alt Hacker NewsOne would think something like the study of the square root of negative 1 would have little utility beyond a hand-wavy appeal to understanding, but it turns out to have major practical applications ranging from electrical engineering to 3d rotations.
I tend to believe we're still likely at a very primitive level of technology. It's difficult to imagine this primarily because of unknown unknowns. For instance at the end of the 19th century it was believed that science had basically been finished, and all that remained was working things out to ever higher levels of precision to wrap up a few curiosities.
Then along come relativity, quantum mechanics, and more - all completely revolutionizing science, and upending centuries of false beliefs. I think we're far more likely to be at the end of the 19th century again, rather than at the end of science, or anywhere near such. And if that's the case, then it may well be that seemingly academic discoveries of today, perhaps ultimately achieved by an LLM, are what help unlock the next great revolution.
Yes, but what will truly advance science is new discoveries that require curiosity and creative thinking, and LLMs are not the technology for this.
As impressive as these mathematical results are, all they represent is known mathematical techniques being combined in ways that humans had not explored. The AI is not inventing areas of math, or even creating it's own conjectures. There will be more of these to follow, but eventually the pace will slow as the space of what can be done with known mathematics is explored.
What is needed for creativity, for new scientific discovery, is more than an auto-regressive LLM assembling someone else's Lego blocks in novel ways - it will require an AI not just built to predict but one that is also built to expand the frontier of what it knows. Prediction failure that results in curiosity and exploration, not hallucination.
We'll get there eventually.