Prisma founder Sean Cole on training lab-grown human neurons to do token prediction and beat silicon on nonlinear tasks

Speaker 5: friendly IPO.

Speaker 1: With great customer service. You ready for this next one, Jordy? You're gonna love this company. They put a brain in the vat, basically.

Speaker 4: Really?

Speaker 2: Yeah. Here's Sean Cole. Hi.

Speaker 1: It's on Parasma. He's the founder. This is his first time on the show. Sean, how are you doing?

Speaker 3: Hi, guys. Thanks for having me.

Speaker 1: Welcome to the show. Is Brannin in the Vat appropriate or is that derogatory? Should I stay away from characterizing your company that way?

Speaker 7: I think it's pretty good.

Speaker 3: I think it's pretty good.

Speaker 1: Okay. Take us through it. Introduce the company. Introduce yourself.

Speaker 3: So I'm Sean. I'm the founder and CEO of Prisma. Yeah. And we are training brain cells for compute. So previously, I think you guys might have seen, you know, we made brain cells play doom. Yeah. And we just launched, like literally just launched, and we got these cells to do token prediction. So kind of the basis for language modeling.

Speaker 2: It's amazing. Okay. So how many cells do you need? Because many people have said that I have seemingly only a few brain cells. And so how many brain cells do you actually need in order to do x token prediction?

Speaker 1: I can predict tokens all day with just a few brain cells. It's no problem. So seems like an easy easy job for you.

Speaker 3: Not that many. So we're we're renting some brain cells out. We're using about 200,000 brain cells to do token prediction, and it's good enough.

Speaker 1: Are these literally, like, donor brain cells from cadavers, or are you growing Your brain cells. Stem cells? Like, where are they coming from? Walk me through the full process. What's your supply chain?

Speaker 5: I think you got

Speaker 3: it exactly. So we're using serotonin stem cells, we're taking these stem cells, differentiating them into different types of human neurons, and putting those in a dish that we can stimulate and get responses from, basically.

Speaker 1: Okay. And then you said you're able to do next token prediction. I imagine that you're not anywhere near the frontier. How are you actually benchmarking? Like, what is capable? Because this is probably all predicated on a a very extreme exponential kicking in at some point. But where actually are we in terms of progress?

Speaker 3: Yeah. I I think somebody gave a pretty funny example, which is it's currently at this current stage. The token prediction is basically, you know, is this a hot dog or is this not a hot dog type of token prediction. It's pretty basic.

Speaker 2: Well, can you say say you're alive? And then it says, I'm alive. And you go, oh.

Speaker 3: Yeah. We we could get it to do that.

Speaker 1: You could. Okay. I

Speaker 3: think fundamentally what we've done is we've proved that these types of sequential context based architectures of electrical stimulation Mhmm. Are a viable architecture on these biological substrates. We had a specific example where it was a nonlinear task. So let's say, you know, that to refer to context early in the sentence. Mhmm. And if we used a linear decoder, so traditional silicon, it could only maximally get 75% mathematically. Mhmm. But we got above that, so we got 78%. We beat silicon on this very constrained task Sure. Because it was able to model these kind of nonlinear dynamics in in context.

Speaker 2: Mhmm. How long do the do the cells actually last before you need to replenish them?

Speaker 3: So I think currently they last six months. But of course, I think human brain cells last far longer, know, like a hundred years, eighty years. So I think the goal obviously is to kind of extend them for as long as possible.

Speaker 2: Yeah. My brand You were kind of getting at this but but So not not why why why do this? I have some ideas of why you might. But seems like kind of a hassle so you got to have a good reason. Is it is it like idea What do

Speaker 1: you against silicon? Okay?

Speaker 5: Yeah. I'm assuming I'm

Speaker 2: assuming like there's like it like energy efficiency Yeah.

Speaker 1: Yeah. Play it out. If this goes the way you want it to go, is there actually a benefit over just a huge data center or something in space with the solar panel on it? It feels like the current chip stack, the AI stack is is pretty efficient. We've squeezed out a lot of the inefficiencies of being a human potentially.

Speaker 3: Yeah. I think that's where it's really interesting because, you know, regardless of how efficient we make silicon, like Mhmm. Silicon we have right now is is supremely efficient, but we still haven't solved the efficiency aspect of it in terms of power efficiency. I think human brains are extremely power efficient in comparison to silicon. I think there are things that we can harness there on top of other stuff like sample efficiency. Human brains learn very quickly compared to silicon, much fewer examples. And continual learning is free on biological substrates because they keep learning over time. Yeah. But I think continual learning is something that has to be expanded on in in the AI space. We're still figuring out what's the optimal approach to use. But, yeah, massive massive energy efficiency.

Speaker 1: How do you actually think about that energy efficiency though? Because I've seen I mean, was that that news of, like, $5,000 worth of SOL tokens solved a bunch of math problems. When I think about like not even the salary of a mathematician, but I just think about the food that goes into generating the calories, generates the energy, that generates the theorems from a mathematician, you're way above five k. So Yeah. It feels like the models are actually pretty efficient, but what do I what am I getting wrong?

Speaker 3: I think if you think about, like, the cells as growth, like, humans, you know, we are expensive. We have to feed ourselves and all that. I think with Doom previously, we showed that it's possible for us to inject information, essentially, force these cells to learn much faster than a human would learn. So you don't have to learn the basics of language ADCs. We can tell it's just predict whether, you know, this is a hot dog, this is not a hot dog, something like that, far quicker. So we can skip that kind of like prior human learning building phase that would be very expensive normally. Mhmm. So that's kind of the stuff that we are approaching it with.

Speaker 2: Yeah. Is it possible that golden retriever brain cells could be better at long running tasks like chasing a ball?

Speaker 3: Oh, I think that human brain cells for now, empirically, they are the best.

Speaker 7: Oh. However

Speaker 1: Shot's fired.

Speaker 7: Well, you know, we've we've tested it, but Okay.

Speaker 1: We've tested you put a golden retriever in the van. Oh. Uh-oh.

Speaker 3: We tested rat neurons. Right? So they're they're rat neurons and then they're

Speaker 1: human Woah. Brain Golden Golden retriever, rat, these are not comparable animals. Let's let's give it.

Speaker 7: Yeah. Yeah.

Speaker 1: What is the what is like the business going to look like over the next decade? Because I imagine at the end of all of this, there's some sort of business model where like you're selling intelligence. But in the in the short term, there's some venture capital that comes in through Y Combinator. Congratulations, by the way. I do. But but the the what's the middle step? Is it partnering with biotech companies? Is it NSF grants or government funding or something like that or partnering with the university? Like, do you keep the lights on and keep the flywheel going? I imagine you can raise more money off of scientific breakthroughs, but I imagine that there'll also be an economic, a commercial flywheel here even before you're, you know, selling the work.

Speaker 3: Yeah. I think something that we're really looking at right now, which is great because, you know, we've just started, is Yeah. With the kind of exponential increase in AI capabilities, we're gonna start building our lab from scratch to be automated. We want to automate the stem cell research to differentiate them into neurons, the optimal compositions, the optimal kind of, like, spacing on the electrodes, let's say. So I think that the automation aspect of our lab can easily be branched out into different things like drug testing. Mhmm. And I think that could be significant revenue in the short term to push this all the way to make sure that we get brain cells to be the the fundamental substrate for compute.

Speaker 1: Very cool. Well, congratulations. What a

Speaker 2: fascinating company. Thanks for taking

Speaker 1: the time and have a great rest

Speaker 2: of your We gotta we gotta introduce Sean to the guy that we had on yesterday. What was