Why AI Viral Score Fail
Ok-Masterpiece-7614 · reddit · 2026-07-12
The author's team built a model to generate "viral scores" for short videos, aiming to optimize content based on predictions.
However, a late-night review revealed that the score barely changed when titles, edits, or subtitles were altered. Because the model primarily looks at faces, motion, and audio, it fails to see:
- Curiosity-inducing "hooks"
- Subtitle information (many watch on mute)
- Character dynamics, storytelling, and "shareability" factors
The team ultimately decided to use it as an early filter to weed out obviously bad clips, rather than a creative target. Real judgments still rely on user reactions and post-launch data. The author warns that once you start "writing content for the score," you optimize for the wrong things.
More from Apps
- Karpathy says a 10-minute voice ramble can make LLMs understand your intent better — karpathy · 2026-07-22
- Grok launches Grok for Excel, letting users build financial models inside spreadsheets — azed_ai · 2026-07-22
- Grok launches inside Outlook with email drafting, calendar management, and connectors — SpaceXAI · 2026-07-22
- OpenAI Codex & ChatGPT Work Agents Hit 10 Million Users — Polymarket · 2026-07-22
- YC-backed TrustAI says agents made unauthorized changes in production systems — ycombinator · 2026-07-22
- OpenAI Codex and ChatGPT Work users are getting a usage reset in the next hour — kevinkern · 2026-07-22