With those test parameters for how long it would take a human to complete the same work, it fits a similar pattern to METR; i.e. at "humans would take 11.5 hours" (Figure 4, median) you're pushing your luck for any success with all but the most recent models*, and METR is testing software where AI has the possibility of fully automating a lot of its own tests.
Even more recent models than they tested, like Opus 4.5, are only 50% successful for tasks that take humans 5h20m: https://metr.org/time-horizons/
Assuming the bubble doesn't pop/WW3 doesn't start first (IDK, 25% and 5% respectively?), and if trends continue (???), I expect a similar paper this time next year to show something like 50% success at automation of similar tasks.
* which they didn't test, I don't blame them for that because this field moves too fast
Or they've determined that micromanaging it is circuitous and increases their dependence on tech giants, so it's a bad deal given that they also need to know the work well enough to verify it anyway.
Sounds about right.
With those test parameters for how long it would take a human to complete the same work, it fits a similar pattern to METR; i.e. at "humans would take 11.5 hours" (Figure 4, median) you're pushing your luck for any success with all but the most recent models*, and METR is testing software where AI has the possibility of fully automating a lot of its own tests.
Even more recent models than they tested, like Opus 4.5, are only 50% successful for tasks that take humans 5h20m: https://metr.org/time-horizons/
Assuming the bubble doesn't pop/WW3 doesn't start first (IDK, 25% and 5% respectively?), and if trends continue (???), I expect a similar paper this time next year to show something like 50% success at automation of similar tasks.
* which they didn't test, I don't blame them for that because this field moves too fast