David Sussillo, a computational neuroscientist who has spent almost two decades working at Google Brain and Meta Reality Labs, along with an academic appointment at Stanford University, has had a unique vantage point to view the changing relationship between neuroscience and artificial intelligence (AI). Sussillo talked to The Transmitter about the two fields’ historical roots, how their relationship has changed, where the greatest opportunities lie for neuroscientists, and what each field might gain from the other.
This interview has been edited for length and clarity.
The Transmitter: Neuroscience and AI have a long history. Given the rapid advances in AI, how is that changing?
David Sussillo: AI arose from early studies of the brain. Warren McCulloch and Walter Pitts, way back in the mid-1940s, basically came up with the idea of an artificial neuron from studying and then abstracting the electrical properties of biological neurons. Neuroscience really did originally provide the intuition and some of the guiding principles for AI. But that’s no longer true today.
Somewhere along the line, interspersed through the ’80s, ’90s and 2000s, AI made some real jumps. They took the connectionist idea from neuroscience—that it’s the connections between neurons that create intelligence rather than just the intrinsic properties of those neurons—and just started running with it. For decades, there was a common ground: Let’s try to build artificial neural network models and then try to understand them. This benefited both the fields of neuroscience and AI.
But it wasn’t until about 2012 that artificial neural network technology started working. So the very science-y questions about the function of these networks became especially interesting and relevant.
TT: Can neuroscientists help us understand AI?
DS: That promise is why they hired me at Google Brain in 2014: because I made the argument that if this technology functions, you’re going to want to understand it. I argued that neuroscientists in particular are well suited to understand it, because we’ve been trying to understand brains for a very long time. And to the extent that you think artificial neural networks and brains have anything in common, one type of understanding may influence the other.
However, I don’t think the neuroscience program has been overly successful in understanding AI, for a number of reasons. The first is that it’s still early days. Neuroscience hasn’t solved the brain, either, so our ideas and theories are nascent and still developing. Further, AI technologies can change very quickly; they are driven as much by engineering principles as by scientific ones. It’s very hard for science to keep up with the industrialization of neural network technologies. Finally, large AI systems likely require a number of disciplines to fully understand their function. But neuroscientists and people from other technical disciplines can absolutely get hired to do this kind of work. It’s a profitable use of time, and I think people should continue to try it.
Honestly, it’s not crazy to think that we’ll have deep theories of AI before we do of the brain, simply from a consideration of experimental limitations. We don’t have access to the whole machine in neuroscience. These limitations don’t exist in AI—they have the entire machine right there. You can run dream experiments, and you can run them quickly.
With large language models, we’re moving past deep learning—that single-pass, put-in-an-input, get-out-an-output model. Now these AI systems have so-called chain-of-thought reasoning. They think in language. Now you can ask much more cognitive-science-esque questions, which relate to what in machine learning or AI would be called the network’s inference mode. The thing is already trained up, and now you just want to understand how it’s thinking. There are now new and interesting opportunities for neuroscientists and cognitive neuroscientists that weren’t present when I was first getting into the game.
TT: So is this now more the domain of cognitive neuroscientists?
DS: It certainly gives them a huge opportunity. The kind of testing and behavioral paradigms that are interesting to think about with humans have now become interesting to think about with these AI models. It’s still interesting and hard to think about how the neural network does its thing, but now you’re actually studying how it’s thinking.
This is just such a large-scoped problem. The statistical physicists are there; the mathematicians are there; the neuroscientists are there; the cognitive neuroscientists are there. And, of course, the AI scientists and engineers are there. It really is a large enough problem that all these disciplines can and should help. It evokes the cliché image of everyone touching the elephant and thinking they’re seeing something different.
It’s pretty dope, if I’m being honest. We’re not just talking about neural networks and deep learning anymore; we’re talking about real AI systems. The opportunity for neuroscientists of all stripes to get involved is probably larger.
TT: Neuroscience trainees today are more interested in getting into the AI field. More and more graduate students are doing internships at AI companies, and many go into industry after earning their Ph.D. Should there be changes in the core graduate education itself to address that?
DS: Here I have to be careful, because I don’t know what the very latest curriculum for educating neuroscience Ph.D.s is, but I definitely have opinions, as I’ve done a lot of hiring. I’m a huge believer that these internships are a good idea. Neuroscience has historically been an extremely academic field, largely because outside of molecular biology, you really couldn’t make any money doing it. That may be changing. But for all of that, allowing these students to go sample what a computational neuroscience program enables them to do is hugely valuable—not least just to sample the culture of industrial research. These internships also de-risk the candidate for the company, which makes these companies far more likely to hire from nontraditional fields. It’s a win-win.
