Import AI 445: Timing superintelligence; AIs solve frontier math proofs; a new ML research benchmark

Import AI · 5 months ago

Facebook developed Kunlun, an optimized recommendation system that improves computational efficiency for ad targeting, and discovered scaling laws that predict performance gains predictably across different compute investments. The system achieved 37% Model FLOPs Utilization on NVIDIA B200 GPUs, up from 17%, and delivered a 1.2% improvement in Meta's ad metrics when deployed. With predictable scaling laws established, Meta can now invest additional computational resources in recommendation systems with greater confidence in the returns, similar to how scaling laws function for language models.