Test finds frontier models poor at labeling ASMR sounds

A developer had eight frontier models label 480 ASMR clips; the best, Gemini 3.1 Pro, reached only 0.325 F1 on trigger sounds. Physical triggers like paper rustling were nearly missed, while whispering scored 0.70–0.79. The author advises checking label learnability before training on model-generated tags.

• Survey finds leading AI models converge on the same "controversial" take about humans2026-10-22
• Claude Opus 5.5 and GPT-6 Launch Spark Price War2026-09-23
• Meta's 816B Muse Spark Open Scores 39; 180B Qwen3.8-Flash-Next Scores 402026-10-23
• Anthropic's Jack Clark tells policymakers the singularity has begun2026-10-22
• Claude Opus 5.5 vs GPT-6 Astra: new AI models explained2026-10-20