> Leading AI companies have an ongoing internal competition for compute resources. With AI becoming commercialized, the competition intensifies because allocating more compute for research means less is available for deployment.
> I’d be curious to see load graphs for these companies. Are experiments run at night when fewer people use ChatGPT? These are the types of strategies companies likely adopt when managing limited resources.
I was surprised to see Lennart say this—it's pretty well known that research compute at top labs does not directly trade against inference compute so elastically. Each R&D team has some compute budget decided in advance based on the limited resources, and they always make sure to queue enough experiments to have ~100% utilization. Each new researcher added to the team must have ideas better than the average existing researcher in order to improve progress.
This is why adding AI R&D automation does not necessarily speed up AI research much, because labs are bottlenecked on research with compute and ideas. And the automated AI researchers are not likely in the short-term to provide ideas better than the average AI researcher.
> Leading AI companies have an ongoing internal competition for compute resources. With AI becoming commercialized, the competition intensifies because allocating more compute for research means less is available for deployment.
> I’d be curious to see load graphs for these companies. Are experiments run at night when fewer people use ChatGPT? These are the types of strategies companies likely adopt when managing limited resources.
I was surprised to see Lennart say this—it's pretty well known that research compute at top labs does not directly trade against inference compute so elastically. Each R&D team has some compute budget decided in advance based on the limited resources, and they always make sure to queue enough experiments to have ~100% utilization. Each new researcher added to the team must have ideas better than the average existing researcher in order to improve progress.
This is why adding AI R&D automation does not necessarily speed up AI research much, because labs are bottlenecked on research with compute and ideas. And the automated AI researchers are not likely in the short-term to provide ideas better than the average AI researcher.