Machines With Intelligence: Part 2

Last time I discussed that a machine has more purpose than generality, that intelligence has unbounded generality, and that intelligence designs machines in response to positive market signals that it’s worth it for us to do so. Something only alluded to in the espresso machine example: the barista.

The barista is an intelligence more efficiently working over a finite set of limitations to produce “better” within the machine’s capability. A person mounted between the dials and the drinker. That’s not a machine and it’s not using unbounded intelligence. It’s a hybrid, and the hybrid is where I need to expand this understanding. What implications when you can easily mount intelligence inside the machine?

We Already Live With These

Turns out we didn’t wait for permission. Even in the world of pre-AI ease of access, the design was already in machines. The autopilot has flown most of every airline flight you’ve ever taken, for decades. Notice what it is: an envelope around operating the machine. It holds heading and altitude, it manages the climb, and it operates inside hard limits it cannot leave, with a pilot holding dials and a disconnect under their thumb. Nobody prompts a plane.

Your car has been running the same play in slow motion. Cruise control was a dial. Then adaptive cruise gave the dial eyes. Then lane-keeping gave it opinions. One degree of freedom at a time, each one earned before the next was granted: a machine acquiring intelligence by installments, edges intact the whole way.

And today, prepare to be blown away, and to be careful.

The Autonomy Dial

Fine, you say, so what’s the design question? Here it is: the most important control on a machine with intelligence is a new dial, and the dial sets how much judgment the intelligence inside is allowed. Andrej Karpathy calls it the “autonomy slider,” and argues we should be building partial-autonomy products (“Iron Man suits” rather than Iron Man robots) where the human grants agency by degrees. Notice the details in his answer. The governing control on the hybrid is itself machine-shaped: visible, adjustable, reversible. Turn it up, watch what happens, turn it back.

Robots are the hybrid at full stretch: a machine body with an intelligent interior. Jensen Huang stood up at CES this year and declared that “the ChatGPT moment for physical AI is here, when machines begin to understand, reason and act in the real world.” Maybe so. But watch the two shapes it can take. The robot vacuum is machine-first: intelligence budgeted to one job, don’t fall down the stairs, cover the floor. The humanoid is generally out looking for a task, the replicator on legs, and it inherits every replicator problem from the last essay. Describe what you want. Calibrate “slightly.” Wait your turn behind everyone else’s heart’s desire.

Herbert Already Voted

Science fiction ran this experiment before we could. The backstory of Frank Herbert’s Dune is the Butlerian Jihad, humanity’s war against thinking machines, and it ends in a commandment: “Thou shalt not make a machine in the likeness of a human mind.” Herbert even has the Reverend Mother explain the grievance: “Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.”

Read that as a design review instead of a prophecy. Herbert’s society’s objection wasn’t to machines (Dune is full of machinery), and it wasn’t to intelligence as human intelligence evolved to replace computers. It was to the unbounded hybrid: judgment delegated wholesale, with no dials, no visible settings, no disconnect under anyone’s thumb. And look what his remedy bought. The prohibition didn’t make the design pressure go away; the universe just relocated the machinery into people. Mentats trained as computers. Navigators as route processors. Ban the hybrid and you end up building it out of humans; the hybrid was never optional in this experiment, only its design was.

So take the commandment and turn it half a click: make machines in the likeness of a task, not in the likeness of a mind.

Here We Are Again

The hybrid worth building looks like the ones already in your life. Intelligence mounted inside machine edges: a fixed task, set formats, a watchable process, autonomy on a dial, and a way out under your thumb. The plane, the watch, the vacuum, and every one of the background business machines from the last essay, with a little judgment installed where a brittle rule used to be. That’s the version that gets adopted, because it’s the version people can trust without a leap of faith. Trust arrives one dial position at a time.

The hybrid to squint at is the mind-shaped one: the general interior with no envelope, the prompt box on legs. Not because it can’t work, but because it hauls every impedance from the replicator along with it and calls that an interface.

I hope you see where this landed. The barista was the answer the whole time: intelligence with dials in front of it and an envelope around it, judgment where judgment pays and machinery everywhere else. In the rush to make every machine intelligent, the likeness worth keeping isn’t the mind’s. It’s the task’s.