Every A.I. system "learns from humans." That fact explains almost nothing about how any of them behave.
It is a common and reasonable-sounding dismissal: self-driving cars work because they learn from human drivers; chatbots work because they read what humans wrote. True, as far as it goes. But it goes almost nowhere — because the data is the one thing nearly everyone has in common.
Tesla, Waymo and every major A.I. lab can access roughly the same raw material: billions of miles of driving, most of the public internet. If data alone determined behavior, every autonomous car would drive identically and every chatbot would give the same answers. They do not.
This ingredient list stays fixed for the rest of the article. Everything that follows changes behavior without touching it.
A neural network is millions to trillions of numbers called weights, which determine how much each signal matters. Below is a toy driving brain. The scenario never changes; only the weights do. Drag the sliders.
You have just changed a driver's entire disposition without showing it a single new mile of road. Now imagine finding the right values for a trillion of these dials at once, all interacting. That search — training — is where the real work lives.
The same holds for language models. Here, three models trained on an identical corpus complete an identical sentence. Because their weights settled in different places, their instincts differ.
Same data in; different behavior out. "They read the same books" tells you nothing about what each learned to do with them.
Even after training ends, behavior is not fixed. What sits in front of the model — the context — steers it. Below, one model handles one request under different circumstances.
The knowledge did not change; the framing did. Data, weights and context are three separate dials — and only the first is "learning from humans."
How large is the space of possible minds that can be built from the same data? Drag the number of weights and watch the arithmetic get out of hand.
If each weight could take just 10 distinct values, the number of possible configurations would be roughly:
The data is the library. Training is the nearly impossible search through that space for one configuration that behaves well. Two teams with the same library routinely end up in different corners of it.
A closing thought for the skeptic at the dinner table: two chefs who shop at the same market will not cook the same dish. The ingredients were never the hard part. ■