John Jumper, who shared the 2024 Nobel Prize in Chemistry for his work on AlphaFold, announced this week that he is leaving Google DeepMind after nearly nine years to join Anthropic. Jumper led the AlphaFold team just six months after finishing his PhD, and the system he built has since predicted the structure of more than 200 million proteins, reshaping biological and medical research worldwide.
The move lands days after Noam Shazeer, co-lead of Google's Gemini models, left for OpenAI — two senior departures in quick succession that underline how thin the pool of elite AI researchers has become. Analysts say smaller, faster-moving labs like Anthropic and OpenAI can now offer a kind of focused, low-bureaucracy environment that's increasingly hard for larger firms to match.
For Anthropic, landing one of the architects behind AlphaFold strengthens its scientific bench at a moment when the company is otherwise consumed by a regulatory fight over export controls on its frontier models. A Nobel-winning scientist choosing to join anyway is being read across the industry as a vote of confidence.
GM's Factory Zero plant in Detroit cut more than 1,000 workers to indefinite layoff, then installed roughly 50 new Fanuc-built cobots on the same assembly line — machines that work cage-free, right next to people, attaching body panels as vehicles move down the track. UAW Local 22 has filed formal grievances over both job displacement and safety.
The economics are blunt: after the 2023 UAW contract, GM estimated labor costs would rise roughly $500 per vehicle, and a Wayne State University professor notes that assembly already takes 50–70% fewer labor hours than it did in the 1980s. Local 22's president called the timing "a gut punch."
GM frames the cobots as a safety and ergonomics upgrade. The UAW calls it the opening move of a fight that will shape contract talks in 2028 — and a preview of how automation and AI are about to collide with organized labor well beyond the auto industry.
Z.ai (Zhipu AI) released GLM-5.2, a 744-billion-parameter mixture-of-experts model with a 1-million-token context window and fully open, MIT-licensed weights on Hugging Face — meaning unrestricted commercial use, no registration, no usage caps. It now ranks first on the independent Artificial Analysis Intelligence Index among open-weight models.
The model's architecture trick, IndexShare, reuses sparse attention indexers across transformer layers to cut compute costs at long context lengths while holding accuracy. Independent benchmarks place it competitive with GPT-5.5 on coding tasks at a fraction of the price — though the full weights run to roughly 1.5TB, so "running it locally" still means a serious GPU budget.
The release lands at a pointed moment: it shipped within days of a US export-control directive that pulled Anthropic's Mythos- and Fable-class models from foreign access, leaving an opening at the top of the open-weight field that GLM-5.2 was quick to fill.
Sakana AI has moved Sakana Fugu from beta into commercial launch: a trained "conductor" model that decides, request by request, whether to answer directly or assemble a coordinated team of expert models — planner, executor, verifier, synthesizer — behind one OpenAI-compatible API. None of the internal complexity reaches the developer's code.
The flagship variant, Fugu Ultra, benchmarks near Anthropic's top-tier models on demanding reasoning and coding tests, including a result that edges Claude Opus 4.8 on Humanity's Last Exam. Sakana frames the pitch explicitly around sovereignty: frontier-level capability that routes around any single vendor's export restrictions, since the system can swap providers without the integration changing.
The work builds on two of Sakana's own ICLR 2026 papers — TRINITY and the Conductor — on learned, natural-language model coordination rather than hand-coded routing logic, positioning orchestration itself as the next architectural frontier.