Can Human Neurons Really Play Doom? The Science Behind Wetware

Intro — did a dish of human neurons really play Doom?

0:00It's called wetwware engineering and it's effectively neurons, human neurons plus electronics. So you cut out the middleman. The learning mechanism is literally minimizing surprise. If on the other hand it misses, it's going to get 4 seconds at 20 volts of highly chaotic electrical activity. So now I can very easily ask who is actually doing the learning. Correct? Is it my reinforcement learning network or is it the actual biological neurons that I have? >> Doesn't count. Hello internet. This is your captain speaking Lester Nar joined as always by my co-host and our resident PhD Krishna Chowdery. And today we're

0:43going to be talking about a topic that sounds like it's something straight out of science fiction, which is did a dish full of human neurons successfully play Doom? the video game. Now, this is something that has been hyped in the news over the last two to three weeks with everyone trying to understand is

Artificial neural networks vs biological neurons

1:04this real or is this hype? We're going to dive into the origins of how this came about through a research paper in Neuron out of Cortical Labs in Melbourne, Australia, published in 2022, and look at that paper versus the recent news. And we might even touch on some of the other big fundamentals that made this concept possible. We are going to learn about the science from the ground up today because this is from first principles.

Wetware engineering — cutting out the middleman

1:51So in our world of AI, we are very much used to artificial neural networks, right? And these are biomimetic AI in the sense that they mimic biology, right? Biological brains became artificial neural networks. The subject of this deep dive today is about a possible transition from this biomimetic AI to straight up biiomputing. It's called wetwware engineering and it's effectively neurons, human neurons plus electronics. So you cut out the middleman. You don't even make any artificial neurons. You're using now biological neurons to help you in

2:33computing. Specifically, there's an Australian company called Cortical Labs. They were recently in the news because

Biological neuron 101 — dendrites, spikes, and action potentials

2:39their wetwware computer could play Doom and it was all hype. Right. I've seen it literally everywhere. Yeah, >> the tech bros are going crazy. >> Yeah, all the doomsday people are like, "Should we be doing this?" Right? >> So, you know, I naturally asked, "How much of this is real and how much of this is hype?" >> And turns out there's a lot of both. >> Okay. >> And that's what we're going to get into. >> Okay. >> Okay. So, let's do a little bit of brief background. Let's talk about the biological neuron, the one that is the substrate of our own nervous system in our human brain. A biological neuron is a cell, a biological cell. We see that on the left

3:21here. It's got a cell body. It's got dendrites which are where other neurons come in and give you input. And in the cell body, what you do is you integrate, meaning you sum up all the inputs. And if all of the incoming inputs from other neurons exceeds some kind of threshold, then you fire what's called an action potential, which is, in other words, a spike. It's effectively saying my neuron went from off to on because all of the other neurons that were talking to me, there were enough there were enough neurons that were talking to me with enough input that I decided to go from off to on. When I do that, I go downstream and the action potential goes

4:03downstream to other dendrites of other neurons and that's how we get a biological neuron. Now, in the artificial neuron, very much similar. You've got you've got incoming neurons and their activity. The GPU of Nvidia sums that up and then fires if it it fires like, you know, an output to downstream neurons. if that output ex if the input exceeds a certain threshold. Right? So we've basically taken the biological thing and we've turned it into a very simple math equation. Now obviously >> in that approximation we're not doing it complete justice because in biological

4:44neurons there's structure, there's proteins, there's all sorts of little things and it's not really just straight up integrating fire though that's the simplest kind of model that you can create. But it turns out for a lot of the AI that we see, I mean, it's doing pretty well, right? >> So even though we sort of have a simplification of this representation of biological neurons in our artificial neural network systems, >> even that simple approximation in and of itself has uh valuable uh productive outcomes. Yes. >> We don't have to replicate the whole of the brain. Yes. in order to get some

5:25benefits from the concept. >> Exactly. And the reason why is because of the neural network. The idea that we can connect a bunch of these neurons into a giant network where everything is connected to one another. Not everything is connected, but you know, you can have layers of connections from one to the other. So in the biological case again on the left, you've got one neuron that's connected to the other. So the output of one neuron is connected to the dendrites of the other through something called a syninnapse. That's the gap in between neurons that let the neurons talk to one another. And in the artificial case, you've got hidden layers of neurons. And from one layer to the next, you've got inputs going in. Again, each neuron is going to integrate, meaning sum up and then fire

6:06if it exceeds some threshold. So you've got these nonlinearities and then that's going to create the output, which is going to create the next output, which is going to create the next output. And if you have trillions of parameters and you know millions and billions of neurons then you can actually approximate any function you want including let's say thought and that's how you get large language models and

Why brains and GPUs aren’t built the same way

6:26all sorts of very very cool things. So that's how we've mostly done artificial neural networks. Right now how are how is the brain different from these artificial neural networks? That's something that we should focus on here because these guys are trying to use biological neurons. If the artificial neural network is just that good, why would we want to try to use biological neural networks? There is an advantage that cortical labs and other people doing this technology say the idea is silicon models which are the things in Nvidia and things like that they use the vonoman architecture of computing. The biological brain is

7:07just the biological brain. The vonoman architecture is the following. Your processor is separate from your memory. >> Mhm. >> Right. The memory is where you store stuff. The processing is where you compute. Those are two separate things in a conventional computer. Right? You've got your like RAM and then you've got your CPU. And over the years, there's a huge gap between how fast you can compute things in your CPU or your GPU and how fast you can access the memory where you store stuff in the brain. Both are colllocated. >> Okay? >> They're the same thing. They're all stored in the synapses, the weights of these things. The memory itself is stored in the weights. And the way that

7:49you process stuff is you go from one neuron to the other in this neural network. But in a artificial neural network, there's that difference. Mhm. >> Right. >> Mhm. >> So the idea is that perhaps this gives the brain an advantage. Another advantage that the brain has is that it's massively parallel and it only uses like 12 to 20 watts of power for 100 billion neurons. On the other hand, state-of-the-art GPUs and neural nets use gawatt of power for computing, right? So there's a clear difference also in the amount of energy needed to create that computing. So the the the the differences between the two, one is efficiency. We've created a simplified

8:31abstraction of what the brain does >> and it takes a lot more energy to even do some of the things that the brain does. So it's a highly inefficient >> as compared to the brain naturally. Yeah. Artificial intelligence is highly inefficient or highly energy consumptive. >> Yes. Um the also the way in which the brain is structured, it can do significantly larger amounts of parallel processing uh which enables that energy efficiency in part. >> Yeah. And also like each neural each neuron itself is like highly efficient as a computational unit, right? It doesn't require a lot of energy. We

9:11think you know our brain does consume 20% of the energy of our body but still that's not actually a lot when you think about things right so perhaps the idea is we can use this natural biological wetwware >> in place of our vonoman architecture hardware the GPUs that take up all that energy that's the argument okay but in order to do that we need to be able to read and write to biological neurons >> right and also where do these biological neurons exist >> yeah yeah where do they come from They're usually in our brain, right? So, let's get into how we actually read and write to real biological neurons in a dish. This was actually um pioneered in 1972 by Thomas G. Pine at Caltech when

