Intelligence lives in the network

Six papers, from how humans got smart to how machines now carry culture. Read apart they look like six subjects. Drawn in one notation they turn out to describe one machinery from six angles. The claim for the next ten minutes is simple: intelligence is a property of the network, not the node, and the machine we just built now sits inside that network. Click any mark to jump straight to its paper.

This is the map for the whole talk. Seven questions, asked of every paper: at what level does selection act, what varies and copies, who is wired to whom, what feeds back, how aligned are the interests, which institutions tilt it, and who is in the system. A full dot means the paper carries the argument on that layer, a half dot that it is present, a dash that it is absent. Columns H, W, K, D, B, P are the six papers in order.

The driver of human innovation is the collective brain: how many minds, how well connected, how different from each other. Two natural experiments make it concrete. Patenting per person roughly doubled in the two decades after a rail line reached a county. After the 1924 Johnson-Reed Act cut immigration, patenting fell by about two thirds in the fields those immigrants had contributed to, and native-born inventors in those fields fell 62 percent. And connection has an optimum: in lab transmission chains the worst learner in the connected group beat the best learner in the less connected one. Partial beats full.

William Muir selected hens two ways. Pick the single best layer in each cage and her edge turns out to be aggression: she pecks her cage-mates down, and flock output falls. Pick the best whole cage, and output rose 160 percent in five generations, with calmer birds. Same hens, same trait measured, opposite result, because the level of selection changed. 280 empirical studies document multilevel selection, 180 lab and 100 field. Ostrom's design principles read as a way to move selection up a level.

The strongest role for AI here is not answering for the group, it is helping the group reason together: an agent in the room, and a technology that organises what the group produces. Comparative judgment is the mechanism: people cannot rank twenty things, but they can compare two, and among twenty people there are over a thousand possible trios to draw on. Ten people comparing three spatula handles yield scores near 0.5, 0.35 and 0.15. Masked identities take status out of the deliberation. The brain runs on about 15 watts, a reminder how cheap this computation can be.

Of 144 distinct two-player, two-option games, the vast majority contain some common interest; pure conflict is the rare corner. Yet AI research over-studied that corner: 10 to 15 percent of machine learning and reinforcement learning multi-agent citations went to conflict papers, against about 2 percent in economics. The agenda the paper sets is cooperative AI, built on four capabilities: understanding others, communicating, making credible commitments, and the institutions that hold the rest in place.

Culture runs on three engines: variation, transmission and selection. Machines now act on all three at once. AlphaGo Zero played 4.9 million games against itself in three days and generated move 37, a play a human would choose about 1 in 10,000 times. It crossed into human play, players recombined it with their own moves, and after 2016 human decision quality, flat for decades, stepped up. The risk the paper names is a monoculture of a few shared models, or a Tower of Babel of private ones.

Recommenders and users form a reinforcing loop. People choose, their choices become training data, the machine retrains after a delay, and its suggestions shape the next choices, each pass narrowing what the next can learn from. The social outcomes, filter bubbles and concentration, are ones neither side intends. Leonia is the small, concrete case: in 2017 navigation apps routed highway traffic onto its streets until residents could not leave their driveways. Managing this takes law and politics, not only better code. So the question: who owns the means of recommendation?

Here is where the six papers meet. Engagement ranking is Muir selecting his best hen: it rewards the single most engaging post, the within-level winner, and that pull beats the between-level health of the whole. The same move that raised a hen's output while wrecking the flock can raise clicks while coarsening the public square. The papers do not fully agree: some are hopeful about better AI and mechanism design, others look to law and redistribution, and only two rest on decades of data.

Four things to take away. Measure the health of the whole, not only engagement per item. Keep networks modular and diverse, because partial connection beats full. Build provenance and verification into any AI that helps a group deliberate. And push for researcher access to platform data, because the loop that decides what a billion people see is studied least. The notation is yours: Isotype after Neurath and Arntz, the full kit at the concept page, the six papers listed under sources.