OpenAI's Math Breakthrough, Meta's Employee Tracking, and New Codex Upgrades

OpenAI's Math Breakthrough, Meta's Employee Tracking, and New Codex Upgrades

OpenAI's reasoning model disproves an 80-year-old math conjecture, Meta faces backlash over employee tracking for AI training, and OpenAI updates Codex with new agentic capabilities.

Research & Breakthroughs

OpenAI Disproves 80-Year-Old Math Conjecture

OpenAI's reasoning model has successfully disproved the Erdős unit distance conjecture, a discrete geometry problem dating back to 1946. Unlike previous attempts that relied on literature searches, this proof was generated by a general-purpose reasoning model and has been verified by external experts, including Princeton mathematician Will Sawin. The proof utilizes algebraic number theory to demonstrate an infinite family of point arrangements that exceed previous grid-based expectations.

This milestone is significant as it demonstrates AI's ability to make original contributions to complex scientific fields, moving beyond mere benchmark performance. The result suggests that general-purpose models are becoming capable of autonomous scientific discovery, a shift that could have implications for fields ranging from physics to engineering.

Google Advances AI in Scientific Discovery

Google has published research in Nature detailing its 'Co-Scientist' tool, which utilizes Gemini to automate hypothesis generation in biology labs. The system employs a 'tournament of ideas' approach, where agents propose, critique, and rank hypotheses to identify promising leads. In a recent Stanford liver-fibrosis project, the tool reportedly identified a drug lead that significantly reduced lab-tested scarring signals.

This development highlights Google's strategy of targeting the scientific method layer of AI, leveraging its extensive stack of specialized databases and models like AlphaFold. The company is currently rolling out access to the tool for individual scientists, aiming to accelerate the pace of discovery in complex biological research.

Industry & Business

Meta Faces Backlash Over Employee Tracking

Leaked audio from a Meta all-hands meeting has revealed that the company monitored employee activity across internal tools, including Gmail, GChat, and VSCode, to train its AI models. Mark Zuckerberg explained that the AI learns by observing 'really smart people' at work. This disclosure coincided with the company's announcement of layoffs affecting 8,000 employees.

The juxtaposition of using staff as training data for their own replacements has sparked significant controversy. This incident highlights the growing tension between corporate productivity monitoring and workforce stability, raising questions about whether such data collection practices will become a template for other companies in the industry.

SpaceX and Anthropic Ink $1.25B Monthly Deal

SpaceX's public S-1 filing has revealed a major infrastructure deal: Anthropic is paying the company $1.25 billion per month through May 2029 for access to compute capacity across SpaceX's Colossus and Colossus II data centers. This agreement underscores the massive capital expenditure required for frontier AI development and the critical role of specialized compute providers.

This deal is part of a broader trend of AI labs securing long-term, high-capacity compute partnerships to maintain their competitive edge. As the demand for training and inference resources continues to scale, such massive infrastructure commitments are becoming a defining feature of the AI industry's financial landscape.

Product & Tools

OpenAI Updates Codex with Agentic Features

OpenAI has introduced a new wave of upgrades for its Codex agentic assistant, focusing on enhanced desktop and web interaction. New features include 'Appshots,' which allows Mac users to attach open app windows to a thread, and a 'Goal mode' that enables the agent to work toward persistent objectives over extended periods. Additionally, the update includes 'Locked computer use,' allowing the agent to operate even when the device is locked.

These tools are designed to improve the agent's context awareness and ability to handle complex, multi-step tasks. By enabling Codex to interact more natively with desktop environments, OpenAI aims to close the gap with competitors and improve the utility of its agentic assistants for developers and power users.

Policy & Regulation

California Issues Executive Order on AI Workforce

California Governor Gavin Newsom has signed an executive order directing state agencies to develop policies to protect workers from AI-driven job displacement. The order mandates the launch of a dashboard to track AI's impact on employment within 90 days and requires agencies to review potential updates to the WARN Act for faster layoff notifications.

This move comes as the state, home to a significant portion of the world's top AI companies, faces increasing pressure to address the economic implications of rapid AI adoption. The order also tasks agencies with reviewing how unions are negotiating AI adoption and exploring ways to direct AI revenue toward public benefit, marking a significant step in state-level AI regulation.