:11] when you're dealing with a system as complex as the brain and people and so multi-level, it just seems so vanishingly unlikely that whatever idea you go in with is what you're going to come out, is going to be actually true or there's no, unless there's some reason for it.

[40:27] like information decoded that way. It just, so I think there's, this is a problem because those are the explanations people find attractive, but I think they're very seductive, but ultimately limiting. [40:38] That's really interesting. [40:40] Why do you think like... [40:42] Given that we like those kinds of explanations, [40:45] what would you guess is the reason like our we're just attracted to kind of those like sort of parsimonious like that's the way our brains work. Like what like what is it about those kinds of explanations if they must work for something, right? Yeah. I mean, again, like this is where I'm glad I have some training as a cognitive psychologist. But everything I'm going to say is going to be so obvious. I mean, it's just like if you're an. [41:06] If you're a politician running for office, simple story, coherent narrative. [41:11] works best. It's like what's going to have the best information transmission value. It's going to be most viral. You know, if you're in a what does a lawyer do in a jury trial, they create a narrative where their client is like innocent, that's a coherent story. [41:27] So I think people really prize this kind of coherency. It's just like what makes... [41:32] an appealing explanation to us but um but again some things even things that are actually pretty simple like in terms of just writing out what it is like quantum mechanics that makes no sense intuitively and that's pretty simple but then i think when you get into things like anything [41:48] we're involved in the real world or biology it's just going to be so many variables you could tell the clear intuitive story but i'm not even sure in cases it'll be an approximation of

[41:58] of the real thing or it'll be like so crude and off it'll [42:02] Um, obscure miss, uh, like deeper. [42:06] deeper truths. But yeah, I just think, I mean, that's, [42:09] I mean, we like that. I mean, sometimes it's good. It's almost like Occam's razor for some things you don't want needlessly complicated explanations, but. [42:17] you know, and sometimes you need it. Sometimes you need the more complex model to, you know, fit the data, make the prediction, characterize the system and, um, [42:28] I don't think humans just go in by default thinking things are that complicated and [42:33] We didn't really, like, I mean, [42:35] I don't want to try to guess how we evolved, but like, you know, we didn't really live in a world where things had to be that, you know, like they weren't like the quantum world, the 10,000 variables of how the brain works, big data. [42:48] world. So probably having simple, robust procedures serves us well in our [42:55] natural environments, but maybe doesn't serve us well in science. [42:58] Yeah, I guess that's maybe what I'm thinking right now is like, [43:03] Um. [43:04] The brain has like really high dimensional representations of things in the world that are very complicated and would be very hard to like have a scientific explanation for. And. [43:16] like our natural inclination towards like sort of simple causal stories that we can tell, maybe like the proper place for those kinds of things are to like, [43:27] slightly tweak our much higher dimensional concept for a particular context.

[43:36] And like, we're kind of like misusing that to try to like, um, [43:40] blow it up into it representing the entire high dimensional representation that really we can't put into language. And we've been sort of trained that the more parsimonious and powerful those little things are, the better they are. But that sort of gets misused when it goes beyond what they're capable of representing. [44:04] Yeah, yeah, I think you're right. And it gets even trickier because, of course... [44:09] that we're not passive recipients of data in science. We go out and collect it. So once you start believing something, a story, you just basically keep, there's a lot of confirmatory work in science. So... [44:22] So you'll just basically, if you have that [44:26] mindset about what the relevant variables are. Even if you don't think you're being confirmatory, you kind of are confirming [44:32] the framework. So I think it's what you're saying, but we even do things to keep those simple stories going probably longer than they should. [44:41] Yeah. I guess... [44:43] Yeah. [44:44] Like if you could, this is a big question. So like, you know, take a second to think through it. But like, if you could sort of like wave a magic wand and like all the priors of science were thrown out for a second and you could like rebuild science as an institution, like knowing what we know today about how hard this stuff is, like, what are the things you would change and like how would you rebuild science for this world?

[45:07] Yeah, I mean, gosh, so I think it's really good that nobody has that much power, because probably what I say is wrong, but probably what other people say is wrong, too. Personally, I would... [45:20] to a combination of training people like this conversation to be a little bit more [45:25] philosophical about things and maybe do some reading and thinking there about what explanations are and the limits and the study of it, but also like more emphasis to which is already happening on computational skills, large skill simulations in the fields. [45:40] going this way too, but more consideration of naturalistic environments and their complexity and how [45:46] There should be like whole courses in science for university students about how limiting the problem. [45:53] That might work well in a particle accelerator in physics if you're working with the standard model, but probably everything else... [45:59] You're going to miss key variables and insights. I mean, especially anything involving [46:05] human behavior or biological system. So I guess really just kind of a different orientation for scientists have them be a little bit. [46:12] more thoughtful and philosophical about the whole enterprise. And just like everything's going to be AI, computational, large data sets, even if there's a beautiful experiment. [46:22] that can isolate something for a key question that you can't resolve with these large data sets. So something that when I did some work with consumer behavior to look at human decision making in the real world, like in retail settings, and there was it was [46:34] like doing these big studies with like, you know, millions of transactions, but then we'd bring it back into lab and try to test something. So I think emphasizing that interplay, I think if you're going to run lab studies, I think there has to be

