Diffusion Starts in the Dark

The crux of AI right now is not to get it to do things in some frameworked, structured way but to figuring out what useful things it should be doing at all. It is fair to say that, the execution story is reasonably solved. The application story isn't at all.

AI application is mostly about content generation (code, emails, blog posts) is where AI delivers undeniable value because the task maps directly onto what the models do well. That's settled! The next layer (business process, operations, decision support) remains largely unexplored. This is not because people aren't trying, but because the useful applications in these domains haven't been discovered yet. Intuition tells us they exist. Intuition isn't a product roadmap.

The diffusion of innovations curve explains why this is so hard. Transformative applications look blindingly obvious in retrospect. Nobody looks at email and thinks "what a surprising invention." But before it existed, it wasn't on anyone's roadmap. It was stumbled upon. The gap between "this is possible" and "this is useful" gets crossed through experimentation, not strategy decks.

You can't think your way to AI's next killer applications. You can't derive them from first principles. The problems that are genuinely AI-native, not just AI-enhanced versions of things we already do, require contact with messy, real-world processes before they reveal themselves. It's a dark room. The only way to find the light switch is to move, bump into things, and adjust. Building a better flashlight framework while standing still doesn't help.

The industry's bias toward infrastructure and tooling is understandable but those are legible problems you can spec and ship. Discovery is messy and mostly consists of things that don't work. But it's the only path to the applications that matter. The content generation phase was the easy part. The only way through the hard part is to keep bumping around the corners until something clicks.