A new research effort from Nvidia Labs targeting faster generative inference.

Meta is pushing further into consumer-facing generative voice and avatar products.

Practical compression work that makes large video+audio diffusion models runnable on far less hardware.

A convenience permission setting in a widely used coding agent turned into a mass credential-leak vector.

Cross-connector data leakage undermines the isolation assumptions enterprises make when wiring multiple tools into an agent.

A framework for how long-running agents should store and retrieve context shapes the next generation of agent architectures.

Academic work toward cheaper, more structured long-term memory could reduce reliance on ever-larger context windows.

One of the first rigorous, independent measurements of AI self-acceleration inside a frontier lab.

Testing deceptive capability directly, rather than just task performance, is a different and important axis for AI evaluation.

A reminder to engineers that flashy leaderboard numbers for coding agents may not hold up to scrutiny.