[Solo] Neuro AI

Okay, Amazing. Yeah. So I was supposed to do a talk with Avicil, that didn't happen, but in the last hour I put this presentation together and it's gonna slap. So I think this conference is just an AI conference with a lot of crypto people, right? Yes. Great. So this is not about crypto, it's about neuro, Neurotech. And I still think it will hit in this audience because it's really interesting technology and so I'm going to run with it actually, before I dive in. Yeah, my name's Casey. I actually. Well, I did my master's research partially in neuro at Johns Hopkins Applied Physics Lab and then I got into crypto and AI and was at Google for a lot of years. And then I went back to Harvard to study more of neuro, but dropped out for Paradigm and then now run my own frontier tech fund called Topology. And we invest in a lot of creative crazy like neuro and AI and crypto and other sci fi things. But long story short is I spend a lot of my life trying to learn a lot about neuro and know enough to teach you guys. But I'm not an expert on this and if you are working in neuro, come find me after. But we'll start with what neurotech is. So neurotech is generally technology that allows you to either read from or write to the brain. And today there's a handful of technologies that we're using to do this. The oldest and most used one and the one I used in my research is eeg, which uses electrical signals. But right now, like the cool kid on the block is something called ultrasound and everyone's using that and they use it for both reading and writing. Functional ultrasound to read and then TFOs, transcranial focused ultrasound to write. There's also like other things like fmri, which you guys are probably familiar with in the medical field. But the problem with FMRI is it's really clunky and expensive and so it's not really brought into the startup sphere. But the takeaway from this is like there's a lot of ways to read and write ultrasound. Very cool and talked about and used in startups right now. But EEG is still probably the best, although it does have really bad spatial resolution, which is a problem. And then just to not make this super nerdy and technical, but instead just like jump to the good part around like what neurotech is actually going to unlock for the world. I think that when we combine neurotech and AI, everything is going to change. I think neuro is like the most underrated technology spoken about. Everybody's like, you know, shilling AI these days, but neuro is like the next AI and this is what's going to happen. So the first, which I'll tell you is already happening and I'll explain is thought decoding. So just being able to think something and then a machine tell you what you're thinking without verbalizing it. The next is neuro and mood modulation. So like having an actual solution to depression instead of an SSRI which not only doesn't work, actually makes symptoms worse over a longer time horizon. And things like having happiness on demand, which is super sci fi esque, but something that like nudge is working on in this way of having like blissful states on demand, then brain to brain calm. So eventually replacing language. Right now that's how we all communicate. But it's a really lossy function as we know. And I think it's totally reasonable to think that in the future we'll just directly stream thoughts to thoughts and then eventually the most crazy thing is going to be digital humans and having this like whole body and emulation going on where we're actually uploaded into machinery. And I know that sounds really nuts, but we'll explain it all. So basically on this list the first one which is these are ranked in order of kind of like insanity and the least insane one is the thought decoding and that's where we are today. So the most iconic paper in this field is something called Mind i2. I invested in the author and he's amazing portfolio founder. But basically what they did is created this system where you decode brain activity using FMRI and then you translate that into the latent space so you embed or vectorize those signals and then you use deep learning models like clip, not their own generative model but models that already exist. And then you reconstruct the images. So this is pulled directly from the paper if anyone wants to go read it. But so this is a human subject that saw these images and then this is a old school brain diffuser. It's still like an AI model that did this. But this is like the attempt to recreate these. It's like pretty mid, it's pretty good. But then Mind Eye 2, you know, just blows everyone out of the water and is like, I mean this is so much more similar than this and these two. And the only reason I'm even calling out brain diffuser is because even in just two years or even actually less than that, like 18 months. AI is accelerating so quickly and we're seeing that progress even displayed in Neuro. And so we were starting to kind of get this unlocked, but now with the latest diffuser models, with the latest foundation models, we're really, really starting to be able to reconstruct images in a way that's super uncanny. Does that make sense? Yes. Cool. Okay, so, yeah, now we're going to skip to the less known stuff, which is like the whole brain emulation. This is actually an image of a fruit fly's brain emulated. And this is a good kind of quote from somebody that's probably more reputable than me, but I'll read it. So mind uploading has been science fiction, but now mind uploading for a fly at least is becoming mainstream science. Basically, people funded this over the last decade and it finally started to work where now we have this digital fly. And what they did was take the different brain states measured in a variety of ways, and then try to create this like digital fly that acts like a real fly without anything other than the emulation, which is like really, really crazy. If you take a moment to grow, rock what that means and it works. And so the company that's doing this, there's two. So Sam Altman's really into this one. It's called netocomb. I have my own problems with this company, but the other one is Ian, who I'm actually walking with, the founder on Friday with, and both of them are working towards this emulation task of how do we upload first flies, the next one's going to be mice and then the final one will be actual human beings. And this is only enabled because of AI, because without AI, we wouldn't be able to really map or verify all the neural connectivity in any efficient manner. And so I think it's a really concrete example of how AI is actually unlocking Neuro. And then those are like two. So so far we've talked about the thought decoding with mindeye and then whole brain emulation with something like eons. Then there's this kind of like third category that I'm really excited about. This is a request for startups, so if anyone wants to start it, come find me. I think like a lot in our social circles we talk a lot about the data wall and like how we've consumed all the data on the Internet. And so when people are always talking about what's next. And so the most common answers I hear are like synthetic data or private data, like your messages or stuff that Vaughn is working on if Anna's here or not right now. Anyways, I think that neural data is probably the next category of data that's gonna just totally rock our world. And when you think about why we don't have it, it's because not only is it really, really hard to get, and there's a ton of issues with all the modalities I listed out, but even when we do get it, we have no idea what it means. So we can look at an EEG signal, we can look at an FMRI signal, We have really genuinely have so little understanding of how our brains work and so there's no utility to it. And so then what we think at topology is that there's this huge opportunity to actually pair brain reading with real world context. So this is the pitch. I think you could right now with the technology and this wasn't even possible last year, but you could go out, you could get a wearable and you could start like recording your physical world and voices and have all these different inputs and then pair that with, with EEG or a different modality that's reading your brain. And then for the first time in the world you can have a dictionary of basically like brain state and then environment. So to make this more concrete, it could be like you're looking at a baby and you record that as like, you know, baby looking brain, then you do that across a community and then from then on you have that mapping and then from there you can actually, I mean then we finally understand how the brain works. Because right now we're taking like a bottoms up approach of like what does this neuron do and what does that pathway do? And it's not working and it hasn't been working. And so now because of VLMS and vision language models, we can go more tops down. I think it would work. And so that's my request for startup. Then the last thing is this idea of kind of like bringing your own brain. Chris Dixon, for the people who are in the crypto sphere here, kind of made this canon or kind of like saying popular in terms of like read, write, own. And I think that as we get closer to this future of Neuro, it's so obvious that this is our most precious data in the world or most precious asset. And so to think that like Elon's just going to take it with neuralink is so ridiculous. Like obviously we're not going to allow that and like every way that people feel about Sam with OpenAI, that's just going to be multiplied with Elon. And brains given like neuralink's the only real player right now. And so I think it's not crazy to feel that in the future we're going to have this concept of like having your own data, owning your own brain data and like bringing it around places. So there might be some crypto overlap in this sphere, but it's really early. That's that. Okay. I don't know if anyone has questions, but I really wanted to stick to under 10 minutes so I went fast. Yeah, Wait, you got to go a lot louder.

