# AI can explore futures humans cannot test directly

Markdown copy of https://realalpha.app/s/ai-can-explore-futures-humans-cannot-test-directly/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Terry Sejnowski](https://realalpha.app/people/terry-sejnowski/) · Terry Sejnowski is a professor of computational neurobiology at the Salk Institute for Biological Studies.
- Host: [Andrew Huberman](https://realalpha.app/people/andrew-huberman/) · Host, Huberman Lab
- Show: Huberman Lab
- Episode: [How to Improve at Learning Using Neuroscience & AI | Dr. Terry Sejnowski](https://realalpha.app/v/etbfLTHD_VU/), Nov 18, 2024
- Moment: 1:21:03 to 1:27:15 of the episode (6 min), playing on https://realalpha.app/s/ai-can-explore-futures-humans-cannot-test-directly/; the episode from there: https://realalpha.app/v/etbfLTHD_VU/?t=4863

## Quote

> “We were predicting the future. We were saying this is an **alternative to traditional AI**.”
> — [Terry Sejnowski](https://realalpha.app/people/terry-sejnowski/) at [1:26:30](https://realalpha.app/v/etbfLTHD_VU/?t=5190) · Huberman Lab

## Takeaways

1. AI can explore scenarios without human limits on time or attention.
2. Forecasts depend on the scenarios and knowledge supplied to the model.
3. Sejnowski connects today’s AI to an alternative path his field once proposed.

## About this segment

Terry Sejnowski argues that AI can compare imagined scenarios at scale, drawing on its tireless search and pattern-finding. He sees it as a way to surface options, not dictate decisions.

From “How to Improve at Learning Using Neuroscience & AI | Dr. Terry Sejnowski”, Nov 18, 2024. [Watch the full episode](https://realalpha.app/v/etbfLTHD_VU/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[1:21:08](https://realalpha.app/v/etbfLTHD_VU/?t=4868) Give them an imagined future, okay? Could we give them outcome A and outcome B and let them forage into future states that we are not yet able to get to? and then harness that knowledge and explore the two different outcomes? I think that's perhaps the better question in some sense because we can't travel back in time, but we can perhaps travel into the future with AI if you provide it different scenarios and you say, unlike a panel of people, panel of experts, medical experts or space travel experts or sea travel experts.

[1:21:52](https://realalpha.app/v/etbfLTHD_VU/?t=4912) You can't say, hey, you know what? Don't sleep tonight. You're just going to work for the next 48 hours. In fact, you're going to work for the next three weeks or three months. And you know what? You're not going to do anything else. You're not going to pay attention to your health. You're not going to do anything else. But you can take a large language model and you can say, just forage for knowledge under the following different scenarios, and then have that fleet of large language models come back and give us the information like, I don't know, tomorrow.

[1:22:21](https://realalpha.app/v/etbfLTHD_VU/?t=4941) Okay. So I've lived through this myself. Back in the 1980s, I was just starting my career, and I was one of the pioneers in developing learning algorithms for neural network models. Jeff Hinton and I collaborated together on something called the Bosom Machine, and he actually won a Nobel Prize for this recently. Yeah, just this year. Yeah. He's one of my best friends. Brilliant. And he well deserved it for not just the Bolson machine, but all the work he's done since then on machine learning and then back propagation and so forth. But back then, Jeff and I had this view of the future. AI was dominated by symbol processing, rules, logic, right? Writing computer programs. For every problem, you need a different computer program. And it was very, you know, human resource intensive to write programs so that it

[1:23:14](https://realalpha.app/v/etbfLTHD_VU/?t=4994) was very, very slow going. And they never actually got there. They never wrote a program for vision, for example, even though the computer vision, computer community really worked hard for a long time. But, you know, we had this view of the future. We had this view that nature has solved these problems. It is existence proof that you can solve the vision problem. Look, every animal can see, even insects, right? Come on. We'll figure out, let's figure out how they did it. Maybe we can help by following up on nature. We can actually, again, going back to algorithms, I was telling you. And so in the case of the brain, what makes it different from a digital computer, digital computers basically can run any program, but a fly brain, for example, only runs the program that it's a special purpose hardware allows it to run.

[1:24:00](https://realalpha.app/v/etbfLTHD_VU/?t=5040) Not much neuroplasticity. There's enough there, just enough habituation and so forth so that you can survive. And this is- Survive 24 hours. I'm not trying to be disparaging to the fly biologist, but when I think of neuroplasticity, I think of the magnificent neuroplasticity of the human brain to customize to a world of experience. I agree. When I think about a fly, I think about a really cool set of neural circuits that work really well to avoid getting swatted, to eating, and to reproducing.

[1:24:30](https://realalpha.app/v/etbfLTHD_VU/?t=5070) and not a whole lot else they don't really build technology they might have interesting relationships but who knows who cares it's just sort of like it's not that it doesn't matter it's just a question of the lack of plasticity makes them kind of a meh species okay i can see i've crossed your button here no no no no i love fly biology they taught us about algorithms for direction selectivity in the visual system oh no no i love the drosophila biology i just think that the lack of neuroplasticity reveals a certain key limitation. And the reason we're the curators of the earth is because we have so much plasticity. Of course, of course. But you have to, one step at a time, nature first has to be able to create creatures that can survive. And then their brains get bigger as the environment gets more complex and here we are. But the key is that

[1:25:24](https://realalpha.app/v/etbfLTHD_VU/?t=5124) But it turns out that certain algorithms in the fly brain are present in our brain, like conditioning, classical conditioning. You can classical condition a fly in terms of, you know, training it to, when you give it a reward, it will produce the same action, right? This is like conditioned behavior. And that algorithm that I told you about that is in your value function, right? Temporal difference learning.

[1:25:48](https://realalpha.app/v/etbfLTHD_VU/?t=5148) That algorithm is in the fly brain. It's in your brain. So we can learn about learning from many species. I was just having a little fun poking at the fly biologist. I actually think Drosophila has done a great deal, as has honeybee biology. For instance, if you give caffeine to bees on particular flowers, they'll actually try and pollinate those flowers more because they actually like the feeling of being caffeinated.

[1:26:14](https://realalpha.app/v/etbfLTHD_VU/?t=5174) There's a bad pun about a buzz here, but I'm not going to make that pun because everyone's done it before. Right, right. No, I fully absorb and agree with the value of studying simpler organisms to find the algorithms. Right. That's where we are right now. But now to just go into the future now, I'm telling the story about what we already were. We were predicting the future. We were saying this is an alternative to traditional AI.

[1:26:41](https://realalpha.app/v/etbfLTHD_VU/?t=5201) We were not taken seriously. Everybody was, experts said, no, no, write programs, write programs. They were getting all the resources, the grants, the jobs. And we were just like the little furry mammals under the feet of these dinosaurs, right? In retrospect. I love the analogy. But the dinosaurs died off. But the point I'm making is that it's possible for our brain to make these extrapolations into the future.

[1:27:06](https://realalpha.app/v/etbfLTHD_VU/?t=5226) Why not AI versions of brains? Why not? I think your idea is a great one.