9:54he developed the first multi-elerode array. Here's the idea. So, neurons produce electricity when they talk to one another. That's how they talk. They use ion channels which are little holes in their membranes, the boundary between the inside and the outside of the cell. And when those little membrane proteins, those channels open up or close, that facilitates the transfer of ions like potassium ions, chlorine ions, sodium ions from the inside to the outside of the cell. Ions are charged atoms in some sense. And so when they move around, they're going to create electricity. They're going to create electric fields. And we can sense that if we have an electrode array, which is just a bunch

10:36of tips of metal that is connected to some kind of amplifier. And as each neuron produces a little electric field blip, I can sense that with my electrode array, right? I can also supply electricity to then influence the behavior of the neurons near my electrode. Like if I've got a little metal tip here and I supply electricity, that is going to change the electric field around that metal tip, which is then going to influence the ions and thus the proteins on that neuron that are nearby that metal tip. And so I can now toggle between on and off based on that stimulation. So I've got a read, which is when the neurons fire, they're going to release a little

11:16electric field, you know, spike that I can read. And then I can write by supplying electricity and that's going to influence the neurons around it. >> I have a quick clarifying question. Yeah. When you talked about the spike just now that you're referring to this the concept you talked about earlier of the action potential. Yes. >> This sort of like surge of of of of electrical signal. >> Yeah. >> That then the neuron utilizes to do its next action or its next step. And so when we have the ability, we have the ability to provide that uh energy or source material necessary for the neuron to trigger its action

11:57potential artificially. It's not only uh cap it's not only a capability that is >> u not reproducible artificially. >> Yes. Exactly. Yeah. It's something that we can we can provide a signal to whatever metal contact is near that neuron and then you know that's going to create a change in electric field which is going to then drive ions around that neuron. It's going to drive the proteins around that neuron and then the neurons will be sensitive to that change in the electric field, right? They're going to think that it's another neuron somewhere that's like doing that. >> So we've to to to just make a fine point on it. We can translate into the language that these neurons speak using

12:38this kind of methodology. >> Yes. Exactly. Yeah. So now we can read and write. >> Yes. >> Okay. So can we do biomputing? Well, early biomp computing actually started in the late 1990s, early 2000s. There was this guy Steve M. Potter. He was at

Early biocomputing — rats, robots, and flight simulators

12:51the um Georgia Tech, Georgia Institute of Technology. He developed this thing called a hybrid robot. He used cultured cortical neurons from rats on a multi-elerode array to actually do just to control like a robot. Okay. So again, this is a little plate that has a bunch of silicon electrodes. Let's say it could be something else. I don't actually know what what but today we use a lot of silicon in our electrodes obviously. Um, and he could grow rat cortical neurons on that electrode like in a petri dish and then sense and then write, you know. >> Mhm. >> So, you've got to read in and then uh

13:33write you can listen to what's happening and you can talk to it to tell it stuff. >> Yeah. Yeah. So, this this was him controlling a robot in 2004. This was in the news. That was pretty cool. Thomas de Mars cultured 25,000 rat cortical neurons in a petri dish and that acted as an autopilot for an F-22 flight simulator. >> This was all over the news as well back then. And so that's pretty cool, right? Like almost 20 years ago, more than 20 years ago, 30 years ago is when we're doing this stuff. So how come now is when we can actually play video games? Okay, the reason for that is one, we've had a huge

14:14improvement in how many electrodes we can fit inside our multi-elerode array. Back then we could only fit about 60. Now we can do thousands. So we can sense a lot finer detail. >> The the bandwidth is greater. >> Yeah. Yeah. It's like the density of the electrodes is literally greater. We can make them smaller. We can make them more sensitive. So we can now sense individual neurons rather than global populations. Uh the other thing is that we need to keep these neurons alive right in the petri dish. And we've had a tremendous amount of scientific innovation in how to keep neurons in a petri dish alive for an extended period of time. Right.

14:54>> Mhm. So now with all of that we can get to Cortical Labs which is a new startup that came out of Australia that is doing this stuff that is trying to make neurons in a pre-tradition play pong and now most recently play Doom but before that allegedly allegedly we'll see right we'll see >> before that we are we are just so grateful and excited for so many of you that continue to join us on this journey of curiosity and discovery. For those of you who are returning welcome back. We greatly appreciate your attendance at our multi-week now uh show release schedule. As we mentioned in our previous episodes, we are now testing

15:36going from just doing one weekly episode to multiple with the rundown being its own episode type. So if you're looking at your podcast feed or your Spotify feed and you're like, "What's going on?" You're getting more of us more often, which has been in high demand. And if you do like what you hear, whether you've been here with us for a long time or you're a new listener, one of the greatest ways that you can support the show is to give it a like, share, comment, a follow on the platform that you happen to discover us on or listen to us on. And if you want to go above and beyond, we do have a donation portal at the website ffpod.com/donate

16:16where you can support the show with a recurring or one-time donation. We are a twoman operation who've set up this unbelievable production through the ingenuity of the both of us. And it is really a pleasure to be able to talk to you guys about the wonders of science multiple times a week. And we just again really, really, really cannot express how important it is to fight the billionaire algorithm to just engage in any way that you can.

Cortical Labs and the original DishBrain Pong paper

16:46Even if it's not financially, it's really helpful for us to get to more people so we can talk about the wonders of what the researchers and scientists all over the globe are providing us in terms of the future of humanity. And with that housekeeping note, we can jump back into Cortical Labs. All right, so now let's talk about Cortical Labs. This is the subject. This is a startup out of Melbourne, Australia. I just recently watched the Formula 1 Grand Prix in Melbourne. Um, Ferrari. No, it didn't. Mercedes is >> Was that the one where uh uhlair and Hamilton were battling each other? >> No, that was the Chinese Grand Prix very recently. Anyways, I don't want to talk about it. So, um, Cortical Labs founded

17:29in 2019 by Hongwen Chong. The company gained international attention in 2022 because of a paper that we're going to cover. It was in the journal neuron and it was describing this thing called dish brain which is a system in which in vitro neural cultures. So that's in vitro meaning in a petri dish learned how to play the game pong in a closed loop environment. Okay. So that is the paper that we are going to be talking about because that is the bedrock upon which their technology is built and it'll help us understand how exactly they did doom. Okay. And it's actually very cool the some of the some of the stuff that they've done. Okay. So, first

18:09of all, you got to get neurons for your wetwware engineering. >> Yes. >> How do you get neurons? Well, first you get primary mouse cortical neurons. Those are fine to get. You know, mice are very much a animal that we use in biological research. So, that's not that hard. Human cortical cells on the other hand are hard to come by. So instead of actually, you know, waiting for a patient who has epilepsy, let's say, to have a little part of their brain excised and get that kind of cortical neuron, you never know with those kinds of samples like, okay, it's a brain that has epilepsy. That's not a wild type type brain. It's not a neurotypical, as we should say, type brain. So instead, what what they did was they generated

18:50humaninduced pur potent stem cells. This is something that we cover a lot. It won the Nobel Prize because it's a technology where we can take somatic cells like stuff from the biopsy of a of skin. We can reprogram them to become stem cells which are blank slate cells that can be anything that they want and then we can give them a cocktail of chemicals to differentiate them into cortical tissue. So we take skin cells, we make them a blank slate and then we make them cortical neurons. I just want to make a quick note that this was uh I think this was the the Yamanaka Nobel and this was really important because prior to that embryionic stem cell

