Out of Distribution
As i keep tinkering with LLMs for all kinds of things, there's this sort of muscle memory i'm developing where i feel that all of the alpha now, would be in building things that are out of distribution of the models' training data. Now this itself might be over in a few months (years?) if the hype bros are accurate that the labs can build continual learning where their model weights can constantly be updated / improved in a continuous manner rather than the current paradigm.
But if that isnt the case, there is genuine value in identifying the gaps, and creating products that meet that.