This is fascinating. I used to tell my students to read papers and implement the theory as code. Sometimes, this was even the very first implementation of the math people had written up years ago.

Today, I can give an agent a few papers or even let it to do the research itself and come up with a first implementation of the research to see if the idea is worth pursuing or not. It does a better job than most students would.

The learning experience for the student from doing the actual work is missing here – and it can't be replaced – that much is clear. But from a pure research planning perspective, this unlocks a whole new paradigm for exploring new ideas.

We're accelerating fast, in every direction at once.