Bayesian Brain, Scientific Method, and Models [Dr. Jeff Beck]

by Zubin Garamani. Um and he was explaining dir the Derchlei process prior. This is when the Chinese restaurant process and all that stuff was like relatively new. And his explanation of it It so resonated with me in in terms of like, oh my gosh, this is the algorithm that summarizes what what the scient how the scientific method actually works.

Right. You get some data, right? You get then you get some new data and you sort of say, Oh, how is it like the old data? And if it's similar enough, then you sort of lump them together and then you sort and you build theories and you properly test hypotheses in the fashion. That's that that's that's the essence of the Bayesian approach is It's about explicit hypothesis testing and explicit models of the

in particular generative models of of of the world conditioned on those hypotheses. It I I believe it is it is the only right way to think about how the world works. And it's the and and and it encapsulates the in the structure of the scientific method.

I mean, if I'm being perfectly honest, what actually convinced me the brain was the the brain was Bayesian had a lot more to do with behavioral experiments done by other people. My principal focus was on well, how does the brain actually do this? So I'm referring to experiments, you know, showing that like

Humans and animals do optimal cue combination. We're surprisingly efficient in in terms of like the information that comes comes using the information that comes into our brains with regards to again these low-level sensory motor attacks.

Um and one one piece of information is more reliable than the other. And the degree of l reliability changes on a trial by trial basis. So you never know a priori that like say the visual cue as opposed to the auditory cue is gonna be the more reliable thing. And yet nonetheless, when people combine those two pieces of information, they take into account the r the relative reliability on a trial-by-trial basis.

And that means that they're optimal in a sense. Now that we have to like be super careful with our words. They're relatively optimal because they're not actually using 100% of the information that the computer. Like your the visual information that you use, you don't use a hundred percent of the information that the computer provided you, right? But you know, there is some loss between the computer screen and your brain, mediated in principle by but it the system behaves as if

Right. It has optimally combined those two cues. It has taken into account uncertainty. This also is because it's with like how we really do think about the world. Like we take into account uncertainty all the time in our decisions. Right. You know this if you're driven in the fog. You're aware of this. The ninety ninety percent of what the brain does is decide what to ignore. Yeah. Right. And'cause if if we didn't.

Right. We'd be screwed. Right. We get we receive an insane amount of information, most of which does not even we don't even bother to process. Right. So yeah.

Right. You lose fidelity, right? It's it forgets. It requires constant input simply to maintain this understanding of like the low level statistics of the visual world. Right. Without input, like you're you know, you're So the question is that is that using all the information or is it just using like the low level information? And w it's information that we don't like directly perceive, but it's still con that is still but it's still definitely being used in a sense.

When it comes to you know, you know, but it's but what it what is it being used for? It's being used to track these sort of low-level statistics that that we sometimes need but don't always need. And so this is why I say that that like, you know, when we say context matters, you know, you can think of that in terms of like we were able to flexibly switch between tasks.

Which means having a lot of resources and having a lot, you know, maintained and having them still be in good working order just in case we need them. Right. And this is why like these self supervised or unsupervised learning approaches that are ubiquitous for like starting you know, getting your L L Ms to give you, you know, sort of your reasonable prior over over uh language.

Is the sort of stuff that your brain is definitely doing. So in the sense it is using everything, but it's not really using ever all of the information that's present.

Don't be ridiculous. Why? That was the m you know, it at some point in the Middle Ages it became humors, right? Because fluid dynamics was like the you know, w was the kind of technology know, the technology that was like the most advanced or uh technology that took advantage of Of water power was like the most advanced technology that we had. Now the most advanced technology is computers. So duh, that's exactly how the brain works.

I think that it's, you know, it it it has causal structure, right? But also it's a convenient class of models, right? So like the good example is large language models, right? So they're all uh you know the most most but not all are auto-regressive in terms of their predictions. All right, well why? Like why is it auto re oh, it's'cause it's mathematically convenient. It's a compact way to like take the past and make a prediction about ab about the future.

Um does it mean that that's actually the way language works? No, I don't think it's actually the way language works, but it's a it's a computationally convenient model. In physics we have like there are in fact like ver like momentum is a good example. Like why why do we need momentum in order to describe that we don't you know we don't observe momentum directly, right? We only you know, you're just looking at videos, you just you know the position of the ball.

Right. Um, you know, y you want to infer the velocity, well, you just take the difference between two adjacent positions and then that gives you but you don't ever directly observe like the momentum. And this is, you know, in a m mechanical, in a mechanical uh setting. So why did we choose momentum? Well, we chose momentum because that's the variable that that you know that if we knew what it if we knew what momentum was, now everything is Markovian.

Right. Everything is it's a f now there's like a simple like causal model that describes how the world works. We picked that model because we we picked that particular hidden variable because it's what rendered the model causal. Does that mean that's how the the universe works? Or was that just a a a a a computationally convenient choice? I'm gonna stay agnostic on that one, but I do like that it's a computationally convenient that ended up working.

with with momentum and and Markov models. We chose to have that hidden variable because it's the thing that made the model simpler, right? It made the calculations easy. Now we can just like go forward in time, just make predictions in a in a in a totally like iterative fashion. That's what makes causal models great. The other thing that makes causal models great is if you do ever intend to serve, you know. Act or behave?

Right, y then you still need to be you you know, you need to be able to um predict the consequences of your action. The the more tightly linked your actions or your affordances are to the things that causally impact the world, the more effective those actions are with respect to your model. But hopefully also with respect to reality. And so we we prefer causal models, you know, in part because they are relative, you know, relatively speaking, simpler to execute.

