Anthropic CPO Mike Krieger: Building AI Products From the Bottom Up

Anthropic CPO Mike Krieger (ex-Instagram co-founder) argues AI product development must be bottoms-up, compute allocation is the defining strategic variable, and most AI-native apps are built wrong.

• Over 70% of Anthropic’s internal pull requests are now Claude Code-generated, up from over 50% recently.
• MCP originated from two engineers noticing that Google Drive and GitHub integrations shared nothing in common despite doing the same thing — a bottoms-up fix, not a top-down protocol mandate.
• Compute allocation — between product inference, known training runs, and exploratory research — will be the most consequential strategic decision at every AI lab.
• Krieger is skeptical of the SSI/Ilya model of withholding models from market: Anthropic built Claude 3.7 Sonnet the way it did because of real-world market feedback.
• Biggest product problem keeping him up: Claude is still too hard for first-time users — results depend heavily on knowing the right workflows, unlike Instagram’s zero-learning-curve UX.
• AI coding makes misalignment meetings disproportionately costly — blocking what now represents 4-8x the engineering output raises the organizational pain of slow decisions.
• Krieger floated a ‘token portability’ concept: letting users spend their $200/month Claude Max subscription on third-party products to bootstrap new app adoption.
• Most AI-native apps make the same two mistakes: stapling AI onto a GUI sidebar instead of redesigning around it, and failing to expose app primitives to the model so it can actually act.

2025-05-09 · Watch on YouTube