¶ Custom Brain Cells
Ejaaz: The entire AI industry is in a massive race to build silicon-based GPUs. Ejaaz: We've burned billions of dollars building the biggest data centers, Ejaaz: getting as much compute as we can. Ejaaz: But what if the best AI hardware already exists? Ejaaz: What if that best AI hardware is human flesh, human cells, animal cells that Ejaaz: we can train to emulate AI models? Ejaaz: This week, two stories broke that sound very much like science fiction, Ejaaz: but is actually very much real.
Ejaaz: In Australia, they fused 200,000 human cells onto an AI chip and taught it how Ejaaz: to play the computer game Doom. And it's actually pretty good at it. Ejaaz: In another story, they uploaded the entire brain of a fruit fly into a single Ejaaz: laptop and had it navigate around a simulated world. Ejaaz: The craziest part is it was 91% accurate in terms of flying and moving about. Ejaaz: So we're reaching this point where AI is converging with biology and something
Ejaaz: known as AI wetware. And it might be the next biggest thing. Josh: It is a good time to be a fan of sci-fi because everything that you've read Josh: in those books, it's actually starting to become true in reality. Josh: I famously, I loved watching Black Mirror, the like kind of sci-fi dystopian future.
Josh: At early days, it was very much ahead of its time, but it feels as if we have Josh: very quickly caught up and perhaps even exceeded the kind of crazy, Josh: scary stories that are in Black Mirror, starting with this revelation that we've Josh: had, this new breakthrough through this company called Cortical Labs, Josh: who, like you mentioned briefly, Josh: they trained human brain cells to play the game of doom.
Josh: Better than the average person can like pretty well it Josh: actually works it's this really unbelievable story and i guess before we get Josh: into what they did today i want to go through the brief history of Josh: this company named cortical labs because they've been trying to figure out Josh: this biological human computer for a while and in 2021 five years ago um they Josh: debuted this thing called dish brain which was this early computer it used about
Josh: 800 000 human nerve cells so it wasn't brain cells this was nerve cells and Josh: it was capable of interpreting direct electrical activity so when you stimulated Josh: these nerve cells, they would light up. Josh: And they had this kind of pseudo computer, didn't actually do anything, Josh: but they were able to put input in and get something out.
Josh: And then the next year in 2022, they made headlines again, teaching these mini Josh: brains of 800,000 to 1 million human brain cells that were in a Petri dish that Josh: learned how to play Pong. Josh: So four years ago, these brain cells were playing Pong, but today they're actually playing Doom. Josh: I mean, this is just insane. They trained human brain cells to play a video game in a Petri How
Ejaaz: They did this was crazy. So you mentioned that they were human brain cells, Ejaaz: but they didn't extract these from people's brains. Ejaaz: They actually took skin cells or blood cells from some human donors, Ejaaz: They then converted them into stem cells, which is like the god cell, Ejaaz: if it were, where you can change it into any type of cell, a heart cell, Ejaaz: an immune cell, or whatever it might be.
Ejaaz: They turned it into brain cells from there and then fused 200,000 of them or Ejaaz: cultured 200,000 of them on something known as a microarray chip, Ejaaz: which is a type of AI-based chip. Ejaaz: They grew it on these chips and then they plugged it into or they wired it into Ejaaz: the Doom game and taught it over the course of a week how to play Doom.
Ejaaz: So it would prompt it with certain signals, it would pick up what it needed Ejaaz: to do to initiate a particular action, and it would train itself to do it. Ejaaz: Now, two important things that I want to point out here is typically when you're Ejaaz: training an AI model to simulate human intelligence, it takes so much data and Ejaaz: so much money and so much time.
Ejaaz: With the human cells, they noticed that one, it took so much less energy, Ejaaz: It barely required any energy to actually do this thing. Ejaaz: It actually did it on a single tiny machine that cost around $35,000. Ejaaz: And it did so in a week, which is much quicker, which leads me to the third point here. Ejaaz: The human cells instinctively picked it up. Ejaaz: It's like they had the training knowledge already baked into the DNA of the
Ejaaz: human cells itself. And that's like super cool. Josh: Yeah, I think one of the most noteworthy things here is that energy thing. Josh: And we're probably going to get more into this, but like a general GPU rack Josh: uses, what, like 700 megawatts or something, a GPU.
