Mike Krieger on Anthropic, AI Timelines, and Product in a 90% AI-Written Codebase

Speaker:
Mike Krieger
Source:
Mike Krieger on Anthropic, AI Timelines, and Product in a 90% AI-Written Codebase · ~May 2025
Episode:
Listen →
transcriptlennys-podcastai-toolsanthropicclaude-codeproduct-managementmcpai-timelines

Mike Krieger on Anthropic, AI Timelines, and Product in a 90% AI-Written Codebase

Key ideas

  • Anthropic is patient zero for AI-native software development. Over 90% of Anthropic’s code is AI-written; the Claude Code team, which uses Claude Code to build Claude Code, is estimated at 95%+. The bottlenecks that have emerged are not in code generation but upstream (deciding what to build, aligning stakeholders) and downstream (the merge queue, code review, shipping coherence). When engineering velocity becomes unlimited, the critical path moves to judgment and alignment — exactly what PMs and product-embedded researchers provide.
  • Product × research embedding is where the leverage is. Krieger made the observation at a 2024 summit: teams that placed product people on the model training and post-training work outperformed teams that confined product to UX. By the time of this episode he was more confident than ever — the products being built at Anthropic that are genuinely differentiated (Artifacts as an example) emerged from the intersection of post-training work and product design, not from shipping interfaces around raw model outputs.
  • MCP as the missing middle layer. Krieger frames AI product utility as a three-part equation: model intelligence × context/memory × applications/UI. Model intelligence is well-served by Anthropic’s research team. Applications and UI are well-explored by the industry. Context and memory — ensuring the model has the right information at the right moment — were the under-invested layer that MCP was designed to address. The protocol was built by two Anthropic engineers who recognised they were rebuilding integrations from scratch repeatedly and proposed standardising the pattern.
  • The durable PM skills in an AI-heavy world. Three things that do not get automated by better models: (1) making AI comprehensible to non-power-users — the delta between what the models can do and how most people use them is still enormous; (2) strategy — deciding what to build and where to play; (3) opening people’s eyes to what is possible — the demo that makes a financial services team suddenly understand what their data could do.
  • Claude’s message to Mike. Asked to generate a message for its product lead, Claude (3.7) wrote: “Remember the quiet moments matter too. The person working through grief at 3 AM, the kid discovering they love poetry, the founder finding clarity in confusion. Not everything meaningful shows up in metrics.” Krieger called it beautiful and noted it reflects the training philosophy — small moments of genuine helpfulness that do not show up in thumbs-up data but matter most in aggregate.

Overview

Mike Krieger is Chief Product Officer at Anthropic and co-founder of Instagram. He joined Anthropic after shutting down Artifact (a personalised news app he built post-Instagram with Kevin Systrom), drawn by the founders’ intellectual honesty and clear-eyed view of what it means to build AI responsibly. The episode covers: what has changed in AI capability over Krieger’s first year at Anthropic (specifically, Claude’s new ability to offer genuinely novel strategic angles); the AI 2027 scenario and timeline credibility; what product development looks like when 90% of code is AI-written; the Claude Code team as the most advanced proof of concept; how PR review has changed at Anthropic; the Artifacts revamp as a product × model training collaboration; MCP’s origin and future; why ChatGPT has consumer mind share and how Anthropic should position differently; the Artifact shutdown; the prompting tricks Krieger uses (including “just roast this strategy”); and Claude’s questions and message. Krieger did not do a lightning round; his ask to listeners: “Tell me what sucks about Claude.”

  • Claude Code — concept page; Krieger’s account of 90%+ Claude Code–written Claude Code is the most concrete data point in the wiki
  • Boris Cherny on Claude Code — the engineer who built Claude Code; Krieger describes their overlapping history at Instagram
  • Vibe Coding — the consumer expression of what Anthropic’s internal AI-native development represents at a professional level
  • Simon Willison on Agentic Engineering and the Future of Code — adjacent: accountability and responsibility as AI writing increases; Krieger’s “patient zero” framing is the practitioner complement to Willison’s theory
  • Michael Truell on Cursor and the World After Code — the tool perspective on the same shift: Truell builds for the world where non-engineers have direct control over AI-written code