> In my view, the resolution of this thought experiment would shed light on many of the claims being made about the ability for AI to accelerate science. Recently, there have been multiple claims of AI systems making scientific discoveries or proving theorems. It is usually difficult to assess whether these are incremental or significant, and the degree to which an AI generated the insight

Specialization and abundance of information have created siloing, less interdisciplinary work. To make breakthroughs in physics, you have to spend the better part of your life studying things like string theory or quantum field theory, and then you still don't know anything about other fields.

Scientific insights, I bet, will come from AI sifting through the data as well as the knowledge and make novel connections that any one human or even a group of humans would struggle to make, if make at all. I can imagine how humans can then build on these new insightful connection to make further leaps that AI can then build off of again.

One thing that puts a limit on the pace of scientific progress is the verification and experimentation parts. While AI can theoretically make discoveries authonomously in fields like mathematics, for physics, biology, medicine, and many other fields, you need to interact with the physical world (experiments, trials), so the breakthrough loop can't accelerate "infinitely".

Robotics can address some of that, but there will remain physical limits, I anticipate.