CHM Live | Anthropic's Boris Cherny, Creator of Claude Code

CHM Live | Anthropic's Boris Cherny, Creator of Claude Code

by Computer History Museum · View the channel on PSI

Humanities, Social Sciences & Thought AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Computer History Museum 👥 179K 📅 September 26, 2026 ⏱ 71 min 👁 4K 📄 expert opinion 🧭 2026-10-06
Available in: English (current) Français

Keywords

Claude CodeAI agenttool usemodel overhangmechanistic interpretability

Summary

In this CHM Live conversation, Boris Cherny, creator of Claude Code at Anthropic, shares his journey from early coding to leading the development of a revolutionary AI coding tool. He recounts his teenage exploits with TI-83 calculators, his entrepreneurial ventures and time at Meta, and his motivation for joining Anthropic to shape AI safety. The discussion centers on the concept of model overhang—the gap between raw model capabilities and what current products allow—and how Claude Code was built to close that gap by enabling models to use tools. Cherny explains that Claude Code is essentially a chatbot with tools, like file reading, writing, and bash execution, which evolved into an agent that can autonomously perform tasks. He discusses the importance of tool use, the evolution of agents from simple scripts to complex sub-agents, and the development of memory for persistent context. He highlights the MCP (Model Context Protocol) as a standard for integrating external tools. The talk also covers safety measures, including evaluations, mechanistic interpretability, and prompt injection detection, which are integral to Claude Code’s deployment. Finally, Cherny touches on the business growth of Claude Code, noting a period of stagnation before widespread adoption. Throughout, he emphasizes the empirical, social-science-like approach to building AI products, contrasting with traditional engineering.

209 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information is high for those interested in AI product development, offering direct insight into how Claude Code was conceived and scaled. Cherny’s narrative is compelling and provides a clear explanation of technical concepts like tools and agents for a general audience. However, the argumentation is largely one-sided, presenting Anthropic’s perspective without critical examination of potential drawbacks or competing approaches. The absence of concrete data or independent validation limits the robustness of his claims. The reasoning is consistent within the frame of AI safety and capability development, but it relies heavily on personal anecdote and industry narratives.

Scientific Rigor, Source Quality, Title Accuracy

The talk is a personal account with no references to external sources or studies; the only source is the speaker’s own experience. The title precisely matches the content—a live interview with Boris Cherny about Claude Code. The presentation is well-structured and engaging, but the lack of citations or discussion of alternative viewpoints reduces its scientific rigor. The audience seemed receptive, but no critical questions were raised, leading to a unchallenged promotion of Anthropic’s achievements.

189 words

Title / Content Match

The title accurately reflects the content: a live interview with Boris Cherny, creator of Claude Code, discussing its past, present, and future.

Quality & Reliability

8/10

The speaker is the founder and head of Claude Code, offering unique insider perspectives. However, the talk is largely anecdotal and promotional, with no external citations or critical examination of risks. Information is accurate in context but potentially biased due to the speaker's role.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Contribution & Novelties

This talk provides first-hand insight into the development of Claude Code, a significant AI product, and reveals the design philosophy of Anthropic’s tool-use paradigm. It introduces the concept of model overhang and details the practical application of mechanistic interpretability for safety. The discussion on memory and sub-agents offers a rare glimpse into current AI agent capabilities.

Pour aller plus loin :

  • Model Context Protocol — A standard for integrating external tools with AI models, directly relevant to Claude Code’s extensibility.
  • Prompt injection — A security vulnerability discussed in the talk, with implications for AI safety.
  • Interpretability — The field of making AI decisions understandable, including mechanistic interpretability mentioned by Cherny.

110 words

Radar Profile

The radar profile shows high scores in quantity and fiability, reflecting the substantial insider knowledge and the speaker's authoritative position. Quality and technical level are moderate, indicating a good balance between depth and accessibility, but not overly technical or evidence-heavy. Overall, the talk is informative but relies on personal perspective.

Reliability 8/10