19:31research was highly controversial >> um and it was the only way to acquire stem cells >> and it really unlocked the ability to actually be able to do this type of work >> by removing the sort of moral ethical conundrum. So now the point being any human cell you can reprogram to something else. Yeah. That wasn't that's a you know that's a huge that's a huge foundational discovery. Exactly. >> Um that that now cuz I I know the sort of stem cell piece for some folks is like they that >> historical point cuz some people still think oh aren't we still getting those from embryos? >> No. Yeah. This is just your skin cell becoming a stem cell. It's pretty insane. Now, of course, with the pleoprotein stem cells, it's not a

20:13complete blank slate, right? The embryionic stem cell is like the gold standard because those things are built to be blank slates, right? So, but this

Building the culture — human stem cells, astrocytes, and 800,000 neurons

20:22this comes pretty close for literally 90% of applications. So, it's it's done a huge service for the biomedical industry because for 90% of applications, I don't need to worry about that. That's still a huge unlock. I just want to be very clear here. So let's talk about the culture that they made. They made 800,000 neurons co-cultured with aststerytes. This is pretty incredible here. >> Okay, >> usually when we make stem cells, right, and we make a culture of cortical tissue, >> you're going to have all neurons. >> That actually is not ideal. Okay, >> for certain applications it is like when we were discussing the ALS story and we wanted to create um an ALS mini brain on

21:05a petri dish. We wanted it to be all motor neurons, right? Because then that's how we test the drugs and so on and so forth. In this case, we want like a live computing unit >> that kind of mimics the human brain like or at least a small chunk of the human brain. And the human brain is not just neurons. The human brain has astroytes. So if we look over here, what they've been able to do is grow neurons along with aststerytes and all of the supporting cells. >> Okay? So the blue that DA um DAPI, >> yes, >> those are marking the neurons, >> the GFAP and the red those are marking supporting aststerytes and those are

21:46critical for long-term functioning. Okay. >> Okay. So this is sort of a secret sauce that allows those cultures in that petri dish to survive for a lot longer. So now they can learn over a longer time scale and you know actually now we can take neurocomputing seriously. >> If you have a professional sports team with these high high skill athletes, >> they're nothing without the training and medical staff. >> Exactly. that enable them to continue to perform over a longer periods of time given the amount of stress and and things that are going on. >> Exactly. And so these are all these supporting cells which are crucial to creating a comput like you know a little

22:27piece of computation on that petri dish. It's not just all neurons. You got to have sort of a mimicked brain environment and they were very successful at doing that. So now that's the side of the cells. We've we've created a really nice culture with 800,000 neurons. That's a lot of neurons, right? Okay. Now, let's go into the multi-elerode array. >> I'm sorry. I just want to interrupt you really briefly. Is is so the 800,000 neurons is that would that be comparable to saying like the GPUs in your Nvidia

Multi-electrode arrays — reading and writing living neurons

22:55cluster as like the corollary in an artificial system? >> Yeah. Yeah. It's like the number of um it's the number of transistors one can think of maybe you know it's like I mean no it's not it's not quite because you know it's not one transistor to one artificial neuron but in your artificial neural network you have you know billions and billions of neurons with trillions of weights in between them here you've got 800,000 neurons it's still small but it's incredible what they're able to do with it >> understood >> you know but that's the idea you know okay so that's our neuron piece now let's get into How are we going to sense them? That's our multi-elerodes, right? Those are the little metal leads that let us read the neurons and then write

23:37to the neurons. So, here they did a highdensity multi-elerode array using CMOS technology. 26,000 platinum electrodes. What you can see over there, all of these little rectangles is an individual electrode, right? So, that can be individually tuned to have voltage, you know, 5 volts, -2 volts, whatever you want it to be. it can read the the electric field around it and it can write an electric field around it. Okay. So, and you can see the the sort of this is a I think it's a electron microscope and you can see the cell bodies and all the dendrites and all of the connections that are just sort of growing through that multi multi-elerode array. So, when I have a connection like

24:18a little axon that's connected from neuron A to B, I can follow that signal with my electrodes. Very very cool. Right. It looks like a map of LA from the sky. >> Yeah. Yeah, it really does. Like you the Spanish are very good at gridding and that's what that's what this looks like. If if you want if you're in Barcelona, it also looks like that, you know. >> Um >> and out of the 26,000 platinum electrodes, >> there's only 1024 independent channels >> that are active. >> Okay, >> so they can pick and choose which ones they want to read through, right? So that that's again a huge thing. And one of the incredible things is that the latency between reading and then making a computation to figure out what to write, right? Because I'm going to read

24:59the the the activity from the neurons, then I'm going to uh I don't know, I guess like see what the game state of of Pong is or Doom is and then write back to the neurons and give them feedback. They got it down to 5 milliseconds. That's such a huge >> which is which is quite a small latency. Yes. >> If you really think about it, >> that is inc. And and if we're trying to mimic the brain, we need that latency as close to zero as possible. >> Exactly. Because for us, our latency is around that. Right. >> Right. Um one of the common things that we think about is like our our eyes are about 60 frames per second is the top level of how many, you know, little pictures it takes every second and

25:41relays back to our brain. And so 60 frames, let's just say 50 frames per second cuz I want to make my math simpler. Yes. Right. So 50 frames per second that's about 20 milliseconds >> is the is the refresh rate >> of my eye to the brain. Right. And this is going down to 5 milliseconds. So it's it's even lower than that. It's that's the speed at which neurons talk. The action potential which is this thing that I was telling you about how when the neuron turns off it actually immediately turns off right after it turns on and off and that action potential takes about one millisecond. So getting it down is mimicking the sort of computational latency that the brain already has and what these neurons in

26:21your dish are used to. >> Yes. >> Right. So you don't you don't have them like thinking harder >> than they need to. >> Than they need to and being which makes it more energy efficient. >> Yeah. So now let's look at the dish brain protocol. How do we play Pong with our video game >> with our DBP? >> Yes. Okay. So over here we've got in the middle our petri dish that has on the bottom of it this multi-elerode array. These are little tiny electrodes that are going to sense the neuron culture that is on top. The neuron culture on top is getting bathed with electrolytes with glucose to keep it alive. Um they can keep it alive for as long as 6

The Pong setup — sensory cortex, motor cortex, and encoded game state

26:58months. And as these neurons in my petri dish turn on and off, my multielerode array is going to sense that. It's going to read it in. It's going to sense the game play and it's going to read it out into the neuron. Now, what we don't want to do is completely jumble up our readin and our readout. >> That was literally going to be my the orchestration problem there is non-trivial, >> right? Yeah. Because we've got a bunch of electrodes and the brain is not really an all readin, all readout, right? There are regions in the brain like the occipital lobe that has the visual cortex. Um there's the prefrontal cortex that makes decisions. There's the motor cortex that actually outputs my

27:40hand when I'm like let's say using a joystick right so there are segregated spots in my brain for where I'm reading in that'll be my visual cortex and where I'm going and reading out that would be my motor cortex. So they decided to create an artificial visual cortex and a motor cortex on the chip itself. Okay. So at the very top on the right hand side what you can see is the electrode layout schematic. Okay. >> On the very So you've got your rectangular multi-elerode array that has 1024 >> electrodes that I can actively read out and read in from. On the very top I'm going to make a rectangular box that's in the upper half. And eight of those

