A Few Objections to Bilibili's Recommendation System

My Bilibili recommendation feed is saturated with things I have no wish to watch, and at the moment this leaves me distinctly displeased:

  1. The interview-question videos should know when to stop. Can one really claim to understand concurrent programming while discussing application-layer spinlocks without even considering interrupts? As for producing linked-list code that cannot determine whether a cycle or an intersection exists, that points to a painfully weak engineering foundation. If I encountered such a collaborator on a team, I would either make them repair their data-structure fundamentals or simply ask them to leave.

  2. Videos about algorithm problems are no more appealing to me, especially when they promote the idea that one need not learn something because it is “not tested” or “rarely tested.” That argument steps almost deliberately into my zone of irritation. What I want is a systematic education grounded in Discrete Mathematics and Its Applications and Introduction to Algorithms, then supplemented by training in parallel algorithms, not a mere accumulation of problem types and solution templates.

  3. Videos on modern C++ and systems details are even more likely to trigger my nearly severe standards of judgment. I have very little interest in the surface form of a given syntax; what I want is an analysis of its evolutionary trajectory and internal value, not a parade of flashy features or another heap of scattered, unstructured facts.

  4. When I finally encounter a video that seems interesting, I click in only to discover that it is a re-uploaded piece of content, now wrapped in an AI-generated machine-translated voiceover. I leave immediately. What, precisely, is the point of consuming machine-translated YouTube replicas on Bilibili?

  5. Why is there such an overflow of videos on research methodology and academic topics, especially AI? I do not have a supervisor to placate. The platform cannot classify me as a graduate student merely because I occasionally clicked on a few related videos, and then proceed to push an endless stream of ambiguous deep-learning content and paper explainers at me.

  6. Some of the math videos are genuinely interesting, but what the system mainly recommends to me is graduate-entrance-exam and college-entrance-exam mathematics, to the point that I begin to suspect the algorithm is, in some sense, “insulting” me. I have no interest in piling up complex integrals. Even if one pushes hand-computation techniques to their limit, they still will not outperform symbolic computation; yet when asked to discuss the history of mathematics, proof, or abstraction, such content often has nothing to say.

  7. Then there is a class of videos that use phrases like “985/211” or “high school student does X” as their title-level selling point. I have little interest in these either. The value of a person’s thought and knowledge does not derive from their title, and the viability of their claims should be tested jointly by reality and logic. As for narratives that use age to dramatize talent, they are frankly rather dull; those people have probably never experienced how uncomfortable a genuine mismatch between age and cognition can be. Instead of spending that time reinventing wheels in isolation, they would do better to admit their limits and look at the problem from the shoulders of their predecessors.

  8. Just when I finally scroll to a gaming video or a female creator I might actually want to watch, I find myself surrounded by cybersecurity content. My interest in cryptography and network security is genuinely limited, and professional matters should be left to professionals. My taste, or more precisely my instinctive judgment, evaluates things from the perspective of performance and systems, not security. I may be able to do some high-performance computing and systems programming, but I am not suited to becoming a security expert.

I admit that I have indeed used Bilibili too little to train my account profile properly; nevertheless, its personalization still appears far too crude. It may be able to infer “what kind of person I am” and “roughly which direction I face,” but it cannot go on to match sufficient content depth or information density. It remains mechanically trapped at the level of tag similarity. The signal-to-noise ratio is too low, and the resulting experience is naturally poor.