Right, in in in a simulation form, but also because they they point directly to well, where should I intervene? You know, where should I go in and and you know, and how should I choose my like series of actions that will give me the desired that will lead me to the desired uh conclusion or goal.

They were in poor health because they were alcoholics and they smoked a lot more than the rest of the population, right? So you do need to do that kind of intervention to discover a causal relationship. However, right, the causal relationships that we care about are the ones that mesh with our affordances, right? If you know identifying a microscopic causal relationship is super, that's great.

Right. But unless you have really tiny tweezers, it's not very helpful. Right. What you need to do is you need to identify the causal relationships that are present. in the domain in which you are capable of acting. We care about the causal relationships at the macroscopic level because that is where we live. We live in the macrosc at the macroscopic level. Most of our actions are at the microscope. Now

One of the best things about humans is our ability to extend the domain of our affordances with technology, right? We have like nuclear power. Because what we did was we acquired the ability to take tweezers, you know, at that l at that scale and like, you know, make these things happen. Right. We figured out how to take advantage of causal relationships at that level.

Not because we have those abilities, but we were able to create the tools that that that gave us access to that space. It all depends on what the problem it is that you're trying to solve. And the causal relationships that you always care about will be the ones that are related to the actions that you are capable of performing. Now that said, there's clearly a great advantage in understanding the microscopic causal relationship.

Right. If for no other reason, then that might lead to us discovering a way to expand our affordances into the m into you know into another aspect of the microscopic domain.

That you know, again, back to affordances, right? So, you know, we need to understand causal relationships at the scale that we can manipulate, right? That's what that's what matters most. Right. Because that allows us to ha uh have effective actions in the world in which we actually live. To the extent that we care about other scales, right, it's it is because simp we simply wish to expand, you know, our domain of influence. Right.

Um downward causation, if discovered, downward causation is what justified your macroscopic assumption. So what do I mean by that? I mean that like suppose I'm in the following situation, I got a bunch of microscopic elements. And they're all doing stuff and I'd like to draw a circle around them and call that a macroscopic object. Now, I am justified in doing so if that particular description of the macroscopic uh at the macroscopic level has the downward causation property.

Right. It's it's it is a way of sort of saying, oh, that was a good you drew that circle you drew, that was a good circle. right, because it summarized the behavior of the system as a whole, right, in a way that rendered the microscopic behavior irrelevant to further for further consideration.

That's what makes it a good macroscopic. I can just write down a simple equation and it says, you know, pressure, volume, temperature, right? They are going to do this over time. And like taking any little microscopic measurement becomes like totally irrelevant, right? But what made it useful wasn't just that the the microscopic measurements are irrelevant.

Right. It's that I had an equation that describes how it would have behaved, you know, that's also fairly accurate. So I have a nice determin you know, relatively deterministic model that's you know, that that is at the macroscopic level. Right. And so when you talk about like intentionality, I think it's it, you know, yes, it can be used as an explanatory variable, but it's only good to the extent that we understand how that intentionality changes over time. Right. It's a long term prediction.

And this is why like, you know, the jurisprudence example made me really uncomfortable because it's sort of like saying, well, you know, what you're kind of doing is you're saying this is a bad person. And I don't know how we would necessarily like identify like that intentionality, except in a very indirect way.

Right. That is that that then they're stuck with but then, you know, because it's only good as a macroscover variable, if it we can make predictions about the how that variable changes over time, and we're not doing that. We're saying that you're stuck with it. Right. And I just that's why it sort of makes me a little uncomfortable.

that in some way gives you this sort of uniform mathematical framework that's widely applicable to a huge number of situations. It has a lot of sort of things that are baked into uh the how we think about the world is kind of like baked into it. And so it can be applied in a whole bunch of different areas. And Carl spent a lot of time. basically evangelizing various different aspects of the scientific community. It's like, oh look.

you can apply this to epidemiology, you can apply this to the social sciences, you can apply this to physics, you can apply you know, and just sort of in and, you know, wrote a seri this is one of the reasons I think he's so prolific is because he's basically, you know, written variations on the same paper.

Right, but just applied in different domains. And he did this and this was intentional, right? Because he wanted to show that this is a unif is a nigh uniformly applicable mathematical framework. And I think he's largely right. As a result, there's all these people from all these different communities that have been pulled into his sphere that think about the world very differently. And it makes for some very entertaining conversations at the pub.

I often ask people in the business like what what what changed? Like what's, you know, what's what you know, why did we have this like massive explosion in um you know in AI development over the last several years? Um, and I get three there there are three common responses and I agree with every single one of them. Autograd, right, the transformer, but why the transformer is something that I I often disagree with with people about.

uh the transformer architecture um and just the a ma the uh the the the ability to scale things up in a manner that we haven't haven't really seen before. I actually um the reason why I say transformer comes with an asterisk is because a lot of the things that Transformers have that that people believe that the Transformer enabled um I think really m resulted more from scaling. And my the point, you know, the point of evidence that I like the site is like Mamba.

Mambo, which is a state's which is a traditional state space model. It's basically a common filter, but like on steroids. So they scaled it way up and yet and now it's, you know, got they've, you know, Mistral has their very nice like coding agent and it works pretty darn well. Right? They got a lot of the same functionality with a completely diff with a you know a completely different architecture simply by virtue of scaling. So Transformers get a

Get an asterisk. I think that the biggest thing was Autograd, right? And Autograd turned um the development of artificial intelligence.

um from being uh something that was done by like carefully constructing your neural networks and then writing down your learning roles and going through all that painful process that it was t took forever. And they turned it into an engineering problem. It made it possible to experiment With different architectures, different networks, different nonlinearities, different structures, different ways of like getting your memory in there and different ways, and all this fun stuff.

that allowed people to just start trying things out in a way that we couldn't do it before. And then we what did we did? We we suddenly discovered, oh, it turns out backprop does work. I mean, when I was a young man Like backprop was considered a non starter for two reasons, right? One is it's not brain like, which is true, right? The brain does not use backprop.

