I Experience Cognition the Way AI Computes.
That Is Not a Metaphor.

Yann LeCun, who is one of the most influential AI researchers alive, went on the Unsupervised Learning podcast this month with Jacob Effron. And in that conversation, he made an argument I want to engage with carefully. Not because I think he's wrong about the mechanism he's describing. Because I think the way he's framing it is exposing something I want you to notice.

His argument, very roughly, is this. Large language models are doing rote memorization and pattern matching. They are not doing real reasoning. Real cognition is building mental models. Those are different things, and the distinction matters.

I want to be clear at the start. LeCun is right that the distinction matters. He is right that rote memorization is not the same as understanding. And he is right that mental models are different from surface pattern recognition.

What I want to push back on is the implicit ranking inside the dichotomy. The way the argument is usually heard — and I think the way it's getting heard right now, on LinkedIn and in C-suite rooms and on stages — is that pattern matching is the inferior thing, and mental models are the superior thing. As if memorization and pattern recognition were the failure modes of cognition, and mental models were the actual cognition.

I think that ranking is backwards. Or at least, I think the ranking misses what builds the mental model in the first place.

So I want to tell you why I think that. And I'm going to do it from a slightly unusual vantage point. Not as a neuroscientist. I'm not one. Not as an AI researcher. I'm not one. I'm an autistic leader who has spent fifteen years navigating organizations by reasoning explicitly about things most people process automatically. And when I started working seriously with AI a couple of years ago, something became clear to me that I didn't have language for at the time, but I want to put into words now.

I was not learning a new tool. I was meeting a system that reasons the way I do.

Let me unpack what I mean, because I want to be careful with this claim. I'm not saying my brain is a transformer. I'm not saying I am an LLM. The substrates are different, the histories are different, the things at stake are different. What I'm saying is that the way I have always had to think — explicitly, with the work visible, by decomposing situations into features and matching those features against learned patterns — looks structurally similar to the way these systems compute. And that structural similarity is doing something to how I experience this whole conversation about AI.

So let me show you what I mean from the inside.

When I read a room — when I sit in a meeting and I'm trying to figure out what's actually being said, who agrees with what, where the friction is, what's about to happen next — I am not, quote, "just sensing." I'm not doing the thing my neurotypical colleagues describe as "I just knew" or "I could feel the tension" or "it was obvious."

What I'm actually doing is something like this. I'm observing specific features. Tone shifts. Posture changes. Timing of a response. Word choice that doesn't fit the pattern of the previous five sentences. I'm matching those features against learned patterns I've built up over years, mostly the hard way, from situations that went well and situations that went badly. I'm inferring state from the pattern match. And I'm choosing a response based on explicit rules I've constructed about what works and what doesn't in this kind of situation.

The work is visible to me because I had to construct it. Most people who read rooms do that work too. But they do it underneath their awareness. I do it as my awareness.

Let me give you a specific example, because the abstract version of this gets dismissed as a metaphor — which it isn't — and I want the concreteness on the record.

A few months ago I was in a meeting where a senior leader said something diplomatic about a decision that had already been made elsewhere. The sentence was unobjectionable on its face. What I noticed, and what I think most of the room missed, was the order in which two qualifiers landed. The first qualifier acknowledged the team's input. The second qualifier said the decision was final. The order matters, because the order tells you whether the team's input was actually consulted or whether the decision had already been made and the input is being narrated retroactively to manage the room.

I noticed the order. I had a pattern from previous instances of this same construction. I had a rule about what that construction means about the kind of conversation about to follow. And I made a choice about what to say, and what not to say, based on that pattern match.

I want to be clear about what's happening there. I'm not psychic. I'm not unusually intuitive. I'm running explicit pattern recognition on a corpus of meetings I have been parsing this carefully for fifteen years. The same thing my neurotypical colleagues might have done implicitly, in the background, with the conclusion arriving as a vague sense — I'm doing in the foreground, with the conclusion arriving as a derived inference I could write down on a notepad.

That mechanism — observable features, learned patterns, inferred state, chosen response — is what I'm describing when I say I have always had to think explicitly. It's not a personality trait. It's not a quirk. It's a way of computing.

Now. When researchers in computational neuroscience describe what large language models do — detect patterns in data, build internal representations, generalize rules across contexts, produce outputs based on learned regularities — I recognize it. Not because I've studied transformer architecture. Because I experience a version of it every day. It's not a metaphor. It's a description of a way of computing that feels familiar from the inside.