Being eyeball suffices.

I'm going to bring this to you. I'm going to bring it to you.

Human eye actually have macular the high first solution and perforation. The first line is fixed single wavelength fixed focus brains because they got too much memory. So you can capture the signals to it, but you cannot duplicate the human being because we remove our eyeball all the time. This per usual have no resolutions whatsoever.

And what's the question or what's the.

Just the thought you cannot using the first five model. Even image generation, they're not compatible.

Why?

Because first I don't have no eyeball to move. There's thousand of tiny pixels in single input. You have color. We have eye wood for we just scan across the room. This is a memory to remember what they stand before such they image. So what you get in the head, the peripheral ratio is all generated in.

The head using like the timing doesn't make sense. Yeah. Can we talk about it more after?

Great talk. Very interesting. From a VC perspective or just in general, what are some of the top startups or products for non invasive and invasive brain computer interfaces that you're seeing come up and also if you can maybe share some stuff on the open source side of things.

Yeah, yeah. The problem like one of the reasons I wanted to talk about this like one big problem is that VCs don't invest in this. And the reason is because there hasn't been a large outcome. So they feel like it's risky. Right. Because like what has really worked. There's like BlackRock and Neuralink. We gotta go this way. Okay. Some of the really cool startups if you want to Google 70s so like Aeon Systems doing evolution. Algeria is doing functional ultrasound for reading. Nudge is doing transcranial focused ultrasound for writing. Forest is doing another ultrasound lab which is on the open source side. They're doing like a nonprofit run by our friend named Sumner, which is super sick. Stella, any more that come to mind? No, I can send you a full list. We have a doc with it internally.

So like, what kind of timelines we talk about for this being a practical reality? Because there's obviously like the Utah Array, which is like a horrible device, and now we have neuralink.

Yes.

And like.

Well, neuralink's still bad for a lot of reasons.

Yeah, it's too local and whatever.

Yeah, it actually kills a lot more than it creates.

And like what kind of timelines are we talking about this being ubiquitous and scalable? One modesty.

Hardware. Yeah. I mean, I think we would have to sit down and talk about which use cases because I think certain things are going to get unlocked. Like Algeria is creating a de aberration algorithm which allows you to finally just like read with ultrasound without having splattering of signals. If that unlocks, then it is a domino effect for a lot of use cases. And I think that could happen in the next two years. If you're talking about like the human emulation thing, I mean, we're a decade away, I think. I mean, things go faster, but like we don't even have it working in a mouse yet and we still need them to raise like $50 million to even do it in a MOUs. Not a lot of people want to give the money because, I mean, it's a crazy thing to do. But I do.

For the open source community though, would you think this needs to be like, well, capitalized, highly focused, private entity?

I mean, I would love to answer and say I think it can be open source. The problem is so much of this is hardware restraints too. And so it's like it's not just a software problem, you know, and like you need access to labs and certain, only certain labs have certain things that we would need to really run these experiments to ground. And so I think it's probably a combination of private and public. I wish the open source community could run with it. I just don't think that's feasible. But push back if you think I'm wrong. No. Okay, cool.