And the other one was a vanishing gradients. Oh, you'll never solve the vanishing gradients problem. And it's like, oh, it'll always be unstable. And they and yet nonetheless, once we turn it into an engineering problem, sort of playing around.

with tricks and hacks and certain kinds of non-s and relues and this and that. We discovered that, oh no, in fact, like there are ways around this. You just, you know, we just, you know, weren't gonna discover them by like playing with equations. We had to actually start it, we turned it into an engineering problem.

As soon as it got turned into an engineering problem, c you know, th that's what enabled the hyperscaling, which is what led to all of this, all all of this, you know, these great developments over the last several years. What got lost in the mix though? Was the notion that that there's more to artificial intelligence than just like function approximation? We got really good function approximation.

But that's not the only thing you need to develop like proper AI, right? You need models that are structured like the brain is structured. You need models that um you need you need models that are structured like how we conceive the world is structured. Certainly if you want to have models that think the way we think. And that that got lost in the shuffle. And we're starting to see, you know, as as as we're starting to see the limitations and the faults and flaws of of of these approaches.

um i w and starting to see them not living up to the hype, which I think is like now it's standard that like like AGI is no longer I don't know if we read the other day, at least according to you know the experts in the field at the top at the uh of of the best companies in the business, like AGI is no longer like a huge priority, right? And that they're they're they're dialing back the rhetoric surrounding that.

Um, in part because I think that they've begun to realize that like just function approximation isn't going to deliver. That was just hype, right? We do need to do something different. We do need to start get you know bringing in what we know about how the brain works. Right, if we're ever going to get to something that is a human like intelligence.

And that was the starting point for us, you know, about a year or so ago is that we were sort of like, yes, let's do the same thing for cognitive models. Like let's talk about, let's take what we know about how the brain the brain actually works. Let's take what we know about how people actually think about the world in which they live.

and start building an artificial intelligence that thinks like we do by incorporating these principles. And this means this means basically creating a you know a modeling encoding framework. For building brain-like models at scale. And that's like the critical element because obviously scaling was a big part of the solution.

And right now, most of the work in the active inference space, as I'm sure you're aware, is not at scale. There's very little like active inference work that is active inference at scale. Most of the models are like relatively small toy gridwordy type models. Um and part of the reason for that is that, you know, it is in fact difficult to scale Bayesian methods. Now that also has now begun.

Right. We now have a lot of great um mathematical tools and a lot of great frameworks for approximating Bayesian inference. You'll never do it exactly, or approximating Bayesian inference. Um which I believe is how the brain works, right? Bayesian brain and all that. that allows us to build these kind of structured models that that that that are structured both after the brain how the brain is structured and how the the world that we live in is actually structured.

hence the the the this notion that what we need to build get the ne to the next layer of of AGI and I also don't like that term and don't intend to use it very often. Um but what we need to get to the next level right is is is this um uh is this framework um that allows us to build the uh the kinds of models that we know people actually use.

and just make them bigger and more sophisticated and and and and so on. And then take advantage like hyperscaling Bayesian inference is part of it, but also like it's you know, constructing models um of the world as it actually works. The way the world actually works.

Right is what is i is i it you know, provides us with the structure of our own thinking. Right. The atomic elements of thought is how I like to phrase it, um are models of the physical world in which we live. And the physical world in which we live is a world of macroscopic objects.

that you know, um that have specific relations and interact in certain ways that we understand, right? Um, you know, uh looking around the room for a good example, right? You sit on a chair, right? That's an example of a relationship. It holds you up and all that fun stuff.

Um and those are the kinds of the you know, that understanding of the physical world was necessary for us in order you know, for us to have in order to survive. Dogs have it too, right? It's language isn't what make you know, it isn't isn't all that special, right? Well it's it's actually quite special. But um uh those are the models that form the th that that understanding of the world in which we live.

Is where we get our the models that form the the the the the the models that form the atomic elements of our thoughts, out of which we have composed more sophisticated models that have allowed us to do all this great systems engineering, build this great technology that we've got.

So that's what we want to do, right? Is we wanna is is we're focused on building cognitively inspired Models that are based on our understanding on the way the world in which we live actually works, because we believe intelligence must be embodied. Building a framework for for putting those models together and experimenting with them at scale, all in a m approximately Bayesian way because we've

It's not just about putting your AI into a robot. It's about giving that giving the robot a model of the world that is like our model of the world. A model that is object-centered. It's dynamic. It's a it's largely causal. Right. Um it's it you know, that's that's that's the big difference. And I think that the the the sort of sparse structured models

Is another sort of key differentiating component. Like when you think about how like a transformer and LLM work, a transformer takes every word in the document. And says, now how does this word relate to every other word? And it does it many, many, many, many times. Same thing with like your your your generative um uh vision language action models. They operate in pixel space.

They are microscopic models. Now, yes, do they have an implicit notion of sort of macroscopic? Yes, they must because they work, right? But it's implicit. And it's not implemented with the kind of sparse structure that actually exists in the real world and in our conceptualization of it.

And that's the thing that we are bu that we are we are saying, no no no look like if we want an AI that thinks like us, right, then we are gonna build models that are structured, like both like the real world is structured. They have this sparse causal macroscopic structure to it. And so should our models and so should and then and and the only way to do that is not just to like put a robot in the real world, but to put a robot with a model that is structured in that fashion into the real world.

