I think it’s plausible current AI mostly accelerates the picking of low-hanging fruit downstream of having access to some level of compute. The things that could be achieved in 2 years with current compute are achieved faster, but the things that could be achieved in 15 years won’t happen meaningfully faster. And so the effects of gaining more compute are more important overall than the details of how quickly current AI expresses them as novel AI capabilities (while compute still keeps scaling rapidly, until 2028-2030 or so).
That is, if the next milestone that qualitatively improves AI (which might include becoming immediately RSI capable) would be achieved in 2029 without AI assistance (on the R&D side), maybe it’s achieved in 2028 with AI assistance, or in early 2028 with superexponential progress or whatever. But if it would instead only be achieved in 2035 without AI assistance, it’s still only achieved in 2034-2035 with AI assistance, because it’s bottlenecked by things current AI is unable to meaningfully help with, which aren’t low-hanging fruit that becomes available with a new level of compute.
The surprising capabilities of current AI translate to a surprising TAM, and so a surprising amount of compute in the near future. It’s still only maybe 10x higher than otherwise (by 2031-2035), which is not that much considering the likely 1500x growth over 2022-2028, where it would otherwise mostly stop if revenues plateaued at about $100bn per AI company. This somewhat increases the chance that the next milestones of qualitatively better AI are reached in the near future rather than in a more distant future (a hard-to-predict but plausibly nontrivial number of years after the scaling of compute mostly stops and the low-hanging fruit is all picked). But the involvement of AI in this change from observed surprising capabilities (according to the model I’m sketching) is more importantly financial rather than through AI itself helping with AI R&D.
The things that could be achieved in 2 years with current compute are achieved faster, but the things that could be achieved in 15 years won’t happen meaningfully faster
I think it’s plausible current AI mostly accelerates the picking of low-hanging fruit downstream of having access to some level of compute. The things that could be achieved in 2 years with current compute are achieved faster, but the things that could be achieved in 15 years won’t happen meaningfully faster. And so the effects of gaining more compute are more important overall than the details of how quickly current AI expresses them as novel AI capabilities (while compute still keeps scaling rapidly, until 2028-2030 or so).
That is, if the next milestone that qualitatively improves AI (which might include becoming immediately RSI capable) would be achieved in 2029 without AI assistance (on the R&D side), maybe it’s achieved in 2028 with AI assistance, or in early 2028 with superexponential progress or whatever. But if it would instead only be achieved in 2035 without AI assistance, it’s still only achieved in 2034-2035 with AI assistance, because it’s bottlenecked by things current AI is unable to meaningfully help with, which aren’t low-hanging fruit that becomes available with a new level of compute.
The surprising capabilities of current AI translate to a surprising TAM, and so a surprising amount of compute in the near future. It’s still only maybe 10x higher than otherwise (by 2031-2035), which is not that much considering the likely 1500x growth over 2022-2028, where it would otherwise mostly stop if revenues plateaued at about $100bn per AI company. This somewhat increases the chance that the next milestones of qualitatively better AI are reached in the near future rather than in a more distant future (a hard-to-predict but plausibly nontrivial number of years after the scaling of compute mostly stops and the low-hanging fruit is all picked). But the involvement of AI in this change from observed surprising capabilities (according to the model I’m sketching) is more importantly financial rather than through AI itself helping with AI R&D.
If you haven’t seen Tom Cunningham’s new blog post on what I think is a very related thesis, you may be interested