We release WildAI: 83B tokens with AI, topic and format labels, plus all 800 models and code. 📄 arxiv.org/abs/2609.40295 💻 github.com/pangramlabs... Joint work with Ben Glickenhaus, @kthai1618.bsky.social, John Wieting, @miyyer.bsky.social, @max.pangram.com, and @bradley_emi
- Pangram labeled 27.5% of June 2026 web tokens as AI-generated, rising to 31.1% by August, after FineWeb quality filtering.
- For high-human-budget runs AI tokens raise loss almost immediately; data-starved runs see an early benefit that saturates and reverses.
- The authors' new scaling law predicts effects on models up to 3.6x larger with 41% lower error than existing laws across all AI ratios.