Essay argues AI adoption is bottlenecked by users unable to see what models can do

The bottleneck is not intelligence

An essay published on mhacevedo.com in early September and picked up on Hacker News's front page makes a pointed claim about AI adoption: the technology's biggest obstacle is no longer capability but visibility. Users cannot tell what a model will do for them until they actually try it, and most never get that far. The author labels this a "discovery problem."

The mechanics, as the essay describes them, are stubbornly simple. You learn what a button does by pressing it, and you learn what a prompt produces by writing one and running it. When every capability sits behind an empty text input, the range of possible uses stays invisible to exactly the person who would benefit from them.

The fixes we have are partial

The essay surveys two familiar mitigations and finds both wanting. Template libraries hand users something runnable, removing the need to invent a request from scratch, but relevance then becomes the open question: a generic template may have nothing to do with a given person's daily work. Context-aware systems that know something about the user can tailor suggestions instead of offering things that matter only in general, which helps. According to the author, neither approach closes the gap.

An ant in the canyon

To explain why, the essay borrows an analogy it credits to computer scientist Alan Kay. An ant at the bottom of the Grand Canyon perceives the sky as a narrow strip of blue between rock walls, while a person standing on the rim sees the full expanse. It is the same sky, but the two observers hold entirely different notions of what exists. The essay's point is that the ant is not deficient in any way; it simply lacks the vantage point, what the author calls the "axis of possibility."

Where fluency creates the divide

The essay grounds this in a concrete contrast. Watch a non-technical marketing professional work for an hour, and someone fluent in agents and tool use will immediately spot a dozen tasks that could be automated, delegated or rebuilt, including tasks the marketer has never even attempted. Reverse the setup, though: put the most capable AI tool in the world in front of that same marketer and the result is a blank prompt and no idea what to type. The intelligence is present; the means of perceiving it is not.

The burden sits on the wrong side

The essay's sharpest framing concerns responsibility. A system this advanced still expects the user to arrive already knowing what to ask for. In the author's view, the work of uncovering what is possible should fall to the system, which ought to reveal its own abilities gradually, contextually and in ways that match a user's actual work rather than generic use cases. Current products largely do not do this, and the essay closes on that note: somehow, the interface itself has to make the wider horizon visible to the person standing at the bottom of the canyon.

Why it matters

For anyone building AI products, the essay reframes the competitive frontier. If users cannot discover use cases, model quality is effectively stranded, because capability goes unused not due to being missing but due to being invisible. That turns onboarding flows, template relevance and contextual suggestion from marketing niceties into core product engineering problems, and it hints that the next round of differentiation may come from interface design rather than raw model performance. The piece is an opinion essay rather than research, and it offers no adoption data, but as a diagnosis of why powerful tools stall with mainstream audiences it is a practical lens: build for the person staring at the empty input box, because that describes most of the market.