Nobel Prize winner John Jumper is leaving Google DeepMind for Anthropic after nine years at the London-based AI lab.

Jumper co-led the AlphaFold team that revolutionized protein structure prediction, work that earned him the 2024 Nobel Prize in Chemistry alongside DeepMind colleague Demis Hassabis. The breakthrough allowed scientists to predict how proteins fold from their amino acid sequences, accelerating drug discovery and biological research.

The departure comes as DeepMind faces commercial challenges in monetizing its research advances. The lab has struggled to sell coding tools to enterprise customers, despite technical breakthroughs in areas like protein folding and mathematical reasoning.

Jumper's move to Anthropic represents a notable talent migration between the two frontier AI labs. Anthropic, founded by former OpenAI executives, focuses explicitly on AI safety research while building commercial products like the Claude assistant.

The timing coincides with broader industry shifts as AI labs compete intensely for top research talent. DeepMind has seen several high-profile departures in recent years as researchers move to startups and rival labs offering equity upside and different research directions.

Multi-agent orchestration emerges

Separately, Japanese AI lab Sakana unveiled Fugu, a multi-agent system that behaves like a single model. The system can decide whether to handle requests directly or coordinate a team of specialized models, managing delegation and synthesis automatically.

Users interact with one API endpoint while Fugu orchestrates multiple expert models behind the scenes. The approach represents a shift toward "orchestration models" that combine multiple AI systems rather than scaling individual models.

Meanwhile, AI coding workflows are evolving from prompt engineering to "loop engineering," where developers build systems that repeatedly prompt, evaluate, and refine agent outputs until achieving measurable goals. The approach treats AI interactions as iterative processes rather than one-shot prompts.

https://threadreaderapp.com/thread/2068862070062485867.html