The spiky advantage of AI at shallowly stitching together a bunch of fields of math while being bad at definitions and conjectures is a core thing we'll have to navigate in trying to semiautomate alignment theory at Resolution:
1. Hire a bunch of human theorists who are world-class at definitions. 2. Accelerate the lower levels of the stack with machines. 3. Try to arrange to not be too bottlenecked on (1).
As a bonus, "AIs can stitch together multiple fields" makes it likely we can cut down the overhead of noticing idea overlaps between different teams at Resolution: if you're in Learning Theory and want a tutorial on some aspect of Complexity Theory, the models can be the first stop.
A subtlety is that making automation go well across the ramp up to superintelligence will require a mixture of research on
1. Short-term hill-climbing on the performance of automation tools 2. Medium-term reasoning about the obstacles we'll face as the models grow in strength, building on https://arxiv.org/abs/2605.06390.
These two kinds of research overlap, but the whole point of hill-climbing research is fast iteration on metrics, which doesn't predict what walls we'll hit in the future. So will need separate subteams and researchers working on (1) and (2), with somewhat different research taste and skill.
I asked Dario 3 years ago why AIs haven't been able to use their vast knowledge across so many fields to connect two known ideas into a new discovery.
It seems like AI did exactly this in the way it disproved Erdos' conjecture aobut the unit distance problem by cleverly onnecting together ideas in discrete geometry and algebraic number theory.
Now that AI has been able to use its knowledge across multiple fields to come up with new ideas, what is the next benchmark?
@3blue1brown proposed one during our interview:
"Good mathematicians prove theorems, great mathematicians come up with conjectures, and the greatest mathematicians come up with definitions."
