Google DeepMind Loses Two Top AI Researchers to Rivals in One Week

Nobel laureate John Jumper leaves Google DeepMind for Anthropic AlphaFold co-creator and 2024 Nobel Prize winner John Jumper announced his departure from Google DeepMind to join Anthropic, marking the second high-profile exit in days after Gemini co-lead Noam Shazeer left for OpenAI. The moves intensify the talent war among frontier AI labs as they race to build next-generation systems. 12

Noam Shazeer, transformer co-inventor, leaves Google for OpenAI ahead of IPO Shazeer, a co-lead of Google's Gemini models and co-author of the seminal 'Attention Is All You Need' paper, departed less than two years after Google paid $2.7 billion to bring him back from Character.AI. His move to OpenAI comes as the company prepares for a public offering, signaling a major talent coup. 34

OpenAI hires former Trump White House AI policy official Dean Ball Ball, who helped publish America's AI Action Plan, will lead a new 'Strategic Futures' team at OpenAI focused on catastrophic risk, recursive self-improvement, and frontier AI policy. The hire bolsters OpenAI's policy credentials as it nears an IPO. 3

OpenAI's o3 model solves 18 rare disease cases that specialists missed A study in NEJM AI found that OpenAI's o3 Deep Research model reanalyzed 376 previously unsolved pediatric rare disease cases and established diagnoses in 18 of them, yielding a 4.8% diagnostic uplift after human experts had failed. The model revisited stored genomic data that had been deemed uninterpretable. 5

OpenAI research shows RL on beneficial traits generalizes alignment gains OpenAI published a paper demonstrating that reinforcement learning on realistic scenarios targeting traits like honesty, corrigibility, and metacognitive transparency produces broad improvements across dozens of alignment benchmarks, with gains persisting under adversarial pressure and generalizing beyond training domains. 6

Anthropic's Claude Opus 4.7 autonomously outperforms human teams in robotics tasks In an update to Project Fetch, Claude Opus 4.7 operating without human assistance was about 20 times faster than the fastest human team at completing robodog tasks from a year ago. The model still struggled with precise physical manipulation, but the results show a pattern of AI progressing from assisting humans to acting independently. 7

Google DeepMind publishes layered security plan to contain rogue AI agents Google DeepMind's 35-page technical report outlines a security framework that treats AI agents as potential insider threats, borrowing from cybersecurity practices like insider-threat prevention. The plan acknowledges that alignment may never be fully solved and adds defense-in-depth to catch adversarial behavior. 8

Anthropic launches live Artifacts in Claude Code for shareable visual pages Anthropic introduced Artifacts in Claude Code, enabling users to generate interactive, shareable web pages such as PR walkthroughs and dashboards that update in real time from coding sessions. The feature is in beta for Team and Enterprise plans and targets engineering and SRE workflows. 9

OpenAI adds credit usage analytics and spend controls for ChatGPT Enterprise New admin tools provide granular breakdowns of credit consumption across users, products, and models, along with flexible spend limits. The updates help enterprises track adoption patterns, manage costs, and distinguish valuable usage from activity needing review. 10

Anthropic's Claude experiences brief global outage, services restored within an hour Users reported blank responses and connection errors across regions on June 18, with over 2,100 outage reports at peak. Anthropic identified and resolved the issue within about 45 minutes, attributing it to growing demand pressures on its compute infrastructure. 11