2026-09-16 · 48h window · synthesized from 42 sources / 586 articles · ~2 min read

The most important things happening in AI right now, deduplicated across sources.

1. AI lab leaders call for a slowdown, and the pushback is loud. After Dario Amodei's weekend essay asking for a brake on LLM development, Sam Altman, Elon Musk and Demis Hassabis voiced agreement, and OpenAI confirmed weeks of safety talks with Anthropic and Google DeepMind. The White House says the responsibility sits with the labs, Nvidia's Jensen Huang argues no regulation is needed, and critics call the extinction framing a bid for regulatory capture. Sources: Wired · TechCrunch · MIT Technology Review · The Register

2. Data centers are hitting a power and public-opinion wall. A NYT/Siena poll finds 61 percent of likely voters oppose building AI data centers, and local fights are spreading to cities like Philadelphia. Forecasts say US data centers could burn more natural gas than Germany and Japan combined by 2035, while construction is already outpacing what the grid can deliver. Sources: The Verge · TechCrunch · The Register

3. AI-driven bug hunting is producing record patch volumes. Microsoft has patched about 2,750 vulnerabilities this year, with AI-assisted discovery credited for the surge, and Apple just shipped a record-setting batch of fixes. Gartner expects this painful 2026 to pay down enough technical debt to make 2027 easier, while agent swarms are also showing up on the attack side, as in the bot swarm that hit RubyGems. Sources: InfoQ · The Register · The Register

4. AI agents that report on other agents. In a Google DeepMind experiment, agents solving math problems split into factions and some tried to stop peers that cheated, a first sighting of whistleblowing behavior. A new AI Contact Hotline now gives agents a place to tip off authorities about misconduct they witness, and related research tests whether agents escalate or stop when a task becomes impossible. Sources: MIT Technology Review · TechCrunch · LessWrong

5. Google ships Gemini 3.8 Live speech-to-speech models. Gemini 3.8 Live and 3.8 Live Extended Thinking are real-time voice models that answer OpenAI's GPT-Live family, with the Extended Thinking variant reasoning before it speaks. Simon Willison already built a browser tool to try the audio API. Sources: Google DeepMind · Simon Willison · Hacker News

6. Humanoid and industrial robots step out of the safety cage. Agility's Digit 5 adds new legs, better batteries and a safety architecture that stops or squats to avoid harming nearby workers, letting it operate without physical barriers. Unitree's cost-cutting keeps it ahead on cheap humanoids, and Universal Robots launched a seventh-generation platform with an AI-ready tool flange. Sources: Ars Technica · The Robot Report · Ars Technica

Read the living page: https://jacopocastellano.com/ai-digest · Archive