AI agents shift infrastructure to continuous learning loops

Cognition AI's Devin now supports the full software development lifecycle, from planning and writing code to responding to production issues. Cognition runs always-on training across distributed data centers spanning multiple countries, targeting 99.99% GPU uptime to prevent entire training runs from failing. Early access to Nvidia's Vera Rubin platform lets researchers adapt model architectures for better price-performance. CoreWeave's newly announced Forge platform connects inference, observation, data curation and reinforcement learning, including a hot-loading service that updates model checkpoints in live deployments without redeployment.