Dr. Alex Wissner-Gross on the first multi-behavior fruit fly brain upload — and why mind migration to the cloud is coming in 5-10 years

employees instant resolution for access requests and password resets. And without further ado, Dr. Alex Wisner Gross, welcome to the show. How are you doing?

What's happening?

Doing really well. Thank you, John. Thank you, Jordy. Thank you so much for taking the time to come chat with us. Uh congratulations on all the virality. For those who uh haven't followed along so far, can you introduce yourself and the project broadly and then I have a ton of questions, but I'll let you kick it off with an intro.

Absolutely. So, quick background on on myself. Originally from New York, undergrad MIT, studied physics, electrical engineering, computer science, math, PhD physics, Harvard.

Was there anything study? You studied everything.

Didn't study the humanities.

Okay. considered considered going to Cambridge to study chemistry and biology for a while. Just started it was told physics is a game for the young. Okay. So stayed in the US did my PhD in physics focusing on AI and nanotech. I've started invested in advised managed dozens of companies at this point. I've had a number of exits and at this point most of my time is focused on smoothing out the singularity as as you guys were just discussing. So most recently, as you were gesturing, a company that I helped found, Eon Systems Public Benefit Corporation, just announced what we've been characterizing as the first multi-behavior upload, if you will. More on that in a second, of a fruitly, which we think is a major step forward for the field. And and quite frankly, in an era when I I do another podcast, the Moonshots podcast with my friend Peter. I I talk all the time about tiling the earth with compute. Right now, all of this AI infra buildout that we're doing to the tunes of trillions of dollars of capex, all of this is going to artificial minds. And Eon is playing, I think, a very important role in doing some early pioneering experiments and developments to try to level the playing field of the future. Not just artificial minds playing on all of this trillions of dollars of infra and data centers, but enabling emulations and if you will uploads of human minds, nonhuman animal minds at the moment, starting with fruit flies and aspirationally working towards mice and humans and getting getting the brains of a variety of human and non-human animals to operate in the cloud. So walk us through the fruitfly experiment uh why it was so exciting for you.

It's really interesting if you think back remember back with the launch of chat GPT that was just a side experiment that that wasn't the main focus of OpenAI at the time. Some have characterized that as an unhobling if you will. GPT3 the the leadership of OpenAI thought at the time was sort of the the main entree if you will. Similarly within Eon, I I don't mind saying this was more of a side project that the we thought it would be somewhat interesting to to the world, but I I think the the level of response and virality probably even took most of Eon by surprise. So this this is a video demonstration of an internal project to take a bunch of building blocks that were already sitting around for the past 1 to two years and put them together for the first time. So one of these was a paper that uh senior scientist at Eon Philip Shu published in Nature in 2024 demonstrating that just from an emulation perspective that you could take the fly wire which was a wellunded project to capture the conneto of the fruitfly brain that you could emulate certain circuits within the fruitfly blame fruitfly brain. We've had that for a year and a half or so. Another project, Neurommechfly version two. It's an excellent project to to be able to uh formulate simulations in realistic mechanical environments of fruit flies. Another project from uh illeal to formulate coordinated motor actuation of fruit flies and then AI. So in in summary, Eon was really the first to put all these building blocks together that were sort of lying around. This is actually tiny subset of what Eon's working on. Eon is working on scanning the connetos of of human and non-human animals at scale and achieving wholebrain emulation which is the the holy grail of the field. There have been many projects in the past that have attempted human wholebrain emulation. Eon is leveraging advances in AI, expansion microscopy, a variety of other techniques, but it was really just putting the pieces together. So, we're quite quite honestly astonished by the the level of interest in just this one demo.

I love it. I have a million questions. Let's start with uh uh IO like the inputs and outputs like what are you simulating? When I pull up the sims that is somewhat of a simulation of a human the you know a basic for loop over a bunch of random probabilities to say are you hungry? If hungry eat, go over to the stove, cook food. Uh that's obviously nowhere near the level of detail at which a human operates. Um is there an abstraction layer here where the output is quantized to things that you would think a fly would do like move a uh like move a wing or or are you operating at the level of like choose to fly or how at what level of abstraction is the flies output happening? Yeah. Uh this is what I was gesturing at a moment ago. So there's a paper by Özil at all that bundles motor actuations in terms of highle representations. Got it.

So a lot of people sort of the the people who got perhaps most excited in the neuroscience community about this video, some of them I think didn't bother to read Ersdale's excellent paper. Uh so a lot of this is why I'm saying a lot of the pieces were sort of lying on the ground waiting for us to pick up and these abstract motor do you need a system prompt like do you need like an like an initialization function or once you once you uh synthesize the conneto synthesize the connections in the brain uh things just start firing do you have to trigger something to to start the process of consciousness or whatever you call whatever the fly is doing.