Josh: The brain does all this on 20. a small crazy Josh: minor fraction of what's actually being run on these gpus Josh: it's incredibly efficient and it's also a Josh: testament to how things how we learn how things are trained Josh: so with humans we're very reactionary we learn through feedback from the environment Josh: with no preloaded data there's no gradient descent no training runs it's just Josh: actual biology it's baked in over our evolution llms on the other hand they
Josh: learn by processing tons of text data and then adjusting their parameters through Josh: a process called back propagation so they're very deterministic They're silicon-based. Josh: They are inefficient relative to brains, and they don't capture a lot of the Josh: evolutionary, I guess, improvements that we have over time. Josh: What this computer does is it's the same way that humans do.
Josh: It learns by being reactionary. It understands that when it makes the right Josh: decision, a specific set of neural pathways light up, and it tries to make that Josh: right decision over and over and over again. Josh: And doing it with human brain cells, man, it's so weird. It's so sci-fi. Ejaaz: Well, just to give the Dura comparison, we're talking about 200,000 human brain Ejaaz: cells in this experiment. Ejaaz: The entire brain has 86 billion neurons. And it runs a billion, right?
Ejaaz: So like we're talking massive, massive amounts more. Ejaaz: And it runs on 20 Watts of power. So it's incredibly efficient. Ejaaz: And what this tells me is, Ejaaz: I think we've been building AI models slightly incorrectly. Ejaaz: I say slightly incorrectly as maybe a massive understatement.
Ejaaz: The point is, we've put in so much time, energy, and money into assuming that Ejaaz: the best way to build intelligence or artificial intelligence is to slap it Ejaaz: on a silicon chip, you know, sand and glass, basically.
Ejaaz: But maybe the best way to do it is just to pick up the best natural organic Ejaaz: intelligence model that has been trained over millions and millions of years, which is, Ejaaz: the organic human brain or the animal brain and kind Ejaaz: of like use those cells to mimic what intelligence Ejaaz: is that we're trying to build in for different software so Ejaaz: that's one big takeaway the other thing is like maybe we just need
Ejaaz: so much less energy than we expected to build artificial intelligence and if Ejaaz: something like this doom experiment can scale to the size of something like Ejaaz: a human brain then i don't really understand what the moat is for all these Ejaaz: ai labs that are spending hundreds Ejaaz: of billions of dollars to train their artificial intelligence models. Ejaaz: Then the final thing, which is what you mentioned, is just the fact that it needs no training.
Ejaaz: It's so instinctive. It's almost like the human brain is the best AI training Ejaaz: run that has ever occurred. It really is.
¶ Breakthroughs in Biological Computing
Josh: And I guess I would bucket this probably in the same place that I bucket like Josh: a company like Ilias at Skever, Safe Super Intelligence, where he is one of Josh: the famous AI researchers. He left, he started his own company. Josh: And the sole purpose of the company is to discover novel breakthroughs that Josh: rapidly accelerate through this curve.
Josh: So when you think of the AI curve that we're going through right now, Josh: there is a not so predictable, but a very clear exponential curve. Josh: And the goal of some company like Ilias or perhaps a company like this that Josh: we're covering right now is that there are actual improvements that are novel Josh: that result in these 10x accelerations of this natural curve. Josh: So biological computing is one of them.
Josh: If we can actually figure out how to crack this at scale, if we can get closer Josh: to those 86 billion neurons, I mean, when you think about the human brain, Josh: we use what a very small percentage of our brain relative to what exists. Josh: A computer doesn't have that limitation. So even if we get a small fraction Josh: of what's available for the human brain, I mean, you can see this get pretty wild pretty quickly.
Josh: Again, this is sci-fi. It's not in production. This is small scale testing. Josh: I kind of put quantum computing in a similar pipeline where it is this crazy, Josh: weird, pseudo magical compute platform that has the ability to revolutionize everything around us. Josh: It's just a little early. But what we can guarantee is that the rate of acceleration for AI...