28:21electrodes I'm going to use as my quote visual cortex. It's really quite a sensory cortex, I should say, because that's what I'm >> writing into, right? The sensory cortex is getting information. When when I when I play a video game, the sensory cortex is through my eyes and through, I guess, sound sometimes >> getting information about the state of the game. >> If you talk to Call of Duty players, sound is as equally important with their high quality headphones to sneak up on you as visual. I'm I'm just being annoying. I mean, I when I was playing Halo, I used to sometimes have to like play on mute because it was like 2 a.m. and my parents would like like, "What are you doing?" So, I I'm actually quite

29:03good at playing Halo on mute. So, I don't know about you guys, but anyways, the idea is there's a sensory cortex that is getting information from the brain. So, I'm going to just arbitrarily make that top half of my electrode the sensory cortex, and that's where I'm going to write information to my neurons. And the idea here is we're segregating it from other section or sections such that it only has one thing to worry about. >> Yeah, exactly. It's kind of like I mean I don't want to give too much agency to these neurons in a dish, but what I'm saying is that those neurons in that upper half of my petri dish >> are acting like the sensory neurons. They're going to sort of understand

29:43that, hey, I'm getting information about something. And I think this goes back to the point you brought up earlier, which is proximity creates stronger neural connections. And so because you're putting them in proximity and having a similar input, >> Yeah. >> the theory is, you know, they will naturally strengthen around that shared >> Yes. idea. >> Very good. Yeah. Exactly. Those neurons in that region, they're all connected to one another, and so they're going to somehow figure it out. We're going to get into how, but that's the idea. They're going to somehow figure it out. And if actually if you if we if we put that back. Yes. The bottom half of that petri dish has those up and down arrows. Yeah. >> That's my motor cortex. >> Okay. >> Okay. So neurons in the down region in

30:26that little block. >> If those fire that's going to trigger my pong um paddle, I guess. It's a paddle, right? It's like a digital paddle. >> I know what you mean. Yes. >> You know what I mean, right? Like that it's going to it's going to trigger the paddle to go down. And if the neurons in the up region fire for those electrodes, then it's going to trigger the neurons to go up. And sometimes you will have to calculate a difference between the up and down because, you know, maybe there's more >> neurons in the down region versus the up. So I also have to normalize based on like how many neurons there are cuz I don't want to just artificially keep going down if there's randomly more neurons in that region. So there's a bunch of mathematical tricks you have to do but at the end of the day segregated

31:07regions one set of neurons is going to fire if the culture wants to go down and one set of neurons is going to fire if that petri dish culture is going to want to go up. Yes. >> Okay. So now we've now we figured out our read and our right mechanism. The right is the sensory cortex at the top eight electrodes and our read is getting it in. And I think an interesting point about this is that geographical or or locationational uh the where it's located and how it's segregated like form is function or form is following function in this context. >> Exactly. Yeah. Now there's one other thing that I want to touch upon which is how exactly that writing mechanism >> Yeah. >> is being implemented. Right. Because

31:47what do we need to give this petri dish of neurons information about? Yes. What is all the information we need to give it? Well, we need to give it a state of the game. right? Which is where is the paddle and where is the ball >> right in relation to that. >> Yes. >> Now for us when we play pong, we've got a very nice visual system. So we see where the paddle is, we see where the ball is and when we also see how it's moving. And so we can sort of >> all of our billions of neurons in our brain can like understand spatially what's going on because we understand space because we've had so much pre-training and we can figure it out. Yeah, >> this is a neuron in a petri dish. It does not have eyes. It does not have

32:27senses. We are creating the senses. So, how are we going to do it? Well, >> they use something called rate and place coding. This is something that is very commonly observed in the human brain. >> Okay. >> Okay. For example, let's take the example of the human cookia which is in our ear. That's how we sense sound, we sense frequency, and we sense loudness. Right? I've um I've touched upon the cookia before. The idea with the cookia is there's like this tube, a snail-like tube that is embedded in our ear. And what that tube is, if you were to unravel it, there would be neurons that

33:07are very thick at one end and there would be neurons that are very thin at the other end. This thing gets wrapped up into kind of like a shell shape, right? But if I hear sound of a certain frequency, that is going to excite a neuron along this axis that is at exactly the right thickness. Right? If the sound is very low frequency, like a bass sound, that's going to excite the thicker neurons because they're bigger and the wavelength of the sound is bigger. So, it's it's kind of just like, you know, the resonance is happening. And if the sound is really high, then it's going to >> excite the smaller neurons in my cookia. >> Yes. >> So that's a place coding that tells me

33:49the frequency. >> Okay. >> Okay. That's my place because my place along the axis is telling me the frequency of the sound. >> Yes. >> The rate coding is how loud is that sound. >> Ah, there it is. >> Right. Because if it's very loud, then that neuron is going to fire at a very high frequency. If it's not that loud, it's going to fire at a low frequency. So I'm getting two pieces of information with place and rate. The place is telling me the frequency. The rate is telling me the loudness. Right? And this is something that is very common in biology. Where the neurons are firing tells us one piece of information. And how much the neurons are firing tells me another piece of information. And those are simultaneously present when

34:29biological neural networks create computation. And just to continue to bring this back to the artificial comparison, I think what's interesting here is this is >> different than just being zeros or ones. Yes. >> Because you actually have two levels of of input. You have the the the place which you could uh let me not make that direct comparison, but the point is it is not as simple as a binary system. Yeah. >> It's more of like of a of a twostate modulator. >> Yeah. No, I think I think you're on to something here in that the neuron is kind of an analog system, right? In that sense. Now, some people will argue that the neurons are still digital because

35:10they are sending out spikes and if you were to condense time and discretize it into smaller and smaller segments, then in each time window, the neuron is either active or inactive. Right? And the idea of a high rate is just there's more time windows that it's active, right? So, there is there is a really healthy debate in the neuroscience community about whether to interpret neurons as analog or digital. Okay. Right. And I don't know where I fall on that. Okay. Um >> that's fair. >> But it's it's it I think it's a fair point that you're making like it's a very important point. >> Right. Right. Right. is like is like that that is that is still up for grabs >> because the answer to that will also then be deter will be relevant to like

35:52our conversation about LLMs versus other models that we had in the last episode in terms of if >> anyway you get the idea. >> Yeah. Yeah. So to be clear I just want to reiterate here what we're doing is using something that is in place and something that the brain is already used to which is place and rate coding. In the idea of the cookia, the place along this axis was the frequency and the rate at which that

Place coding + rate coding — how the neurons “see” Pong

36:19neuron fired was the loudness. >> In this case, >> Mhm. >> how are we actually um if we go back to photo 14. Yes. Actually, >> how are we actually going to use that sensory cortex, the eight electrodes that we're trying to write in our information? How are we going to use place and rate coding to inform my petri dish brain about the state of the game? That's the key right? >> What they did was they said something very interesting. The um where the ball is or where the um sorry where the paddle is. >> Yes. >> Is going to be my place.

