AI summary by TwinMind
Anthropic CPO Mike Krieger: Building AI Products From the Bottom Up
Sequoia Capital
23m 58s
Introduction and Background
- Mike Krieger: Chief Product Officer at Anthropic, former Instagram co-founder
- Brief stint as Sequoia founder for "hot week" with Instagram
- Discussion focus: AI content creation, product development at Anthropic
AI Content Creation Philosophy
- Story and human connection remain central regardless of medium or AI usage
- AI as another tool in storyteller's toolbox
- Majority of future content will be AI-generated, making "AI vs human" distinction less useful
- Important questions shift to derivation, provenance, citations
- Blockchain-style solutions more feasible when entire pipeline is digital bits
Product Development Framework at Anthropic
- Traditional product principles still apply: solving real problems, meeting user needs
- Key difference from Instagram era: shift from top-down 3-6 month planning to bottom-up creativity
- Best AI products built very close to models, capabilities only clear late in process
- Examples: Artifacts started as research prototype, taken by designer/engineer to production
MCP (Model Context Protocol) Origin Story
- Emerged from pattern recognition after implementing Google Drive, GitHub integrations
- Third integration would have been another bespoke implementation
- General pattern: do things three times, then identify abstractions
- Started with two engineers prototyping, not top-down protocol planning
- Made truly open for broader adoption beyond Anthropic
- Now involves enterprise considerations: authentication, identity management, Exchange servers
MCP Future Directions
- Two key areas of excitement:
- Taking actions beyond context retrieval (Zapier actions, GitHub integration)
- Agent-to-agent interaction protocols (still early to standardize)
- Internal discussions about agents hiring other agents, economic implications
Coding Products Development
- Over 50% of Anthropic's pull requests are Claude-generated (likely over 70%)
- Focus beyond benchmarks: generating code people enjoy working with
- "Vibe coding" has natural limits for large team codebases (100+ developers)
- Code review challenges: Claude reviewing Claude-generated code ("turtles all the way down")
- Architectural oversight questions when AI can overpower tech debt with rewrites
Second-Order Effects of AI Coding
- Makes organizational inefficiencies extremely painful
- Alignment meetings now block equivalent of 4-8 hours of work instead of 1 hour
- Product organization strain with faster code generation
- Models not yet helpful with organizational decision-making and alignment
Internal AI Usage at Anthropic
- Uneven distribution even within AI lab
- Non-technical disciplines (sales) using for meeting prep
- Personal usage: thought partner for strategy docs, planning, performance reviews
- Shared visibility breaks down AI stigma (Slack integrations, public channels)
- Performance review drafts generated publicly during review season
- Comparison to early Midjourney shared usage patterns
Future Roadmap and Agent Development
- Goal: models working autonomously for hours at a time
- Requirements: memory, advanced tool use, organizational onboarding
- Beyond just model capabilities: verifiability, logging for hundreds of agents
- Won't play all parts of stack but enable through models and building blocks
Product Challenges and User Experience
- Biggest concern: products still too hard to use effectively for first-time users
- "Hold it the right way" problem - incredible results possible but not intuitive
- Far from Instagram's simple "take a photo" user experience
- Large gap between model capabilities and user accessibility
AI Scaling and Compute Considerations
- Compute allocation critical topic across all labs
- Trade-offs: RL vs customer use cases vs next pre-training
- Profitable products directly compete with research compute capacity
- SSI's non-commercialization approach as natural experiment
- Anthropic believes market feedback essential (37 Sonnet example)
Research vs Product Balance
- Products should leverage unique research capabilities, not just API-buildable features
- Artifacts example: fine-tuned into model
- Product pods should include applied AI and fine-tuning team members
- ~10% of research team doing product-oriented work
- Considering more fine-tuned product-oriented model versions
Business Model and Positioning
- Users comfortable paying for multiple AI services currently
- Cloud Max launched with coding token integration ($200/month tier)
- Exploring "bring your tokens to other products" concept
- Potential for AI service bundling (avoiding "cable bundle" terminology)
Agent-to-Agent Communication Standards
- Internal prototypes of agents talking to each other
- Unsolved research problem: discernment around information sharing
- Identity management for agents, longitudinal memory requirements
- Auditability at scale for enterprise deployment
- Questions about agent naming, anthropomorphization
Common Application Layer Mistakes
- AI-light products adding AI as sidebar/secondary surface
- Insufficient exposure of application primitives to models
- Models can't access core functionality ("can't do that for you Dave")
- Need to rethink core building blocks for AI-native design
- GUI-first approach limits model's primary user role
