Transcription
In the name of, you know, throwing out, kind of new unproven ideas. Right? So this is, you know, this is this is just my my my conjecture. We've done we've done some we've done some computational modeling of it, which which I initially this was a, a, a a very, clever, student that I work with named Lakshwin who did did some models with me. And, I initially thought it was a bug, and then I realized that, no.
Actually, this is this is this is the feature. The idea is this. Imagine so so we've been working for a long time on a concept of, competency among embryonic parts. And what this means is basically the idea that, there are there are, homeostatic feedback loops among various cells and tissues and organs that attempt to reach specific outcomes in anatomical morphospace despite various perturbations. So the idea is that if you have a tadpole and you do something to it, whether by a mutation or by a a drug or something, you do something to it where the eye is a little off kilter or the mouth is a little off, all of these organs pretty much know where they're supposed to be.
They will try to minimize distance from other, landmarks, and they will remodel, and eventually you get a normal frog so that so that they will they will sort of, recover the correct, anatomy despite starting off in the wrong position or even things like changes in the number of cells or the size of cells. They're really good at getting their job done despite various changes. Right? So, okay. So they have these competencies to to optimize specific, things like like their position and and their structure and things like that.
So, so so that's so that's competency. Now, now here's here's the interesting thing. Imagine that you have, a species that, has some some degree of that competency. And so you've got an individual of that species comes up for selection. Fitness is high, looks pretty good, but here's the problem.
Selection doesn't know whether the fitness is high because his genome was amazing or the fitness is high because the genome was actually so so, but the competency sort of made up for it, and now everything kinda got back to where it needs to go. So what the competency apparently does is shield information from evolution about the actual genome. It makes it harder to pick the best genomes because your individuals that perform well don't necessarily have the best genomes. What they do have is competency. So what happens in what happens in our simulations is that when if you start off with even a little bit of that competency, evolution loses some power in selecting the best genomes, but it but but where all the work tends to happen is increasing the competency.
So then the competency goes up, so the cells are even better at and the tissues are even better at getting the job done despite the the bad genome. That makes it even worse. That that that that that makes it even harder for for evolution to see the best genomes, which relieves some of the pressure on having a good genome, but it basically puts all the pressure on being really competent. So you've got this so so basically what happens is that, the the g the the genetic fitness basically levels out at a really suboptimal level. And in fact, the the pressure is off of it.
So so it's tolerant to all kinds of craziness, But the competency and the mechanisms of competency get get pushed up really high. So in many animals and but there are other factors that sort of push against this ratchet, but it becomes it becomes a positive feedback loop. It becomes a ratchet for optimal performance despite a suboptimal genome. And so in some animals, this sort of evens out at a particular point. But I think what happened in Planaria is that this whole process ran away to its ultimate conclusion.
The ultimate conclusion is the competency algorithm became so good that basically whatever the genome is, it's really good at creating and maintaining a proper worm because it is already being evolved in the presence of a genome whose quality we cannot control. So so in computer science speak, it's kind of like, and Steve Frank put me onto this, analogy. It's kinda like what happens in RAID arrays. When you have a nice RAID array where the software makes sure that you don't lose any data, the pressure is off to have really, really high quality media. And so now, you can tolerate you can tolerate media with lots of mistakes because the because the software takes care of it in the in the rate and the and the and the architecture takes care of it.
So so, basically, what happens is you've got this animal where that, that runaway, feedback loop went went so far that the algorithm is amazing, and it's been it's been, evolved specifically for the ability to do what it needs to do even though the hardware is kinda crap. And and and it's it's incredibly tolerant. So so this has a number of implications that that, to my knowledge have never been explained before. For example, in every kind every other kind of animal, you can you can call a stock center and you can get mutants. So you can get mice with, with kinky, kind of kink tails, you can get flies with red eyes, and you can get, chickens without toes, and you can get, you know, humans come with various, you know, albinos and things.
