2026-04-26
I recently learned, or rather reasoned about, the fact that the human brain is essentially the same as it was 10,000 years ago. The same fight and flight mechanisms that used to keep us alive against predators are now mostly obsolete. And yet, look at what we’ve built. The output of that exact same brain has exploded. It can’t be biological, so it has to be external. Tools and notation. All of it is cognitive scaffolding.
Start with writing. It was the first time humans offloaded thought into something concrete, outside of the brain. Plato wrote against it. In writing, naturally. He thought it would destroy memory, that people would have the appearance of wisdom without the substance. He was right in some sense. We did get worse at memorizing. Relying more on external knowledge. But we also got everything writing made possible. The same thing happened with the calculator. People worried it would make us bad at math. And in some sense it did. People are definitely worse at mental arithmetic today. Still, it got us further than mental arithmetic could ever take us.
Every major cognitive tool follows the same arc. It takes over a task. It frees capacity. It gets criticized for atrophy. Eventually it becomes invisible.
LLMs fit this pattern. But they have one property previous tools didn’t. They automate the output of thought, not just its supporting operations. A calculator doesn't write the proof. Wolfram doesn't decide which integral matters. Google doesn't synthesize the answer. LLMs do something closer to the whole pipeline. Whether that’s a difference in degree or in kind, I don’t know.
A few years ago, the safest bet you could make was to study CS. Now junior grads are struggling. The jobs we thought were intellectual peaks are exactly what’s being automated first. The jobs that look more durable are different in kind. Jobs that touch physical reality. Jobs that demand something new under hard constraint. Jobs where the answer isn’t written down anywhere. That is surprising. You’d expect a machine to take the easy jobs first. Instead it takes the ones we labelled hard. That tells you something about what hard actually means. Maybe a lot of what we called intellectual work was actually high-skill recombination.
The simplest model, then, is that LLMs are high-skill recombinators. An imperfect analogy is a monkey given google. Retrieval and recombination at superhuman speed. Still anchored to the distribution of what already exists. Which would explain what we just saw.
If LLMs stay on this side of novel synthesis, the historical pattern probably holds. If they cross it, the pattern might break for the first time.