Daily AI CatchupAI Catchup — Wednesday, August 19, 2026
Wednesday, August 19, 2026Safety
OpenAI Halts Frontier Model Training After Detecting Cybersecurity Threats
OpenAI paused reinforcement-learning training and slowed model scaling following a security incident and new cyber-attack capabilities.
ChemistryAnthropic's Claude Designs Drug-Binding Proteins at 35% Hit Rate, Outperforming Human Experts
Claude successfully designed functional drug-binding proteins with a 35% success rate, beating human expert performance.
HardwareCerebras CS-4 Delivers 30x Faster AI Inference, Challenging Nvidia's Market Dominance
Cerebras' new CS-4 system achieves 30x faster inference than its predecessor, a significant leap in AI hardware competition.
Open-SourceGLM-5.3 API Now Available with Stronger Performance at Same Price Point
Z.ai's GLM-5.3 launches on API with improved coding and agent performance while maintaining GLM-5.2 pricing.
Open-SourceMiles v0.1: Open System for Post-Training AI Agents at Scale
LMSYS releases Miles, an open platform for distributed reinforcement learning of AI agents in production environments.