You can, you can, there's, there's always mutants that you can get. Planaria, there are no there are no abnormal lines of planaria anywhere, except for the only exception is our two headed line and that that one's not genetic. That one's that one's bioelectric. So so isn't it amazing that that that nobody has been able to despite despite a hundred and, you know, I don't know, a hundred and twenty years of experiments with planaria, nobody has isolated a, a a line of planaria that is anything other than a perfect planarian. And I think this is why.
I think it's because they have been actually selected for being able to do what they need to do despite the fact that the that the that the hardware is just very junky. And and so so that's my that's my current that's my current current take on it. And and and really, it puts more kind of more emphasis on on on the algorithm and the decision making among that cellular collective of what what what, you know, what are we gonna build and what's the algorithm for for making sure that we're all working to build the correct thing. So if you translate this idea into computer science, a way to look at it is imagine that you find some computers that, have, hard disks that are very very noisy. And, where the hard disk basically makes lots and lots of mistakes in encoding things and bits often flip and so on.
And you will find that these computers still work and they work in pretty much the same way as the other computers that you have. And, there is an orthodox sect of computer scientists that thinks, it is necessary that, every bit on the hard disk is, completely reliable or reliable to such a degree that you only have a mistake once every hundred trillion copies. And you can have an error correction code running on the hard disk at the low level that corrects this and after some point it doesn't become efficient anymore. So you need to have reliable hard disks to be able to have computers that work like this. But, how would these other computers work?
And it basically means that you create a virtual structure on top of the noisy structure that is correcting for whatever degree of, uncertainty you have or the degree of randomness that gets injected into your substrate. David, Dave Eckley has a very nice metaphor for this. Do you know him maybe? Yeah. I know.
Yep. Yeah. He's a, I think, beautiful artist who explores complexity by tinkering with computational models, and really find his work very inspiring. And he has this idea of best effort computing. So he in his view, our own nervous system is a best effort computer.
It's one that does not rely on the other neurons around you working perfectly. But, make an effort to be better than random. And, then you stack the improbabilities empirically by having a system that evolves to measure in effect the, unreliability of its components and then stack the probabilities until you get the system to be deterministic enough to do what you're doing with what you to do with it. Right? So you, if you have a system that is as in the planaria inherently very noisy, where the genome is an unreliable witness of, what should be done in the body, you just need to interpret it in a way that stacks the probabilities that is evaluating things with much more error tolerance.
And, maybe this is always the case. Maybe there is a continuum, maybe not. It's also possible that there is some kind of phase shift where you switch from organisms with reliable genomes to organisms with noisy genomes and you basically use a completely different way to construct the organism as a result. But it's a very interesting hypothesis then to see if this is a radical thing or a gradual thing that happens in all organisms to some degree. Yeah.
What I also like about this, description that you give about how the organism emerges, it maps on, in some sense also in how perception works in in our own mind. At the moment, machine learning is mostly focused on recognizing images. So or individual frames and the you feed in information frame by frame and the information is actually disconnected. For a system like, DALL E two is trained by giving it several hundreds of millions of images. And they are disconnected.
They are not adjacent images in the space of images. And a baby could not probably learn from giving 600,000,000 images in a dark room and only looking at this introduced the structure of the world from this, whereas DALL E can, which give gives testament to the power of our statistical methods and hardware that we have that far surpasses, I think, the combined power and reliability of brains, which probably would not be able to integrate so much information over such a big distance. For us, the world is learnable because its, adjacent frames are correlated. Basically, information gets preserved in the world through time and we only need to learn the way in which the information gets transmogrified. And these transmogrifications of information means that we have a dynamic world in which the static image is an exception, the identity function is a special case of how the universe changes.
And we mostly learn change. I just got visited by my cat and my cat is, has difficulty to recognize static objects compared to moving objects. So it's much, much easier to see a moving ball than a ball that is lying still. Yeah. And it's because, it's much easier to segment it out the environment when it moves.
