- Google DeepMind is losing AlphaFold lead John Jumper to Anthropic. Jumper shared the 2024 Nobel Prize in Chemistry for AlphaFold, so his move adds to a highly visible wave of senior AI talent leaving Google.
- OpenAI is adding Noam Shazeer and Dean Ball as it prepares for a possible IPO. Shazeer is a Transformer co-author and former Gemini co-lead; Ball will lead a new Strategic Futures team focused on frontier AI policy and internal governance.
- Anthropic’s Fable 5 and Mythos 5 remain caught in a U.S. government ban story. TechCrunch frames the episode around developer impact, Anthropic’s brand, and the broader politics of model safety and market positioning.
- Artificial Analysis’ AA-Briefcase benchmark shows how far models still are from real knowledge work. Even the top performer, Claude Fable 5, fully satisfied all criteria on only 3% of tasks built from messy Slack threads, emails, meeting transcripts, and large exports.
- OpenAI researchers report that small amounts of “beneficial trait” reinforcement learning can make models safer. Training on traits such as truthfulness, humility, corrigibility, fairness, and concern for human well-being improved 44 of 53 evaluations and made harmful steering less effective.
- ServiceNow’s MosaicLeaks benchmark highlights a privacy problem in deep research agents. When agents combine private documents with external web search, their outbound queries can leak fragments of sensitive information; optimizing only for task success can make that leakage worse.
- Hugging Face argues that software should be benchmarked for agentic use. The point is not just whether an agent eventually gets the right answer, but how many tokens, tool calls, seconds, and detours it needs—and whether better CLIs, Skills, and examples help.
- The Hugging Face PEFT team says LoRA is dominant, but not automatically the best choice. Their recommendation is to compare parameter-efficient fine-tuning methods under the same conditions and track not only accuracy, but also VRAM, runtime, checkpoint size, and forgetting.
Bottom line: today’s AI news is less about a single breakthrough model and more about the control layer around AI—talent, governance, safety, privacy, and whether agents can actually do useful work reliably.