Frontier LLMs solve fewer than one-third of 87 multi-GPU CUDA benchmark tasks, though some generated kernels still outperform public reference implementations.
Post-training migrates from monolithic RL pipelines to decentralized specialist systems merged through on-policy distillation into a generalist student—a scaling pattern that resolves capability conflicts across domains.
New AI models can apply the same technical capabilities to either cybersecurity patching or attacks on critical infrastructure – countries must now invest in defensive measures.
Anthropic offers Fable 5, a Mythos variant with safety filters for public use, while Project Glasswing participants gain access to less restricted Claude Mythos 5, accompanied by new federal rules controlling frontier AI models.