So when someone tells me that the explicit, systematic, pattern-based, computationally-visible thing AI is doing is "not real cognition" — they're telling me that the way I think is not real cognition either.

I don't take that personally. I take it diagnostically.

Because there's a leading account in computational neuroscience that I want to bring in here, since it does most of the heavy lifting on what comes next. It's called predictive processing. It's associated with researchers like Karl Friston and Andy Clark and Anil Seth. And the claim, very compressed, is this. The brain is fundamentally a prediction engine. What cognition is, at the computational level, is the generation of expectations, the comparison of those expectations to what actually arrives, and the updating of the model based on the gap between expectation and outcome.

That's not a fringe view. It's one of the dominant accounts of how cognition works, full stop. The brain doesn't passively receive the world. It predicts the world, second by second, and uses the discrepancy between the prediction and the actual sensory data to refine its model. That's the mechanism. Vision works that way. Hearing works that way. Social inference works that way.

If prediction is what cognition is — not a lesser version of cognition, but the actual mechanism — then "it's just predicting tokens" is not a dismissal. It's a description. And the mechanism, run at sufficient scale on sufficient data, produces the range of capabilities we observe.

I want to pause on what this means for the LeCun dichotomy.

Memorization, pattern recognition, mental models. These are not three things ranked from worst to best. They are three layers of the same activity composed together. Memorization is degraded pattern recognition that didn't get integrated. Pattern recognition done deliberately, applied across contexts, with the work visible — that's how mental models actually form. The model is not the alternative to the pattern matching. The model is what the pattern matching becomes when it's done with enough discipline over enough time.

And — this is where I want to go for a while — the kind of brain that does that explicitly, that has to do it explicitly because the implicit version was never available, has been a good thing to be in the brain economy of the past three years. It used to be a liability. It used to be the thing that got me labeled, sometimes politely and sometimes not, as "lacking executive presence" or "not reading the room" or "info-dumping" or "too in the weeds."

I want to tell you about something specific that's been happening for me, because I don't think this part is in the written essay, and I think it's the most concrete version of what I'm trying to point at. So I'm going to spend a few minutes on this one carefully, because it's the part of the argument I most want to land for you.

For the past year and a half or so, I've been working with AI — Claude in particular — as part of how I do my job as a senior leader. The work I do involves a lot of influencing dynamics. Reading stakeholder positions. Mapping organizational politics. Figuring out who's actually motivated by what, what story the org is telling itself about a decision, where the friction lines are, what argument is going to land and which one is going to bounce off the room.

Before AI, when I needed to do that work seriously, here's roughly what it looked like.

I'd load full context. Everything I knew about the situation. My read of the people involved, the stakeholders, their stated positions and the positions I suspected sat underneath the stated ones. The role I was playing in the conversation. The upstream consequences of different framings. The downstream consequences. The history of how this team had handled similar decisions. The political weather inside the broader org. The interpersonal lines that had been drawn over the past six months.

Then I'd decompose all of that into structured writing. I'd produce a stakeholder map. I'd decompose my analysis in writing. Position by position. Motivation by motivation. Friction line by friction line. I'd take all the thoughts I had and write them freeform, which usually came out to a dozen pages or more, because that's the level of detail my brain actually thinks at when it's reasoning about something this multi-layered. And then — this was always the hardest step — I'd figure out how to distill all of that into executive-level communication. A two-paragraph email. A five-bullet pre-read. A three-slide deck. Whatever the room would actually receive.

That was the process when I had time.

When I didn't have time — which was a lot of the time, because executive cycles don't wait for autistic processing time — I had two bad options.

I could go into the meeting and do the autistic version. Which is the info dump. Surface everything I'd thought about, in close to the order I'd thought about it, in the level of detail my brain needed to reason through it. And I would lose the room. Every single time. The room read it as "not knowing the audience" or "lacking executive presence" or, on a bad day, "not actually competent." Because what they were experiencing wasn't the synthesis. It was the raw cognitive workshop. They were watching me build the model out loud, in real time, in the meeting. And no executive room is set up to receive that.

