It's never been clear to me what it will look like when AI is able to solve tasks of an 8 hour time horizon, but no more than that.
It seems like at some time horizon T, an orchestrator should be able to chunk out an arbitrary-length project into T hour chunks of labor. At which point, the system can do much longer than T hour tasks...
At a certain point, I'd expect the time horizon to just go vertical with respect to some set of well-specified tasks (excluding compute-heavy tasks)
I'd weakly guess most tasks could be decomposed into 8 hour tasks.
If it's not the case that there's some time horizon where we see escape velocity for well specified tasks, then what would it even look like for a human to "unblock" a week long ball of complexity
Neel Nanda told me he thinks that his mech interp MATS fellows are already going way faster due to AI assistance. The AIs are particularly helpful for that kind of project!
It's never been clear to me what it will look like when AI is able to solve tasks of an 8 hour time horizon, but no more than that.
It seems like at some time horizon T, an orchestrator should be able to chunk out an arbitrary-length project into T hour chunks of labor. At which point, the system can do much longer than T hour tasks...
At a certain point, I'd expect the time horizon to just go vertical with respect to some set of well-specified tasks (excluding compute-heavy tasks)
I'd weakly guess most tasks could be decomposed into 8 hour tasks.
If it's not the case that there's some time horizon where we see escape velocity for well specified tasks, then what would it even look like for a human to "unblock" a week long ball of complexity
I worry that AI companies will not be completely open with the level of automation and speedup. The transparency hasn't been great so far.
When can we expect to see vibe coded mech interpretability 😀?
Neel Nanda told me he thinks that his mech interp MATS fellows are already going way faster due to AI assistance. The AIs are particularly helpful for that kind of project!
These models and its assumptions are far removed from the current empirical reality of AIs effect on software engineering. It’s far more limited.