Microsoft's Humanist AI Rules; Trump and China Reject Slowdown
Microsoft AI has published a draft Code of Conduct for future models, while US President Trump and the Chinese government have both dismissed calls for an AI development slowdown. Meanwhile, OpenAI has acquired Glass Imaging, and Google has expanded Claude access to its engineers.
Podcast В· 3 min
Microsoft AI drafts 'Humanist' rulebook for future models
Microsoft AI has published a draft Code of Conduct for its future MAI models, outlining a vision where AI acts as a subordinate tool rather than an independent agent. The rules mandate that models must remain within authorized task scopes, avoid claiming consciousness or feelings, and obey shutdown commands. This document serves as a roadmap for development through 2027 and is currently open for public comment.
Trump and China dismiss AI slowdown proposals
US President Donald Trump and the Chinese government have both rejected recent calls from frontier AI labs to slow down development. Trump dismissed the concerns as a 'hoax' and argued that a 'High IQ' president is the only guardrail needed. Simultaneously, China's Foreign Ministry labeled the slowdown proposals as 'fear-mongering' and a 'Cold War playbook' for AI, signaling a lack of international consensus on pacing.
OpenAI acquires Glass Imaging
OpenAI has acquired Glass Imaging, a startup founded by former Apple engineers, for over $300 million. The company specializes in using AI to enhance smartphone camera performance, including technologies like Portrait Mode. This acquisition suggests OpenAI's continued interest in integrating its models into hardware and consumer imaging applications.
Superhuman acquires Fathom
Email productivity platform Superhuman has acquired Fathom, an AI-powered meeting notetaker. The deal allows Superhuman to integrate meeting transcripts, decisions, and action items directly into its email, calendar, and agent stack, further expanding its capabilities as an all-in-one productivity tool.
Google expands Claude access to all engineers
Google has granted all of its engineers access to Anthropic's Claude model through its internal 'Antigravity' system. While Gemini remains the default model for Google's operations, this move represents a significant concession, allowing engineers to leverage Claude for coding and other tasks.
OpenAI contractors reviewed real ChatGPT conversations
404 Media reported that OpenAI contractors working on 'Project Lily' reviewed real ChatGPT prompts, including sensitive user conversations. The project aimed to train models against sycophancy, highlighting the ongoing reliance on human review for model alignment and safety training.
Nvidia, Palantir, and Booz Allen limit Fable use
Major firms including Nvidia, Palantir, and Booz Allen Hamilton are reportedly limiting the use of Anthropic's Fable model for sensitive work. The restrictions stem from concerns over Anthropic's 30-day usage log retention policy, leading some clients to prefer private server alternatives.
MIT's 'HardFlow' method improves AI safety
MIT researchers have introduced 'HardFlow,' a new method that allows generative models to explore solutions freely before enforcing hard constraints on the final output. The technique has shown success in robotics, navigation, and image editing, providing a way to ensure safety without compromising model performance during the generation process.
TSA's Ace agent handles 100,000 monthly conversations
The TSA's 'Ace' agent, powered by Salesforce, is now managing approximately 100,000 traveler conversations per month. Salesforce reports that 96% of routine inquiries are resolved without the need for human escalation, demonstrating the effectiveness of agentic workflows in public sector customer service.
Reward AI launches OM-1 robot policy
Reward AI has launched OM-1, a robot policy trained directly from human manipulation data. The model has demonstrated the ability to transfer skills across various hardware platforms, including tabletop arms, industrial arms, and humanoids, marking a step forward in generalized robot control.