TechnicalInsightful

Boris Cherny: Building Claude Code

Y Combinator

Boris Cherny, creator of Claude Code, discusses how Anthropic's Opus 5 model represents a major leap in AI capabilities, requiring a fundamental rethinking of how to build AI products. Rather than accumulating complexity, successful builders must regularly delete system prompts and code to "unhobble" the model, allowing it to tackle increasingly complex tasks like rewriting entire codebases and maintaining software automatically.

Summary

Boris Cherny shares insights from building Claude Code, Anthropic's most advanced coding harness. He explains that Opus 5, just released, achieves a 30% score on Arc AGI—a massive improvement from previous single-digit performance. The model can now run for weeks or months continuously and demonstrates unprecedented resistance to prompt injection attacks through a combination of alignment research and mechanistic interpretability techniques that detect injections at the neural level.

A counterintuitive finding: Anthropic deleted over 80% of Claude Code's system prompt when upgrading to Opus 5. Each new model generation requires ablation studies—methodically removing every instruction to understand what's actually necessary. This stems from the reality that each model behaves differently with slightly different capabilities and "personality." Rather than treating AI development like traditional software engineering with careful upfront design, Cherny advocates treating models as "living creatures" requiring empirical, scientific approaches.

Cherny introduces the concept of "product overhang"—capabilities that models possess but products fail to elicit. The original Claude Code solved this by giving Opus 3.5 the simplest possible harness to write entire files and features, rather than limiting it to single-line autocomplete like competitors. Examples of current product overhang include language-to-language codebase rewrites (a bun team member used Claude to rewrite a 100,000+ line JavaScript runtime from Zig to Rust in 11 days), image generation via OpenCV, and complex multi-stage orchestration tasks.

For spawning thousands of agents productively, Cherny describes dynamic workflows—an algebra for agents allowing sequential and parallel execution within a sandbox. Anthropic now uses Claude to maintain itself through daily routines: cleaning dead code, unifying duplicate abstractions ("abstraction police"), deleting obsolete tests, and shipping experiments. A notable example: Boris's two-week ongoing prompt to rewrite Claude's Electron desktop app in Swift while pixel-by-pixel comparing outputs.

On skill development, Cherny emphasizes that successful Claude users share an empirical mindset—forget prior knowledge, observe where the model struggles, and iterate. He advises students to learn computer science practically through real problems, combining technical skills with product sense, business understanding, and user empathy.

Key Insights

  • Anthropic deletes over 80% of Claude Code's system prompt with each major model release and rebuilds it incrementally through ablation studies, removing lines one-by-one to measure impact, because each new model generation has fundamentally different behaviors and capabilities
  • Opus 5 demonstrates unprecedented prompt injection resistance through three-layer defense: well-aligned model training (3 years of research), a mechanistic interpretability-based classifier that detects neurons lighting up during injection attempts, and an auto-mode classifier
  • Product overhang describes situations where models possess capabilities that products prevent them from expressing; the original Claude Code solved this by removing scaffolding and giving models write access to write entire files and features, capability that existing products only elicited at single-line autocomplete level
  • A bun team member used Claude to rewrite a 100,000+ line JavaScript runtime from Zig to Rust in a single 11-day dynamic workflow run with steering—a task that would have taken traditional engineers over a year
  • Anthropic now uses Claude to maintain its own codebases through daily routines spawning hundreds to thousands of agents per day, including dead code deletion, test writing, duplicate abstraction unification (abstraction police), and experiment shipping—work previously requiring dozens or hundreds of engineers

Topics

Opus 5 capabilities and improvementsProduct overhang and unhobbling modelsSystem prompt ablation studiesDynamic workflows and multi-agent orchestrationPrompt injection resistance through mechanistic interpretabilityEmpirical vs. theoretical approaches to AI developmentModel elicitation and capability discoveryClaude-driven software maintenance automationTest-time compute scalingBuilding products on top of AI models

Transcript

[0:00] All right, Boris, we're so excited to have you here, the creator of Cloud Code. Thank you. >> It's great to be here. >> Fresh off the press, you guys just shipped Opus 5 yesterday. >> Yes. And it seems that model performance [0:31] keeps accelerating. You guys got took Arc AGI 3 to 30%. Which is incredible. >> Yes. >> And for context, before the best score was in in the low single digits or low teens, right? >> What can Opus 5 do now that it couldn't versus a previous version? Yeah, there's um there's a lot that goes into every new model and there's a lot of new capabilities that we teach and uh [1:03] get…

Full transcript available for MurmurCast members

Sign Up to Access

More from Y Combinator

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.