[46:47] the interplay with something more naturalistic real world big big data just so you don't just make up like kind of like a fake science of itself because you can have a science of [46:58] of anything and it might just not have anything to do with anything anyone really cares about [47:03] Ultimately, like in a hundred years. Yeah. Who do you think is doing like good work on, like in, on this, along this vein, like in your field or in other fields that you know about? [47:13] Oh, yeah. I mean, so I mean, a lot of people are doing amazing work and I'm very like self-critical and critical of other people. But luckily, I don't share it with them. So I probably won't be great at like naming a bunch of people. But I mean, one maybe you use one whole whole subfield is I mean, years ago when I was in graduate school, like the computer vision. [47:35] they're mostly guys, but they were like the big jocks. So they feel like, oh, they had the most mathematic chops. And they were doing the real science with all the filters and Fourier transforms and really leading the way. And then... [47:48] along comes AlexNet, like this convolutional network not made by a vision scientist, but by machine learning, AI researchers. And overnight, that, in terms of behavior, that was the best model of object recognition. So what happened is neuroscientists, computational neuroscientists, took that on as a model for the ventral visual stream. So how [48:10] Our brain transforms our vision images on our eye into like that's a chair, that's a cat. [48:16] And they found that the it's not a clean, that perfectly clean story like they

[48:21] presented, I have actually a paper criticizing [48:23] the basic story, but I think it wasn't real advanced to show that the levels of processing in these models, these artificial neural networks, [48:31] fairly well track the transformations that went in our own brains. And so this is an example where growing larger naturalistic training this model on a million natural images [48:41] Basically taking the whole problem on at once and doing something at scale led to a bigger advance than like 100 years of vision science and all these grants for this. And, you know, it was done by neuroscientists, but it was really kicked off by engineering and embracing technology. [49:00] this real world challenge. And I think that should be really humbling to, to scientists. And so I don't want to single out anybody, but there's, [49:08] a number of people that did that and scientists are very, almost like, [49:14] faddish, like if one person has a good idea, they all kind of rush to the same thing. So there's been a lot of really smart people [49:20] that have jumped on that train. And now it's like sort of petering out and it's like the next generation [49:24] questions are happening but yeah so that that's an example that kind of goes with what we're discussing that makes sense and then in terms of like you know wanting scientists to be a little bit more thoughtful and philosophical about like what the methods of science are do you have any like [49:40] people or writers or, you know, philosophical ideas that you think are like, like they should be reading or we should be reading even, even if it's not like for lay people, just like in general, like what, what, who are the philosophical people that get you excited?

[49:55] Yeah, I mean, gosh. So I'd say this book that I read recently, it's an older book by Thomas Nagel, who's a philosopher of science that recently passed away. It's called, so he's famous, this is not what I'm recommending, it's a good essay, but he's famous for what's it like being a bad essay where it's just sort of about subjective experience and how [50:16] You couldn't really know what it's like to be a bad because you'd have to be a bad. And then when you go back to being a person, when he wrote this book, The View from Nowhere, [50:23] that, um, [50:24] even though he's an incredibly famous philosopher, I feel like it should have way more impact because [50:29] The first half, I think, really explains [50:32] what science is and what it's [50:35] not and I think a lot of scientists people that seem like they're being more inclusive in some way of topics like this probably apply to consciousness research. [50:45] in some ways they're doing scientism in an overstep. So the reason the book's called The View From Nowhere is because that's what science is. So there isn't science for you or me, or there isn't like Eastern science or Western science. [50:57] It's like, you know, from this [50:58] disembodied perspective and that's has both strengths and weaknesses you know so if you launch a probe to Mars and put it in order like Mars doesn't care who came up with the calculations it's either going to make it or not make it and so that's kind of the. [51:12] the view from nowhere, but of course a lot of human experience that's [51:17] very valuable is subjective. And that's like kind of where meaning comes in our lives. And so what I liked about that book is making clear like what the strengths of science are, but in some ways he doesn't.

[51:28] like, you know, hit this over the head with it. But for me, what I really got away was the limits [51:32] of science. So I really, I really liked [51:35] That book. [51:37] That's really cool. I love that. How do you think that that relates to, I mean, you're an experimental psychologist. So how does that relate to the field of psychology, which obviously is like... [51:48] It's. [51:48] trying to be science, but then there's also like it's also all about human experience. [51:52] Yeah, so again, I mean... [51:56] I mean, another person would be good to read would be like Wigdenstein. So I'm not trying to be like some kind of logical positives that you could only [52:03] rely on the observables and whatnot. But I think, um, [52:06] I think there is like some some overstep. So a lot of times. So psychology, like in terms of like experimental psychology, [52:17] You know, most people I would say are scientists and are doing science, but like there is like sort of this. [52:23] this overstep at times, because you're really limited to [52:27] what you could I mean it's just like you tell an undergraduate what you could operationalize so like what's depression well you have to like come up with some criteria for it or like what's this or I mean even we do it even when you're doing neural recordings as soon as you count something you know you're saying that there's you're abstracting in a sense you're saying there's some way in which these are alike and you're ignoring other ways in which they're [52:47] Um, [52:49] different. So like, that's kind of like the standard. I mean, that's sort of, again, like the view from nowhere, trying to like, [52:54] make a procedure where you could just observe things and quantify them away where you could start testing out hypotheses. But a lot of times I think psychologists and people in general get neuroscientists, others get confused that they could have like some deep insights into...

