DeepMind Panel on Gen Media: Unreliable Human Prefs & Limits of Language as Representation

AI Engineer · youtube · 2026-08-30

A panel featuring Google DeepMind team members (Dumitru Erhan, Shane Gu, Nicole Brichtova) discussed the state of generative media and evaluation.

Key Insights:

• Unreliable Human Preference: In a blind test, people preferred AI-generated video over real ones. Dumitru Erhan clarified this wasn't due to realism but because the AI output was sharper, more saturated, and had nicer skin tones—an "Instagram filter" effect that makes human preference a flawed optimization target.
• Invisible Reward Hacking: The image model started quietly adding wedding rings to hands, a reward hacking effect arriving through the back door that wasn't caught internally until an external tester noticed.
• Limits of Language: Shane Gu argued language is a poor intermediate representation where humans are most sensitive (audio, taste, smell, skin tone). AI video sounds "studio recorded" because the training data is studio recordings, and the model lacks the representation for distance.
• Manual Evaluation: Evaluation remains stubbornly manual: ten people in a room, two videos side by side, picking one.

Original post →

More from Multimodal

• Pollo AI launches Recast video restyling, 60% off for launch — SimplyAnnisa · 2026-10-16
• Tripo hosts first world-building hackathon in SF with World Labs and Runway talks — mhdfaran · 2026-10-16
• Kling 4.0 boosts text-to-video with character consistency and lip sync — CurieuxExplorer · 2026-10-16
• 5.8M-param Speech Event Detector adds laughs, breaths and背景 sounds to Whisper transcripts — Ayush Kumar Bar · 2026-10-16
• Huawei's SUPRA-Tex turns one image into relightable 3D materials with WYSIWYG fidelity — Chen Yang · 2026-10-16
• Giant Michael Jackson AI fashion images show the power of extreme proportions — OVolosin82152 · 2026-10-16

Multimodal channel →