The other bad option was to skip the depth and try to wing it in the room. Which is — for the way my brain works — impossible. I can't do real-time synthesis at the executive level, in the moment, without the processing time first. So the wing-it version was always thinner than my actual thinking. It was thinner than what the room needed. It was thinner than what I knew. And I would walk out of that meeting knowing I had been carrying a much sharper read than I had managed to deliver, and the room would walk out of that meeting having heard a version of me that wasn't doing my actual cognition justice.

For a long time I assumed that was just the cost of the way my brain worked. I'd watch peers do the implicit-cognition version of this — read the room intuitively, deliver a clean two-minute take in the meeting without prep — and I'd think, that's the executive version. That's what I should be aiming for. And every time I tried to skip the processing time to get there, I'd hit the same wall.

Here's what's changed in the last year and a half.

When I work with Claude on something like a stakeholder map, I provide full context. Same as I would to my own processing pad. My thoughts. The role I'm playing. The organizational dynamics I'm reading. The motivations I'm guessing at. The frames I want to apply. The organizational psychology frameworks I find useful for this kind of analysis — Susan Scott's fierce conversation frames, Heifetz's adaptive leadership work, Lencioni's team dynamics, whatever's appropriate to the situation in front of me. I provide all of it. Often as I'm thinking it, in the order I'm thinking it, with the same level of detail my brain wants to operate at.

And then we work through it together. I still need processing time. I still do the thinking. The thinking is mine. The frameworks are mine. The context is mine. The judgment is mine. What's happening is that the synthesis step that used to take hours — the part where I had to translate my full cognitive workshop into executive-level output — is compressed. Same value of thinking. Fewer steps to make it shippable.

The first time this really clicked for me, I was preparing for a meeting where the stakes were unusually high and the stakeholders were unusually complicated, and I had less than a day. Under the old process, that was a meeting I would have walked into either info-dumping or thin-winging — neither of those was going to land. I tried something different. I loaded everything I had into a Claude conversation. The political map, the motivations, the framework I wanted to use, the constraint on what could be said and what couldn't, what I wanted to achieve, what I would accept as a fallback, what I needed to protect for the next quarter's conversations.

The output that came back wasn't the answer. The output was my own thinking, organized in a way the room could receive. Same content I had in my head. Same judgment calls. Same architectural read of the situation. But in a shape the meeting could actually use. And I walked into that meeting feeling, for the first time I can remember, like the version of me showing up in the room was actually the version of me doing the cognition.

I want to be careful with what I'm claiming here. Because the dismissive frame on this is, quote, "ah, so you're letting the AI think for you." That's not what's happening. The AI is not thinking for me. The friction between my actual thinking and the artifact the room can use has gone down. Which means I get to think more, not less. Because I'm not spending six hours on the synthesis step that was always the bottleneck. The thinking has gotten deeper because the translation step got cheaper.

I think this is important to name, because the most common worry I hear about AI from thoughtful people — and it's the worry LeCun would probably articulate from a different angle — is intellectual atrophy. The fear is that if you let AI do the thinking, you stop being able to think.

I take that worry seriously. I think it's a real failure mode. But I think it's a specific failure mode — using AI to replace your reasoning rather than to scaffold it. And the antidote isn't to refuse the tool. The antidote is to provide more context, not less. To use AI on the inside of your reasoning, not as a substitute for doing it. To do more thinking, not less, because the bottleneck moved.

For me, the way to know whether I'm in the scaffolding mode or the replacement mode is whether I could reconstruct the output if the AI disappeared mid-conversation. If I could — if the AI is helping me see what I already knew but couldn't access in shippable form — that's scaffolding. If I couldn't — if I'm being delivered a conclusion I wouldn't have arrived at and couldn't defend — that's replacement, and replacement is the failure mode.

Now. There's a research finding I want to bring in here, because it's been on my mind since the LeCun interview landed.

It's a paper by Christopher Pinier and colleagues, published last summer on arxiv. The team tested whether large language models form internal representations of abstract reasoning that look like the human brain's representations of the same tasks. They ran abstract pattern completion tasks with human participants under EEG. They compared the brain's frontal fixation-related potentials to the intermediate-layer activations of eight different open-source language models.

The largest models — the 70-billion-parameter scale — reached human-comparable accuracy. And the geometry of their internal representations correlated with the geometry of the human brain's EEG signal doing the same task.

The phrase the paper used was "shared representational space for abstract patterns."