But it turns out people do something which is very different. People are creative. People can solve novel problems. They can't, it's not just about mining old problems and figuring out where I can move some words around and get a pro and get an answer that looks more or less right. Right, we actually are capable of creating, we're capable of inventing new things. The way that we invent

I think is exemplified by via like systems engineering. Right. How does systems engineering work? Well, I I you know, I know how a you know, I I know repeat you know I I know how an air airfoil works to create lift, I know how a jet engine works to create thrust. Right. And I could take those two bits of information to invent something brand new, which is an airplane. That kind of systems engineering was predicated upon having this sort of model of the world that was

Relational, right? Here's the wing. I can put a jet on you. I can like st I don't know, you don't staple it on. I'm sure you use rivets or something. Right. I know how to put things together. I know how to construct new relationships and new objects.

An AI that that is designed for systems, that is designed to do systems engineering will have a object-centered or system-centered understanding of the world and will know how those all of the objects relate so that it can sort of start experimenting with different ways to combine. It's absolutely i you know, w without that, the only thing you will ever be able to do, right, is just

retool old solutions for new purposes. And it won't even and even that is I think is a generous interpretation of what a purely predictive model is gonna do. Right. So this is how I like to think about like You know, the the principal advantage of taking this object centered approach, right, is that it enables systems engineering.

something. And that something is not just the data that it saw. So example, vision language model. A vision language model is like is a way of grounding the visual model in the linguistic space. And this is the approach that we're taking. This is what Langchain does, right? It's all about taking models, and everything becomes a blank language model.

Right. Uh you know, vision, uh, you know, audio, you know, whatever. Everything becomes s and what what you're when you do that, what you're doing is you're saying that that you're grounding all of your models in like a common linguistic space so that they can communicate with one another.

Right via language. Now, why did we choose language? Well we chose language because like o honestly, I think it's because we wanted models that we could talk to, right? We wanted we wanted a model that like, you know, it was really all about the inner making the interface convenient for us. And which is great. That's it's totally something you want. But it begs the question, what's the right domain in which to ground your model?

Now I like grounding models. Like so we also use the phrase like you know, what are those with ground truth. And of course ground truth is the thing you made up and a priori said was ground truth, right? So what's ground truth? What is the what is the right do you know domain in which to ground models in order to get them to think like we do? That's the relevant question.

And so my my view is is that if you know again if you want AI that thinks like we do, you need to have it grounded in the same domain in which we are grounded. And we are grounded in this domain, right? This is why i the embodied bit is such an important thing.

Um we want models that um that are are grounded in the physical world in which we evolved. And the reason for this is because that is the world that provides us with these atomic elements of thought. A single cell like lives in a in a soup. Right. And it has uh you know, and it it it it's you know, whatever model it has of the world to the extent that it has one or it behaves as if it has one.

Um that model is it is the model of its environment, right? If it didn't understand the environment in which it lived to some extent, right, then it would it wouldn't be able to continue to exist and function in that environment. So you can sort of say that a cell has a model that's grounded in chemistry. Right, of the chemistry of the soup in which it lives.

You know, when we talk about like that that is a prerequisite for its survival. Now we talk about like mammals and bigger animals and things that live in the macroscopic world that includes other animals. Right. And you know, and all that stuff. So what's that model? What's what's what's the wor the the the you know well at the very least we can say that whatever models we have, a significant subset of them are grounded in that world.

Right. And that world we know has properties that that we can understand. It's it is object centered, it's relational, it's all this you know, all this stuff. Um and so The ground the the the the the the grounded bit is more about like properly grounded. grounded in the domain in which you know in in which we are grounded as a route to con to to creating you know AI you know an uh a AI models that in fact think like we think.

Right? That's the grounding that we that that that we're particularly focused on. If you had to choose the domain in which to ground your models, what would you choose? Right. I don't think language is the right one. Language is an incredibly poor description of both our thought processes and reality. I tell the story all the time, right? So

You ever you you ask any cognitive scientist or psychologist who's done some experimental work with humans, right? The you know, you you put them in a i in a chair, you make them do some tasks, you carefully monitor their behavior, you look at what they did.

Right. And then you have a nice way of and then you you know, that informs your theory of that behavior or or however that works. Then you and y you know, and if you do the experiment well, you have a very good model of how they made whatever decisions they made throughout the course.

And then you go back and you ask them what it why did you do what you did? And they give you an explanation. It sounds totally reasonable. It also is completely inconsistent with an accurate model of their behavior. Self-report is the least reliable form of data, right, that one gets out of a cognitive or psychological experiment.

And so we don't want to rely on that. We don't want to ground our models in what we know is an unreliable representation both of the world and of our thought processes. Right? We want to ground it in something that's a good model.

uh mathematical framework that you can throw up many problems. And a lot of what has gone on in the active immersive community over much of the last 20 years has been demonstrating that um it's act it's it's like uniformly applicable. So there's been a lot of bread breath

And not a lot of depth. Right. And part of, you know, and I and I think that, you know, of course, like you know, that that's appropriate, like given, you know, you if you really want to make the the argument that everyone should be using this, you shall see in this in this domain it works on your like toy examples.

But the people doing that right, it's kind of, you know, the active image community has had this has this habit of showing like, like, see like, oh, this basically like I I'm I can handle this like psychological phenomenon, I can model this cognitive phenomenon.

Oh and look, like it's a good post hoc description of this neural network's behavior and things like that, right? They're they've been showing that but they've they've never really sat down and and and s and like tried to tackle any really big, really hard problems.