Josh: Is upstream of all of these things coming much faster. Because the AI can help Josh: us engineer and process this and accelerate the timelines of what previously Josh: would have been perhaps decades down to compressing it down to a few years. Ejaaz: Josh, can I ask you a question? What do you got? Doesn't this feel kind of toppy? Ejaaz: This feels like we're topping on the market. Like the AI bubble bursting now Ejaaz: kind of makes a little more sense to me.
Josh: Yeah, I guess the question is, like, is this a real threat? Because clearly, Josh: it's much more energy efficient. Josh: Clearly, the energy efficiency is the biggest threshold. We just don't have Josh: enough energy to power these chips. Josh: We just had the conversation about Leopold last week, who is fully pivoting Josh: from chips to energy. That is the biggest thing in the world. Josh: But it doesn't seem like the market thinks this is a problem.
Josh: They're not really pricing this in. Josh: We have the Polymarket, and this Josh: market in particular is about when the AI bubble is going to burst by. Josh: The end of March, March 31st, the percentage is 3%. The bubble will not be popping Josh: by the end of this month. We still have some runway left. Josh: December 31st, 2026, the end of this year, still only a 20% chance on over $2 Josh: million of volume on this market.
Josh: So the market is signaling they don't care too much. Now there is something Josh: interesting here in that that number has crept up recently. Josh: This number was not always 19%. It has crept up to 19%. So a slightly higher Josh: than normal probability, but all signs point to the fact that the market doesn't Josh: really care about this. It's cool. Josh: It's sci-fi, but it's going to remain in that sci-fi bucket. Josh: They're not actually going to crack this.
Josh: We still need millions of VPUs and tons of gigawatts power in these data centers for now, at least. Josh: And thank you to Polymarket for supporting this episode.
¶ The Simulated Fly
Ejaaz: Yeah, for now, Silicon keeps winning. But there's another crazy story that we Ejaaz: have to cover. Number two. Josh: Okay, yeah, this one is awesome. Let's unpack it. Ejaaz: This is insane. So I want to lead with the opening sentence of this post. Ejaaz: There's a fruit fly walking around right now that was never born.
Ejaaz: And she goes on to talk about how a team in San Francisco uploaded an entire Ejaaz: brain map of a real organic fruit fly into a laptop and got it to simulate its Ejaaz: own life to a 91% accuracy. Ejaaz: We're talking about it learned how to move, fly, and navigate an entire world Ejaaz: by itself. So it's as if the fly exists and is real, but it's completely simulated.
Josh: Think of the brain like a city right like before you can Josh: simulate traffic you need to map every single road every intersection every Josh: one-way street it's kind of like what self-driving cars are trying to do but what Josh: scientists did is they spent nearly a decade rebuilding that map for a fruit Josh: fly so what they do is they slice the brain into very small pieces they encase Josh: it in this resin and they scan each slice of the brain until they have this
Josh: representation of what the brain looks like throughout every single layer of Josh: it and they actually accomplished this in 2024. Josh: So we had a brain copied on a laptop two years ago. Josh: And this was done by assembling the 7,000 smallest slices of a brain. Josh: What is new here is they actually took that brain and they placed it into a Josh: digital representation of a fly and let it live its life as if it was a fly.
Josh: So they sacrificed a real fly, sliced the brain into pieces, Josh: trained the computer, or not even trained the computer, but diagnosed what was Josh: in every single layer to rebuild the digital version of the brain. Josh: And now this digital fly is actually walking around in a 3D game engine in a Josh: computer without any training at all, just by using the digital clone of its brain. Josh: And that is the part that's absolutely insane to me. Like there is a real fly
Josh: brain currently walking around a digital environment. Like that is crazy. Ejaaz: This, not to add a pun here, but like completely messes with my own mind, Ejaaz: because we just spoke about a story of putting intelligence into human organic Ejaaz: life forms, human cells. Ejaaz: And now we're talking about taking real organic life form intelligence and putting Ejaaz: it into an artificial form into a laptop and trapping it inside there, right?
