C (4:33)

Go this way and turn left and then straight.

A (4:35)

Okay, thank you very much. I visited her lab at MIT last fall.

B (4:40)

How are you guys?

A (4:41)

Hi, I'm Dina tupperwest.

B (4:42)

Nice to meet you.

A (4:43)

Very nice to meet you, too. The first thing I noticed when I walked into her office was the unusual decor. On a shelf above her desk sat two jars. I squinted to make sure I was seeing it right. They're just brains in jars?

B (4:58)

Yeah.

A (4:59)

Inside each jar was a brain, a human brain. And are they special kinds of brains?

B (5:05)

No, they're mostly people who donate their brains for dissection for, like, medical classes,

A (5:10)

which is one way scientists have tried to understand the brain for generations. You have to wait until someone dies to study it. Then you can slice it and map it and compare one brain to another. The problem with that, of course, is that it doesn't necessarily tell you much about how thinking works in real time. And that's the puzzle Ev has spent most of her career trying to solve.

B (5:34)

I'm a neuroscientist at MIT at the McGovern Institute for Brain Research.

A (5:40)

Though everyone on Earth has a brain, many of its secrets remain out of reach. The human brain contains about 86 billion neurons, each one connected to thousands of others. It has a quadrillion synapses. That's a 1 with 15 zeros. The result is a network so complex, scientists still don't fully understand how thinking actually works. But Nev is trying to change that. Her research focuses on how the brain processes thoughts and language, two things that people tend to assume are one and the same.

B (6:19)

It's very easy to conflate them, because that's what an effective communication System does. It allows us to share thoughts. Right.

A (6:27)

For centuries, philosophers and linguists from Plato to Noam Chomsky argued that human thought requires language. That is, our thoughts are not just communicated with words, they're actually formed by them. But Ev wasn't convinced. She had a hunch that maybe thought and language aren't the same thing at all. So she set out to test it, which means putting people inside this, A complicated, expensive MRI machine. So I'm gonna lift you up now. That's the sound of me getting prepped for an FMRI. At EVs lab, FMRIs, or functional MRIs, measure how active a specific part of the brain is.

C (7:17)

So in this task, basically, you're just gonna see some passages appear on the screen.

A (7:22)

To test if thinking actually does require language. She identified what part of the brain lights up when you hear a story read to you. Then Ev had people solve non language problems, like math equations, for example.

B (7:35)

And then we ask, do these language regions, are they active when you're doing these things and they're not active, they're basically silent. They're working about as much as when you're looking at a blank screen.

A (7:46)

After hundreds of MRIs, Ev reached a surprising conclusion, Something that overturned that long standing assumption.

B (7:54)

In humans, language and thinking are separate. They rely on distinct parts of the brain.

A (8:01)

And since the brain has a separate language network, it meant scientists could map it and maybe someday even repair it for people with language processing disorders. But this kind of research takes time, a lot of time, just to get to this initial discovery. It took 15 years, in part because studying language is hard. Most other scientists can use animals as stand ins.

B (8:27)

I've always been jealous of my colleagues who study like, vision and motor control because they have this abundance of animal models like mice, you know, some rats. It used to be a much wider

A (8:37)

array, much cheaper, much simpler. But you can't use rats to study human language because, well, they don't speak it. Then something unexpected happened. Not in neuroscience, but in technology. When we come back, Ev finds something inside these models that starts to blur a line we thought was pretty clear. The one between how humans think and how machines do. Stay with us. This show is supported by Human Rights Watch. There are more displaced people in the world than at any time since World War II. The great unrooting is a limited series that tells this epic story through the eyes of a young man from Myanmar. Where do you go when you have to flee? What do you take with you? What if they don't want you when you get there? It's a story of flight and survival, of climate change and social media, of borders and passports and hope. The great unrooting from Human Rights Watch Wherever you get your podcasts the Wired

D (9:54)

newsroom is known for award winning reporting on how technology shapes our world. On WIRED's Uncanny Valley, we take that curiosity even further. Each week, journalists from Wired break down the biggest stories in tech while speaking directly with the people building, challenging and reshaping the future. Is the AI boom sustainable? How do you protect your privacy in an age of constant surveillance? Uncanny Valley tackles the questions driving today's tech debates and lighting up your group chats. Listen to new episodes every Thursday. Wherever you get your podcasts,

A (10:28)

it is called Internet. I use the World Wide Web information superhighway.

D (10:33)

Cybersecurity.

A (10:34)

Why do things go viral? Click here. ChatGPT the revolutionary new language model developed by OpenAI. It also announced this week. ChatGPT can now quote, see, hear, and speak to you.

B (10:52)

I'm interested in how language works, and suddenly these large language models started producing language that was incredibly well formed. And so of course me and a lot of people in my group got really excited.

A (11:10)

Interacting with AI can often sound a lot like a human conversation, in tone, in flow, which raised an intriguing possibility for ev. If machines can generate conversation like ours, could they be used to study language itself and bypass all the usual complications of studying the brain? EV decided to test that. She fed a set of language tasks into a language model, and then she gave the same material to human volunteers.

B (11:41)

And then you have some metrics that allow you to compare how similar those things are. Say, you know, a few thousand sentences, and then see how these things are represented is similar.

A (11:52)

So now two systems, side by side, same input, same task. And what she found took her breath away.

B (12:01)

We and a few other groups have found that when people process sentences, those representations are actually quite similar to what you see inside LLMs.

A (12:15)

Large language models work by predicting what words should logically come next. EV says. Our minds do something surprisingly similar.

