Reverse Engineering Cognition

A Second Intelligence

A second intelligence is emerging.

Not a tool. Not another species. Not a digital god. Intelligence — another instance of the same thing we have — and it's getting better, much faster than we are.

By intelligence I mean something quite specific: the ability to take in information, spot patterns, compress them into useful models, and then use those models to achieve goals in messy, uncertain situations. Whether that constitutes “real” thinking in some deeper philosophical sense — whether it is reasoning or mere pattern-matching, understanding or sophisticated mimicry — is less relevant here. From an operational standpoint, the distinction does not matter. The capability is rising.

Right now, cognitive work is becoming both absurdly cheap and dramatically more capable. In 2020, GPT-3 cost $60 per million tokens. Today you can get models far more powerful than GPT-4 for $0.10–$0.50. A 100x–600x collapse in price while capability has soared. Generation, reasoning, synthesis, coding — all of it is racing toward being basically free and far more powerful.

And it isn't just white-collar intelligence. Humanoid robots are starting to move, grasp, and work in the real world.

In every previous industrial revolution, the answer to displacement was straightforward: we moved up. We left the fields for the factories. We left the factories for the offices. We left routine work for knowledge work, strategy, and creativity. There was always a higher rung on the ladder.

But this revolution is different. It is eating both ends of the ladder at once. Cognitive work is collapsing in price and soaring in capability. Humanoid robots are starting to operate in the physical world. For the first time, there may be no obvious “next level” to climb into.

So what actually remains for humans? What becomes the new scarce resource? What is worth doing when almost everything can be done better and cheaper by machines?

The honest answer is that nobody knows yet. But the shape of the question matters, because most attempts to answer it are still trapped in the old frame — humans versus machines, one ladder, one winner. That frame is the problem.

The real relationship doesn't have to be adversarial or hierarchical. It can be symbiotic. Two systems, different in kind, more valuable together than apart — but only if each remains itself. A symbiosis where one side absorbs the other isn't symbiosis. It's absorption.

For that symbiosis to actually work, something essential must come from the human side. That something is agency — the capacity to set goals, exercise judgment, and decide what is worth pursuing. It isn't a relic of the pre-AI era. It's the human half of the equation.

But this raises another question:

Which parts of human cognition can be deduced by a sufficiently powerful intelligence, and which can only be reached through observation?

That is the question that decides whether humans end up augmented or absorbed.

The Demon and Its Collapse

In 1814, a mathematician called Pierre-Simon Laplace had a beautiful idea. He imagined a vast intellect — later called Laplace's Demon — that knew the exact position and velocity of every single particle in the universe at one instant. With that knowledge, it could predict the entire future with perfect certainty. If the universe was just an enormous clockwork machine, then in principle, everything — including human thoughts and decisions — was fully knowable.

It was an elegant vision. It just didn't survive contact with reality.

In the 1890s, Henri Poincaré was studying the three-body problem in celestial mechanics when he stumbled into something disturbing. Even in simple, fully deterministic systems, tiny differences in starting conditions could produce wildly different outcomes. What we now call chaos. The Demon's perfect knowledge suddenly became practically useless. You could know the rules perfectly and still have no idea what would actually happen.

Then quantum mechanics arrived in the 1920s and broke the machine at an even deeper level. At the smallest scales, nature isn't deterministic at all. It's probabilistic. Measurement itself changes the system. The clean, predictable clockwork gave way to fundamental uncertainty.

And in the early 2000s, Stephen Wolfram formalized the concept of computational irreducibility. Even in systems governed by very simple rules, there is often no shortcut. To know what the system will do at step 10,000, you frequently have to simulate every single step. There's no faster formula. No clever compression. The universe simply doesn't always offer us the courtesy of prediction.

Human minds inherit the universe's messiness. They are chaotic, noisy, and computationally irreducible. Not impossible, but expensive. In principle, with godlike resources, you could simulate any individual mind perfectly. In practice, the combination makes high-fidelity prediction from first principles brutally expensive — often more expensive than simply watching real humans over time and learning from what they actually do.

The Demon doesn't survive. But a different path does: through careful empirical observation, we can build increasingly accurate approximations of how minds work — not by deducing them, but by observing them.

What Observation Makes Possible

Once cognition becomes empirically legible — captured not as output but as process, in real situations and at real scale — two things become possible at once.

The first is that foundation models begin to understand humans differently. Today's systems are trained almost entirely on what humans produce: text, code, images, the finished surface of thought. The reasoning underneath — the timing, the hesitation, the sketching, the self-correction, the moves a person considered and abandoned — is largely absent from the training signal. A model trained on output learns to produce more output. A model trained on the process of cognition learns something else: how to meet a person where they actually are. How to teach without lecturing. How to explain without flattening. How to operate inside a real organization with real context rather than generic plausibility. This is the layer that current AI lacks, and it is the layer that empirical observation of cognition produces.

The second is that humans begin to understand themselves differently. For most of history, the question of how learning actually unfolds, how decisions actually get made, how expertise actually develops — these were matters of folk knowledge, philosophical speculation, or controlled laboratory abstraction. None of them captured cognition in the wild. With cognition instrumented at scale and over time, something rarer becomes available: an empirical science of how minds actually work in the situations where they matter. And that self-understanding is what makes genuine augmentation possible. Not faster output. Not bigger memory. Sharper thinking. Clearer learning. More accurate self-knowledge.

These two axes are not separate. They are the same substrate seen from two sides. The data that lets a model understand a person better is the same data that lets the person understand themselves better. Augmentation is what happens when both sides of that loop close.

What this eventually opens up is harder to predict, but the shape is becoming clear. Once cognition is no longer a black box, new categories of work become possible at the interface between humans and machines. Hardware that reads state and intent rather than just commands. Interfaces that operate as extensions of the mind rather than another surface to interact with. Systems that belong to the person using them, run locally, and consume the kind of energy a body can actually carry. None of this is buildable on top of generic models trained on generic output. All of it becomes plausible once the foundation is in place.

The infrastructure we are building today is small. The future it serves is not.