Comment by energy123

Not at all. Some work is the combination of short duration, bottlenecking in nature, and serial.

Example: You're merging 3 branches, and there's a minor merge conflict.

You only need 15k tokens to fix it, so it's short duration. And it's bottlenecking. And it's a serial task.

This belongs on Cerebras or whatever.

Once fixed, go back to slower compute.