Attention is all you have
AI made it nearly free to test whether an idea is worth pursuing. It also made it nearly free to walk away from the one you already committed to.
"Attention Is All You Need" - the paper that showed machines could get by on attention alone. The irony is that the tools built on top of it have made it harder than ever for people to do the same.
What else got cheaper
An idea you had this morning can be a working demo by lunch. For testing whether something has legs, that's as good as it gets. Ideas that would have stayed in notebooks can now be held, turned over, tried. The barrier between thinking and testing has mostly disappeared.
But the same speed that makes testing cheap also makes avoidance cheap. When the project you're working on hits the hard middle - past the prototype, not yet near the finish - a new idea is one conversation away from becoming a working prototype of its own. Building to test an idea is different from building to avoid the project you were already on. They look the same from the outside but lead to very different places.
The hard middle
There's a specific moment in every project where the work stops being exciting. The architecture is decided, the prototype works, the interesting problems are solved. You already know what the thing should do and roughly how to build it. What remains is refinement, edge cases, the slow work of making something good enough that other people want to use it. This is where most of the value gets created and where most of the fun disappears.
When a new idea arrives in that moment - and it will, because ideas are cheap and plentiful - you have a choice. You can stay with the thing that's no longer new but almost real, or you can spend an evening making a new thing that's exciting and definitely not real. Before AI, the second option cost enough time that it registered as a detour. Now it costs an evening and produces something that looks like progress.
What separates exploration from displacement is what happens next. Exploration tests what was built, learns something, and decides whether to continue or discard. Displacement moves on to another new thing the following week, and the week after that.
AI can help with the middle. It's good at the mechanical parts, the boilerplate, the things that are tedious but well-defined. What it can't do is make you show up for it.
The decision to open the same project on a Tuesday morning, when your head is full of ideas you could start instead, is still entirely yours.
AI can help with the middle. What it can't do is make you show up for it.
The question that matters each morning is not what you can get done. It's what you're staying with.
Attention is all you have
The transformer paper showed what machines can do by attending - focusing on what matters in a sea of input and ignoring the rest. The tools that followed gave us the ability to create at speeds that would have seemed absurd only recently. What they didn't give us is the ability to stay with what we've created past the point where it's novel.
That part is still on us. It always was. It's just that the alternatives got cheaper.