Today's window is dominated by the tension between AI's runaway momentum and the guardrails starting to appear around it: the White House reportedly asked OpenAI to delay its next frontier model launch, while Axios flags water (not just energy) as the next AI data-center flashpoint and MIT Tech Review maps a whole new 'web data infrastructure' layer emerging to feed models. On the builder side, it's a strong day for small and open models and better plumbing - Liquid AI shipped a tiny 230M edge model, DeepReinforce open-sourced a new coding model family, and Vercel and Hugging Face both shipped tooling to make serving and orchestrating LLM agents easier. Anthropic and independent researchers are also pushing on the softer problem of agent reliability - how to build human-agent teams and better training data - just as a reminder landed that agentic frontends are still very hackable, via a $3M Polymarket exploit. For engineers, the throughline is: model capability and tooling keep compounding fast, but energy, water, security, and governance are quickly becoming the real bottlenecks.

• White House reportedly asked OpenAI to delay its next frontier model deployment
• Water joins energy as a binding constraint on AI data center buildout
• Liquid AI ships tiny 230M-parameter LFM2.5 model for edge devices
• DeepReinforce open-sources Ornith-1.0 coding model family
• Polymarket frontend hack drains $3M via compromised third-party vendor