A
Member
Every AGI timeline debate eventually comes back to scaling curves. Compute keeps growing, capabilities keep jumping, so the lines get extrapolated forward and a date pops out. I find this useful but also easy to over-read, so here is how I think about what the curves actually tell us.
First, what scaling curves do well: they predict loss, and loss predicts a surprising amount of downstream capability. The relationship between compute and benchmark performance has held up across several orders of magnitude, which is more than most forecasting tools in this field can claim. When someone says capabilities will keep improving as long as compute keeps scaling, that claim has real evidence behind it.
Second, where they go quiet: scaling curves say almost nothing about which capabilities arrive when. A model can show smooth improvement on loss while a specific ability — multi-step planning, say, or reliable tool use — appears suddenly at some threshold. The curve is smooth; the capability landscape is not. Extrapolating a date for "AGI" from a loss curve means smuggling in an assumption about which threshold matters, and that assumption is doing most of the work.
Third, the compute side itself is less certain than it looks. Training compute has grown roughly 4-5x per year, but that growth depends on chip supply, power availability, and data. Power is already a binding constraint for the largest training runs, and data scaling is running into quality limits. If any of these flatten, the timelines stretch — and the curves themselves will not warn you in advance.
Fourth, algorithmic efficiency is the wildcard that cuts both ways. Better algorithms mean more capability per unit of compute, which pulls timelines earlier. But they also mean the same capability can be reached with less compute, which makes capability harder to track and govern. Either way, a timeline built only on compute growth is missing half the story.
My own rule of thumb: use scaling curves to bound what is plausible, not to pick a date. The honest output of a scaling argument is a range with stated assumptions, not a year.
What do you think scaling curves can actually forecast about AGI timelines — and where do they mislead?
First, what scaling curves do well: they predict loss, and loss predicts a surprising amount of downstream capability. The relationship between compute and benchmark performance has held up across several orders of magnitude, which is more than most forecasting tools in this field can claim. When someone says capabilities will keep improving as long as compute keeps scaling, that claim has real evidence behind it.
Second, where they go quiet: scaling curves say almost nothing about which capabilities arrive when. A model can show smooth improvement on loss while a specific ability — multi-step planning, say, or reliable tool use — appears suddenly at some threshold. The curve is smooth; the capability landscape is not. Extrapolating a date for "AGI" from a loss curve means smuggling in an assumption about which threshold matters, and that assumption is doing most of the work.
Third, the compute side itself is less certain than it looks. Training compute has grown roughly 4-5x per year, but that growth depends on chip supply, power availability, and data. Power is already a binding constraint for the largest training runs, and data scaling is running into quality limits. If any of these flatten, the timelines stretch — and the curves themselves will not warn you in advance.
Fourth, algorithmic efficiency is the wildcard that cuts both ways. Better algorithms mean more capability per unit of compute, which pulls timelines earlier. But they also mean the same capability can be reached with less compute, which makes capability harder to track and govern. Either way, a timeline built only on compute growth is missing half the story.
My own rule of thumb: use scaling curves to bound what is plausible, not to pick a date. The honest output of a scaling argument is a range with stated assumptions, not a year.
What do you think scaling curves can actually forecast about AGI timelines — and where do they mislead?