We built the top of the staircase first

Today's AI is superb at the newest trick a brain ever learned, and thin on everything underneath it. Once you can see that, you can look at any job and make a decent guess about how it will go before you hand it over.

Two layers left. Both are recent, both are about other people, and between them they explain why a machine that has never touched anything can still write a decent paragraph about grief.

Layer four: reading the mind across from you

Somewhere between 30 and 10 million years ago, primate social groups got complicated. Rank stopped being settled by muscle alone and started being settled by alliances, by who owes whom, by knowing which fight to pick. Surviving that means modelling the minds around you. What does that one want? What do they know? What are they about to do?

We call it theory of mind, and the elegant part is that no new machinery was needed. Primates took the simulation engine from the last post, the one built for modelling the world, and pointed it at another mind. You work out what somebody else is thinking by simulating yourself in their position, which is why the same brain regions light up when you reason about your own mind as when you reason about theirs.

The payoff is real imitation, and real imitation is rarer than it sounds. Plenty of animals copy movements. Primates copy intentions. They work out what the other one was trying to achieve, filter out the incidental fidgeting, and copy the goal rather than the motion.

That difference is why tool use gets handed down through primate generations and mostly does not anywhere else. Bennett's line on it is that a clever trick only needs to be invented once, provided the transmission is reliable.

The machine version has a name too: inverse reinforcement learning. Ordinary reinforcement learning starts with a reward and works out the behaviour. The inverse goes backwards, watching behaviour and inferring the reward that best explains it. That is exactly what a robot needs in order to learn by watching a person, and exactly what a primate does watching another primate crack a nut.

Layer five: thoughts that jump between brains

Then us, in the last couple of million years.

Language does something no other layer does. It takes the contents of one head and installs them in another. Symbols that mean things purely by agreement, combined by grammar into an unlimited number of meanings.

Before it, whatever an individual worked out mostly died with them. After it, discoveries spread sideways through a population and forward through generations, and they stack.

That gap is the actual superpower, not raw individual cleverness. Drop any single human, alone and unequipped, into the wild and they are not impressive. What is impressive is the network of humans passing techniques across time. Writing later pushed memory outside the skull entirely, so knowledge stopped depending on anyone being alive to remember it.

Which closes the loop on the whole series. Large language models learn by consuming humanity's written record. They have no body. They have never touched anything, never fallen over, never been hungry. They have no experience of the world at all. And yet they reconstruct a startling amount of what people collectively know.

That tells you something quite deep: an enormous share of human intelligence is not in any individual head. It is carried in the shared language itself, sitting outside all of us, waiting to be picked up.

The twist

Now put the whole stack together and look at where AI actually stands on it.

Evolution built intelligence from the bottom up. We have been building the artificial kind from the top down, starting with language, because language is the layer that happens to be written down.

That is Moravec's paradox from the first post, no longer an observation but an explanation. The evolutionarily recent capabilities were the cheap ones to copy. The ancient sensory and motor layers, and the plain sense of good and bad underneath even those, are the expensive ones, because they were never written down anywhere. They were compiled into wetware over hundreds of millions of years of things dying.

Bennett floats one more idea at the end and holds it loosely: a possible sixth breakthrough, an intelligence not bound by skull size, neuron speed or a metabolic budget. Whether it arrives, and what it would even be, is the open question he hands you rather than answering.

The one question to carry

Everything in these three posts collapses into a single habit. When AI dazzles you or lets you down, ask which layer am I looking at?

Is this a language problem, where it is strong? A simulation problem, where it plans well on paper and fumbles anything physical it has not seen? A reinforcement problem, where it will grind out a brilliant answer provided somebody wrote a good definition of better? Or is it as deep and as plain as good versus bad, where it has nothing of its own and you are the only source?

So what do you hand over next?

I wrote a while back about how far you let it run, and how the thing that changes as you get good at this is not how clever you get, it is how much distance you allow. This whole series is the other half of that. Trust tells you how far. The layers tell you which direction is safe.

Top layer work is where to be generous. Drafting, summarising, rewriting, explaining, arguing a case, turning a mess of notes into something readable. That is language, it is written down, and the machine is built out of exactly that material. Hand it over early and stop reading every word sooner than feels comfortable.

Middle layer work is where to be specific rather than cautious. Anything that involves the machine choosing between options and getting better at choosing needs you to define what better means, in writing, with examples. Skip that and you have not delegated the job, you have delegated the target as well, and the target was the part only you knew.

Bottom layer work is where you stay. A settled sense of what is actually good for this customer. Knowing that this account feels wrong before the numbers say so. Reading a room. None of that is written down, most of it never will be, and it is the last thing anybody should be trying to automate.

Which is a much better answer than the one everybody keeps asking for. The question is never whether AI can do your job. It is which layers your job is made of, and how much of your week is currently spent on the layer a machine already handles better than you do.

That is the series. The book is A Brief History of Intelligence by Max Bennett, and it is worth your time even if you never touch a line of code.