Smarter AI Still Has to Touch the World to Learn Anything New

No matter how smart artificial intelligence gets, it will still have to go out and touch the world to learn anything truly new. That is the argument of investor Cliff Sosin, who says we mistake intelligence for something like the height of a tree, which can keep growing forever, when it is more "like the roundness of a ball," which tops out at perfect.

In his telling, intelligence is a gadget for guessing. We start with a thin scattering of facts about a complicated universe, Sosin says, "and we use intelligence to interpolate between them," to color in the blank spaces. The smarter the system, the better the coloring gets.

In some places, that coloring can be perfect. In mathematics, where answers can be worked out by logic alone, more brainpower gets you all the way to right.

But most of the questions people care about do not work that way. They involve what mathematicians call complex dynamic systems, systems with many moving parts that feed back on each other over time, like the weather, like cream mixing into coffee, or like crowds of complicated people interacting.

Those systems cannot be solved with a tidy equation. They have to be simulated step by step, and every simulation has a little error in its starting measurements and in each step, because the steps have to be rounded off. The error compounds until the forecast drifts away from reality.

The classic example comes from early weather forecasting on room-size computers. Researchers ran a forecast, then ran it again to check the work. A technician had loaded the second run with starting values rounded off by one decimal place. At first the two forecasts matched, then they diverged, until one forecast called for warm sunshine while the other called for a cold downpour. In principle, a butterfly flapping its wings in Africa could set off a hurricane in North America, and that is why this sensitivity is called the butterfly effect.

That is why Sosin says that to exactly simulate the whole universe, you would need the whole universe. Simulation gives useful approximations in many cases, and fails in many others where the future is fundamentally unknowable.

The same limit applies to discovery. A very smart model will be wonderful at finding connections inside what we already know, noticing that three obscure papers, read together, imply something no person noticed. Sosin expects a burst of such finds.

But a guess that lies outside all known data is different. For a truly new idea about physics or biology, he argues, a superintelligent model may do somewhat better at guessing than the smartest person, but its odds are still low. What the model can do is dream up clever experiments worth doing. Past a point, progress stops being about thinking harder and starts being about doing the slow work of testing reality.