I feel pretty uncertain about this sort of modeling in general. It feels very sensitive to assumptions and inputs. If it were really hard to get the model to put any significant probability on TAI this century, I’d take that as an update (similarly with the model making TAI soon look very very likely). But for most middling values I’m not personally inclined to base too much on them.
Yes—this needs to be said again, and again, and again. And then people need to consider how valuable arguing about the details of these models really is.
And yes, I think that it’s incredibly valuable for people to have done thinking about this in public, but the difference between 25% and 75% probability of AGI in a decade is a tiny rounding error for this type of modeling compared to the uncertainties and approximations, and the fact that we’re talking about a loose proxy for an upper bound anyways!
Yes—this needs to be said again, and again, and again. And then people need to consider how valuable arguing about the details of these models really is.
And yes, I think that it’s incredibly valuable for people to have done thinking about this in public, but the difference between 25% and 75% probability of AGI in a decade is a tiny rounding error for this type of modeling compared to the uncertainties and approximations, and the fact that we’re talking about a loose proxy for an upper bound anyways!