B (12:24)

So the human brain is definitely a very predictive kind of a system, always

A (12:28)

guessing, always anticipating what comes next. For brain researchers, this was a massive breakthrough, because that bottleneck we started with begins to loosen. All those limitations that used to slow Ev and her colleagues down suddenly started to fall away, because in a sense, she had found a new kind of lab rat. Except this one lived inside a computer, which meant she could test ideas that you can't test with real human brains. It was faster, cheaper, and much more precise. You could pause it, probe it, change One tiny piece, and then watch what happens. Andrea de Varda is a postdoc in EVs lab. Originally from Italy, he works with her doing that kind of digital dissection.

C (13:16)

You can also look inside the model and extract activations from individual neurons, which is something that is very hard to do in humans.

A (13:24)

Inside these systems are what researchers are calling digital neurons, not real cells, but close enough that you can study them. And unlike the neurons buzzing around in your brain or mine, these can be measured, isolated, even altered.

C (13:40)

With LLMs, you can destroy certain components that deal with certain cognitive capacities in the model.

A (13:47)

Right? So you delete one little part of your model and see if it changes language. And if it does, maybe that hypothesis is correct. And if it does nothing, maybe you need to rethink your hypothesis. Is that the idea?

C (13:59)

Yes, that is the idea.

A (14:01)

So instead of guessing what the brain is doing, you can test it directly over and over and over again. Which means scientists can do something they've never really been able to do before, run controlled experiments on the kinds of processes that usually stay hidden inside our brains. Ev and her team can now start to build a more detailed map of how the brain does what it does. But language is just the first step. The bigger question is still out there. Can these systems help us understand thinking itself? For a long time, Ev was convinced the answer to that question was no.

B (14:41)

I mean, so in a lot of early models and even a lot of current LLMs like reasoning tasks, they just clearly don't reason like humans, at least not at first.

A (14:52)

Then researchers began building something new.

B (14:56)

So there is now a class of what's called large reasoning models, which, in addition to language, are also trained, for example, on a whole bunch of math problems or other kind of reasoning problems. And they're trained in a way.

A (15:06)

These models don't just answer. They show their work. They slow down. They break problems into steps. Andrea showed me how. He rolled his chair up next to me, opened his computer, and started typing.

C (15:19)

I can ask it a very simple question. For instance, how much is 25 minus 4?

A (15:30)

A green cursor blinked on his screen, and then he hit enter.

C (15:34)

And then the model, as you see here, will generate many, many tokens to solve these questions. So in this case, for instance, for this very simple question, it generated 264 tokens, right?

A (15:45)

Tokens are like the model thinking out loud, one tiny step at a time, line by line by line. In this case, 264 steps to solve a simple subtraction problem. Then Andrea tried something harder.

C (16:00)

How much is 24 divided by 7 plus 32 multiplied by 9.

A (16:13)

This time there's a long pause, as if the computer is thinking harder.

C (16:17)

Then the model would generate way more tokens to Right.

A (16:20)

So it filled the entire screen with lines of code.

C (16:25)

Exactly. Exactly. And this is the internal thinking of the model.

A (16:35)

Messy, step by step. Slower when it's harder. What Ev and her colleagues discovered was striking problems humans solve quickly. Machines solve quickly, too. And harder problems take longer for both. Which raises a bigger possibility. Maybe what Ev was discovering wasn't just about language anymore. Maybe it was about thinking. In the same way. Large language models can be used to study language. Large reasoning models may be a way to study human thought. A place where you can test ideas about how we reason without waiting years, without needing a human subject at all. Which opens up an entirely new laboratory for neuroscience and new possibilities for brain research that could answer questions like how humans solve problems, how we learn what intelligence actually is. And some of those answers might surprise us. But the implications don't stop there. Because the same kinds of models scientists are now using to understand the brain are also being used to understand us. What we want, what we feel, what we buy, how we decide, what we'll do next. It doesn't just observe us. It can anticipate us. For centuries, the brain has been the most mysterious machine we know. Closed off, hard to access, impossible to experiment on directly. Now we may have built another one, one not made of cells, but of zeros and ones. A version you can finally examine from the inside. And by studying it, we're not just learning how machines work. We're learning the patterns underpinning our own thinking. What makes us predictable, what makes us human. And maybe the gap between how we think and how machines do is smaller than we imagined. Erica Gaeda and Sean Powers Field produced the story. This is Click Here. We'll be back on Tuesday. Have a great weekend. Click Here is a production of Recorded Future News and prx. Today's show was written and produced by Megan Dietre, Sean Powers, Erica Gajda, Zach Hirsch and Casey Giorgi. It was edited by Karen Duffin and Sarah Covedo and fact checked by Darren Ancrum. Original music is by Ben Levingston with additional music from Blue Dot sessions. Our staff writer is Lucas Riley. Our illustrator is Megan Goff, and our sound designers and engineers are Jake Cook and Jesse Niswonger. Find us on X or Facebook at Click here. Show or leave us a voice message at 661-5ch. Talk. Sometimes we'll turn those moments into reporting, sometimes into a conversation, and sometimes into a future story you'll hear on this show. I'm Dina Temple Raston, and thanks for listening.

B (20:12)

Support for this program comes from Recorded Future. In cybersecurity, the biggest risk isn't what can be seen, it's what gets missed. Recorded Future analyzes billions of signals to help organizations stay ahead of threats. Recorded Future Know what matters? Act first.

E (20:27)

Looking for more of the cybersecurity and intelligence coverage you get on? Click here. Then check out our sister publication, the Record from Recorded Future News. You'll get breaking cyber news from reporters in New York, Washington, London, and Kyiv, among others. And you'll see for yourself why it attracts hundreds of thousands of page views every month. Just go to the Record Media.