Because the emphasis has been on evangelism. You couple that with the fact that there is this strong bias within the active inference community towards being as Bayesian as possible. And so of course they also like shun the really hard problems because Bayesian inference

you know, is i has been historically challenging to scale. There have been a lot of developments over the last few years that, you know, um that have come, you know, out of the machine learning community as well, you know, but mostly out of the Bayesian machine learning community. um that that have really made it possible to start scaling Bayesian inference in ways that we we we we really weren't able to do it before.

Um uh and you couple that with a desire, you know, to sort of stop the evangelizing and start solving really hard problems with these methods. Um, and you've got a way to prove that like active inference really can live up to its prompt.

Apparently just that there are all these empirical methods that have been discovered in the last twenty years that just make it massively more scalable and tractable. And is it the same thing here? Like are are there some specific techniques that have dramatically improved the tractability of active infra

Uh my my my current favorite trick is is you know normalizing flows, um, which is a great way of ensuring that you have like access to sophisticated likelihoods but nonetheless result in tractable probability distribution. Um uh uh there's the work um, you know, I mean I've been using

I've been using like natural gradient methods for a very long time, which allow you to like massively speed up gradient inference. And in some situations, completely eliminate the need to do gradient inference and instead like, you know, do coordinate descent and allowing you to take massive jumps in parameter space and not actually lose the the ability to do learning of sophisticated model in in a sophisticated modeling scenario.

Um I also like the fact that like the the natural gradient stuff has been getting some great acronyms recently. Bayesian online natural gradient for bong for short. These guys these guys get me every time. I I wish I was that clever, honestly. But like it's it's but there've been a lot of developments in that space as well.

Um, you know, making you know, uh in in addition additionally, there's been a lot of developments in like rapid sampling methods, uh conditional sampling methods, constraint methods like that that that that that that have really improved things. And I think that like one of the problems again with the active earnings community, you know, historically that that I think is now starting to change.

has been a hesitance to use these sort of certain approximate methods. There's been this this focus on like straight up old school message path. And you know, as soon as you sort of, you know you know, if relax the desire to be as Bayesian as possible, it opens up a lot more possibilities for for scaling this stuff.

Um, you know, so if you want to train you know a vision model to understand like YouTube videos or something you know really complicated like that, you basically take one big model and you train it on a ton of data, right? You just keep training, keep training, keep training. And then eventually it sort of gains this implicit and it does get an implicit sort of sort of object-centered understanding.

Another way to go, right, is is to is to you know train objects in specific domains that are so s these are smaller data sets. Like I'm only gonna worry about like the Zillow problem, like the inside of people's houses.

Right. And that's gonna have a much smaller set of objects, right, that it has to that it has to learn an implicit distribution over. And you can do this with one big neural network, right? You can train and the you know, there's a a really great like You know, Gaussian splatting paper where they trained a massive neural network that is able to like, you know, sort of make predictions about what's going on inside people's houses and some nice language models.

Um but obviously it has it has an understanding that's limited to a house and the objects that are inside a house. If you have an explicitly object-centered mob. Then you end up not just with one model that understands a house. You end up with one model that's actually thousands and thousands of little models.

Each of which, right, you know, um uh sort of explains like a single object or object class within the house, right? So you got like a book model. So all the books come in different shapes and colors, right? But there's just one like book model. And the beauty of doing this is that uh is that that book model.

You have to be a little clever about the how you structure the interactions between these things. But if you're a little bit clever um ab about um how you describe the relationships between objects within this modeling framework, you gain the ability to train a model just on the insides of houses, a model like just on you know just on like um you know part

and park benches and take the objects that were discovered in this space and the objects that were discovered in this space and put them into a combined environment that has objects of both of those kinds and it still works. Right. That's the advantage of the of taking an object-centered approach or a what I like to refer to as the lots of little models approach.

Um you can, you know, let's pretend we're living in a world of rigid bodies, right? So all I need to worry about is like weight and shape of things. And that defines a particular object type. But I also need to know how they interact. And so in in in Newtonian mechanics we have like you know, what we can do is we can take these objects, we can watch them like bouncing off of each other and doing all these sorts of things, and we can quickly infer that like, oh,

Like their interactions are all governed by a s in a single language, which is the language of forces and force vectors. Right. Um that l that that language of interaction right, is really what makes it work. Right. Otherwise we'd just have like, you know, pictures of things. That's all we would have got. What you're empirically discovering is sort of a generalized notion of forces that describe the relationships between things.

And you cons the constraint that you place in order to avoid the problem of things being too brittle, right, is that well, they all have to use the same class of forces together in order to interact. We're stuck with that. But by being flexible. About our definition of what a force is and the having the ability to discover new kinds of forces, not just like literal force vectors. Right. Um uh gives us the ability to sort of generalize um without be without becoming too brisk.

Right. It may very well have been that all of your observations of this object, you know, when it was in a house were like really simple. It was all just it sits on a shelf, right? And so what do you know? Well what you know is that that object sits on a shelf. But you have to be you know, which is one kind of interaction, right? That's just the you know

It has a force pushing down, there's a force pushing up. You don't know anything about like the weight and all this. You have nice, but but if you keep error bars about that, if you keep error bars about the other kinds of interactions that you have seen, but are agnostic, right, about the specific deals for this particular object.

It gives you the flexibility to say, Well, I'm gonna put it in this environment, and I can make some predictions about how it's gonna behave, right? But if I throw a bowling ball at it, I'm going to be making some assumptions about how it might behave. But once the bowling ball hits it, right, I might have to revise those assumptions.

This is the other critical element of the approach we're taking, which is continu which is you have to have some kind of continual learning element. This is something that really doesn't exist in contemporary AI. Right. And you know, you know, when you build your big model, you've spent millions of dollars training it and then you're done. Right. Yes, someone else can come along and fine-tune it a bit for a particular task, right, which is great.

