I think their view point is that AGI "can solve any problem" so the problem they're solving for is "all of them", to which AGI is "the answer".
Some problems are likely to be too big for AGI to solve, many others would be bottlenecked by the same thing that bottlenecks human intelligence - real world evidence/experimentation.
The idea that somehow the right answer is hidden in preexisting data and that a sufficiently complex algorithm would be able to find it is amusing and quit likely wrong.