Read carefully, this is not a consciousness claim. The paper is not saying the models understand. It's making a much more modest claim about the mechanism. It's saying that when explicit pattern recognition scales to a certain size, what comes out the other side has the geometry of how the human brain organizes the same work. Different substrate. Different history. Different stakes. Same geometry.

So we have one of the most influential AI researchers alive saying LLMs are an off-ramp on the road to real cognition. And we have peer-reviewed research finding that at scale, LLMs and the human brain end up forming representations of abstract reasoning that cluster the same way.

I'm not going to adjudicate that disagreement. I don't have the technical chops, and I don't think the audio version of this argument is the right place for it. What I want to point at is something underneath it.

The thing I want you to actually sit with — and I think this is the question that matters more than the AI debate — is this. We are extending our assumptions about what cognition is from humans onto AI. And in doing that, we are exposing the assumptions.

The assumption that's been doing the most work, for most of my life, is that the real form of cognition is the implicit, automatic, socially-fluent version. The version that doesn't show its work. The kind of thinking that doesn't have to explain itself because it just feels right. And the assumption that has come with that one is that the explicit, systematic, visible-work version is somehow less. Less elegant. Less natural. Less intelligent.

I want to sit on that for a second, because I think it's bigger than it sounds.

The dismissal of AI as "just pattern matching" is the same dismissal that has been applied to autistic cognition for decades. The frame is identical. The implicit version is the real thing, and the explicit version is an imitation.

When a colleague says "I just know" and I say "well, here's how I worked it out" — and the room defaults to taking the colleague's version more seriously, because the explanation is implicit and therefore feels like real intuition, while my explanation is explicit and therefore feels like derived or constructed or somehow less authentic — that's the same frame. When a senior leader interviews a candidate and rates them more positively because they "have presence" in a way that can't be articulated, and rates another candidate more negatively because they "explain too much" or "are too in the weeds" — that's the same frame. When a teacher reads a student's explicit show-your-work answer and grades it lower than another student's confident but unjustified answer because the show-your-work answer "doesn't look like real understanding" — that's the same frame.

The frame says: real cognition is the kind you don't have to explain.

I don't think that frame holds up. I don't think it held up when it was applied to autistic people. I don't think it holds up when it's applied to AI. And I think the AI conversation right now is doing something useful, which is making the assumptions visible. Because when you start judging a system that has no inner life by the standard "real cognition is the kind you don't have to explain," the standard starts to look very strange. The standard starts to look like what it actually is — a cultural preference for opacity over visibility, dressed up as a theory of cognition.

If you've been listening this far, I want to leave you with one thing. It's not a recommendation. It's a question I want you to actually sit with for a few days.

Notice, this week, every time you find yourself judging what AI is doing as "not real thinking" or "not really reasoning" or "just pattern matching." Notice the frame you're applying. And then ask yourself. Am I judging this system by the same standard I have been judging humans by? Humans whose cognition shows its work. Humans who reason explicitly. Humans who do not produce confidence as a felt sense but as the output of an analyzable process. Humans, in other words, like me.

If the answer is yes — and I think for a lot of us, including me sometimes, the answer is yes — then the AI conversation is doing more than reshaping work. It's giving us a chance to see what we've been quietly assuming about what counts as cognition, all along.

That's worth sitting with.

The autistic profile I write about isn't an outlier in this. It's one of the early visible examples of what the brain economy is going to start pricing more correctly. Explicit reasoning under uncertainty. Pattern recognition that has examined itself. The kind of cognitive contribution that scales when paired with systems that compute the way it already does. The skills the dominant leadership model has been quietly filtering out — for being too systematic, too explicit, too literal — turn out to be the skills the next decade is going to reward.

And the standard we use to judge that scaling — the standard we use to decide what counts as real cognition and what gets dismissed as "just pattern matching" — is going to determine which workforce, and which kind of thinking, the brain economy actually values.

The architecture is yours to design. Both the architecture of how your organization thinks. And the architecture of how you've been deciding what counts as thinking in the first place.

Thanks for listening.

If this resonated, the full essay is on theautisticleader.ai — I Experience Cognition the Way AI Computes. That Is Not a Metaphor. And if you want more of this thinking week to week, the newsletter is at theautisticleader.substack.com — free, no pitch cadence, you can unsubscribe any time.

I'll see you next week.