But at the end of the day, when you're at the deployment phase, you turn learning on. Um, whereas in this approach we're saying, no, no, no, you've one of the things that's critical, a critical aspect of the way we think about the world and the way we learn about the world, is is that it's continual and it's interactive.

Right. So there's you know and that needs to be true of the objects that we're discovering as well. We've learned classes of interactions, but just because we haven't seen a particular class of interactions previously doesn't mean we say the others never happen. Right. We still allow for that possibility, right, and then do continual learning quick with with rapid updates when we see something happen. We we see you know a new interaction. Now the what makes that work.

Right, is the fact that you've specified that there are certain set of kinds of interactions, some of which you previously observed, some of which you still don't know about and might observe soon, and then you could update your posterior beliefs about whether or not that object interacts in that.

And and then we have the consistency problem because maybe this agent has empirically learned that these two things are a book but the agent over there, you know, just thinks this one thing is a book and then there's how many objects are there? Would it become intractable, you know, like realistically I don't know what you're saying? So

Maybe a texture, color, or something like a fork is an asset, right? Or a little like three-legged stool is an asset. Um, and it has a bunch of properties. But it's basically, you know, that the you know, that are associated with um its shape, color, mass.

All of this stuff. And then it has a set of interaction rules which are like Newtonian forces, force vectors. Then you've got other things like water and sand that have like special rules for them because if you just try to'cause otherwise, you know, you need a macroscopic rule to describe them, otherwise

it the compute would be insane and stuff like that. So it's very similar to that, right? We you know what we do when you take this lots of little models approach, what you end up with is the moral equivalent of a giant list of video game assets. Right. And then when it goes to modeling a particular environment, right, when you find the agents that you're talking about that has this lots of little models model in its head, what it does is is it it um

It sort of looks at the scene and says, oh, okay, I need to worry about these 10,000 little models right now. And that's it. I don't need the rest of it. Right? And then it just sort of operates in that space, running something that looks a lot like a video game series. Right. So it's it's that sparsity.

Um is what makes is what makes this little this lots of little models approach, right? You may have a million little models, but at any given time you only need a tiny fraction of them and you just instantiate those.

that represents things in the world in the game engine. And if it's learned a sparse, robust model library, you could in principle take the same learned model and apply it to a robot in the real world and it would generalize.

But it could also be a result of a situation where the artificial environment just isn't actually a very accurate representation of the real world. Right. And I think it's I think it's largely the last. Right. Is that is that these, you know, the these um I mean uh also coupled with the fact that the robot that the artificial agents internal model looks nothing is not structured like the m the world that it actually is being trained to function in.

I think those are the two big those are the two biggest problems. But it could also be the case, you know, what do you need in order to address those? Well one is you need a good model for the robot's brain that has the structure of the world in which it lives. The other thing is you need a mapping from real-world data to simulated data. And right now what we're typically using is video game engines. Now video game engines are great. I know I certainly enjoy them on a 10-hour a week basis.

Um the uh the the problem with them though is that they weren't trained to be realistic physics. Right. Most of them were were were designed to be plausible. They were designed to look good to the user. Um uh part of this has to do, you know, and and there's a lot of like tricks and hacks and things that are thrown in to deal with the fact that the equations of Newtonian mechanics.

um are very stiff, right? When collisions happen, if you're just a little bit wrong about that, you know, then things can you know, then weird stuff can happen and and non non-physically realistic things can occur.

Um so if if you if you had the ability to construct an environment that had good enough physics, that accurately represented the real world and trained your robots in that domain where they have these models in their heads so they're actually capable of learning, you know, the quote unquote ground truth that you've implemented in the in the simulator world.

then I I I I believe that they will uh generalize better to functioning in the real world. And this is this is absolutely critical, I think, for robotics going forward, if for no other reason than right now. I mean like large language models, all these self-supervised models, um the way that we're currently training robots to like put your put your groceries away and things like that is all by training them to mimic human behavior, right? It's expert trajectory learning.

They're not really learning the physics of their environment. They're learning to mimic human behavior without like crushing the Right. And so with you know you know, if you want them to be able to generalize across domains, across tasks, you need to get rid of reliance on expert trajectory learning. And so that's the and and that only happens when you move to something that is explicitly model-based, with a model that accurately represents the world in which

that has been trained to like manage a warehouse, right? And so it understands things like forklifts and boxes and workers, hopefully, you know, and and all the and it knows how to and then one day, one day something comes along. That's never seen before. It's a cat. Cats don't belong anywhere else. A cat comes along, right? And so the the model has no idea, has never seen a cat before.

Right, because that's the environment in which it was trained. This is the one of the beauties of this approach. So it's the cat comes in the warehouse and it's like, what the hell is this? And like it's you know, it's screwing with my system. And um because we're taking this sort of like, you know, free energy based approach, right? One of the critical elements is that is is tracking surprisal. So when a cat comes along

Doesn't know what a cat is, the surprisal signal goes crazy, and then it says, okay, stop, right? Don't run over the cat, right? Let's figure out what's going on. And what it can do is it can take a picture of the cat. And it can fire it off to a a server somewhere that has a huge bank of models and has been pre-trained on on model selection to a small extent. Um and say and it says, what the hell is this?

And then the big bank of models says, Oh, I think it's, you know, here's like seven or eight things it could possibly be. And it's different kinds of cats, maybe there's a dog thrown in, whatever. And then it ports those little models over to the to the to the to to the warehouse model. And then it sort of does some proper hypothesis testing, watches the cat behave for a little bit. Ah, it's a cat.

