Recommender systems: why algorithms will soon stop choosing and start creating

Think back to the last time you scrolled Netflix for half an hour and went to bed without picking anything. The problem isn't a shortage of films. The problem is the logic behind today's services. Any modern recommender system is, at heart, a very fast and pushy librarian. It knows what is on the shelves and tries to guess which book you will like. But what if the book you need simply isn't on the shelf?

Researchers at Huazhong University of Science and Technology believe it is time to retire the librarian and hire an author instead. In their new survey, the Chinese scientists declare a shift from the old paradigm of "content selection" to a new one: "content generation." This is not a cosmetic overhaul of the algorithms but a fundamental shift. Traditional discriminative models (Discriminative RS) are always limited to the existing catalog. They estimate the probability of a click on something someone else has already made. Generative recommendation systems (GenRS) change the rules: they use the power of large language models (LLMs) and multimodal neural networks to create a personalized answer, or even the content itself, at the moment of the request.

Why does this matter now? We have hit the ceiling of classical machine learning. We used to be pleased that an algorithm took our likes into account. Then image and text analysis were added. But the "cold start" problem, when there is simply no data on a new user or item, has not gone away. Generative models solve it elegantly. They don't need to wait for click history; they understand context and semantics. If the system sees that you are looking for a "cozy cyberpunk-style evening," it will not search for similar tags. It will synthesize a description, pick the visuals and, in the longer term, create a video stream that fits your request perfectly.

The researchers' analysis shows that GenRS is not only about text. It is about deep integration of modalities. Imagine a marketplace that doesn't show you ten similar T-shirts but generates, in real time, an image of an ideal model with your body shape, taking into account your style preferences and current trends. This turns consumption from searching into a process of co-creation with the algorithm. Companies like ByteDance are already looking hard in this direction, understanding that an endless feed will become even more addictive if its content is created personally for each viewer.

Of course, questions of ethics and AI hallucinations remain. If a neural network starts generating recommendations "out of thin air," how do you verify they are accurate? But the scientists from Wuhan are not alarmed. They identify three key stages of GenRS adoption: from using LLMs as ranking assistants to full autonomy, where the AI itself decides what to create and how to present it.

We are at the beginning of the end of the era of "catalogs." The future of the internet is not a warehouse of ready-made files but an endless stream of generation that adapts to every movement of your eyes.

The bottom line: recommender systems are turning from filters into content factories. Expect that soon each of your requests will give birth to a unique digital product that did not exist before you.