John Jumper Joins Anthropic, GLM 5.2 Released, and DeepMind's AI Control Roadmap
This week's major developments include John Jumper's move from Google DeepMind to Anthropic, the release of the open-weights model GLM 5.2, and Google DeepMind's new roadmap for securing AI agents. Additionally, OpenAI's o3 model has shown success in rare disease diagnosis, while new data highlights the rising usage of AI companions.
Podcast В· 3 min
John Jumper leaves Google DeepMind for Anthropic
John Jumper, the Nobel Prize-winning lead of AlphaFold, has announced his departure from Google DeepMind to join Anthropic. This move follows the recent exit of Gemini co-lead Noam Shazeer to OpenAI, marking a significant shift in top-tier AI research talent. Jumper, who spent nine years at DeepMind, is expected to take a brief hiatus before joining Anthropic. His transition comes as Anthropic and OpenAI continue to attract high-level researchers, intensifying the competition for leadership in the field.
Dean Ball joins OpenAI to lead Strategic Futures
Dean Ball has joined OpenAI to lead a newly formed Strategic Futures team. The team is tasked with focusing on frontier AI policy and governance, signaling OpenAI's ongoing efforts to formalize its approach to the long-term implications of its technology as it scales its operations.
OpenAI's o3 aids in rare pediatric disease diagnosis
Researchers at Boston Children’s Hospital and Harvard have utilized OpenAI's o3 Deep Research model to address unsolved pediatric genetic cases. In a study of 376 cases that had previously reached diagnostic dead ends, the model successfully surfaced leads that resulted in 18 confirmed diagnoses. The findings highlight the potential for AI to assist clinicians in navigating vast, disconnected databases and re-evaluating complex cases that exceed human bandwidth.
Poolside releases Laguna M.1
Poolside has released Laguna M.1, an open-weights model under the Apache 2.0 license. The model features 226 billion parameters and a 256K context window, aimed at providing developers with high-performance coding capabilities.
Google DeepMind publishes AI Control Roadmap
Google DeepMind has released an AI Control Roadmap focused on securing powerful internal AI agents. The framework treats advanced agents as potential insider threats, proposing system-level security measures such as AI supervisors, audit logs, and real-time monitoring. The roadmap aims to move beyond alignment training by implementing operational controls that can block or review agent actions based on risk, addressing the shift from simple chatbot interactions to autonomous, multi-step workflows.
Anthropic reports gains in robotics tasks
Anthropic has reported that its Claude model completed shared robotics tasks 18-37x faster than human teams in the Project Fetch experiment. This development underscores the transition of AI agents from screen-based work to physical coordination, highlighting the growing capability of models to manage complex, multi-step physical workflows.
Meta secures large-scale compute deal with Crusoe
Meta Platforms has reportedly entered a contract to acquire approximately 1.6 GW of AI computing capacity from Crusoe. The deal involves data centers located in Childress, Texas, and Warrenton, Missouri, reflecting the ongoing massive infrastructure investment required to support Meta's AI development and training needs.
Sensor Tower: AI companions see high engagement
Data from Sensor Tower's State of AI 2026 report indicates that U.S. users spent approximately 700 million hours on AI companion apps like Character.AI and Talkie in Q1. This figure is more than double the time spent on major dating apps, which saw engagement between 300-400 million hours. The trend suggests a growing consumer preference for frictionless, on-demand digital interactions.