Puts the other models, sends the other models back because it doesn't need them anymore, right? It's figured out that this is what it is. And now it's incorporated understanding of the cat into the system. This is the beauty of taking an explicit this is another beauty of taking an explicitly object-centered approach. It gives the model the ability to be

Um to uh to to to know what it doesn't know. This that comes from the active inference component. Know what it doesn't know. When it doesn't know it, it can go phone a friend. That's another way to describe it. The friend will respond by saying, oh, it's a cat, and then they can take the model of the cat, incorporate it into its warehouse model, right? And now it understands that.

This is really great from a compute there's a huge compute advantage to this, right? If we had started with one big model that already knew what all a cat was, think of how many parameters that would have. It'd be huge. This model is very frugal in the sense that it only knows it only needs to know two things. What it needs to know about the environment in which it in which it exists.

Right. And when it sees something it doesn't know, and then it can just go pull it. So that's the idea is that you have this massive bank of models, but when you instantiate a particular if for a particular use case, you don't need them all. Yeah. Right. You just need the ones that are relevant to that environment. But you could but

These cape these models are continuously tracking surprise or uncertainty. And when it sees something it doesn't know before, it's smart enough to say, I don't know what that is.

and it's like, you know, three hundred pixels squared or something, you know, that that would be a challenge for Bayesian inference. So I'm thinking could you could you uh just use a like a vision language transformer or something and use that as part of the Bayesian framework or Could you even use deep learning models as a way of bootstrapping the knowledge acquisition in the in the Bayesian framework?

For sure. Um and you know, that's that that's a that's a great example of of of one where we're taking sort of like, oh well like why wouldn't we use this if it's compatible with our framework?

And you're learning a mapping from that distribution onto the thing you actually are observing, the thing you care about. So in this case it could be an image. In fact, I actually don't think we should call them diffusion models. It's a normalizing flow.

The diffusion is it's it's it should be referred to as a diffusion training protocol for a normalizing flow. So to some extent we will be using some of those tricks as well. You can say, yeah, we're gonna use diffusion models if you can make me roll my eyes and say that.

And they're and their the reward functions that we specify for these artificial agents um are definitely not the same as our reward functions. Now there's a few exceptions, it's like like like go. Right. Any game where you either win or you lose, the reward function is obvious. Um, but in general, like In complicated situations, reward functions um

uh it's not so obvious what the reward function should actually be. You know, I know that there's this definite belief that like reward is all you need, and I and and and there's some truth to that. But the question is, well, where'd your reward function come from? Now from a philosophical perspective, you know There there is no normative solution to the problem of reward uh function selection. I always say barring divine intervention.

And that which is just another fancy way of saying that like your values and my values might be different and it's really difficult to say whose are better, right? Um from a practical perspective, right, um a a situation that I like to point out is like if you're talking about self driving cars, obviously you'd like to penalize your self driving car if it drives over a squirrel.

Right, but if it had to choose between a squirrel or a cat, you probab most people would want it to choose the squirrel. And the way you would do that in RL model is you say minus 10 points for a squirrel, minus 50 for a cat. Where do those numbers come from? It's completely ambiguous, right? It's relatively arbitrary. They're kind of sort of made.

Um and so relying on on arbitrarily selected reward functions like seems like a terrible idea. We also know that like things can go horribly wrong and I know everyone's sick of this example, but like You know, you alwa you know, when when you when you rely on reward, you're you're effectively like, you know

making wishes from like a malevolent genie or you know, you run the risk of saying, hey, hey, Skynet, N world hunger, and it's like no problem, kill all humans, right? Those those are if you don't specify your reward functions very carefully, you can get very degenerate behavior. So the goal of alignment in an RL setting is to get you it is to get it would be to somehow get my reward function, or perhaps humanity's collective reward function, right, into the AI agency.

This is really, really, really hard. It's really really really hard because measuring reward functions is really really really challenging. Um the approach that we're taking is we're saying, well, so how do people actually do this? How do how do we as humans um construct alignment? Well the first thing we do. is we try to figure out what other people's reward functions are. The problem of reward of reward function identification is conflated by the fact that people have different beliefs.

Which is what we can observe other people doing, is a combination of of their beliefs, right? And their and and their and their reward function or their value. Um and so just sort of like taking those, you know, so the problem of course is that you only observe people's actions. Like there's a difference of opinion about what to do, right? And so you want to figure out why. And it could be because your beliefs are different or it could be because your values are different.

The only the but it but it it's ambiguous. You can't tell. It's math mathematically it's not even possible to separate these two. Belief and value are fundamentally conflated when all you observe is action or decision. The way that we solve this problem as people is we talk about our beliefs. Right. I tried to I ask you, well, why do you think this is the action? And then you tell me, like, oh, well, it's because you know, this fact, this fact, and this fact.

suggest that if I do this, then this will happen, right? And then I can go in and say, ah, I see. So maybe the reason for the disagreement in our beliefs or in our in our decision, right, is because you're not aware of this fact. And I'd forgotten about this fact.

And so what we do is say, well, let's incorporate all of these things together and then sort of see and then sort of see and then you would still say like, well, I still think we should do X. And I'm like, no, it's still definitely Y. And we continue this conversation until we both e until each of us has a very reasonable model. of the belief formation mechanism that the other person has. At which point, the only cause for disagreement is a disagreement about the reward function.

You know, what would happen if I did X, Y, and Z? And then it just sort of tells you, well, this is the this is the ultimate outcome. And then we're like, oh, okay, well then maybe A, B, and C were better choices, right? And things like that. Um that eliminates them from the consider from the consider from from uh participating in the val in the you know, in the actions sorry, that that prevents them from using their reward.

