Mike Krieger on Anthropic's AI-First Approach

Mike Krieger, co-founder of Instagram, transitioned from his Chief Product Officer role at Anthropic to an individual contributor position in AI engineering. He cited a growing "FOMO" (Fear Of Missing Out) from observing others develop with AI models while his own work was restricted to strategic oversight. This shift reflects a broader industry trend where experienced technical leaders, including former CTOs, are returning to direct engineering, embracing the dynamic and innovative environment of AI development.

Engineering practices are evolving from detailed task breakdown and delegation to defining desired end-state outcomes for AI models. AI models are increasingly capable of completing complex tasks autonomously and providing explanations for their decisions and trade-offs. Krieger notes that this advancement means AI models are becoming more sophisticated, necessitating a change in management style from explicit instruction to broader goal setting.

Krieger advocates for teaching users to be more demanding and 'unreasonable' in their interactions with AI to push the limits of its capabilities. He observes that early AI product designs often overly constrained models, limiting their "degrees of freedom" and potential. Providing AI with direct access to tools, such as virtual machines and bash execution, enables more comprehensive and effective problem-solving, fostering greater innovation.

Krieger provided a practical example where he used AI to port a large codebase from Python to TypeScript. The AI model identified a superior deployment strategy involving Bun, prompting the migration. He remarked that such an undertaking, which would have been considered impractical in the 2010s, is now feasible and efficient with AI assistance.

Krieger successfully utilized AI to port an entire Python project to TypeScript by establishing a dynamic workflow. This process involved the AI verifying and iteratively double-checking the code, reading both versions, and repeatedly refining the conversion. The workflow was completed over a weekend, resulting in a fully ported version by Monday.

At Instagram, during the transition to Python 3, Krieger's team developed MonkeyType to aid in adding type hints. MonkeyType captured runtime type data—specifically, the types actually used in production—and mapped these back to the types within the codebase. This method allowed for a more data-driven and incremental approach to code conversion.

Krieger emphasized the critical importance of identifying clear boundaries for incremental changes during large-scale migrations. He warned against the risks associated with attempting to overhaul entire systems overnight, a strategy he likens to "boiling the whole ocean." Gradual, well-defined steps are crucial for managing complexity and reducing potential disruptions.

Krieger shared two enduring pieces of advice for infrastructure scaling: thoroughly pre-measure all potentially relevant metrics to prevent ambiguity during outages, and thoughtfully implement knobs and feature flags. These practices enable engineers to make informed decisions and dynamically adjust system behavior at runtime. Such capabilities are essential for managing load changes efficiently, sometimes within seconds, ensuring system stability and performance.

Modern development heavily relies on feature flags to manage segmented user experiences and facilitate iterative development. Krieger highlighted the utility of flags in enabling different user groups to interact with varying product versions. He also mentioned the role of "Claude code usage" for interactive elements or specific iterations, indicating a sophisticated approach to managing thousands of toggles for dynamic application behavior.

Anthropic's product teams operate without traditional top-down management structures. The role of a lead, or "engine manager," primarily focuses on ensuring that each team member is assigned to tasks they are most enthusiastic about. This flexible structure remains fluid, adapting and solidifying only as a product gains significant traction and requires more defined organization.

Krieger acknowledged that Claude currently faces limitations due to disconnected product surfaces, hindering seamless user interaction. The objective is to eliminate artificial distinctions between different design elements and integrate them into fully functional applications. Achieving this requires persistent data capabilities and the ability to share artifacts across various interfaces, enabling a more cohesive user experience.

Krieger invoked the "Microsoft Word problem," where an abundance of features leads to low individual usage and increased complexity. He mentioned a recent decision to remove the 'Styles' feature in favor of 'Skills' to streamline product offerings. He stressed the importance of eliminating artificial distinctions between chat, code, and co-work sessions, which often fail to interoperate effectively and prevent delegation between them.

Krieger discussed the transformative role of AI models in simplifying the experimentation process for startups. While AI may not directly solve ideation or taste issues, its core value lies in significantly accelerating development and iteration cycles. This capability allows startups to move faster and overcome initial limitations, unlocking a new generation of entrepreneurial potential.

Krieger asserted that large platforms, such as Google, are often constrained by their existing integrations, offering an advantage to startups. Startups can succeed by maintaining a "laser-obsessed" focus on a particular vertical, industry, or specific user group they understand intimately. This deep, specialized knowledge creates a competitive moat that major labs are unlikely to replicate, enabling smaller entities to outperform larger ones in niche areas.

Krieger affirmed that small teams, typically 4-5 people, intensely focused on a problem, consistently outperform larger, more complex organizations. He emphasized that the ability to write code was never the primary limiting factor for startup success. Instead, true success hinges on a profound understanding of users and the capacity for rapid, iterative execution based on those insights.

The financial sector continues to be a key area for vertical AI development. A significant challenge in this domain is finding the optimal balance between the inherent flexibility of AI models and the critical need for verifiable, immutable data. Enterprises in finance demand robust audit logging and clear data provenance to ensure compliance and accountability. The art of system design in this space involves enabling agentic workloads without compromising these essential regulatory and operational requirements.