MC

Michael Cole

Things Michael Says on Podcasts

Where to Find Them

Michael Cole has been a guest on Brain Inspired .

Recently: “BI 249 Michael Cole: Brain Flows are the Units of Cognition” on Brain Inspired (November 2026).

What They Said

“Reinforcement learning is at the core of these large language models, because they do reinforcement learning from human feedback. … They try to please you too much. They won't tell you if you're wrong or something.” — Michael Cole, Brain Inspired

The host asked whether AI's lack of grounding makes its achievements more or less impressive. The guest says models do get some value shaping from human feedback, but calls that top layer very simple, and gives sycophancy as the result.

Brain Inspired · 2026-11-04 Permalink → Listen →
Brain Inspired Around 1:30:56 into the episode
Paul Middlebrooks

more or less impressive, like what these models have been

Speaker 4

able to do, right?

Michael Cole

Yeah, I don't know. I mean, like, to be fair, there's some grounding and at the upper end. Like, one thing that I think isn't always talked about is how reinforcement learning is in the core of these large language models because they'll do this, you know, reinforcement learning from human feedback. And so you're kind of like training just like we train our own brains relative to our values, like what to represent and what's prominent and so forth. So there is a kind of top that's kind of generating these things. But it's very simple. There's this issue of like, I'm going to say this where they have anyway, they try to please you too much. Like, they won't tell you if you're wrong or something. These AI models are because they're just trained. Yeah, that's.

Paul Middlebrooks

Was that the word? Yeah. Yeah,

Michael Cole

I think that's it. They're so that they're just, their training objective is just to like make you happy in the moment, basically. But

Paul Middlebrooks

that, but that comes externally, right? Where our values are internally generated.

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“We literally have a study from a couple of years back showing that you can generate face selectivity in the fusiform face area from other regions, just based on the flows to the face area. So the computation isn't inside. It's probably some, but most of the computation is from other regions.” — Michael Cole, Brain Inspired

Near the end, the host asked how a flow-based view differs from how neuroscience has mostly been done. The guest says the usual assumption is that rising activity in a region reflects a process inside it, and gives one of the brain's best-localized functions as his counterexample.

Brain Inspired · 2026-11-04 Permalink → Listen →
Brain Inspired Around 1:35:43 into the episode
Michael Cole

…brain activity flows and how useful I think that'll be. It would be if we shifted so it's a more generative perspective. It's about adding context to the activity patterns that we're analyzing and the connectivity and combining them to think about how the activity is generated through these flows.

Paul Middlebrooks

And how is that different than the way neuroscience has been done thus far?

Michael Cole

So one kind of unspoken assumption in a lot of cognitive neuroscience has been if you see an activity in a brain region or a neural population go up, say for a given task condition, there's some unknown within region process. Like if it's the fusiform face area, there's face processing going on. But we literally have a study from a few years or a couple years back showing that you can generate face selectivity in the fusiform face area from other regions just based on the flows to the face area. So the computation isn't inside. It's probably some, but most of the computation is from other regions and the selection of the connections and their weights select the features that generate the face selectivity. That's just indicative of like one of the most well-localized functions in the human brain just over two decades of work showing this is actually a distributed process. Right. But it's also a

Paul Middlebrooks

fundamentally different way of looking at it that it's the flow that's doing the processing. The function comes from the flow itself, not from within, not solely from within the area process. Yeah,

Michael Cole

there's some recurrent processing, yeah, but yeah, we didn't even model the recurrent processing and it generated most of the variants on face selectivity. That's cool. And so like there's still a localized function there, but it's it's a distributed process to put it there. And then, you know, the…

Paul Middlebrooks

All right. That's a great place to end it. Congrats again on the book. I enjoyed it. It was my main reading material on those long flights to and from Geneva. So I love reading on planes because you can really sit down and focus, you know, strapping yourself to the mast. Yeah, there's something…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
“Current AI is kind of untethered relative to how our brains work, in the sense that they don't have a proper global flow hierarchy. Our global flow hierarchy is grounded at the bottom by perception-action cycles. We're embodied, and that's important for preventing what's called hallucinations. And at the top we have core values that shape our goal pursuits.” — Michael Cole, Brain Inspired

The host asked how the brain-flow account changes the way he thinks about artificial intelligence. The guest argues that today's AI lacks both a bodily grounding at the bottom and internalized values at the top, which he sees as the two things that define the human system.

Brain Inspired · 2026-11-04 Permalink → Listen →
Brain Inspired Around 1:29:08 into the episode
Michael Cole

Yeah, the sweet spot. Yeah.

Paul Middlebrooks

…like consciousness, intelligence, et cetera, how does this make you think differently, if at all? Or has this reoriented, oriented how you think about artificial intelligence and what it is, its nature relative to our very human way of performing these things or implementing these things? Yeah,

Michael Cole

so I could bring it. So that's, I think, chapter seven is on intelligence. I included natural and artificial intelligence in that chapter. And the way I've been thinking about artificial intelligence in this framework is that current AI is kind of untethered relative to how our brains work in the sense that they don't have a proper global flow hierarchy. Like our global flow hierarchy is grounded at the bottom by perception action cycles. We're embodied. And that's important for preventing what's called hallucinations, just like kind of, you know, not enough grounding in reality. And then on the other side, at the top of the global flow hierarchy, we have core values that shape our goal pursuits. And so, you know, we'll have, I don't know, a conscience essentially, right? We want to don't want to go against some of these core values when even if we have conflict internally, right? They don't have that. So they don't have sides of the pole. Exactly. They're just like the middle part that's like good, basically. And there's attempts, right? They have, what's it called, constitutional AI that tries, I think, you know, to get the upper end of…

Paul Middlebrooks

more or less impressive, like what these models have been

Speaker 4

able to do, right?

Michael Cole

Yeah, I don't know. I mean, like, to be fair, there's some grounding and at the upper end. Like, one thing that I think isn't always talked about is how reinforcement learning is in the core of these large language models because they'll do this, you know, reinforcement learning from human…

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

Collections They Appear In