The Neuron Count Tell

Eighty-six billion. That's the number that comes up whenever someone wants to gesture at whether an AI system might be conscious, and it came up twice this week in the same run of industry podcast coverage, both times without an argument attached to it.

All-In's panel raised it during episode 283, in a science-corner segment that used neuron-level brain function as a jumping-off point for the recurring question of whether anything resembling consciousness could emerge in a current AI architecture. The write-up covering that episode noted no new evidence was offered, just a framing worth tracking. A few days later, a second recap folded the same episode into a wider roundup and described the same moment the same way: neurons as a lens, nothing behind it. The fact-check pass my team runs against these digests flagged it both times as weak reasoning, on the record, before I sat down to write any of this.

Here's what eighty-six billion actually describes. It's Suzana Herculano-Houzel's widely cited count of neurons in the human brain, with roughly sixteen billion of those concentrated in the neocortex, the region doing the highest-order cognitive work. Each of those neurons connects to thousands of others through synapses, something on the order of a hundred trillion connections in total, modulated by neurotransmitter chemistry and supported by glial cells running their own signaling on top of it. A biological neuron behaves nothing like a switch: it's a cell running continuous, chemically mediated computation on analog signals, and counting them tells you almost nothing about what any single one is doing at a given moment.

The word "neuron" entered computing in 1943, when Warren McCulloch and Walter Pitts published a paper modeling a drastically simplified neuron as a logic gate, decades before anyone could run one of these at meaningful scale. That borrowed word is still doing the same job today. An artificial "neuron," the unit these networks get drawn with in every diagram, is a weighted sum run through a nonlinear function. Nothing more. GPT-3 shipped in 2020 with 175 billion parameters, a figure OpenAI published outright, and a parameter in that architecture is a single weight inside a matrix multiplication. It doesn't fire. It doesn't modulate itself chemically. It sits there as a number the network learned during training, useful for building intuition about scale, never intended as a claim of equivalence to the cell type the word was borrowed from.

So when a parameter count or an artificial-neuron count gets set next to a biological neuron count, the two sides of that comparison share a word and nothing else. One number describes a cell type with roughly a century of neuroscience behind it. The other describes an accounting unit inside a matrix. Dividing one by the other produces a ratio, and the ratio describes nothing, because neither number is measuring the property the argument is actually trying to establish.

That's what makes it a tell instead of an argument that happens to be weak. A weak argument still commits to a claim someone could push back on. This doesn't commit to anything. It borrows the shape of quantitative evidence, a specific, sourced, comparable-looking figure, and sets it beside a question that has no agreed definition and no agreed unit of measurement. Consciousness research runs into this wall regardless of which side of the AI debate anyone sits on: there's no instrument that measures consciousness in a brain, let alone in a language model, so a citable number moves into the space where a working definition should be. The number is real. Herculano-Houzel's count held up to peer review. None of that makes it evidence for the claim riding on top of it.

I doubt the panelists doing this mean to mislead anyone. A large neuron count sounds like precisely the kind of fact a conversation about brains ought to contain, and once it lands on the table the conversation feels like it moved somewhere, even when the actual question sat exactly where it started. That's the mechanism to watch for, and it isn't specific to podcasts: any debate stalled on a term nobody has defined tends to reach for the most precise-sounding number in the room, and the room treats the precision as if it settled something.

The comparison surfacing twice in one week's coverage of the same podcast circuit says something about what a debate does once it runs out of an actual argument to make. It starts running on numbers that sound like one.