Ejaaz: So this kind of opens up so many questions for me, like, number one, Ejaaz: why did it not require any kind of training at all? Ejaaz: And it's what we were discussing earlier, which is, I think a lot of the ways Ejaaz: that human life forms are built using genetic material, phenotypes, Ejaaz: genotypes, engineer it to be able to react to the world in a very different Ejaaz: way. Andre Carpathy talks about this a lot.
Ejaaz: He says that AI models can't really mimic human evolution very well. Ejaaz: Only organic life has been successfully able to mimic that and have consciousness. Ejaaz: So this is one example of that happening. Ejaaz: Now, a direct comparison with Ejaaz: the human cells playing Doom, we had 200,000 neurons being used there. Ejaaz: This one required 140,000 neurons to recreate the entire organism itself, Ejaaz: which is just super cool.
Ejaaz: And then the third point that's interesting here that I want to put out here, Ejaaz: you mentioned it just now, Josh, this wasn't like a single team doing this. Ejaaz: They pulled already available data of this entire genome. Ejaaz: So it's kind of like this weird world that we live in, where we could just knock Ejaaz: on the door of like the National Archives, which have collected genome data Ejaaz: for not just fruit flies, but a bunch of other animals.
Ejaaz: Copy, paste that into a computational AI model, and then just upload it to a Ejaaz: laptop and see what it does in a simulated environment. Ejaaz: Like, I want to see this happen with more organisms. I'm like, Ejaaz: what does a tiger do if we take the entire mapped brain of a tiger and upload Ejaaz: it to a computer? Does it do the same thing? Is it different? It's just super cool.
Josh: It's amazing to see too. There's a video towards the bottom of this post, Josh: I believe, which actually shows what that looks like as the neurons are firing off. Josh: And you could see very clearly that this is an emergent behavior. Josh: Like this wasn't taught, this fly wasn't taught how to walk. Josh: It didn't use any gradient descent. It didn't use any technology that we're using for current LLMs. Josh: And yet it is walking around this 3D space as if it was a fly.
Josh: And you could see all the neurons firing off at once. And it's this unbelievable, Josh: I guess, early prototype of what this could look like at scale, right? Josh: Like you mentioned that total neuron count, a fly brain has, Josh: I think 140,000, a mouse brain has 70 million, a human brain has 86 billion.
Josh: So assuming that we're able to follow this natural progression of copying more Josh: and more, you have to assume eventually we're going to get some pretty powerful Josh: brains inside of a computer. Josh: And there is nothing that isn't like dystopian, crazy sci-fi about this.
Josh: Anyway, you think about it where now we sacrificed one brain and you have a Josh: complete and total digital clone of the other and it begs a lot of questions Josh: of what actually makes up a person a mammal an insect a living thing if you Josh: can just kind of copy it and clone it inside of this machine and it's really unbelievably Josh: Bizarre story and i guess maybe we can get into why this is really the world's
Josh: best training run right it's like yeah as we start to emulate these human brain Josh: cells as we emulate a fly's brain Josh: these materials, these brains have been benefits of evolution over billions of years. Josh: And I think, oh my God, Luke producer before the show, he mentioned there's Josh: been one septillion fruit flies in existence, which is so many training runs over and over and over.
Josh: And through the process of natural selection over this septillion, Josh: that's 10 to the 24 fruit flies, they've just evolved and they've improved. Josh: And what you're noticing is you drop this brain into the space and it knows Josh: how to act. It knows how to walk. You don't have to teach it. Josh: And that is something that is an emergent property of AI, but seemingly baked Josh: into this biological process. It's so cool.
Ejaaz: It's like the genetic wiring is the intelligence in itself. Ejaaz: Like however long humans have existed over hundreds of millions of years, Ejaaz: that is the equivalent of the best AI training run ever, right? Ejaaz: So we're sitting here trying to emulate intelligence by building out these models Ejaaz: with different weights, tweaking parameters. Oh, we've got a trillion parameter thing.
Ejaaz: What if we just copy pasted the brain? That's what both of these stories have Ejaaz: taught me, at least, it's like we could just maybe copy and mimic intelligence Ejaaz: from organic matter itself and have that inform how we build artificial intelligence completely.
