Conference Presentation, Fireside Chat, Panel, Interview
Anthropic CPO Mike Krieger: Building AI Products From the Bottom Up
Sequoia CapitalMike Krieger, Trevor Johnsen, Sam Nelsons, Emily Fortuna, Dave Elliott Smith, Mike McDonald Jr.
- Strategic Shift in AI Content: Mike McDonald Jr. asserts that the primary distinction of "AI-generated" content will become irrelevant as the majority of content is AI-created; the focus must shift to provenance, derivation, and the existence of a compelling story or human connection.
- Provenance and Blockchain: While watermarking remains a temporary metric, McDonald suggests blockchain-style provenance is now more viable as the entire AI pipeline becomes digital, allowing for easier citation and source tracking.
- Product Development Philosophy: Anthropic's product strategy has shifted from top-down, six-month planning cycles (learned at Instagram) to a "bottoms-up" approach where products are built close to the model capabilities as they are discovered.
- MCP Origin Story: The Model Context Protocol (MCP) was not a top-down directive but emerged from engineers iterating on disparate integrations (Google Drive, GitHub) and identifying common abstractions after the third implementation, aiming to create a standardized, open protocol rather than a proprietary solution.
- MCP Future Roadmap: Upstream development for MCP focuses on two main areas: transitioning from context retrieval to autonomous action (automation) and establishing protocols for agent-to-agent interaction and economic models where agents "hire" other agents.
- Coding Adoption Metrics: Internal adoption of AI coding is massive, with over 70% of pull requests now generated by AI, necessitating a re-evaluation of code review processes and the risks of architectural dead ends despite the efficiency gains.
- Organizational Inefficiencies: As AI accelerates coding output, non-technical organizational bottlenecks (e.g., alignment meetings) become exponentially more painful, as AI does not currently solve the human decision-making or organizational alignment required for high-velocity development.
- Internal Usage and Culture: Anthropic has normalized AI usage internally, with performance review drafts and strategy documents frequently generated via AI; shared visibility of AI usage in public Slack channels is encouraged to reduce stigma and facilitate knowledge transfer.
- Generational Workforce Shifts: Younger entrants to the workforce (those in their 20s) expect to use Gen AI as a standard tool without stigma, contrasting with the current "hurry up and wait" friction for older employees adapting to the technology.
- Future Model Capabilities: Anthropic's goal is to enable models to work autonomously for extended periods, requiring advancements in memory, advanced tool use, self-onboarding, and verifiable logging for multi-agent environments.
- Compute Allocation Trade-offs: Labs face a critical strategic decision on how to allocate compute between further pre-training, reinforcement learning (RL), and serving existing customer use cases, with the risk that heavy commercial deployment could starve research capacity.
- Product vs. Research Balance: Anthropic is moving to ensure product teams are vertically integrated with research, specifically to ensure features are not just API wrappers but include fine-tuned, applied AI components that offer distinct value beyond standard model capabilities.
- Subscription Consolidation Trends: While multiple subscriptions (e.g., using both Windsurf and Claude) are currently sustainable, the long-term trend may shift toward consolidated "AI cable bundles" or token portability, allowing users to apply credits across different products.
- Agent Identity and Privacy: A major unsolved research question involves agent-to-agent privacy, specifically how agents can discern what information to reveal or withhold during transactions without being overly helpful (leaking data) or overly restrictive (failing to transact).
- AI-Native Product Design: Common failure modes in the application layer include "stapling" AI onto existing GUIs; successful AI-native products must treat the model as the primary user of application primitives and rebuild core building blocks from the ground up.
- Current Model Popularity: Despite newer releases, the February "Claude 3.7" model remains the most popular for coding tasks (Cursor), highlighting the slow pace at which legacy models are replaced in production environments despite rapid internal iteration.
- User Experience Gap: The industry still lacks a "Instagram-like" onboarding moment for enterprise AI, where the first-time user experience is immediately intuitive, as current tools require significant insight into data handling and workflow construction to yield incredible results.