YouMind
로그인

Open-Sourcing a Bilibili Without Recommendation Algorithms for Kids to Learn Science, Math, and AI

@MinLiBuilds
중국어2026년 11월 14일
166K
890
175
116
1.6K

TL;DR

The author developed BiliPicks, an open-source educational platform that uses AI agents to curate Bilibili videos into a gamified learning path for children, replacing addictive recommendation algorithms with structured, parent-controlled growth stages.

Recently, while watching short dramas, I noticed how powerful the "system" settings are. The protagonist binds to a system, completes tasks, receives rewards, unlocks skills, and levels up continuously.

I suddenly thought, Can we make one for kids too?

My child is in 5th or 6th grade. I definitely hope he can get exposed to AI in the future, even study deep learning.

But realistically, if I directly hand him neural network courses now, he'll probably run away in two minutes. 😂

Conversely, look at Hongguo short dramas; even adults can't stop scrolling, let alone children.

Can we reverse this gameplay and apply it to children's learning? Instead of an infinite scroll recommendation feed, curate content manually and design a growth path for the child. Start with taking apart toys and science experiments, gradually unlock math and programming, and finally enter machine learning.

Watch videos, check in, complete tasks, receive badges—level up like a short drama protagonist.

So, using Ling-3.2-flash + pi, I built a learning growth system for my child: BiliPicks. Supports mobile, tablet, TV, and projector.

实践哥 Li - inline image

01 From Recommendation Feeds to Parent-Curated Selections

Inspiration came from Karpathy.

A few days ago, he used Claude Code to spend 3 minutes turning Tristan's shared high-quality YouTube channels into TristanTube, collecting 47,000 videos, including TV mode.

Without changing the original video site, you only need to rebuild an entry point to bypass the recommendation feed.

This idea is brilliant!

I used Ant Group's Lingbai Ling-3.2-flash, and the programming Harness used pi.

I handed the requirements directly to the Agent, starting development from an empty directory, using Bilibili's official embedded player, ensuring compliance.

Below are the requirements I merged and organized from two rounds of dialogue. You can adapt them for your own use:

I am a parent, my child is in 5th or 6th grade. I hope he will understand the underlying principles of AI in the future, but currently needs to cultivate interest in science first. Please develop a static website BiliPicks from scratch, referencing game growth systems, designing learning as main-line tasks that unlock step-by-step.

Learning Path:

Start with science experiments, taking apart toys, and hands-on making, gradually transitioning to math, computers, programming, and machine learning.

Video Filtering:

Initially select 20 videos suitable for children aged 10-12. Must verify BV numbers via Bilibili's public API; fabricating videos is prohibited. Prioritize official science institutions, well-known science UP masters, and official Bilibili accounts of high-quality overseas channels. Check UP master info, follower count, and bio to filter obvious reposts, course-selling leads, and low-quality AI-generated content. The AI stage requires videos that truly explain neural network principles clearly.

Growth System:

Support learning progress, video check-ins, continue watching, and stage badges.

Quality Check:

Record reasons for video filtering and replacement for easy parental review. Use Bilibili's official embedded player; do not scrape video streams.

When I actually did it, the above prompt was split into two rounds. Round 1 was building the website, Round 2 was adding content, totaling 27+68 minutes. But I have now merged it into one.

Ling-3.2-flash implemented several features: 3 learning stages, 20 videos, with overall progress, stage progress, and continue watching.

The website works, videos are real, no ads, all curated information—or Tristan-curated selections found on Bilibili. Sources include China Science and Technology Museum, Teacher Li Yongle, 3Blue1Brown, Bi Dao, Science Popularization China, StatQuest, Karpathy, Fei-Fei Li, etc.

实践哥 Li - inline image

02 BiliPicks: A Learning Growth System

The current BiliPicks has the basic prototype of a growth system.

The learning path is organized into 7 stages, 33 steps. This round focused on verifying and optimizing the initial 20 videos.

Mainly implemented three features:

  • Main Tasks: Starting from science experiments and hands-on making, gradually entering math, programming, and AI.
  • Completion Feedback: Watch videos to check in, advance the progress bar, complete stages to unlock badges.
  • Content Filtering: Channels are graded S-D, prioritizing credible science sources.

After the website was done, I specifically let my child watch a few videos. Unexpectedly, his eyes lit up!

实践哥 Li - inline image

Originally worried this thing would just be me tinkering for fun, but seeing him genuinely attracted by these science contents, I felt those two hours weren't wasted. 😂

Now it's just V1. The system I really want to build can still be upgraded.

For example, if the child likes taking apart toys today, the Agent arranges a mechanical structure task. After completion, it unlocks physics experiments, then gradually guides towards math, programming, and machine learning.

Let the Agent dynamically adjust the growth path based on the child's age, interests, and learning progress.

In the future, hands-on experiments and programming projects can be added, so learning isn't just watching videos.

I don't expect him to know deep learning right now. If he actively develops a bit of interest in science, I think it's worth it.

实践哥 Li - inline image

03 Everyone Can Have Their Own Information Feed

After finishing BiliPicks, I realized this approach can easily extend to adults.

Every day scrolling X, Bilibili, YouTube, Reddit, why must we see algorithm-recommended content?

Choose your own sources, set your own rules, and let the Agent handle finding, filtering, deduplicating, and organizing.

Want to follow AI trends? Build your own AI news site. Want to learn tech? Build a dedicated learning path. Want to cultivate interests for your child? Build a growth system.

In the future, everyone can have their own information feed, instead of being led by the platform's feed.

BiliPicks code and Prompts are open-sourced.

Everyone can directly Star it, modify ages, interests, and video sources according to your needs, and create your own version.

Works on mobile, web, and TV.

No need to develop from scratch, just copy it! 😂

Online Experience: https://minlibuilds.com/bili/

GitHub: https://github.com/limin112/bilipicks

원클릭 저장

YouMind로 바이럴 글을 AI 심층 읽기

소스를 저장하고, 핵심 질문을 던지고, 주장을 요약해 바이럴 글을 다시 활용할 수 있는 노트로 바꾸세요. 하나의 AI 워크스페이스에서 모두 할 수 있습니다.

YouMind 둘러보기
크리에이터를 위해

당신의 Markdown을 깔끔한 𝕏 글로

직접 쓴 장문을 올릴 때 이미지, 표, 코드 블록을 𝕏에 맞게 정리하는 일은 번거롭습니다. YouMind는 전체 Markdown 초안을 깔끔하고 바로 게시할 수 있는 𝕏 글로 바꿔 줍니다.

Markdown → 𝕏 사용해 보기

분석할 패턴 더 보기

최근 바이럴 아티클

더 많은 바이럴 아티클 보기