Boris Cherny: How We Built Claude Code

Key Concepts

• Future-Focused Development: Building for LLM capabilities expected in 6 months, not current limitations.
• The Bitter Lesson: Prioritizing scaling the model itself over extensive custom “scaffolding” code.
• Rapid Iteration & Codebase Turnover: Constant rewriting and adaptation of codebases due to rapid LLM advancements.
• Exponential Productivity Gains: Significant increases in engineer productivity through the use of LLM-powered coding tools.
• Evolving Engineer Role: A shift from traditional coding to more generalist roles focused on orchestration and problem-solving.
• AI Safety as a Core Principle: A strong emphasis on developing safe and beneficial AI, particularly concerning potential existential risks.

Building for the Future & The Bitter Lesson

Anthropic’s core development philosophy centers around building tools not for the current capabilities of Large Language Models (LLMs), but for those anticipated in approximately six months. This proactive approach is crucial to avoid being overtaken by competitors leveraging newer models and recognizes that Product-Market Fit achieved today is temporary. This principle is deeply rooted in Rich Sutton’s “The Bitter Lesson,” which argues that scaling the model itself consistently outperforms human-engineered solutions. The speakers advocate for “never betting against the model,” prioritizing its inherent capabilities over extensive custom code (“scaffolding”). This presents a constant trade-off between immediate engineering effort and anticipating model improvements.

Quad Code & Claude Code: Evolution & Impact

The journey began with Quad Code, initially a simple terminal app for accessing the Anthropic API. Through constant iterative development and user feedback, it evolved into a powerful coding assistant. The initial terminal interface, intended as a temporary solution, unexpectedly became a core feature due to its simplicity and ease of use. This iterative process involved trying ideas, releasing them, learning from user behavior, and rapidly iterating – recognizing that not all ideas would succeed. Quad Code’s codebase is in a state of constant flux, with approximately 80% being less than two months old, and tools added and removed every few weeks.

Claude Code emerged as a more user-friendly version, designed for non-technical users. Co-work, a GUI wrapper around Quad Code, was built entirely using Quad Code in approximately 10 days, incorporating safety features like virtual machines and permission prompts. Early use cases included automating Git commands, operating Kubernetes, writing unit tests, and even writing code for the Perseverance Mars rover. Internal adoption was rapid and organic, with all technical employees and approximately half of the sales team at Anthropic utilizing the tools. Examples of usage extend beyond traditional coding, including monitoring tomato plants, recovering data from corrupted hard drives, and applications in finance.

Productivity & the Future of Software Engineering

Anthropic is experiencing dramatic increases in engineer productivity. Data shows a 70% increase after the team doubled in size, and a 150% increase since the introduction of Quad Code, measured by pull requests, commits, and commit lifetimes. This contrasts sharply with experiences at other companies like Meta, where a 2% productivity gain required significant resources. The speakers predict a future where coding, as traditionally understood, will be largely “solved.” The role of the software engineer will evolve towards more generalist roles – builders, product managers – with coding becoming a ubiquitous skill across all functions. One speaker reported generating 20 pull requests per day using Quad Code and Opus, and even uninstalling their traditional IDE.

AI Safety & Existential Risk

A strong emphasis is placed on AI safety, particularly the potential risks associated with increasingly powerful models. Anthropic’s mission is centered around building safe models, and this concern permeates discussions. The discussion touches on Artificial Superintelligence (ASI) Level 4 (ASL4) – a recursively self-improving model – and the potential for catastrophic misuse, such as bio-weapon design. This underscores the importance of responsible AI development and the need for robust safety measures.

Technical Details & Data Points

Key technical terms include LLM, PMF, scaffolding, Quad Code, Claude Code, IDE, pull request, commit, ASL, zero-day exploit, Electron, GUI, and VM. Data points include a reported 1,000x productivity increase compared to Google engineers, 70% productivity growth after team expansion, 150% productivity increase with Quad Code, 80% of the codebase being less than two months old, 70% startup adoption of Claude, and 4% of all public code commits being made by Claude Code.

Conclusion

The development of Quad Code and Claude Code exemplifies a radical approach to software development – one that prioritizes building for the future capabilities of LLMs, embracing rapid iteration, and recognizing the power of scaling the model itself. The resulting productivity gains are substantial, suggesting a fundamental shift in the role of the software engineer. However, this progress is tempered by a deep awareness of the potential risks associated with increasingly powerful AI, highlighting the critical importance of prioritizing AI safety and responsible development. The core takeaway is that the future of coding is not about writing more code, but about effectively orchestrating and leveraging the capabilities of increasingly intelligent models.