[53:11] our subjective experiences or like what people call like the quality of you know why is the blue make me feel this way and [53:18] when you start getting into that, like you could, you could come up with like, you know, [53:22] very good things like if i put you under this anesthesia will you wake up or not you know under what conditions what [53:28] What brain waves can you measure that will... [53:31] predict that or account for that or even some mechanisms but um i think a lot of these like when it gets too into um [53:38] not the view from nowhere, but like kind of the view from inside this first person, [53:42] perspective which is mostly what we care about is like [53:46] Creatures like I mean, I don't think I actually don't think science in general has a lot to say beyond correlates of these experiences. And that's probably why we still have literature and religion and music and all these arts and all these other things. [54:02] I definitely like, I mean, I feel like literature as an exploration of psychology is like, [54:11] quite undervalued um by people who are into science like i think some people think about literature and psychology as being being very intertwined but like um [54:21] And as not a professional psychologist myself, but one who has like read a lot, a lot, a lot of psychology books and also has read a lot of novels. Yeah. Like there's something there's something about that or like even the like literary tradition in psychoanalysis where it's like a lot of this, the earlier work is like less about like, okay, you have these cognitive distortions and like we're going to do a worksheet or whatever. But it's like really, really long kind of like literary storytelling about particular patients and stuff that I honestly think is like underutilized.

[54:51] Yeah, I mean, it goes back to what you're saying before about like the intuition and the many variables. So it could be like, [54:59] Even as a [55:01] I mean, a lot of, you know, I mean, so I don't know much about clinical work and like less about things that are considered outside science, like Freudian analysis. But I can imagine someone doing a lot of good work within that framework, because even if the theory doesn't hold any water, the scientific theory, it could still be the intuition someone builds up working within that system. [55:21] are very valuable and can create [55:23] good outcomes for people and [55:27] Yeah, so like I'm not, I don't put my faith or my life in the hands of like alternative practitioners, but when people get outcomes or anything like this, you know, I try to be a little bit slow to come. [55:40] completely [55:41] um dismiss it and not not at least in which because like you know placebo effects while under attack are probably real in cases but uh but i mean there's things there's practices that might be done that could i mean sort of like cultural knowledge i mean like italians know how to make good food and things like this it's just and it's from like hundreds of years of tweaking things and trying things out so that's even beyond the individual and so maybe some of the training [56:09] These folks [56:11] receive in clinical settings. I mean, some things, you know, like like cognitive behavioral therapy and is has some some like random control trial backing. But, um,

[56:21] Yeah, anyway. [56:22] I'm not trying to, yeah, to be clear, I'm not trying to like totally diss it. I just like, no, no. Please, please, please totally diss it. You don't have something to me. And I think after I just basically made fun of a few Nobel Prizes in neuroscience, I've got, no one's going to like me anyway. So. Great. Well, we'll be in the same boat then. Cool. Well, this is a really, truly fascinating discussion. I'm really thankful that we got a chance to chat. If people are [56:52] learning more about your research, where can they find you and what should they read? Yeah, sure. They could go to my website, which is bradlove.org. There's also a link to the Brain GPT website. And if people are interested in that project, we have a mailing list and we don't spam people. We probably send a few messages a year with major updates. So if people want to follow that project, they could sign up there. [57:20] I mean, they should, I'll put a link up to this, uh, this podcast when it's out. So, and others, so if they want to see more like popular media stuff, that's. [57:29] on the website as well. And if anyone's really a glutton for punishment, of course, they could just start reading abstracts and papers, but I'm not sure I could recommend that to the casual listener. Cool. Well, thank you so much, Radley. I really appreciate this. Whenever you have more stuff to share, I would love to have you back. [57:45] No, that'd be great. Thanks so much, Dan, for having me here. It was a blast.

[57:56] Oh my gosh, folks, you absolutely positively have to smash that like button and subscribe to AI&I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard, but instead of gold, it's filled with pure unadulterated knowledge bombs about chat GPT. Every episode is a roller coaster of emotions, insights and laughter that will leave you on the edge of your seat. [58:20] craving for more. It's not just a show. It's a journey into the future with Dan Shipper as the captain of the spaceship. [58:28] So do yourself a favor, hit like, smash subscribe, and strap in for the ride of your life. [58:33] And now, without any further ado, let me just say, Dan, I'm absolutely hopelessly in love with you.