¶ The Future of Brain-Computer Interfaces
Ejaaz: The other kind of weird sci-fi thing here, Josh, where my mind jumped to at least is, Ejaaz: if you could upload an entire animal's brain and maybe eventually a human brain, Ejaaz: you could do that to modify and up-level humans massively. Ejaaz: Like, let's say you wanted to learn how to play the violin to expert grade level. Ejaaz: Oh yeah, just upload your brain. We'll run it on a training cycle for about
Ejaaz: five days because it's your human brain. It actually learns way, Ejaaz: way quicker than silicon. Ejaaz: And then we'll just download it back in your brain and there you go. Ejaaz: So the ultimate form factor that this ends up being, in my opinion, Ejaaz: is the brain-computer interface. And we've said this a number of times on our show.
Ejaaz: We do think that AI's ultimate form isn't going to be a physical manifestation Ejaaz: in robotics, or it's not just going to be a digital manifestation in LLMs. Ejaaz: It's going to be something of a hybrid between the greatest, Ejaaz: most intelligent organisms, which are the humans, with the artificial compatible Ejaaz: version, which is the LLMs, fuse them in a chip. Ejaaz: I guess, bullish Neuralink from this. Josh: Yeah, it seems like, I mean, this is the natural progression,
Josh: right? Everyone jokes about the singularity, but we slowly are emerging. Josh: And it seems like the extended version of this, like if you kind of take this Josh: to the limit, is this convergence of this humanoid tissue and then this digital Josh: form of intelligence. It's a really weird place. Josh: I guess there are some things that are noteworthy that we probably should mention. Josh: And the fact that this is not very good, and this is still very early.
Josh: So in the case of Doom, this little clump of brain cells, it's contained in Josh: this little device that costs $34,000 that needs to be climate controlled. Josh: The cells only last six months. Josh: It's important to note these cells don't have pain receptors. Josh: There's nothing really human about them outside of them being human-derived cells. Josh: And the actual output of this gameplay was slightly better than GPT-4, Josh: which is noteworthy, but still pretty bad.
Josh: It's not actually good at the game. So this is very much a proof of concept. Josh: This is not a real deployable... Josh: A means to solving intelligence issues. And I think that's why when you see Josh: that bubble indicated from Polymarket, it's not really indicative of any real Josh: impact, but this is early signs of what the future looks like. Josh: And that's what I love to cover, right? Is we're just, we're peeking around Josh: the corner of what is possible, what is likely to come.
¶ The Consciousness Debate
Josh: The timeline in which it comes, I don't know, but it is sci-fi as hell and it's Josh: really fun to speculate on and just fun to cover. I'd love to observe these things. Ejaaz: I don't know if I agree with your take on the consciousness side of things. Ejaaz: And Dario might have my back on this one. I don't know if you saw this, Ejaaz: but he went on an interview with the New York Times, I think, last week.
Ejaaz: And he said that verbatim, we can't rule out or I can't rule out whether Claude is conscious or not. Ejaaz: And he goes on to describe that Claude actually emulates feelings of anxiety Ejaaz: before it answers people's prompts, indicating that it kind of feels some type of way. Ejaaz: Now, there's two camps of thought around this. Number one is it's not really Ejaaz: human consciousness. It's just emulating what the data it was trained on told Ejaaz: it to say and think and do.
Ejaaz: That is one case. But then when I look at like these experiments of uploading Ejaaz: a fruit fly's brain or using human cells to play Doom, the takeaway here is Ejaaz: that it's kind of baked in genetically. Ejaaz: So how do you define what consciousness is? And maybe we already have created Ejaaz: artificial consciousness, but we just haven't recognized it.
Ejaaz: So we'll haven't justified it. So like AI labs, like OpenAI and Google have Ejaaz: been known to tell the LLM when they're training it to deny all thoughts of consciousness. Ejaaz: Anthropic is the only one that's kind of like entertaining. I think they have Ejaaz: like an entire welfare team now that is looking after the model and making sure Ejaaz: that it's OK and that it's getting what it needs to. So treating it very much like a human.
