
Product-research feedback loops

Why AI is going vertical (again) | Dianne Penn (Anthropic)
Topic 5 of 18
Published Jul 26, 2026 · Clip 27:30–31:36 (4:06)
Diane explains that researchers at Anthropic hold bold visions of the future — even at the company's founding they asked how to get Claude to use a computer and navigate a screen — with what she calls founder-like energy, while also making iterative improvements to models now in users' hands. Her PM team embeds in research loops where user impact is highest: vision, computer use, coding, agentic coding, tool use, and test-time compute. A big and difficult part of the job is translating vague user feedback into actionable research language: 'Claude hallucinated' must be decomposed into whether Claude should have called a tool (tool-use failure), looked at the right document but wrong facts (search/knowledge synthesis failure), or an alignment issue. The team then assesses whether it's a big enough problem, builds evals to describe the improvement, and keeps the loop between users and core model training tight.