I think that vertical is actually a step function that is still limited by the human bottleneck of deciding what to build and making sure that it actually gets built.
Give the AI an underspecified task, you get what you asked for and also what you didn't ask for, in good old Genie fashion.
Hill-climbing problems seem to escape this limitation.
But I wonder what will happen when the AIs start acing benchmarks faster than you can write new benchmarks, and when you run out of hard problems to solve because you rely on humans formulating them.
The real vertical will come when the AI gets access to physical world instead of bash; and when the human is finally out of the loop.
Hill-climbing problems seem to escape this limitation. But I wonder what will happen when the AIs start acing benchmarks faster than you can write new benchmarks, and when you run out of hard problems to solve because you rely on humans formulating them.
The real vertical will come when the AI gets access to physical world instead of bash; and when the human is finally out of the loop.