37:00>> Mhm. >> Okay. So, if my paddle is lower on the screen, I'm going to fire the electrodes on the left. I'm going to make the left electrodes give some electricity to my brain. If the paddle is up on the screen, I'm going to give the right electrodes. Okay, so now I've got place. What's the other thing I need to code in? That's the distance to the ball. How close is the ball? How close is it getting? Right? And so, for that, they use frequency. >> They went all the way from four hertz to 40 Hz. Four hertz means it's very far away. 40 Hz means it's getting close. You need to you need to figure out which way to go. >> Right. >> And and the idea here is is we are

37:41basically creating a translation between what we experience as a visual input >> where we can do the rate and place just through our visual input. >> That's right. into an electrical input such that but that is uh regional >> as a means to create an extra layer of like degree of freedom in in in encoding more data in what is a relatively flat kind of system. >> Yeah, exactly. And we need to be able to give those neurons in my petri dish both of these pieces of information. Right. And one thing I do want to be clear here is the neurons in this petri dish are not directly seeing the game like we

38:23are. This is kind of what I'm trying to >> right >> poke at. >> Yes. They're not seeing the game like humans do. Instead, they're sort of seeing the game based on the paddle's point of view in some sense. It's like where am I >> if I'm the paddle and also how far away is the ball, >> right? >> Those are the things and the only res the resolution is only like eight pixels in some sense, right? >> But from that they're trying to make this petri dish play pong. Okay. So the other thing that we need to think about is fine now we've got a read and write mechanism. >> Yes. >> Now we need a training mechanism. >> Right. >> Right. >> Cuz now all of a sudden I mean we're

39:03telling the petri dish the state of the game and we have some way of reading what the petri dish wants to do >> which is move the paddle up or move the paddle down. How does the petri dish know whether it's doing well or not? >> Yes. >> Right. The challenge is actually quite non-trivial because in the brain, how do we get a reward? You know, we we feel happiness, maybe we

Learning by minimizing surprise — predictable pulses vs chaotic punishment

39:27get a cookie literally that is going to fuel dopamineergic neurons in our brain that are going to send signals of dopamine, pleasure, and all this other kind of stuff. So, we've got like emotional signals for reward. We've got other crazy signals for reward. This is a brain in a petri dish. It does not have dopamineergic neurons. It does not have these centers and this complex brain anatomy to have reward. >> Yes. >> So how are we going to do that? The re and this I think is actually the coolest part of the paper to me. >> Okay. >> Okay. It's this idea of a closed loop feedback architecture and the learning mechanism is literally minimizing surprise.

40:08>> Here's what they're doing. If there is a successful hit, meaning the the pong paddle thing hit the hit the ball. Yeah. >> Okay. And we're good to go. What it's going to do is >> all of the electrodes in my multi-elerode array is going to send out a predictable electrical pulse. It's about 10 volts on every single electrode. >> Okay. >> At 10 hertz. So 10 every second. Yeah. >> For about 50 milliseconds. Very short, very highly structured. >> Okay. >> Okay. >> If on the other hand, it misses, >> it's going to get 4 seconds at 20 volts

40:50of highly chaotic electrical activity. >> Mhm. >> And the idea is brains hate that. >> Yes. >> They want to be able to predict the future. Yes. >> Okay. And when you're giving them highly unpredictable stuff, even this naive network of biological neurons >> wants to decrease the amount of times it gets that. >> So if we if we look at the GIF, I just wanted to show one thing. So here we've got the brain coming in. You see that bump in electrical activity everywhere. And now it's about to miss. >> Y >> it's going to miss for 4 seconds. It's just like there's noise everywhere. >> Yes. >> Okay. We're feeding a bunch of electrical noise. The neurons hate that.

41:30Yes. >> Now the it's going to go it's going to actually Oh, I guess it's going to recycle. Let's recycle. >> So, it'll recycle. When it gets there, >> bunch of predictable electrical signal. That's all we're doing. >> Yes. >> Our reward mechanism is simply if you got it right, we're going to reward you with a bunch of predictable activity. >> Yep. >> 10 hertz, 50 milliseconds. And if you get it wrong, four seconds of random chaotic activity. >> And I I think this dovetales with some sort of historical uh science research that's been done around >> the you know when we look at how do we

42:11get ordered systems with entropy and this idea of everything goes to soup and the way in which that works is by minimizing surprise. Yes. this concept it's like a biological it is how life fundamentally is able to instantiate itself >> and that we're basically hijacking that natural kind of concept by artificially

Results — the neurons really did get better at Pong

42:37>> punishing the wrong decision >> with chaos which makes the system work harder >> and it is naturally going to want to not work harder and so that's why it chooses the right decision. It's not, it doesn't really know that it's the right decision. It is optimizing for the less >> en the more energyefficient. >> Yes. >> Pathway. >> Yes. The less surprising. >> The less surprising. >> Yeah. It's really the less surprising outcome is what it wants. Right. And now let's look at the results. >> Okay. >> Okay. With this training paradigm, let's look at the results. The apparent learning is demonstrated in 5 minutes. So here what we're looking at is two sets of box plots. Mhm.

43:17>> The green box plot is what happens from 0 to 5 minutes. >> Okay. >> Um we're looking at the change in rally length, which is like the, you know, the how many times the rally happens for for the pong. It's just like in tennis, like how many times do I keep the ball alive? Um and you can see from 6 to 20 minutes, there's a giant change in the size of the rallies from 0 to 5. So within 5 minutes, this thing is learning quite well. there's a up change in rally and that's for the stimulus for the controls which is silent and no feedback there's no change. Okay, so we're getting a significant change only when they're actually playing the game and they have this feedback. Now, if we go to the next

43:57>> photo, what we're also seeing is that human cortical neurons do a lot better >> than mouse cortical neurons. >> Okay. And that I don't know that makes me feel good that my neurons are better than a mouse's neurons. >> They are better, >> right? So, and that kind of makes sense. We're We've had evolution and our brains arguably are probably better than mouse. >> And the thing is there was an improvement in both cases. It's just our improvement had a >> had a had a big that's that's a good point. Very good. Yeah. It's that ours had a much bigger improvement than the >> compared to the mice, >> right? So now that we have that, let's actually look at this free energy principle that you were alluding to a bit earlier. >> Oh, interesting. Okay.

44:38>> That's what it's called. The free energy principle. This is was it was first pioneered by Carl Fristen and the idea is what you were saying all self-organizing biological systems want to minimize something called the free energy the variational free energy in other sense what we want to do is maximize the information entropy and we want to minimize the surprise and the prediction error okay >> what we can look at is how much surprise there is for what is the probability of a certain state. >> Mhm. >> If the probability of a certain state is very low, there's a lot of surprise, right? And you can quantify that in the same way that we actually quantify

45:18entropy in physics, which is the log of the probability of stuff. >> And effectively what the free energy principle is saying is that biological systems want to maintain homeostasis, meaning like constancy >> over all the random stuff. And we want to minimize that surprise, right? >> And what do we what do we do if we want to do that? Well, we've got some internal model of what the world is. And we are going to tune the parameters of that internal model such that that internal model mimics more closely the world that we live in. Right? That is the idea. It's a stark departure, I should say,

46:00and highlight from traditional reinforcement learning. >> Okay. >> Okay. In tra in traditional reinforcement learning, you've got some algorithm and that algorithm explicitly gives negative feedback through back propagation for wrong trials. Here we are letting chaos itself >> be the negative feedback. >> Yeah. Yeah. Yeah. >> Now, one of the things that's kind of crazy is that the algorithms for reinforcement learning require thousands of epochs of trial and error >> to update these weights through back propagation. Right? Biological systems will inherently self-organize to minimize this free energy and they'll do it very very quickly. They've just got

46:41really nice cellular mechanisms to find that sweet spot. >> This goes back >> in their internal model, >> the energy efficiency piece again. >> Yes. Exactly. Right. And so maybe that's why we want to use the wet wear >> in the first place. Right. So let's go back to this FEP in a dish. Our free energy principle in a dish. Effectively what's happening is you've got predictions and then you've got prediction errors. The predictions are where I want to move my >> paddle. And the prediction error is the chaos that that I get when I get it wrong. Right? I'm trying to change the