Right? So the task of learning on a moving environment, a dynamic environment is much easier because it imposes constraints on the world. And so, how do we represent a moving world compared to a static world? The semantics of features changes. And an object is basically composed of features that are there can be objects themselves.
And the, scene is a decomposition of all the features that we see into a complete set of objects that explain the entirety of the scene and the interaction between them. And causality is the interaction between objects. Right? And, in a static image, these objects don't do anything. They don't interact with each other.
They just stand in some kind of relationship that you need to infer, which is super difficult because you only have this static snapshot. And so, the features are classifiers that tell you how to, whether a feature is a hand or a foot or a pen or a sun or a flashlight or whatever, and how they relate to the larger scene in which, again, you have a static relationship in which you need to classify the objects based on the features that contribute to them. And you need to find some kind of description where you interpret features, which are usually ambiguous and could be many different things depending on the context in which you interpret them into one optimal global configuration. Right? But if if the scene is moving, this changes a little bit.
What happens now is that the features become operators. They're no longer classifiers that tell you how your internal state needs to change, how your world needs to change, or your simulation of the universe and your mind needs to change to track the sensory patterns. Right? So a feature now is an change operator, a transformation. And, the feature is in some sense a controller that tells you how the bits are moving in in your local model of the universe.
And they, organize in a hierarchy of controllers. And these controllers need to be turned on and off at the level of the scene. And they have a lot of flexibility once you have them. They can move around in the scene. They're basically now self organizing, self stabilizing entities.
In the same way as the mouse is moving around in the organism, a feature can move around in the organism and shift itself around to communicate with other features until they negotiate a valid interpretation of reality. That's that's incredibly interesting because, you know, as soon as soon as you started saying that, I was starting to think that the the virtualization that enables right? So it's the the earlier part of which you were saying the virtualization of, the the the information that allows you to, deal with with unreliable hardware and everything. The the the bioelectric, circuits that we deal with are a great candidate for that because, actually, we see exactly that. We see a a bioelectric pattern that is very resistant to changes in the details and make sure that everybody does the right thing under a wide range of, you know, different defects and so on.
But but but even more than that, the other thing that what you were just, emphasizing this, the fact that we learn the delta, right, and that and that we're looking for change. Very interesting. If you if you pivot the whole thing from the temporal domain to the spatial domain, so in development, when we look at these bioelectric patterns, now these patterns are across space, not across time. So unlike in neuroscience where everything is kind of in the temporal domain for neurons, these things, these are static voltage patterns across tissue, right, across the whole thing. So for the longest time, you know, we asked this question, how are these read out?
What, how do cells actually read these? Because, because one possibility early, this was a very early hypothesis, you know, twenty years ago, was that maybe the local voltage tells every cell what to be. So it's like a paint by numbers kind of thing. And every, and and each voltage, you know, ray each voltage value corresponds to some kind of outcome. That turned out to be false.
What we did find is that there and we have computational models of the of how this works now. What is read out is the delta, the difference between regions. It doesn't care nobody cares about what the absolute voltage is, what what what is read out in terms of outcomes for for downstream cell behavior, gene expression, all that. What is actually read out is the voltage difference between two adjacent domains. So that is exactly actually what it's doing just in the spatial domain.
It it it only keys off of the delta. And what is in what is learned from that is, exactly just as as you as you were saying, it modifies, the controller for for what's downstream of that. And there may be multiple ones that are sort of moving around and co inhabiting. I mean, it's a very it's a very compelling, picture actually and way to look at some of the, some of the simulations that that we've been doing about how the bioelectric data are interpreted by the rest of the cells. If you enjoyed this toe clipping, then the full video is linked in the description.
You should also sign up for toe mail, which is, again, in the description and the pinned comment. You'll receive immediate access to all exclusive updates from the Theories of Everything podcast. It also helps me communicate directly with the core Toe community. You'll also receive my top 10 Toes. Think of it as the intellectual version of Quentin Tarantino's obsession.