Google DeepMind shipped Gemini 4 Neon, a cheaper member of the Gemini 4 family aimed at long-running coding and agent work, and opened up Project Genie’s Genie 4 world model. OpenAI published two very different pieces: a detailed engineering walkthrough of how the Codex sandbox constrains the commands agents run on your machine, and a lecture-adapted essay by Dean Ball on whether humans keep the wheel as AI systems take over more decisions. Meta’s engineering team describes how agents now carry most of the work of porting ML improvements across its ads ranking fleet. The security section covers a decade-old NGINX HTTP/2 flaw that reaches code execution without authentication, and the hijack of Microsoft’s agent-framework PyPI package into a credential stealer.
From the labs
- Introducing Gemini 4 Neon — Google DeepMind
- Genie 4 is now available — Google DeepMind
Worth your time
- Inside the Codex sandbox — OpenAI
- The Silent Consensus — Dean Ball
- How Meta Uses AI Agents to Scale ML Innovations Across Ads Ranking Models — Meta Engineering