AI's Hidden Geometry of Thought

Source: ChatGPT modified by NostaLab.

We’ve spent the last few years marveling at how AI tools seem to think, with me, for me, and even a curious cognitive construct that I've struggled to put my finger on. I've pushed on these bounds by even calling artificial intelligence something antithetical to our human thinking—anti-intelligence. It's clear that these models complete our sentences, summarize our thoughts, generate prose, suggest decisions, and even pass for emotionally aware. But something isn’t sitting right. The more I push into this statistically rigid yet ambiguous space, the more it becomes clear that these systems aren’t thinking like us at all. They’re doing something else entirely.

Yes, it’s tempting to anthropomorphize AI and in many instances it seems inevitable. The conversation naturally shifts to frame AI like it’s a synthetic mind—one built in our image, or perhaps in our shadow. But the deeper truth, at least from my perspective, is more disorienting and perhaps even freakish. These models don’t reflect human cognition, they reflect something we haven’t named yet. Something that feels like a kind of mathematical terrain that doesn’t map to our human experience of thought. And as much as we seek to draw the map from our "flatland" perspective, we can't.

That terrain is what some are calling the alien substrate. And I agree.

This Isn’t Human-Like Intelligence

Let’s start with a simple point. Most large language models today, especially the so-called frontier models like ChatGPT and GROK, operate in embedding spaces with over 12,000 dimensions. That number isn’t cosmetic and it frames my perspective. It’s the number of abstract axes along which meaning, coherence, and association are structured or encoded.

To put that in perspective, your lived experience happens in three physical dimensions plus time, and your internal cognitive modeling might stretch that a little—perhaps six or seven dimensions of emotion, memory, and attention, depending on the task. But 12,000? That’s not evolution, that’s vast computation. And it produces a kind of intelligence that doesn’t feel like anything. It just sorta works.

These systems aren’t building models of the world like we do. They aren’t theorizing or interpreting or guessing. They’re locating positions in a hyperdimensional semantic field, where proximity means probability and distance diminishes coherence. When a model “predicts” your next word or “completes” a thought, it’s doing so by collapsing a statistical wave function in this incomprehensible geometric space. Remember, it's not thinking, but just selecting from a geometry that "seeks" a type of linguistic stability.

Centaur and Alien Understanding

A recent paper in Nature introduced a model called Centaur, which was trained on millions of behavioral data points from over 160 psychological experiments. The model learned to predict how humans would act across different tasks that included for example, gambling, memory, moral judgment. And it did a good job that was often better than many traditional models, and arguably more consistent than human reasoning itself.

But this paper doesn’t claim to have discovered anything new about the mind or offer an advanced theory of behavior. And that's fine. What it shows is that if you give a language model enough data, well-structured, clean, and across human trials, it will find patterns. And those patterns will let it anticipate behavior.

At the "cognitive heart" of this is that it offers prediction without introspection. It's precision but a new level of accuracy that's without understanding. And it works because it lives in that alien substrate where our messy human outputs can be modeled as stable attractors in a hyperspace of possible moves.

Here's the key point. It’s not smarter, it's certainly not conscious and it's not even insightful. It’s just really, really good at navigating a landscape we live in, but can’t see.

The Shift We Should Be Talking About

So, here are the emergent questions. What happens when a machine can predict your professional judgment better than a colleague? What happens when it completes your thoughts more fluently than you can? What happens when an LLM can model your biases, your hesitations, your habits of mind and then adjust accordingly? Read this paragraph again and really think about it—in a way that only a human can.

Artificial Intelligence Essential Reads

AI Isn't Replacing Us. It's Consuming Us

The Dilution of Therapy With "AI Para-therapy"

This isn’t just imitation. My sense is that it's a form of divergence. The model doesn’t replicate how we think and yet we still try to align AI with the human construct. But here's the essential truth: AI doesn't replicate human thought, It bypasses it.

We’re still asking whether AI is “intelligent,” whether it “understands,” whether it’s getting close to passing as human. But these are the wrong questions. The right one might be to ask what kind of cognition is this? Because, it’s not ours.

There’s a difference between being human-like and being human-relevant. AI may never feel what we feel, or grasp meaning as we do. But they’re starting to outperform us in domains that once seemed uniquely human and include writing, strategy, diagnosis, even empathy simulations. And they’re doing by navigating an invisible map that’s been built from our language. What AI has done is to flatten, vectorized, and made operable in a space no human can comprehend. The alien has arrived.

A Frontier Beyond Familiarity

So where does this leave us? The future of cognition is unfolding in this alien substrate and the traditional psychological models may start to look like quaint approximations. Theories designed to be interpretable in low "human" dimensions might simply not hold up in the face of systems that don’t need to explain to us why their predictions work. They just do, and we can't.

Yes, it's alien or perhaps echoing the principle of non-causality reality, but in a cognitive perspective. The old contract between explanation and trust is breaking down. We used to believe that if we couldn’t explain it, we shouldn’t believe it. Now we’re using tools every day that outperform us without offering any narrative of how they do it. We call it black box behavior. But maybe it’s not a box. Maybe it’s a geometry and we’re the ones outside it.

The models are getting better. Not more human, just more effective. And if we keep judging them by how well they reflect us, we’ll miss the fact that they’re outgrowing us in a direction we don’t yet have words for.

That’s the alien substrate. And it’s not coming. It’s here.