Right. And you can get that just by like training them to just do good prediction. That's totally great. But that doesn't give us the kind of automation that we really want. Right. What we really want are decision uh you know, artificial dis agents that are decision makers that can act on our behalf. Um and so it's either gonna be human in the loop or it's going to be something like what I propose where we figure out how to like solve the alignment problem in in that fashion.

I have an intuitive feel for it. I assume you work the way I work. I have beliefs and intentions. Uh you know, that's my perspective of myself. And so I conclude the same about you. Um it's uh it's kind of like emergence. Emergence is such a funny concept, right? Is that um there's this whole class uh th this whole like branch of the emergence literature that defines an emergent phenomenon as anything that I didn't predict.

Right, which is a remarkably anthropocentric and I would argue ignorance-based definition of of emergence. And I don't like it. That's for the for those reasons. Um the same sort of thing g you know goes with you know uh you know, I I think this th the sort of converse of that is what's going on here. It's like we we know that these that these algorithms do not have like the capabilities do anything other than predict. And so we don't believe they have intent.

I don't even like that. I don't even like that. Um I am more sympathetic to that, but I I prefer definitions of emergence that are sort of more pragmatic, right? That are sort of like, oh no, an emergent phenomena this is why I like downward causation as like a fundamental feature of emergent behavior.

Uh mostly because downward causation that is it it's not only a nice explanation of when you can you know, a nice sort of fairly rigorous definition of when you can say a phenomenon is emergent, it also comes with a practical tool. It tells you you don't need to model the microscopic phenomenon.

And so what they do is they have a a a field and there are obstructions and they're squares and circles and things like that. And then they have these little creatures that are sort of like a little amoeba-like swimmers, like they've got like fins in the back and everything. And they swim and they'll hit one of these obstructions.

Which will cause them to deform and kind of look like, oh, it's gonna die. That's so sad. And then it reforms and becomes itself again. And so we thought of this as like a really nice abstract environment in which to to test.

the um you know some of the properties of the physics discovery algorithm. Because one of the nice things about the approach we've taken is is that as the thing, as a little swimmer goes and hits hits something, it's possible that it loses its identity when it like deforms into something new. And then reforms into itself. And we wanted to see if the approach we've taken captures that. And it more or less does, right? It hits the obstruction.

Right. The obs i either you know it it it changes its identity into an object of a different type and then reforms and comes out the other side and then regains its identity back.

And I haven't read it in detail but it looks amazing, so definitely look at that. And now you're taking your system and you're applying it in like something like a game of life basically. But you still expect it to work.

Right. And it's even a force that's like kinda like real forces. It's like a weird kind of charged particle thing. Um and so I I still think that you can, you know, it's you know the the approach that we're taking is basically just discovering what are the effective forces between

It's you know we're not worried about the microscopic forces. We don't care. That's the whole point of a macroscopic physics. You know that there are microscopic forces that govern the behavior of the system as a whole. But what you're interested in are the things that that that make predictions on the scale you care about. And so what you're doing is you're discovering the effective rules.

that describe the interactions between not just the particles that make or the pixels that make up the the the little floater or or whatever flyer. Um but the rules that govern its interactions with other floaters or physical objects like obstruct they put obstructions in the domain and things like that.

So do you think in the future we might actually have an AI system which is running inside a cellular automatic?

Find a bunch of discrete objects that go in a certain direction and then you're allowed to like tweak the rules that govern the local interactions until you get something that more or less does that. Um that's just programming in a sort. backhanded way. Um but uh uh I think that's I mean it it it those systems are very interesting because it is remarkable that really dumb simple rule

right can lead to like really interesting sophisticated behavior. But I f the thing that I find interesting though isn't the fact that like s like complicated s you know complicated stuff can result from like simple local rules. What I find interesting, what I'm more interested in rather, is like, well, what are what are the properties of the resulting large-scale objects? Right? How is that related to the small-scale object?

What's the mathematical description of those big things? The things that w the things that are that have emerged? I'm less concern uh less interested in how they precisely emerge. This probably is because of my bias towards taking a a human cognitive approach. Most of the people don't act you, you know, when you look at the game alone.

When most people think, oh, that's really cool, like look at these pretty pictures and all these little creatures and they're doing fun things, they don't really care about the low level rules, right? The thing that captures their imagination is the macroscop the the la high level, the macroscopic level behavior of these things. Um

Um and to some extent the approach that we're taking is compatible with program synthesis, right? We're taking this object-centered description of the world. Um and the reason we're doing that is because we want to automate systems engineering. Well what's systems engineering? Oh that's like taking this object and attaching it to this one, attaching to this one until you get something that does something really cool. Right? Program synthesis.

Right. Is an abstract way of doing that, right? Is that you start with one program, you attach it to another program, attach it to another program, and so on and so forth.

And it it it's great, it has many good properties that it's a program, but it doesn't really make sense to us.

Um, there was a a paper in Nature. This is actually at one of those situations where um, you know, neuroscience is is making interesting statements about machine learning. uh from Tony Zador. And what he had done is they'd taken a whole bunch of neural networks that did a variety of different things and then they came up with a way of genetically encoding them.

Um for the purposes of seeing if like okay, so what's the str so it's like, oh I had to have a layer that did this and then a layer that did this and the And then and th and then I'm what I'm gonna do is I'm gonna like compactly represent the weights in each layer and come up with a representation of that. And then I'm just gonna look at a whole bunch of different neural networks that solve the whole bunch of different problems and say, are there any like patterns?

that are present in these neural networks such that when I have a new problem I'm interested in, I can sort of just, you know, take Something that understands this genetic code, maybe mutated a little, as a way of sensibly, you know, traversing the space of possible neural networks. Until I find the best one. Program synthesis could in principle exploit the same trick. They just need the data set to do it.