Historically speaking, progress is slow, but it adds up.
There's likely a lot of mechanisms to understand and modeling to do. Each time we make advances it becomes more clear what problems remain and what to do next.
AI agents calling third-party logic looks like a very promising direction. As much as possible AI should be a glue connecting various parts that are already shown to work well.
"Another potential reason to give more weight to insiders is their technical expertise. We don’t think this is a strong reason: there is just as much AI expertise in academia as in industry."
Historically speaking, progress is slow, but it adds up.
There's likely a lot of mechanisms to understand and modeling to do. Each time we make advances it becomes more clear what problems remain and what to do next.
AI agents calling third-party logic looks like a very promising direction. As much as possible AI should be a glue connecting various parts that are already shown to work well.
"Another potential reason to give more weight to insiders is their technical expertise. We don’t think this is a strong reason: there is just as much AI expertise in academia as in industry."
What's the basis for this claim?
The AIME graph demonstrates conclusively the plateauing of capability with compute.
Replot the graph with a traditional linear x-axis and you will see that capability has plateaued.
Much, much smaller improvements for each order of magnitude of compute.