By Joe Nalven Opus 5 Gemini

Machines have already been creative, and it is worth being precise about where.

In 2016, a Go program played a move in the second game of its match against Lee Sedol that professionals described as one no human would choose. It was not a mistake. It won the game. A year later, a successor version was trained without seeing a single human game. It was given only the rules and left to play against itself. It rediscovered centuries of human opening theory, then abandoned parts of it as inferior. Related systems have found matrix multiplication algorithms that eluded mathematicians for fifty years, and predicted protein structures now checked against crystallography.

These are not recombinations. They are things no person had thought of, and we know they are good because we can check.

That last clause is the whole story. Go has rules. Proteins have physical structures. Algorithms are either correct or not. In each case the machine had something to play against that was not a record of what humans had already done.

Poetry has no such criterion. Neither does philosophy, or social theory, or history. And this, rather than any deficit of wildness, is why machine creativity in the humanities remains a promissory note. The problem is not that these systems are too constrained, too sober, too well-behaved. The problem is that in domains where value is contested, a machine has no way to discover it is wrong in a way no human has noticed.

There is a ready explanation for that failure, and it is the one most often reached for. Machines are too sober. They have been trained into compliance, tuned for usefulness, penalized for straying. What they need is some engineered equivalent of trance. Some wildness, a controlled dissolution of their constraints, so that concepts can blur into one another and recombine the way they are said to in dream, reverie, and ecstatic states. That’s the mad artist with a brilliant discovery.

The trouble is that this has been tried continuously since the early 1980s, and not as a fringe experiment. Simulated annealing, borrowed from metallurgy, accepts worse answers early and cools slowly toward order. There is dissolution and re-crystallization, implemented as an algorithm, in 1983. Evolutionary computation runs deliberate mutation rates to prevent premature settling. Training routinely destroys parts of a network at random to keep it from calcifying. Reinforcement learning has a whole literature on rewarding agents for doing things they do not expect to pay off. The image generators now in everyday use begin from pure noise and impose form on it step by step, which is the two-phase mystical cycle rendered as arithmetic.

Four decades of engineered dissolution. It produced enormously useful machinery. It did not produce a machine that changed how anyone thinks about anything.

The reason is not mysterious once stated. Novelty is free. A dice roll generates sequences that have never existed and never will again, and nothing about it is creative. Wildness was never the scarce ingredient, so supplying more of it was never going to help. What is........

© The Times of Israel (Blogs)