Promo hype vs paper reality — what the box plots actually show

47:20connections in my neurons on my petri dish such that I minimize prediction error. And that's effectively what these neurons are doing. It's actually pretty incredible when you think about it. We're not giving them a bad and good signal. We're only giving them a chaos versus non- chaos signal. >> Right. Right. >> And that itself just because of the nature of information and the physics of information is enough to like steer them like this is bad and that's good. I think that's like quite deep. >> The universe the universe never chooses violence. >> Yeah. It always chooses to billy when given I I'm being a little facicious,

48:02but >> no, but I think I think I think that's a very good and that's a very poignant statement here. So now let's look at like some of the hype that they created around it. Let's look at some of the promo videos. >> And this was for the Pong. >> This is just 2022. This is we're still on 2022. >> Just confirm cuz I know people are listening waiting for the Doom. >> Yeah. But we need to really understand this in order to get to Doom 100%. >> Okay. So this is um one of their promo videos. Look at how many this this this is doing really well. It's already at like rallies of three. >> Yeah. >> Maybe it'll get to four. >> Yep. Yep. Is it going to get to four? >> Yeah, it got to four. >> Four rallies, right? Like this is it's really it's really doing it. Five. It

48:42was getting to five. Now, let's look at the results from their paper. >> Okay. >> Okay. From the results from their paper, if we look at the left hand side >> Yes. >> Um, figure B. >> Yes. >> That's the average rally length. Yes. >> And look at the y-axis. >> Yep. What is the y-axis on? It's on one, right? Most of their trials are on one. Even on their box plot, the edge of their trials for human cortical neurons, which is the right hand side box plot, >> is going to 1.5, >> right? >> Which is that's the I think it's the third quartile usually for box plots. So, it's like the 75% of data falls within 1.5. So clearly they're like showing a very nice example in their

49:25video right? >> It's still statistically significant. If you look at the human cortical neurons, which is the right box plot in figure B. >> Yes. >> Um from 0 to 5 minutes, the average rally length is at 0.5 and now it's bumped all the way to one. And if you look at the spread of that data, clearly it's highly significant. If you look at all the other metrics, you can see for human cortical neurons, they do the best compared to mouse cortical neurons and all the other controls. So there is clearly an effect, but the cortical labs team are a startup. >> Yes. >> And so, you know, >> as is tradition in startups, not just in Silicon Valley, but around the world, even in Australia, they're going to hype up some of the results. And that's fine.

50:07I still think it's a very cool >> thing that they did. >> Yes. And we just walked through why it's incredibly cool. Um, and if you think that AI is using too much energy, you should be all for wetwware. >> Yeah. Yeah. I mean, it maybe it maybe has a lot of applications. So, now we're going to scale this up. >> Okay. >> We're going to scale it up to complexity level 10,000 because we are now trying to play Doom. >> They built the CL1 platform, Cortical Labs >> first deployable biological computer in the world. >> I like that. >> That's what it looks like. It's a It's a little box that at its center has the petri dish with all of the biomime like

50:48the biological fluids and all the things to keep the neurons active. You've got the multi-elerode array. And you've got a little screen on the side that shows

Scaling to Doom — the CL1 platform and the cloud wetware pitch

50:58you the health of the neurons, the amount of oxygen there is, the amount of CO2 there is, and so on and so forth. So, it can regulate temperature, it can manage gas exchange, it can clear metabolic waste cuz these neurons are alive, so they're going to be creating waste. All of this other stuff. And they can actually survive in that in vitro petri dish for up to 6 months. It's a very long time. Okay. For a sort of little mini organoid to survive. >> So how about the software? This is very cool. So for the for the Pong for the Pong story, they had a customuilt C C++ machine language type thing. Nobody wants to code in C++. I get that you

51:38want the base layer >> Yes. of your software to be in C++ because we want very low latency, right? And C is an extremely fast language. Python is quite slow, but at the high level, I want to code in Python. Okay? And then maybe the Python gets transferred into C. And what they did was they created something called a cortical cloud and a Python API. So now you don't need C++. And you've got a cloud computer of a bunch of CL1 platforms that are each >> like, you know, little tiny >> brains on a dish that I can now just like get into my, you know, on my laptop. I can access it's open open

52:19platform. I mean, you have to like apply and say like this is what I want to do with it. But this is I think very clever from Cortical Labs because >> even IBM like one of the reasons why KisKit which is the quantum computing platform for IBM like their software platform that's used a lot because they've created a cloud platform where theorists and like people who want to do testing of quantum algorithms and things like that they can go into KisKit they can apply for time on the IBM quantum computer. they've got like a dilution refrigerator with that IBM quantum computer and they can like test out their algorithms and like their error correction methodologies and things like that. So democratizing it is a really good way for like a startup or like a

53:00small computer lab to be like, "Hey, we've created the hardware. Why don't you guys go at it?" >> Yes. Right. >> Yes. >> It's very Matrixesque. If you remember the scene where Neo unplugs and he sees all the pods of humans to capture heat for the machines, that's what that effectively wetwware server rack. AWS Nvidia server rack looks like. >> Exactly. Yeah. It's like AWS but for like wetwware >> for wetware. >> It's it's pretty crazy. An AWS for a brain on a chip. >> That is so outrageous. Amazon Web Services for those who don't know is the infrastructure layer that provides the back end for uh a large amount of all of the software applications that you use every day.

53:40>> Yeah, exactly. So with that in mind, now that we've got a cloud computing platform, >> Yes. >> now independent researchers and developers >> can mess around with it. And this is where we get into Sean Cole. He's an independent developer. >> Okay. >> Who wanted to use the CL1 platform >> to play Doom Freedom >> Classic. Okay. And the, you know, one of the I don't know why this is a meme, >> but like whenever you have like a new piece of computation or something, everyone's like, "Can it play Doom?" >> Can it play Doom? >> Right. Like, like even the ESA has like made >> satellites play Doom and things like that. So, I guess it's a meme. And the guys at Cortical Labs knew this was a

54:22meme. Yes. >> And knew that if they could make their little thing play Doom, it would get a lot of hype. And it certainly worked. >> And I just want to give a shout out to John Carmarmac who was the inventor of Doom at ID Software and literally was one of the pioneers that created all the video game like like he that that was the like the 3D platform like that concept and what they did at the time is what created all firsterson shooter style or like all these types of games. >> So is Doom before um what's the James Bond >> James Bond Golden Eye? >> Yeah. I I don't have a a chronological knowledge, >> but I feel like James Bond Golden Eye

55:02was also kind like from what I've I've never played Doom, but I've seen the videos while I was researching this and it looked a kind of like >> it's similar. >> Yeah. >> But it's not Doom. >> Yeah, it's not Doom. I guess >> I I was a 64 guy. I played Golden Eye all the time. I love the music, but but Doom is sort of >> Doom is kind of like this OG. >> It's it own thing. >> Yeah. So, it it is way more complex than Pong. So, you know, one of the things that I as I was like researching Doom and I was like watching some videos of people playing Doom, you know, it's really you're transitioning from a 2D linear game like Pong to now it's kind of like this 2.5dimensional. I wouldn't really say threedimensional because it's like >> first person and like the it's not

