The kind of problems that AI is solving

The AI news cycle is tightening.

Every four months or so a better-faster-greater AI model is revealed.

Every few weeks or so we learn about another problem that AI solved. The reports usually mention how AI solved it in a “days”, while humanity couldn’t for decades or even centuries. This also applies to finding software bugs and vulnerabilities.

In the field of mathematics, there are concerns that “AI is mining” the limited number of interesting problems, making it harder for early-career mathematicians to advance.

In the book “An Intoroduction to General Systems Thinking” the author, Gerald Weinberg, observes that newly developed technologies and methods tend to first solve problems that yield to these technologies and methods.

Therefore, every time you see an announcement that goes like this:

“Our latest model M solved problem X!”

mentally substitute for something like:

“We tried model M on problems A, B, C, …, U, V, W, X, Y, and Z. It was able to solve X!”

Of course, tomorrow’s models might be able to make more progress and solve a few more problems, and the cycle will repeat.

I am rooting for the discovery of a class of problems that LLMs, or deepd neural nets in general, will not be able to solve in principle. This will propel us into a more interesting future, where LLMs are an intermediate tool we use to build even more exciting things and places.