The lunch briefing.
Mid-day check: Andrej Karpathy joins Anthropic, Google I/O kicks off with AI focus, and a US deepfakes law takes full effect.
AI Talent. The AI industry sees significant talent shifts and model advancements. Former OpenAI co-founder Andrej Karpathy has joined Anthropic's pre-training team, aiming to return to frontier LLM research. Meanwhile, Alibaba's Qwen 3.7 Max preview has surfaced in LM Arena, indicating concurrent iteration of two 72-billion-parameter models. Horizon Robotics also open-sourced HoloMotion-1, a 4-billion-parameter robot cerebellum model capable of 300FPS on edge devices.
AI Governance. Regulatory and ethical concerns around AI are coming to a head, with new laws and real-world incidents. The US "Take It Down Act" is now fully in force, criminalizing the distribution of nonconsensual intimate imagery, including deepfakes. Separately, an AI announcer at a graduation ceremony mispronounced and skipped names, prompting apologies and do-overs. Spotify is also launching verification badges for podcasts to help listeners identify and avoid AI-generated content.
AI Infrastructure. The foundational elements of AI, including infrastructure and security, are under scrutiny. A CISA contractor reportedly leaked highly sensitive government AWS keys on GitHub, raising major cybersecurity concerns. In a novel approach to data center solutions, a sub-sea AI data center is now operating in China, seeking alternatives to land-based facilities. These incidents underscore the critical need for robust security and innovative infrastructure to support the growing demands of AI.
Andrej Karpathy, OpenAI co-founder, joins Anthropic's pre-training team
Andrej Karpathy, a prominent AI researcher and former OpenAI co-founder, has joined Anthropic to focus on large-scale training runs for frontier models. This move is seen as a significant gain for Anthropic in the competitive AI talent landscape.
Google I/O 2026 kicks off with focus on Gemini and agentic AI features
Google I/O 2026 is underway, with expectations for new Gemini AI models and advanced agentic AI features to be central to the keynote. Updates for Android 17, including AI-generated widgets, have already been previewed.
OpenAI boss Sam Altman wins legal battle against Elon Musk
The long-running legal dispute between Elon Musk and Sam Altman has concluded with a victory for the OpenAI boss. Musk has stated his intention to appeal the verdict.
Anthropic adds new privacy and security features to Claude Managed Agents
Anthropic is rolling out two new features for Claude Managed Agents, providing users with enhanced control over security and privacy. These additions aim to strengthen the protection of user data within its AI agent services.
Discussion highlights the high cost of AI development and deployment
A Hacker News discussion thread points to the significant financial burden associated with developing and deploying artificial intelligence technologies. The conversation explores various factors contributing to the expense of AI.
AI announcer mispronounces names at Glendale Community College graduation
An AI-powered announcer at Glendale Community College's graduation ceremony mispronounced and skipped student names, leading to an apology from the college president. Many students were offered a chance for a do-over.
Matrix Super Intelligence unveils MATRIX-3 humanoid robot, targets 100,000 units by 2027
Matrix Super Intelligence, a Chinese company, has revealed its MATRIX-3 full-size humanoid robot, standing 170cm tall and weighing 65kg. The company aims for mass production of 100,000 units by 2027.
Alibaba's Qwen 3.7 Max preview surfaces, indicating dual 72B model iteration
Alibaba's Qwen 3.7 Max preview has appeared in LM Arena rankings, alongside an active Qwen 3 Ultra preview. This suggests the company is simultaneously iterating two flagship 72-billion-parameter models ahead of a public release.
Horizon Robotics open-sources HoloMotion-1, a 4B robot cerebellum model
Horizon Robotics has released HoloMotion-1, a 4-billion-parameter robot cerebellum model capable of real-time inference at 300FPS on edge devices. This marks a significant advance in whole-body humanoid robot control.
US cyber agency CISA contractor exposed sensitive AWS keys on GitHub
A contractor for the US federal cybersecurity agency CISA reportedly leaked highly sensitive government AWS keys and plaintext passwords. The data was found in a spreadsheet uploaded to a public GitHub repository.
Spotify launches verification badges for podcasts to combat AI-generated content
Spotify is introducing verification badges for podcasts, similar to a recent initiative for music, to help listeners distinguish human-created content from AI-generated audio. This aims to reduce the spread of "AI slop" on the platform.
Apple enhances accessibility features with more on-device AI processing
Apple's latest accessibility updates for iPhone, Mac, and Vision Pro incorporate more on-device AI processing. New features include AI-powered generated subtitles for uncaptioned videos and enhancements to VoiceOver and Magnifier.
US "Take It Down Act" now fully in force, criminalizing deepfakes
The US "Take It Down Act," which criminalizes the distribution of nonconsensual intimate imagery including AI-generated deepfakes, is now fully in effect. Experts are debating its effectiveness and potential for censorship.
China's sub-sea AI data center becomes operational
A sub-sea AI data center is now operating in China, representing a novel approach to data center solutions. This project explores alternatives to traditional land-based data centers, potentially offering efficiency benefits.