The other thing—this feels like such a dated comment, and it should have been solved 10 years ago—MATLAB systems drive the experimental rigs in a lot of university neuroscience setups, so you end up with all these people who are wizards at MATLAB, and that stuff isn’t used [in industry]. It really holds some students back once they graduate, when everything is done in Python. It’s a bread-and-butter thing, and it should be obvious to everyone, but you’d be surprised how it still isn’t completely solved, from an educational perspective.
TT: There’s more and more funding from nongovernment sources for NeuroAI, from both companies and foundations, some funded by crypto wealth. What does that mean for the field, that these private funders might have more control over the direction of NeuroAI research?
DS: The obvious answer, in a realpolitik way, is that you take the money that’s available, assuming the strings aren’t too attached.
Corporations will often give money in these directions, but in my experience it’s always one-offs or very short programs, because things in corporations change very quickly. If the time horizon of academia is 10 years, the time horizon in a corporation is 3 to 5 years. So you couldn’t build a lab around it—not in the world I’ve seen, and I’ve been at Meta Reality Labs and at Google. But it’s money, and if it comes your way, God bless.
With respect to the crypto billionaires, I’m still not sure what to make of that, though I’ve had extended conversations with a few of them. What’s happening at Astera seems like a win-win for the community. They’ve got a rock star—Doris Tsao—leading it. Everything sounds amazing. So, to the degree that it gets reproduced, it seems like there’s a there there.
My experience is that there’s an education process you probably have to go through—these senior crypto billionaires don’t really understand what’s feasible—but they really want to learn! There’s a lot of translation, like, “Well, I want to upload my mind.” The real task is to be able to translate what may sound like science fiction to somebody in the know about neuroscience. Can you back-propagate that into meaningful neuroscience research in 2026 that’s in line with the objectives of that person and still true to the science? That’s the challenge, and I’m guessing a lot of it is doable.
The other component here, to embrace some of the sci-fi for a minute: It’s not at all crazy that neuroscience—even computational neuroscience—becomes more of a commercial endeavor in the near future. Between the brain-machine interface laboratories and what you might broadly call AI alignment strategies, it’s not crazy to think that neuroscience-based strategies and neuroscientific data could become highly meaningful in these nascent fields. I’m the first to admit I’m speculating here—grossly speculating, in fact—but it’s not crazy, either.
TT: Do you think there are new lessons for AI to learn from neuroscience?
DS: This entire question hangs on timescales. I’m very down on the idea that we’re somehow going to learn something from brains that’s going to change the course of AI in the near future. My skepticism comes from thinking about the limitations of experimental neuroscience—skulls are essentially biological safes; one can record from only a miniscule fraction of the neurons, that kind of thing. I guess the thing that I think would do it is to build a big foundation model of high-quality neural data. You go make a virtual brain—of a mouse or something—and then you’re on parity with one of these AI models. You can interrogate it quickly and run dream experiments. You can really learn something new. But if you don’t do that or some similar moonshot, what’s going to happen is AI will just hire the neuroscientists, cognitive neuroscientists and statistical physicists, and they’ll become AI people.
TT: Do you think the brain will give some insight into developing AI systems that are either more energy efficient or data efficient?
DS: Again, it depends on the timescale. If you say 10 or 20 years out, then I’m open to all of that, absolutely. In the next five years, it seems very unlikely. Take Naveen Rao, a neuroscientist from Brown University. He started a company working on dynamical systems for energy savings in deep-learning systems. That’s neuroscience inspired, but he’s smart enough to know you don’t want to reproduce the wheel exactly. He knows you can’t just reproduce neuroscience—you’ve got to change things up for the problem at hand.
On a longer timescale, I would say I’m effectively certain that neuroscience is going to have a couple more tricks up its sleeve and reveal some things that will be super important. For example, one thing we know nothing about is whole-brain dynamics. In the “let’s understand this little circuit, and then this little circuit” approach, systems neuroscience has made some progress, but we have no idea how any of it works together or coordinates. That’s what you could get out of one of these huge brain-simulation foundation models. That would be pretty cool. So there’s an irony that in the short term, AI will almost surely be helping neuroscience, while in the long term, neuroscience may still unveil some secrets.
