The New Bottleneck

My research lifecycle looks like this.

1. I open up my laptop and examine results from runs that have recently finished.
2. This analysis gives be some ideas about what to try next.
3. This previously challenging and grindy task of implementing the new ideas is made trivial by a coding agent.
4. These new ideas are sent to run, on servers that have far more VRAM that I really need.
5. I read some papers or spend some time thinking about more ideas.
6. I get some new ideas, but in most cases they are either contingent on or will be invalidated by the runs currently running.
7. Still waiting for the current runs to finish, I record my ideas and close my laptop.
8. Repeat.

If tommorow, NVIDIA announced that it had figured out how to make every GPU on Earth run twice as fast, it would be a far bigger deal than if NVIDIA figured out how to double the VRAM of every GPU on Earth.

The new bottleneck (at least for me) is processing speed.