Last week I opened yet another dashboard someone at work had built with AI. It was fine. Clean charts, sensible filters. And I realized I had no idea who was supposed to look at it. I’ve watched dozens of these appear in the last few months, because building one now costs almost nothing. Looking at one still costs exactly what it always did.
That’s the whole problem. We sped up production. We didn’t speed up consumption, and we can’t. Human attention scales with the number of humans, and our brains don’t get any faster. So every new dashboard, report and generated doc competes for the attention of the same people.
What happens to all the stuff nobody has time for? In the best case, another LLM reads it and summarizes it for a human. Follow that round trip. A person has a thought and writes a prompt. A model expands it, padding it with statistically plausible material harvested from the history of human writing. Another model compresses it back down for the reader. What comes out the other end is roughly the original prompt, maybe a little more. We burned the compute to move a paragraph from one person to another.
Notice what survives the trip: only what the human put in. The expansion is exactly the part a model can regenerate at will, which is why the other model throws it away. So why send it at all?
If your reader has an LLM too, send them the prompt instead.
It sounds like a joke. I mean it as a rule. What’s scarce is the new information: an observation, a measurement, a decision and the reason behind it. That’s usually small, and no model can make it up.
Of course, the same logic applies to the dashboard. Building it ahead of time is a bet that someone will have that question later. When building was expensive, it made sense to place that bet once and reuse the result. Now it doesn’t. Generate the view when someone actually asks, and invest in what lasts: the data and the model underneath it.
In fact, many of those dashboards did have a reader: the person who built them. Asking a model to slice your data ten different ways is a fast way to understand it, and that might be one of the best uses of these tools we have. But that value stays with the builder. Keep what you learned, and don’t publish your thinking tool as if it were meant for someone else.
Cost used to do the filtering for us. Nobody spent two weeks on a dashboard without knowing who it was for. That filter is gone, so we have to apply it on purpose, by asking one question before we make anything: who will act differently because this exists?
If the answer is nobody, make it, but please don’t send it to me.