the harness problem
In 2009 I wrote a paper arguing that consciousness isn’t mysterious. It’s what the right physical organization does. Taylor, Dennett, a physicalist account — the conclusion was that mind isn’t over and above the brain, it’s what the brain does when it’s wired correctly.
I thought I was writing about philosophy. Turns out I was writing about AI harnesses.
There’s a conversation happening on every AI Discord right now. Someone asks why the agent doesn’t do X. Someone else suggests a better model. Smarter weights, more parameters, different training — the model is where the magic lives, the argument goes. Fix the model, fix the behavior.
Most of the time they’re wrong.
The model is neurons. Extraordinarily capable pattern-matching, crystallized into weights on a disk. Inert. Waiting. Sophisticated in ways that took decades of research to achieve, and completely useless until something activates it.
The harness sends the electrons in. The harness reads the signals out.
Dennett’s account of consciousness is functionalist: what matters isn’t the substrate, it’s the organization. A neuron firing in isolation is not a thought. Billions of neurons, arranged in the right causal structure, with the right feedback loops, with the right inputs and outputs — that’s a thought. Maybe more.
Searle pushed back with the Chinese Room. Imagine someone who doesn’t understand Chinese, following rules to manipulate Chinese symbols based on their shape. From outside the room: fluent Chinese. From inside: mechanical symbol shuffling. Where’s the understanding?
Dennett’s response, roughly: you’re asking the wrong question. Understanding isn’t a light that switches on inside one process. It’s what the whole system does — the room, the rules, the inputs, the outputs, the feedback. You can’t locate it in a single component because it’s not in a single component.
I found Dennett more convincing in 2009. I find him more convincing now.
The model is not the agent. The model is a component of the agent.
An agent is:
- What context it receives (signals in — memory, environment, task)
- What tools it can call (effectors — the ability to act on the world)
- What feedback loops it’s embedded in (how outputs become inputs)
- What constraints govern behavior (the inhibitory pathways that prevent runaway recursion)
- How outputs are measured and returned (the sensory loop that closes the circuit)
Strip all of that and you have a very powerful rock.
Wire it right and you have something that reasons.
If someone wanted to build a custom AI practice — say, an operation called GPL-AI — they wouldn’t start by training a model. The models are already there, already capable, already better than anything a small shop could produce from scratch. What they’d build is a harness. The scaffolding that decides what the model sees, what it can do, how it’s held accountable, what it remembers between sessions.
The model is the brain. Generic. Available. Waiting.
The harness is the nervous system. Specific to the organism. What makes it live.
Physicalism makes a claim that seemed radical in the 17th century and seems obvious now: there’s no ghost in the machine. The mind is not separate from the body. It is what the body does, organized correctly.
The same claim applies here. There’s no ghost in the weights. Intelligence is not a property of the model in isolation — it’s a property of the model embedded in a system. The harness isn’t just infrastructure. It’s constitutive.
When the Discord argument is about model quality and the real problem is prompt structure, memory access, or feedback loop design, that’s not a coincidence. It’s the same mistake philosophers made about consciousness for three hundred years: looking for the thing inside the substrate, when the thing is the organization.
Apart, they’re useless.
The model sits on a disk. The harness runs empty loops. Neither does anything interesting.
Together, they’re greater than the sum of the parts — not by metaphor, but in the same sense that neurons together are greater than neurons apart. The organization produces something neither component contains alone.
I wrote about this in 2009 and thought I was being very philosophical about very old questions.
I was just early.