No system prompt needed. On the other hand, in in the early days of chat GPT before instruction tuning, there was no system prompt either. In the days of of early language models, no one had thought of prompt engineering. So it's entirely possible that that maybe in the future, some sort of prompt engineering equivalent for for whole brain emulation will be essential. But at the moment there there's no direct analog of that.

Help me compare some of the scales that we're talking about on complexity of a fly, complexity of a mouse, complexity of a human. I imagine that we're on some sort of exponential here, but uh how many orders of magnitude in complexity, size, number of neurons, something like that. Whatever the the metric is, how far away are we from humanity? We're many orders of magnitude away. So the the canonical estimate for the human brain is uh depending on how you count brain cells and there are multiple types of brain cells. Call it order of magnitude 100 to 200 billion cells in the brain. We're many many orders of magnitude

many orders of m many orders of magnitude away from that. So this isn't going to happen immediately. We're not going to get human whole brain emulation

immediately. On the other hand, I I think it's important I I think back to the early days of driverless cars and autonomous vehicles. And what the world is missing right now is, I think, sort of a scale or a framework for starting to think about what that future looks like. We're at the earliest stages of that, call it level one or level two uploading. A lot of people even get triggered by the term uploading. They prefer something else. And I I think we need to start having these conversations in a critical thoughtful way and start to define what is a framework for multiple levels of fidelity for uploading or emulation or maybe we come up with a new term but right now we're starting that discussion and we're trying to do it in a thoughtful responsible way.

So you like the term uploading because you did not uh you did not choose a random structure or an average structure of this particular flies brain. You chose the exact structure of a particular flies brain.

This is a very expensive structure of a very particular fies brain. This was the the structure was again captured through the the Flywire project at great expense.

Mhm. Uh so that's like a clone digital twin type of thing, right? Is that is that

aspirationally? I mean again we're at the earliest stages. I I don't want to oversell. This is just a prototype and quite frankly this is a sliver of everything that is able to put a few of these pieces together build a prototype with relatively small effort suggests that you know we talk in the AI community all the time about overhangs. Many people thought that the arrival of large language models was in some sense a technological overhang that we could have one could imagine a thought experiment. Could we have had large language models 20 years ago if we simply knew what we should have been building? I I think probably

I thought we were very GPU gated on that.

Well, GPUs were being spent on video games for for the first few decades. So, we could have in principle like Marov models. We we knew how to build Marov Bablers. I was building Marov Bablers in the 90s and early 2000s. It just feels like even if you even if you try and port back Frontier models now to a, you know, 1080 Ti, which was a graphics card for gaming, uh, you're going to have a bad time. Well, I mean, we're with the rise of SLMs and some of the the amazing deflationary properties of algorithmic progress in in SLMs, I think we will find ourselves in the world next two years where some of the earliest PCs could probably have hosted non-trivial conversations with LLM. There's just that much algorithmic progress.

Extremely bullish for Apple anyway, not to put it in in hardware terms. Um,

what do you want to do? What do you want to do next? like where do you where do you go from here?

I know it's a I know it's a

more flies or straight to mouse.

It it's such a profoundly interesting question. So Eon is beginning fundraising. Eon wants to to tackle both mice and men

and I I I think the time scale will be determined by both technical capability increases and also fundraising abilities. So I I think it's an interesting time. I I I like to say uh you know on my other pod that the singularity looks like all sci-fi tropes happening everywhere all at once.

And one of these one of these sci-fi tropes is most definitely mind uploading. And I I'm doing my best visav eon systems to push that part a little bit to the left so that the AIs don't have the Dyson swarm all to themselves.

Interesting. uh what uh architecturally what does the neural network look like that you've built for the fly? Is it is it similar to a large language model like transformer-based architecture uh or are you a beneficiary is all you need

to totally unlike the transformer

totally unlike the transformer. So transformer architecture if you subtract off the encoder and decoder layers looks can comes in many variants but the most vanilla variant looks like alternating linear uh and attention layers. So totally unlike the transformer architecture at least if you look at it it looks like more like a little bit more like a graph neural network but really that that's such an AI way of framing it. It looks like a leaky integrating fire Lif model of that that we've had for for decades from the neuroscience world. So it's it's just a graph of nodes that have

leaky entry and fire dynamics and they're firing at each other.

So what does that tell you about the nature of syn uh synthetic human intelligence? like do you think that there is a path to human level AGI that does not involve the transformer and invol instead falls along your path and and your architecture?

Oh, sure. Well, first of all, I I would argue we've had AGI for at least six years now, five to six years at at the very latest since summer since summer of 2020 when OpenAI published their language models or Fshot Learners paper. Yeah. GPT3. I would argue that was an AGI. So I I I think there are many ways my bet is there are many ways to achieve generality of intelligence. I published a paper a number of years ago arguing that not only is intelligent behavior a general process. I I went further and argued that it is a general physical process that you can even formulate intelligent behavior in pure thermodynamic terms. So my bet is intelligence is this very very general effect and lots of ways to implement it.

