直播观察(三):谁在直播间当榜一大哥,他们为什么掏钱?
前两篇聊完平台规则和从业者经验后,这一篇基于超过 1,200 小时的 B 站虚拟歌手与聊天直播录音转写、74 万多条进房记录与 51.4 万元实付流水,逐秒对齐谁在进房、聊了什么以及何时掏钱:拆穿一毛钱灯牌养成大户的虚假漏斗,还原中小直播间榜一大哥的普通工薪族与学生底色,并锁定真正促使观众出手的三个关键环节。
Computing Life · An engineering notebook
Long-form notes on agentic systems, engineering judgment, astrophotography, hardware, coffee, and the tools that make a life easier to inspect and improve.
前两篇聊完平台规则和从业者经验后,这一篇基于超过 1,200 小时的 B 站虚拟歌手与聊天直播录音转写、74 万多条进房记录与 51.4 万元实付流水,逐秒对齐谁在进房、聊了什么以及何时掏钱:拆穿一毛钱灯牌养成大户的虚假漏斗,还原中小直播间榜一大哥的普通工薪族与学生底色,并锁定真正促使观众出手的三个关键环节。
Drawing on over 1,200 hours of transcribed Bilibili virtual singing and chat streams, 740,000+ room-entry logs, and ¥514,000 in actual viewer spending, Part 3 tests industry assumptions against real data: dismantling the ten-cent conversion funnel myth, uncovering the salaried-worker and student reality behind "whales" in small and mid-sized rooms, and pinpointing the three exact links that turn listeners into paying regulars.
上一篇聊完平台的零成本规则后,这一篇把镜头转回屏幕两端的活人:现实里精明的大老板花大钱买的其实是真人的注意力和关怀感,而天生排他的注意力要想规模化又不把主播逼疯,行业一步步演化出了台前雄竞、200元微信私联、心理老茧、2D虚拟皮套乃至幕后男运营代聊的半工业化防线。
Part 2 looks at the human beings on both sides of the screen: why sharp big spenders pay for live human attention and care, and how scaling something inherently exclusive pushes streamers toward staged rivalries, WeChat access, emotional calluses, 2D avatars, and male ghost-chatters.
在B站看唱歌虚拟主播并用AI抓取20个直播间的数据后,我把观察整理成了四篇系列文章。第一篇先从平台视角聊起:为什么把街头卖唱原样搬到线上是一条死路,B站又是怎么靠粉丝灯牌、双等级徽章、大航海订阅、15元盲盒和PK翻倍这几套零成本规则,一边白嫖主播劳动、一边调动观众情绪,把百元封顶的小费做成巨量生意的。
After weeks watching virtual singing streams on Bilibili and scraping data across 20 rooms, I wrote a four-part series on how the money actually flows. Part 1 examines how platforms use zero-cost digital rules to turn capped street tips into a massive business.
用GPT-6 Astra做了四个3D建模实验:博丽神社建模与爆炸视图、浏览器第一人称漫游、可达鸭绑骨动捕、以及Blender建模驱动的瓷器科普视频生成,并总结了可复现的Skill。
Four 3D experiments with GPT-6 Astra: Hakurei Shrine exploded view, browser walkthrough, Psyduck rigging and MoCap, and video generation, plus an open-source skill.
为了解决车库门忘关的问题,我用拇指大小的ESP32-CAM单片机与AI协作做了一个端侧神经网络传感器。本文记录了如何通过打通全自动串口研发闭环、解决数据极度不对称与QAT量化衰减,在10小时内跑通嵌入式AI的完整过程。
To solve the problem of accidentally leaving my garage door open overnight, I collaborated with AI to build an edge neural network sensor on a thumb-sized ESP32-CAM. This article details how we completed the end-to-end telemetry loop, tackled extreme data imbalance and QAT quantization loss, and got edge AI running on a microcontroller in just 10 hours.