LLMs are unexpectedly good at two things: recommending things to people, and forming the questions that reveal what to recommend. These are not separate abilities so much as the same skill viewed from opposite sides. Yet almost no products use it.
I spent this past Thanksgiving in the United States. I asked an LLM to recommend music by Native American artists that I had not heard but might like. I gave it one instruction: ask questions first, until it was confident I would like the recommendation. I named four artists I already liked: Sharon Burch, Black Lodge Singers, Primeaux & Mike, and Joanne Shenandoah. Before recommending anything, the model asked four or five questions. What drew me to those artists? What setting did I want to listen in? Did I prefer vocal-forward music or something more instrumental? Did I have regional preferences? Each question advanced from the answer before it. Within a few minutes, the model had a deeper picture of my taste in that corner of music than any onboarding survey I have ever filled out. And the recommendation was good.
The first time I noticed that this was the same skill was while working on entity resolution at scale: matching records that look like different people because of differences in spelling, address, or transliteration, but in fact refer to the same person. In active learning, the model selects the pairs it is least sure about and asks a human to label them. That is standard practice. The textbook selection methods — margin sampling, query-by-committee, expected error reduction — all choose pairs near the current decision boundary. Out of curiosity, I replaced that selection step with an LLM. Show it the candidate pairs and let it choose which ones to send to the labeler. Then go further: let it write the question itself. In our own work, the information gained from each label increased by roughly an order of magnitude. The pairs the LLM chose were not isolated points on a boundary. They clarified whole rules: for example, how to treat middle-name initials in a particular cultural context. With the same labeling budget, convergence was much faster.
Recommendation and elicitation are the same skill viewed from opposite sides. To recommend well, you have to know what to ask. To ask well, you have to understand the space of possible answers and how each answer changes what should be recommended. LLMs are strong at both because they can hold a generative model of the domain and the user at the same time, and reason in natural language about which question would reduce its uncertainty the most. Classical machine learning can do a narrow version of this as expected information gain over a fixed feature set. LLMs do it across domains, without a schema, in a more open-ended way.
This pattern is still rare in products. Recommendation was formulated as a problem of embeddings and vector databases back when models were not yet good at compositional reasoning, and that pattern became the default. Our intuitions about latency and cost are also out of date. Five turns of elicitation with a small model is cheap. And active learning has never really been fashionable in industry. Its gains are quiet, and hard to see until you work with it.
Much of the discussion around LLMs assumes more: more autonomy, longer reasoning chains, larger context windows. But there is a quieter direction with the opposite shape. In short, targeted exchanges, the model’s job is to ask the right question and then make one good suggestion. The technology is ready. The product patterns have not caught up.