55:43really fully threedimensional, I would say, but there's still it involves complex navigation, threat detection, and all of this other very cool stuff. So it's a way harder problem. >> Yes. >> Than >> up, down, >> how close. >> Yes. Exactly. Right. Oh, you're already getting there. Is like those are the sort of salient things that I need to code in. You know, before in Pong, I needed to code in position of my paddle and where the ball is here. In this case, I need to figure out where what direction the enemy is and how close they are. Anyways, let's get to um >> the stem the cells. We got 200,000 human pluropotin stem cells. >> Yes. >> So this is a fourth of what we had in

56:23Pong. >> So fewer than what we have in >> Pong. Right. Okay. Yep. >> But because of like the cloud computing nature and the fact that >> this is all in Python, the guys were able to map the engine to the tissue in just a week. >> And this is what it looks like. >> Jesus. So, you've got you've got your

How Doom was mapped onto the neurons

56:42chip and it's sensing motor neurons or it's sensing the neurons and then it's playing Doom. >> This is crazy, >> right? >> Yeah. Yeah. Yeah. >> It's pretty cool. >> Yeah. And we're seeing this like like Yeah. The the real >> Yeah. Those are the neurons that are coming in and then there the the petri dish is sort of controlling which way to point and which way to fire. Right. So, let's get into let's get into how actually they did it. Because in Pong, remember, we've got a sensory cortex, we've got a motor cortex. They're going to do very similar things. In this case, the proximity, we want to we want to encode information about the game state, right? So, there's proximity, which is how off

57:26how close I am. >> Yes. >> And there's direction, which is where where I'm trying to fire. They're using place and rate to code that. Right >> now, we want to take the damage. So, taking damage and dying is going to be the chaos. >> And if we shoot someone, that's going to be the predicted >> ah reward. >> Okay. Interesting. Cuz I was wondering how it was going to map from pong to doom. >> Yeah, that makes sense. >> And that's how you do it. And for the motor neurons, um there's specific firing patterns that are going to be move forward. There's specific firing patterns that are going to be turn. and their specific firing patterns for fire a weapon. And the performance after 7 days of training, the agent's kill ratio

58:07was roughly twice that of pure chance. Now, I don't know what that means. And this is where it gets kind of annoying because there's no paper, right? So, there's no methods. They just kind of said that, >> right? >> I'm like, I don't I don't really know what that means. But they're explicitly admitting that the cells played like a beginner who's never seen a computer. So, it's not very good, >> right? But it's still I guess higher than chance. >> Yes. >> So maybe that's something. One of the things that I found kind of interesting was that during the study the agents actually generally settle settled >> for a strategy of survival which is hiding rather than active combat. And this is something that is well known in reinforcement learning right you just

58:47like go for the local minima which is >> I don't want to deal with anything. I don't want any reward whatsoever, even the positive or the negative. Because that positive or negative reward comes from a dense like chain of brittle action, right? I have to do stuff. I have to do stuff and at the very end I might just get chaos and then it might just be bad. So why don't I just like go to a corner and try to hide? I wish they gave the uh the uh nuke testing AI uh paper people the same wetwware to to go to the local minima because as we talked in our two episodes ago all the major model companies Anthropic Gemini Open AAI all used Nukes 90% of the time which

59:27does not sound like what you're talking about. >> Yeah. Yeah. No, this one is just they just want to >> they're like we're just trying to hide. >> Yeah. I'm just going to try and hide. So, now we've covered what Cortical Labs is purporting they did. Yes. >> In their YouTube video. This is where I got a lot of that information. >> Yes. >> Now, let's ask what's actually going on. >> Yeah. >> Okay. Cuz there's a lot less info about the Doom >> game versus the Pong game. The >> Pong game had a full ass, >> you know, peer-reviewed paper in Neuron. This one does not. This one is just >> like blogs is all I got. Okay. Here's my concern, and this is something that um I actually read Tommy Blanchard's

1:00:07Substack on. So, if you guys want to check out his Substack, he's got a lot more um info on it.

The key concern — is reinforcement learning doing the real work?

1:00:13>> Here's the concern. There are eight input channels for Pong, right? But Pong is a very small game. >> I can reasonably put all of that information into eight electrodes. How do I encode everything that's going on in Doom with only eight inputs? >> Yeah. each of which can only have a few possible values like either the electrode is on or off. And the answer is you actually don't. >> You use a reinforcement learning model. >> This is where it gets a little bit fishy. >> Okay. >> Okay. >> Here this is from their this is from their doom neuron. Um like they've got a little >> it's kind of like a GitHub where they have a documentation on what they did. This is the architecture of what they're

1:00:53showing is they've got a computation unit and that is talking to their CL1 computer. >> Yeah. Yeah. >> Okay. In that computational unit on the left right hand side, what you can see is there is a PO policy >> which is a reinforcement learning paradigm >> that is then encoding >> the input of the game that is then going as stimulation to the brain to the to my brain in a petri dish and what I get back is action and that is getting fed back into my reinforcement learning algorithm. So now I can very easily ask who is actually doing the learning >> correct? Is it my reinforcement learning network or is it the actual biological neurons that I have? >> Doesn't count, >> right? It shouldn't really count. And

1:01:34it's not it's not that far away. Like reinforcement learning, there's video games that have like perfected reinforcement learning. Dota 2, OpenAI defeats Dota 2 World Champions. So that's not that big of a deal. Correct. >> So >> the question is, was the software decoder, which is that translation layer, >> Yeah. Was that actually in the the learning algorithm or is the actual neuron doing something? >> Cuz in the pawn case that we already talked about in the earlier direct it was direct and it was maybe incremental but it was demonstrable learning. >> Yes. And and we could very visibly say the neurons are doing the computation. >> The neurons are doing the learning. >> The neurons are doing the learning because there's no funny business going

1:02:16on where like I've got a reinforcement learning agent that okay fine. So now it's playing a video game. What's the big deal? knows all the history of Doom's lore internally. >> Exly. So, here's how they address that. They actually did address it in their um >> frequently asked questions section. They said, "Okay, isn't the decoder doing all of the learning?" What they tested it was the how they tested it was through an ablation study. So, they've got a metric called a breakthrough rate, which is the amount of episodes that scores above a negative 600 reward threshold. Now,600 is pretty low because 1,000 is the floor. I've never played Dune Doom. So maybe somebody can tell me ifgative600 is like a lot or a little.

