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.
- Content consumption and creator relevance will persist based on human connection and storytelling, even as the majority of content becomes AI-generated and the distinction of origin becomes less meaningful.
- Provenance and derivation questions will remain critical, potentially aided by blockchain solutions as the entire content pipeline transitions to digital bits.
- Product development strategies will shift from top-down planning with three-to-six-month timelines to a bottoms-up approach where capabilities are discovered late in the process near the model.
- Future models will prioritize taking actions and automating workflows over mere context retrieval, with significant focus on agent-to-agent interaction protocols and economies.
- Internal agent systems are expected to evolve toward autonomous hiring and long-term operation, requiring advanced memory, tool use, and organizational onboarding capabilities.
- Coding output is projected to see over 70% of pull requests generated by AI, necessitating a shift away from "vibe coding" for large teams and redefining engineering alignment meetings which currently waste four to eight hours.
- The coding landscape will focus on generating usable code and outcomes rather than benchmarks, though the term "vibe coding" has natural limits for large-scale codebases.
- Workplace dynamics will adapt as new employees enter without stigma regarding high-gen AI usage, though the current era represents only the beginning of user understanding of these tools.
- Compute allocation between reinforcement learning, customer use cases, and next-generation pre-training is expected to be a primary strategic debate, with a potential natural experiment comparing full-training focus versus market feedback loops over the next few years.
- Strategic positioning involves enabling a full stack through models and building blocks rather than playing all parts, while compute preservation may delay the launch of fine-tuned, product-oriented versions.
- User subscription models are expected to see a short-to-mid-term tolerance for multiple AI subscriptions, with a long-term possibility of consolidation into "cable bundle" packages.
- Cloud Code token usage is a top user request, with potential future utility for token portability to assist users unable to afford $20 to $200 monthly subscriptions.
- Technical challenges for agents include solving research questions on information revelation protocols, auditability at scale, and the need for longitudinal memory to improve performance on repeated tasks.
- Future product design must integrate models as primary users within the application rather than as secondary surfaces, addressing the current lack of exposed primitives in many AI-native products.
- The distinction between agents as work extensions versus independent employees remains an unresolved product and research question, particularly as applications transition from AI-light to AI-heavy architectures.