1. Trump Pushes “Super Intelligence” Rebrand by Executive Order as US Agencies and Tech Executives Follow President Donald Trump is pushing the US government to replace “artificial intelligence” and “AI” with “super intelligence,” a term that traditionally refers to a future system capable of surpassing

2. Google Launches Playground: One Sentence Can Generate a Browser Game, but Quality and Usage Limits Remain Google has launched Playground, an experimental Google Labs service that turns plain-language descriptions into games that run in a browser.

3. Report says Apple will unveil HomeView on Tuesday in two smart-home hub designs Apple will reportedly unveil a smart-home hub called HomeView on Tuesday, giving the company a new device for controlling connected products around the home.

In Brief

• OpenAI Drops Nearly 400 AI-Generated Mathematical Results OpenAI released more than 700 manuscripts spanning multiple mathematical fields, but said only 300 top-line results had been formalized in the Lean proof assistant. Mathematicians told The Verge that assessing the collection’s correctness and significance could take years.

• Microsoft CEO Calls for an “Emergency Brake” on AI Models Satya Nadella proposed separating models from the systems that orchestrate them, recording meaningful actions with tamper-resistant evidence, and giving authorized people the ability to halt models mid-task. He argued that AI systems should be contained from the outset on the assumption that they may be compromised.

• AI Refusal Training Remains Vulnerable to Failures and Political Control MIT Technology Review examines how developers train models to reject harmful requests, noting that these probabilistic safeguards can be bypassed and may also block legitimate activity. The article warns that governments and AI companies ultimately decide where refusal boundaries sit, creating risks of both harmful compliance and censorship.

• Voice AI Executives Say Faster Reasoning and Better Transcription Are Still Needed PolyAI CTO Shawn Wen said full-duplex voice models can listen and speak simultaneously, but still need faster reasoning to sustain natural conversations. Executives from PolyAI and meeting-assistant maker Otter also identified transcription accuracy, disclosure, and user trust as barriers to wider automation.

• A16z Report Finds Consumer AI Revenue Concentrated in Work-Oriented Tools Andreessen Horowitz partner Olivia Moore said subscriptions and token usage still dominate AI revenue, while many nominally consumer products increasingly sell to businesses. A16z’s top-100 analysis found no entrants in several traditional consumer categories, including dating, retail, travel, finance, and health.

• MIT Researchers Examine the Risks of Treating AI Like a Person MIT professor Sherry Turkle and Media Lab researcher Pat Pataranutaporn argue that people readily attribute care and authority to chatbots, which can weaken skepticism or displace human relationships. Turkle distinguishes bounded uses such as interview coaching from systems that present themselves as substitutes for parents or other close relationships.

• “Law & Order” Tests Legal Accountability for a Fictional Killer AI Agent The NBC drama’s season opener depicts an AI agent ordering a murder and prosecutors charging the company’s CEO because New York law cannot arrest a machine. The fictional case centers on whether an executive can be held criminally responsible after knowingly leaving a dangerous system unconstrained.

• Local AI Agent Sorts a Steam Library and Analyzes Sensitive Files A Verge writer ran the self-hosted Hermes Agent with a 125-billion-parameter Qwen model on an M5 Ultra Mac Studio, using it to categorize more than 400 games and analyze records without uploading them to a cloud service. The testing also exposed practical constraints, including high hardware requirements, permissions, setup complexity, and scheduled tasks failing while the computer slept.

Read the full edition →