Nano Search and the Question of Who Checks AI Search Results

Jiang Wu · 2024-11-27

*Originally published in Chinese on November 27, 2024.*

Nano Search is an AI search product that has models cross-check one another's answers. Here is how I read it, and where I think it runs into trouble.

Where search is going

The scope of search and its entry point

First, search and our understanding of models. In the future, search will no longer be confined to a search product, and search itself may change a great deal. Look at what Sogou's input method is doing with search: search will not live in one particular product or one particular assistant, and it will probably sit at the user's first entry point. Where search will live is a question mark today, and technology may need to evolve further. The second step for search is search as a service, which means connecting search logic with the user's whole service scenario. So what we need to watch is how the user's products and scenarios are defined as a whole.

An effective answer

To understand search, what finally has to be shown is an effective answer for the user. The problem with search engines in the past was that their answers were not effective. Early AI search retrieved web pages, weighted them with a model, structured the content and gave users an answer that looked precise. But the calibration was still left to the user.

What Nano Search adds

Models checking each other, and users checking models

Nano Search takes a step here by letting models correct each other's results. One side is multi-angle analysis of user behavior, and the other is cross-calibration of the output between different models. If this process produces useful data about users, or a useful summary of results, it also trains search scenarios and models to some degree. So the problem it has to solve is getting one step closer to the user.

Perceived effectiveness

The product may have no innovation in form, yet users feel it is more effective, which is probably a stacking of capabilities or functions. Before, users did their own calibration: they searched, got results and checked them in their usual way. Now a model checks once more. The accuracy of the results may not rise by much, but users trust them more.

It is like buying cooked food at a deli. You may know the vendor has already washed it, but if they wash it again in front of you, you feel it is cleaner. In that respect Nano Search has moved the product forward.

Challenges and openings

A traffic business without a stronger model

Of course it faces some serious problems next. The core problem in search is that it turns into a traffic business, and this does not solve the weakness of the models themselves.

Diversion from AI assistants, and the attitude of model companies

The second dimension is that AI assistants now search far more widely. Users who are able to start searching with AI may be diverted in large numbers, and that may happen in the future too. It still moves the category forward one step. As for whether model companies welcome it, I think they do. Model capacity is not even fully used today, so a high-traffic search system that starts calling their models would be a good business for them.

Home · All articles · Topics · 简体中文