> The roadmap to success will most likely start with training or fine-tuning on data from human professionals performing the task, and proceed with reinforcement learning in custom environments designed to capture more of the complexity of what people do in their jobs.
> [...]
> We think this is essentially a data problem, not an algorithms problem.
This is extremely hand-wavy. How are you going to instrument the various thought processes and non-verbal communication that goes into building successful software? A huge part of it is intuition about what makes sense to other humans. It's related to the idea of common sense, but in the software world there's this layer of unforgiving determinism and rigidity that most humans don't want to deal with. I just don't see how AI crosses that chasm.