The Mind Is a Machine

Three Books, One Pattern

I read three books recently. I didn't plan to read them together. They just came up in sequence.

The first was The Psychology of Stupidity, a collection of essays edited by Jean-François Marmion. The essays argue, mostly, that reasoning comes after the decision, not before. We decide something and then generate reasons for it. The reasons feel like they caused the decision. They didn't. There's evidence for this. A lot of evidence.

The second was Algorithms to Live By, by Brian Christian and Tom Griffiths. It takes decisions that feel ordinary, like when to stop searching for an apartment, and shows that computer science has formal solutions for them. Optimal stopping. Explore versus exploit. Caching. The book demonstrates that people approximate these algorithms without knowing they're doing it, and the approximations are often wrong in predictable ways.

The third was The Man Who Mistook His Wife for a Hat, by Oliver Sacks. Neurological case histories. A man who can see but can't recognise faces. A woman who lost the sense of where her body is in space. A man permanently stuck at age nineteen. The cases show something that took me a while to articulate: the things I think of as "me" are actually components. Each one can break independently. When one breaks, the rest keep running. That means the mind has parts, and the parts do specific jobs.

Put the three books together and you get a picture:

1. The reasons we give for our decisions are generated after the fact.
2. Our decisions run on algorithms we can identify, and sometimes improve on.
3. The mind is made of separable components.

If all three are true, the mind is a machine. I mean that literally. Not as a metaphor. If it were a metaphor I could ignore it. It reads as a description.

The Question

If the mind is a machine, can you build another one?

That question has a history. Turing asked it in 1950, framed as a game. Searle answered no with a room. Dennett said the question itself is confused, that what we call consciousness is a bundle of separate capacities, which matches the Sacks cases exactly. Hofstadter wrote about loops. I've read parts of this history. But the version of the question I keep coming back to is narrower, and I think it's the version I can actually do something with.

Not: can a machine be conscious.

Not: can a machine understand.

Those questions are about the inside of the thing. The inside is unobservable. I can't verify an answer to either one, and a question you can't verify is a question you can only argue about. I'm not interested in arguing.

The version I want: can a machine predict what a specific kind of person will do, better than I can guess?

That's it. A prediction problem. Predictions can be checked. You make one, wait, see what happened, count the misses. That's a loop I can run.

We Already Simulate People

Also, we already do this. This is the part I keep coming back to, and it's so obvious that I want to write it down once so I can stop thinking about it.

Focus groups are simulations. A company wants to know how people will react to a product, so it puts eight people in a room for an hour, asks them questions, writes up a summary, and treats the summary as data about millions of people. That is a simulation. It's a bad one. Small sample. Staged environment. One loud participant distorts the whole thing. But companies run them because the alternative, launching blind, is worse.

So "should we simulate people" is not the question. We already do. The question is whether we can do it better. More samples. More control. More dimensions. And if the objection is that a simulation can't capture a real person: fine, but neither does a focus group, neither does a survey, neither does my intuition, and I now have three books of reasons not to trust my intuition.

One detail from the books matters here and I keep almost skipping it. A realistic simulation has to include the bugs. If I build a model of a person and the model is too rational, it's not a better model, it's a worse one. Real people confabulate. Real people follow the group. Real people anchor. A model that doesn't do those things isn't modelling a person. It's modelling an ideal, and ideals don't buy products or reject pitches.

The Sketch

So I wrote down the smallest version I could think of. Not a person. One component of a person, for one purpose.

belief: a number from 0 to 1
update: evidence comes in, the number moves

convincer:    input that moves the number up
dealbreaker:  input that moves the number down
update style: how the number moves
  jump:     moves all at once
  gradual:  drifts
  stubborn: barely moves

That's the whole idea so far. It doesn't capture a person. It captures one axis of one kind of behaviour, the axis that matters for one question. Which is what a model is. You don't model everything. You model the part that answers the question, and you stay honest about the boundary.

The Chinese Room doesn't bother me. I don't care whether the room understands. I care whether its output predicts better than my guess. Those are different questions and I'm answering the second one. The first one can stay with the people who enjoy it.

The Sacks cases keep coming back when I think about scope. The man who can't recognise faces still has most of a person intact. A model of him that matters for a buying decision doesn't need face recognition in it. The machinery is separable. That means I can build small pieces and test them individually. I don't have to solve the whole problem at once, which is good, because I don't think I could.

I don't know if this goes anywhere. I start a lot of things. This one has a shape now though, and shapes are easier to come back to.

Unanswered

• Where exactly is the line between a model of a person and a claim about personhood? If the model keeps getting better, does the line move, or was it always somewhere else?
• If realism means including the bugs, what's the right ratio of correct behaviour to broken behaviour?
• If I simulate a thousand people, am I simulating a thousand selves or one machine with a thousand parameter sets? Is there a difference that matters?
• Does spending a lot of time thinking about people as functions change how I treat actual people? In which direction?
• If the self is narrative, assembled from parts, what exactly am I modelling when I model "a person"? The narrative? The parts? The assembly process?