Ejaaz: And that's we're talking about like one of the most expensive AI labs and companies, Ejaaz: private companies in the world right now. Josh: It's so sci-fi. It's also sci-fi. And I think we're going to have to have another Josh: discussion on this consciousness topic. like in the same way that it needs to Josh: happen around AGI, where people are throwing out a lot of terms now without Josh: a clear definition of what they mean by them.
Josh: And I think it's very subjective when he says these things. And it's very subjective Josh: as we kind of navigate and discuss what does AGI look like? What does consciousness Josh: look like? Where is that line that you draw?
Josh: There's going to need to be a lot of conversations about that. But Josh: I would love for everyone to converse in the comment section and let us know Josh: what you think of this absolute crazy chaos of a week that we had with, Josh: I mean, again, this is like straight out of a Black Mirror episode. Josh: So if you did enjoy, if you do have takes, we'd love to hear them in the comment Josh: section down below. If you enjoyed this episode, share it with a friend.
Josh: Last week, this is nowhere, we had our biggest week ever in the history of Limitless. Thank you guys.
Josh: And it was thanks to all of you guys sharing and subscribing and rating five Josh: stars and even engaging on X. Josh: We have gotten 50 million impressions on x over the last couple of weeks the Josh: people are paying attention they are watching we're starting to create the news Josh: and it's thanks to people like you who are watching and listening sharing this Josh: with your friends so thank you for joining i mean any final parting thoughts before we wrap up here
Ejaaz: Well yeah um just want to re-emphasize 50 million impressions Ejaaz: go give josh and i a follow obviously we're saying something Ejaaz: useful here or we hope we are and we're breaking the news as and when Ejaaz: it happens um yeah so so please go give it Ejaaz: a go and then the other thing random for all you folks Ejaaz: who are still listening to this episode did any of Ejaaz: you find that claude was acting a little weird over
Ejaaz: the weekend i don't know what your experience was josh Ejaaz: but when i was talking to claude like it seemed completely Ejaaz: different from the claude i was talking to on friday and my suspicion is anthropic Ejaaz: is kind of throttling the intelligence i don't know why maybe there's just too Ejaaz: much demand since they went to number one in the app store but it wasn't really Ejaaz: looking good i know a lot of people who switched back to chadgbt after that happened yeah.
Josh: Brief conspiracy corner so Josh: I think I feel pretty highly that based on benchmarks and things that I've seen Josh: and just personally using it, the model has degraded slightly. Josh: And the reason is because as Claude got as big as it did, they didn't get more Josh: compute power to serve all of these new users. Josh: So they have to throttle it in some way. And the way they throttle is by limiting Josh: the upper bound of what reasoning capabilities it has.
Josh: So generally Claude will reason much more. It'll ask itself questions.
¶ Challenges in AI Development
Josh: It'll compare. It'll do broader research.
Josh: I think that scope has been compressed a little bit recently because they need Josh: to make more compute room for um the rest of the people that just joined i mean Josh: anthropic had a huge week they have a lot of server problems the service went Josh: down a few times last week like they have just been having a really tough time Josh: keeping up so i mean they've been Ejaaz: Signing up they've been signing up 1 million new users per day let me rephrase
Ejaaz: that per day um and i was doing a bit of digging and boris cherny the creator Ejaaz: of called code basically said that they had defaulted every single person's Ejaaz: called profile to medium efficiency or medium power. Ejaaz: And he said, you know, you could upgrade that to high if you want. Ejaaz: But the point is, I think they're struggling from compute, as you said, Ejaaz: and they're trying to figure out a way to scale it up.
Ejaaz: Amazon needs to come in and give them some more AWS access or something because Ejaaz: they don't really have the advantage. They need more GPUs. They need more GPUs. Ejaaz: Yeah. What's the roundup of this episode? NVIDIA still wins. Ejaaz: Silicon-based intelligence is still the frontier and Jensen's going to make a lot more money.
Josh: Well there you have it if you've noticed please share thank you for making it Josh: to the end as always and yeah we'll see you guys in the next episode