100時間かけて動画編集AIを作ったら、気が狂いそうになった。

[PR: Surfshark] Access the promo link https://surfshark.com/yurucom or use the promo code yurucom at checkout to receive an extra 4 months of Surfshark VPN! 30-day money-back guarantee! Take advantage of this offer! This is the story of how I almost went crazy after creating a video editing AI. I was completely at the mercy of the AI's bizarre cognitive characteristics! [Table of Contents] 0:00 I tried making a video editing AI 0:37 AI: "I saw a winning strategy" 2:50 Editing without watching the video is creepy 7:30 Speaker identification isn't working well 13:09 The text on the screen is high quality 14:20 The right edge of the screen keeps going off 18:31 "Measuring" is difficult 24:14 Human-oriented content is high-cost for AI 28:46 AI doesn't understand "humor" 37:28 A video editing AI that's actually usable 41:13 Our valued customer, Surfshark [References] ◯OpenAI "Introducing Whisper" (https://openai.com/index/whisper/) →Confirms an overview of Whisper and that it was trained using multilingual, multitasking voice data from the web. ◯arXiv "Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio" (https://arxiv.org/abs/2501.11378) →Confirmed experimental results showing Whisper hallucinations occurring with non-speech and frequent use of phrases like "thanks for watching." ◯GitHub openai/whisper Discussions "Share your hallucinations here #1873" (https://github.com/openai/whisper/dis...) →Confirmed public reports of "Thank you for watching" appearing with silence or traffic noise. 【Join the supporter community here!】 】 https://yurugengo.com/support 【Parent Channel: Yuru Linguistics Radio】 @yurugengo 【Clip Channel (Official)】 @yuru_kiri 【Physical Store Project: Yuru Gakuto Cafe】 @yurugakuto 【Contact Form】 https://forms.gle/BLEZpLcdEPmoZTH4A *We look forward to receiving your messages! 【For Business Inquiries, Click Here!】 [email protected] 【Horimoto Ken Profile】 Graduated from Keio University, Faculty of Science and Technology. Majored in Information Engineering. Makes a living by creating and distributing theoretical content. Twitter → Twitter: kenhori2 note Magazine → https://note.com/kenhori2/m/m125fc452... Personal YouTube → @kenhorimoto 【Mizuno Taiki Profile】 Born in 1995. From Aichi Prefecture. Graduated from Nagoya University, Faculty of Letters. Majored in linguistics. Works as a magazine editor. Author of books including "Linguistic Analysis of 0.2 Seconds of Conversation" (Shinchosha). Also hosts the podcast "Let's Meet in Jinbocho". Twitter → https://x.com/yuru_mizuno Let's meet in Jinbocho → https://open.spotify.com/show/6cYkvDO... 【Sister Channels】 ◯Yuru Musicology Radio (@yuruongaku) ◯Yuru Folklore Radio (@yuruminzoku) ◯Yuru Astronomy Radio (@yurutenmon) ◯Yuru Philosophy Radio (@yurutetsugaku) #YuruComputerScienceRadio