1:02:57But here's the key metric that they said. They said that if I were to use the neurons in the loop, I get 27% breakthrough rate. So my score is above that 600. If it's complete silence, so no input from the neurons. Or if I get put random noise from the neurons, then

Cortical Labs’ defense — ablations, frozen encoders, and what’s still unclear

1:03:15I get only a 7 to 8% breakthrough. So the neurons are doing something. >> Yes. Again, there's not a lot of trials. They're just spitting out these numbers. It's not, you know, I don't I don't see data. I'm just like taking their word for it that they wrote in a blog post that it's 27% rather than 8%. So, from baseline, okay, fine. Baseline is 8% 27%. Maybe I'll take your word for it. The next problem that they addressed in their frequently asked questions is that there's a you know, is it the encoder? Mhm. >> Is it is it the thing that is encoding the neuron state and then feeding it into the reinforcement learning thing because that itself is a neural network. So is that doing all the learning? Well,

1:03:57let's see. So if you look at um photo 30, this is their response. They say that >> you know well >> the brain cells in the loop they're effectively saying that assuming the BL brain cells are static. This is not correct. The policy in the cells are dynamical systems. So the biological neurons have an internal state and that is affecting the reinforcement learning. Now to me all that's saying is like well we actually made it harder for the reinforcement learning agent. >> Yeah. >> That you know it's like okay the neurons are kind of like a noisy controller that I'm like >> using to to to create action to my game. And maybe you know to me it's like okay well that just means my reinforcement

1:04:38learning agent is like >> okay with a noisy controller. The last sentence is what is um key. During testing, encoder weights were frozen and still observed improvements in reward. Which means that I >> I froze the way that I am reading my neurons and still there was improvement, which means that the neurons were maybe doing some type of learning. >> Y >> you know, but there's still the encoder in between even if it's in a frozen state. And so how do you know that it's not attributed to the some base fun some base capability in the encoder even if it's frozen? >> Yes. So that's where I'm >> that's my naive >> that's exactly right. Right. You know it's like in the YouTube video they talk

1:05:18about oh it's like >> you know if the neurons fire a certain way the game is going to turn left. If the neurons fire a certain way the game will move for the character in the game will move forward. And if they want to fire the neurons will fire a certain way. Well, really, not really. If the neurons fire one way or the other, it might just be that the reinforcement learning agent is just like also interpreting the state of the game and just like overwriting everything. So, I don't think it's a very clean test is what I'll be honest about, right? And perhaps they're working on a >> paper that is going to elucidate some of these holes. I certainly hope so, because I do want, right, >> a petri dish game,

1:05:59>> right? I sorry I do want a Petriish brain to play Doom. I think that would be hella cool. >> But I'm I'm I think I think the jury is still out because one, all of this is anecdotal. Yes. >> So there's no hard data. They're not showing actual, you know, box plots about the spread and so on and so forth. There was another thing that I read just briefly that was saying that the the the seed in my neural network actually affects how the gameplay is done, right? random seed that initializes weights and things like that. It's like, well, if it's that fragile, can you really say something about whether the >> neurons are doing anything? So, I don't know, you know, but that's my skepticism. >> And but and this is not to take away

1:06:40from the foundational uh uh discovery that Dishbrain in the Pong context, which is why we took the time to talk about how we got to the Doom version, >> which is like there is a there there. >> Yes. But the way in which they're trying to scale it up for the increased complexity of the tasks they're trying to get the wet wear to do. Yeah. It is not clear that it is purely from improvements at the >> uh >> artificial human pur potent uh stem cell level that's doing the learning. >> It could just be the infrastructure that has now been built around it. Exactly. >> That makes it kind of like an interesting thing. Exactly. But not

1:07:21brains on a chip are doing all of the heavy lifting. >> Exactly. And that's what all of the

The sentience backlash — what the Neuron rebuttal argued

1:07:27that's what all of the news articles are saying, right? All of the news articles are like >> uh brain in a petri dish play doom. >> I don't think anyone's mentioned this reinforcement learning paradigm in the middle right? >> That is kind of the whole game maybe, >> right? And because there's not a paper again like the Pong version which they previously done, the jury is going to still be out on this. And that's why it's really important to have peer-reviewed journals because a reviewer would be like, >> you know, what are you doing? >> We're we're doing that here on the pod, let alone in a professional. >> Exactly. Yeah. Um, one last thing I want to um touch on is there was a lot of heat that the authors took for their 2022 paper because the 2022 paper in

1:08:07their title mentioned sentience, >> learn and exhibit sentience. There was another paper that came out um that was a backlash. Okay, again published in neuron as a rebuttal and they were effectively saying that you can't use sentience, right? Or the idea of perceiving >> because the behavior that you're mentioning and you're observing here when the petri dish plays pong is one of homeostatic feedback stabilization. There's no highle cognitive intent. >> Yeah. Yeah. >> Right. And so you can't really say things like that because then with sentience comes ethical concerns. In the public lexicon, the term sentience means

1:08:49there's some morality now that we have established. >> Yes. >> Onto this dish. And if that's the case, well, what are we actually doing? >> Yes. >> Yeah. And I think we kind of over the course of this episode we've kind of like landed on that same point which is oh this is an interesting fundamental understanding of how you can control neural networks through electrical uh s like electrical read and write architecture to optimize for stability in the system. That's cool. But that's effectively what it's and then we're just connecting that to an external thing we are interpreting and giving value to >> as like oh hit it with the paddle.

1:09:31>> But it it >> fundamentally >> Yeah. fundamentally it's just trying to reduce surprise. But surprise itself is a mathematical construct that is trying to reduce right >> just because of the way that the proteins and the >> and the neurons work when they when they get together. So it's it's a far leap to say this dish brain is sentient because it can sense and respond to an environment like Pong. Right. Right. But this again I don't know this is again this this is I'm we're going beyond my pay grade and I do want to be very clear about that. Right. Because questions of consciousness, questions of sentience is something that I grapple with all the time. I have a a PhD in like neural

1:10:12systems, right? Like one of my PhDs was about how neural networks talk to one another and sleep and so on and so forth. You can create very simple neural systems that sleep and mimic behavior. >> I don't know what consciousness is, but

Wrap-up — real breakthrough, real hype, and what game should it play next?

1:10:24for them to put that in the title is is a bit weird, >> right? And and I think that's what that's what the rebuttal was about, >> which is which is which is I think is fair push back not to minimize the work being done. And just to be clear, right? Like this is not trying to be like, "Oh, you didn't do anything." Sometimes it's pretty cool. Sometimes our reviewers are like, "Yeah, you didn't do anything." Oh, yeah. That's not what we're saying. It's just that there's a nuance about the really cool thing that is being done and then the CNN news headline that says, "Brain on a chip is playing Doom." >> Yeah. >> Which is not >> Yeah. quite accurate. >> Yeah. Exactly. And so, you know, and on this podcast, we're in the business of like just talking about what was done.

1:11:04Yes. >> What were the methods and what were the actual >> findings and the results. And so, you know, it is what it is. They use a reinforcement learning paradigm. I don't know if that changes things. It seems to me that they have a lot more work to do. Yes. To convince someone like me. Yes. >> That it's the neurons that are playing Doom and not some small artificial neural network. Right. >> Not a lot of details. So, I'm looking forward to hopefully >> a peer-reviewed scientific paper. You know, >> if Cortical Labs, if you want us to come over to Melbourne, I think it's Melbourne. >> Yeah, it's Melbourne. uh come in and do some coverage as you prepare for release of a research paper. We are more than happy to join you and have the opportunity to learn more. Uh if you we

1:11:46will send this episode to you to get your feedback as well or corrections. So please let us know um if there's anything that we can do to clarify here. We are f like this is super interesting and we are just curious to learn more and so I'm sure they are already well ahead of us uh on that point. So we'll see what happens. We will cover if there's a follow-up we will cover it again. This was a phenomenal episode around some of these details. I just we could talk about this forever. However, we said we were going to keep our episode shorter, so I'm going to quickly wrap us up here. Uh after that, let us know in the comments. Do we have a comment for the community today for folks to comment on

1:12:26what game do you want to see? >> Oh, that's a good one. Yeah. >> What game would you like to see the brain on a chip play next? Doom was number one. What's number two? I know what my thought is. Maybe we'll share it next episode. >> Yeah. >> I'm your host, Lester Nari, joined as always by my co-host and our resident PhD, Krishna Chowdery. Again, multiple episodes a week. We'll see you later this week.