Agree that Anthropic has a lot to clarify here. Jack Clark said in his most recent podcast with Ezra Klein that one of his priorities this year is better data on internal AI R&D acceleration so I’m hopeful this is coming?
Random ideas as a side note: what if you or AI futures created the survey you’d like to see frontier lab leadership ask their engineers/researchers? obviously the labs would be the one implementing the survey but seems plausible to me that one of bottlenecks is just not quite knowing what to ask/having the time to think about it.
"Current work acceleration" and "fraction of work done by AIs" seem like relatively useless operationalizations because they both ignore the fact that having AIs changes *what work* a company does.
"Serial labor acceleration" and "parallel labor acceleration" are better in this respect, but still not ideal since even with e.g. accelerated labor the speedy humans still won't produce as much code as the normal speed humans with AI since writing all the code by hand is still much more costly than AIs writing the code. (The speedy humans will do more of other kinds of work instead to make the tradeoff worth it.)
> It seems that Anthropic might have their own takeoff speeds / timelines model that differs substantially from current public modeling, produces much less conservative conclusions about the level of concern
I think you meant "more conservative", not less?
Great post by the way. I hope Anthropic will improve their estimates and transparency in the future. (Frankly I'm surprised we got any estimate at all)
interesting post!
Agree that Anthropic has a lot to clarify here. Jack Clark said in his most recent podcast with Ezra Klein that one of his priorities this year is better data on internal AI R&D acceleration so I’m hopeful this is coming?
Random ideas as a side note: what if you or AI futures created the survey you’d like to see frontier lab leadership ask their engineers/researchers? obviously the labs would be the one implementing the survey but seems plausible to me that one of bottlenecks is just not quite knowing what to ask/having the time to think about it.
"Current work acceleration" and "fraction of work done by AIs" seem like relatively useless operationalizations because they both ignore the fact that having AIs changes *what work* a company does.
"Serial labor acceleration" and "parallel labor acceleration" are better in this respect, but still not ideal since even with e.g. accelerated labor the speedy humans still won't produce as much code as the normal speed humans with AI since writing all the code by hand is still much more costly than AIs writing the code. (The speedy humans will do more of other kinds of work instead to make the tradeoff worth it.)
> It seems that Anthropic might have their own takeoff speeds / timelines model that differs substantially from current public modeling, produces much less conservative conclusions about the level of concern
I think you meant "more conservative", not less?
Great post by the way. I hope Anthropic will improve their estimates and transparency in the future. (Frankly I'm surprised we got any estimate at all)
Define “productive”….