There’s a belief that we cannot outsource thinking to these agents but I’d like to push back a little on this. I am not confident in that when I truly debate the other side. There is so much literature out there unexplored fully, new information easily gathered at inference time, and work that can be inspired by other domains of work. The quantity of talented people to try things has been bottlenecked by time/intelligence/motivation. To me, this seems like mostly a money problem now which is really a compute problem. If I could make one prediction in 2026, it’d be that there will be a correction to this. That we will believe these models can invent novel things to a higher degree—that perhaps they will outpace us.

One interesting observation is that over a long horizon, the agents themselves produce a lot of valuable information (often verified as a result of working) and i find this very interesting for the models to utilize that new information to explore further. Some of the time, it will produce results or information that has yet to be seen by humans or re-check information in the event it was incorrectly reported.