Terence Tao on the limits of artificial intelligence

Terence Tao is one of the greatest living mathematicians and he gave an interview to Big Think, where he makes analogies that I found very pertinent and also talks about the limits of artificial intelligence.

The first analogy is about how scientific work works: it is as if you were going on a hike and wanted to reach a certain waterfall. So you need to find your paths and cross difficult stretches. But sometimes, halfway there, you see another waterfall or something interesting. You draw conclusions and end up in places you didn’t even plan before.

With artificial intelligence it is as if you now had a helicopter. If you had seen the first waterfall, you would go straight there, without really understanding the path, the obstacles and what the challenges would have been to get there. You land right on the conclusion, which is great, but it ends up creating other problems.

And he also brings up the whole theory of geocentrism and how Copernicus and Kepler persisted: at first, the heliocentric model didn’t work as well as geocentrism numerically, but they insisted so much, testing several variations, until they landed on the ellipse model. He said that if it were artificial intelligence trying directly, it would stop quickly, because it would see that the heliocentric model produced worse numbers and was not the right direction.

Of course, he is always cautious, saying the models may change. Just as he said that, even working with artificial intelligence as a collaborator, there still isn’t a good fluency of exchange, where you know your work partner and half a word is enough for them to get what you mean, and vice versa.

And he points out, obviously, the challenges that show up for young scientists. We used to give easy problems for a master’s student to solve, to gain reputation, space, to understand the work. Nowadays, those easy problems that are still open will often be solved by artificial intelligence. So the initial research work for a young scientist is increasingly scarce.

I found the way he thinks interesting, how he positions himself and how he communicates. He is certainly an enthusiast, but he recognizes that today there are limits and that it can also harm the scientific community, including, in some way, a stagnation.

An addendum. Talking with my brother Guilherme, he also brings up a possible problem: as the models advance, it will get harder to verify the discoveries. That is the case of the Millennium Problem that the OpenAI team solved, which I mentioned in the note about AGI. Even if we use Lean and the like, it will be harder to find flaws that are not formal math… after all, the models used to get wrong how many r’s there were in the word strawberry. And more: my friend Angelim says the models will indeed find waterfalls along the discovery process, we just need to ask them to pay attention to their own reasoning. I don’t know. I haven’t yet seen a model asking good questions, only giving good answers.