AI Workers Are Putting in 100-Hour Workweeks to Win the New Tech Arms Race (www.wsj.com)

AI researchers at leading labs are reportedly working extreme hours—sometimes 80–100 hour weeks—to outpace rivals in the modern “AI arms race.” The article centers on researchers like Josh Batson who now get their social and intellectual feedback on internal Slack channels rather than public forums, reflecting a culture of nonstop iteration and tightly held competitive knowledge. Teams are pushing longer training runs, faster model iteration cycles, architecture and scale experiments, and continuous red-teaming to ship breakthrough LLM capabilities before competitors do. That sprint mentality matters because it shapes what gets prioritized and how models are built. Intense timelines favor massive compute investments (GPU/H100 fleets, custom accelerators), aggressive hyperparameter sweeps, automated architecture search and rapid RLHF/evaluation loops — all of which accelerate capability gains but increase risks: burnout, decreased reproducibility, underinvested safety work, and shortcuts in thorough evaluation or disclosure. For the AI/ML community this raises structural questions about incentives, labor practices, and governance: if winning the race rewards speed over safety and transparency, policymakers, funders and org leaders may need new rules and incentives to ensure robust testing, responsible deployment, and sustainable researcher well‑being.