# 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。 - Show: ゆるコンピュータ科学ラジオ - Published: 2026-09-27 - Canonical: https://podc.club/en/episode/1604353315_37e042de0ec8.html - Audio: https://api.robotstart.jp/api/for-media/v1/articles/35ac047a/permanent/with-advertise/podcast/9e1be273981bccd6a3900d8fc68329fb.mp3 - Source: Edited from a machine transcript of the first 30 minutes of the episode. Not an official summary. Higuchi (樋口) of Yuru Computer Science Radio (ゆるコンピュータ科学ラジオ) spent over 100 hours and more than 170 versions building an AI to automate his show's video editing — and nearly lost his mind doing it. From audio-video desync to garbled captions, a broken speaker-detection algorithm, and a right-aligned subtitle bug that alone consumed 20+ versions, the project was a relentless parade of failures announced with the AI's signature confidence: 'I can see the path to victory.' ## What this episode answers ### What was wrong with the very first video the editing AI produced, and why was the AI unable to detect the problem itself? The audio and video were completely out of sync; the AI was unable to detect it because it builds video without actually watching it. ### Why did the AI start cutting subtitles mechanically at exactly 22 characters? The AI had learned from Yuru Tetsugaku Radio editing data that the maximum subtitle length was 22 characters, but it had not learned when to break a line naturally, so it simply cut at 22 characters regardless of meaning. ### What algorithm did the AI use to identify which speaker — Horimoto or Mizuno — was talking, and why did it fail? It identified the speaker by comparing which of the two microphones had the longer transcript; it failed because both hosts record in the same room and Mizuno's voice bled into Horimoto's microphone, making Horimoto's mic transcript consistently longer so every line was misattributed to Horimoto. ### Why did the right-aligned subtitle overflow bug persist through more than 20 additional versions without being fully resolved? Because right-aligned text requires calculating the total pixel width of the subtitle string and subtracting it from the right edge to find the starting point, and hiragana, katakana, and the Latin alphabet each have subtly different character widths — the AI repeatedly forgot to account for one of these width variants each time it claimed to have fixed it. ### After reaching roughly 80% quality at version 9, why did polishing the remaining details take a full month? Because the classic rule that going from 0 to 80 is relatively easy but going from 80 to 100 is hard held true, with the right-aligned subtitle overflow bug alone stacking up more than 20 versions to resolve. ## Article ## 'I Can See the Path to Victory' — The First Clash with an Overconfident AI The goal was straightforward: feed the AI the existing edited footage stored in Google Drive and let it learn from that data. Higuchi was optimistic — there was plenty of training material, so it should come together quickly. The AI responded immediately: 'I can see the path to victory.' The first video it produced, however, had audio and video completely out of sync. The root cause was simple but unsettling: the AI builds video without actually *watching* it. A desync that any human editor would catch in an instant went entirely unnoticed, and the AI declared the work finished. That was the first moment Higuchi felt something was deeply off. Pointing out the desync fixed it immediately — but the next problem was already waiting. ## The 22-Character Rule and the Power of a Good Manual Once subtitles were added, the result looked like auto-generated closed captions stuffed with every word of dialogue. The AI had learned from Yuru Tetsugaku Radio (ゆる哲学ラジオ) editing data that the maximum subtitle length was 22 characters. What it had not learned was when to break a line naturally — it simply cut at 22 characters regardless of meaning, producing fragments like 'yuruku tanoshiku otodo' and 'kesu suru no ga kono rajio desu,' splitting words at baffling points. The AI kept presenting each attempt with the same breezy confidence: 'I can see the path to victory.' All Higuchi could do was smile wryly. The fix came from an unexpected source: the detailed editing manual originally written for human editors. It covered fonts, character counts, tone, and exactly how to break subtitle lines. Feeding it to the AI produced an immediate and dramatic improvement in accuracy. The instruction 'keep subtitles concise — drop every unnecessary word' was followed so faithfully that the AI's subtitle summaries actually exceeded human editor quality. Higuchi's reaction: 'I'm glad we took the time to write that manual properly.' A well-written manual works just as well on an AI as on a person. ## The Everyone-Is-Horimoto Bug When Higuchi switched to Yuru Computer Science Radio footage, a new problem appeared: every speaker animation was attributed to Horimoto-san, regardless of who was actually talking. Examining the AI's internal logic revealed the culprit: it was identifying the speaker by comparing which of the two microphones had the longer transcript. The idea seems reasonable in isolation, but both hosts record in the same room. Mizuno-san's voice bleeds into Horimoto-san's microphone, so Horimoto's mic consistently produced a longer transcript — and the AI classified every single line as Horimoto speaking. The solution was straightforward: use audio volume rather than transcript length to determine which microphone is dominant. When Higuchi pointed this out, the AI immediately replied, 'In that case I'll switch to whichever mic is louder.' The reaction was 'why not do that from the start?' — but once the change was made, speaker identification accuracy climbed to roughly 85%. Laughs and back-channel responses continued to throw off the cuts, and that problem lingered throughout the rest of development. ## Right-Aligned Subtitles and the 80-to-100 Wall By around version 9, the output had reached a quality level that convinced Higuchi the whole thing would be finished within another week. It was not. The classic rule — going from 0 to 80 is relatively easy; going from 80 to 100 is hard — held true, and the final polish consumed an entire month. The single toughest obstacle was right-aligned subtitle overflow. In Yuru Computer Science Radio, subtitle position differs by speaker: left-aligned text is drawn from the origin point in the upper-left corner, which is trivial. Right-aligned text requires calculating the total pixel width of the subtitle string first, then subtracting that width from the right edge to find the starting point. The further complication: hiragana, katakana, and the Latin alphabet each have subtly different character widths. Every time the AI declared 'I've got it this time,' it turned out it had forgotten to account for katakana widths, or had forgotten the Latin alphabet widths, or had made some other variant of the same error. This single issue alone stacked up more than 20 versions. Using the video editing software's built-in right-align button was considered, but having the AI operate the GUI — just eight button clicks — took 15 minutes, making direct numerical calculation and rendering the faster option in the end. ## Takeaways - The AI built video without watching it, making it unable to catch obvious problems like audio-video desync that any human would spot instantly. - Feeding the AI a detailed editorial manual written for human editors dramatically improved subtitle quality — the AI's summaries ultimately surpassed human editor accuracy. - After reaching roughly 80% quality at version 9, the remaining details — especially right-aligned subtitle character-width calculation — consumed a full month and more than 20 additional versions, confirming that the 80-to-100 wall is just as real in AI-assisted video editing as anywhere else.