Interesting. Do you so do uh I mean you've mentioned AGI six years ago singularity here and you're smoothing it. Do you have other sort of binary benchmarks that you're looking towards like ASI recursive self-improvement? Are any of these terms useful to you?

Oh, we're already there. I mean, we're already er in the era of recursive self-improvement. All of the frontier labs are are pretty public about it. At this point, I'm looking past the singularity. I spend most of my time on bets for what the posts singularity world looks like. So Eon is one of those bets. Another bet, I have a company, physical super intelligence, that's trying to solve all of physics with AI and doing an amazing job. I just wrote a book with Peter called solve everything arguing that entire disciplines are going to get steamrolled by super intelligence and that what matters now is what disciplines we aim the super intelligence at.

Interesting. What

what what decisions do you make in your personal life in the way that you live

that are based around your beliefs around uh technological progress?

There are a few that come to mind. So one is I'm hoping not to die. It would be a shame if if I died sometime soon and get to miss the most exciting developments.

Same.

That's what gets the applause.

That's great. Yeah, that's a white pill.

White pill. Very good. Brian Johnson thanks me in advance. Uh so uh but I think also I mean there there are so many other angles. Another is not trying to bet against the collective intelligence of the market. I've had a number of friends who who thought they could outsmart all of the AI algo traders and day trade and they they when they inevitably fail, they attribute it to bad luck or or something other than than just fundamentally betting against progress and betting against AI intelligence. So trying not to make the mistake of betting against collective intelligence of civilization and AI capabilities. I I also I I'm I'm trying to make bets assuming that the AI capabilities keep increasing to the point where they're onlogically shocking and trying not to duplicate effort of Frontier Labs. And I I think that's perhaps something I don't see enough of in the venture community. So many new startups being formed now just aren't being ambitious enough. I I have startups that uh uh newly formed or recently formed that are literally trying to grow new islands and new coastlines with AI. That would have been unthinkable years ago. I I have that's with physical AI or or they're just dreaming them up.

Uh it's with AI reaching into the physical world to to steer ocean currents and grow new islands and new coastlines. That would not have been possible a few years ago. that's now starting to become possible. I have a startup that's working on solving interspecies communication starting with dogs.

That would not have been possible a few years ago. So I I spend a lot of my time thinking about what the posts singularity state of the world looks like and how to smooth that out and bring it here sooner.

Yeah. What does the post singularity world look like where you are able to upload yourself? I assume that feels like there's a copy of Alex in the cloud and then you still exist in not in the cloud. How do you think that that interfaces?

In in my mind, if if we wind up having the copy question come about, that's almost a failure mode. I I speaking for myself, I don't want a copy of myself. I I want myself

and Hans Moravec and others have written about this. What I would like to see in the space and what Eon is working toward uh at least aspirationally is this idea of a continuous transfer of consciousness so that it really is you. It's not a copy of you. It's not a lowfidelity faximile of you. It should be a better expanded version of you that's still you. And whether that looks like moving from what Eon's doing right now with fruit flies to maybe replacing a single cell in a brain at a time invasively or non-invasively with a substrate independent or substrate migrated implementation. Some variant of that I suspect is where all of this will go. So that you never have to worry is it really myself or not. It will by construction be yourself. And I'm I'm not sure of the precise timetable, but I I I think 5 10 years from now, I think the world will see market progress on the the problem of not just wholebrain emulation, but also transfer of human intelligence to new substrates.

That's very exciting. This is very

new sci-fi corner.

I love it. Uh

every sci-fi trope everywhere all at once.

Time travel.

Good framing.

Is time travel on the table? We don't know yet. So, it's an interesting question. We don't know whether the physics of our universe are compatible with I think what most folks would conrue as as time travel. There are there are versions of temporal non-locality that are consistent with the physics that we have right now, but not in a useful way that you would that would be worth filming a sci-fi movie over.

What about uh faster than light travel?

We don't know yet. Is same problem in some sense I exactly the same problem as time travel. We don't know whether the physics of our universe will ultimately allow that

because it seems like if we don't get FTL like there's a it's going to be a boring travel to go to another solar system.

I I will say this. I think odds are pretty good that AI in the next few years will tell us whether the the laws of physics of our universe are compatible with that or not and help us solve this. One of the reasons why I helped found physical super intelligence to to discover any of such laws.

Very cool. Well, thank you so much for taking the time to come chat with us.

Yeah, great to meet you. Congratulations on the progress.

Fascinating stuff.

Thank you. A lot of questions.

We'll talk to you soon.

Cheers.

Have a good one.

Let me tell you about Figma. No matter where your idea starts, Figma make qua code, codeex, or a sketch, the Figma canvas is where ideas connect and products take shape. Build in the right direction with Figma. And let me also tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real-time experiences and new