Keep pulling the thread on Boris Cherny.
Opus 5 achieved a 30% score on the Arc AGI 3 benchmark.
Opus 5 is capable of running continuously for extended periods, such as days, weeks, or months at a time.
The Opus 5 model is highly resistant to prompt injection attacks.
Anthropic employs a three-layer defense against prompt injection consisting of a well-aligned model, a prompt injection classifier, and an AutoMode classifier.
Anthropic's prompt injection classifier is based on mechanistic interpretability research from Crystal and functions by detecting the activation of specific neurons associated with prompt injection attempts.
Current large language models, such as Opus 5, are capable of rewriting entire codebases from one programming language to another.
An engineer at Bun used a single prompt within a Claude Code dynamic workflow to rewrite the entire Bun codebase from the Zig programming language to Rust, a task that completed in 11 days.
The Rust version of the Bun runtime, which was rewritten from Zig by an AI agent, is now in production and is used by the Claude Code application.
A Claude agent has been continuously running for over two weeks on a single task to rewrite an Electron desktop application in Swift, using pixel-by-pixel screenshot comparison against the original app for verification.
Dynamic workflows are a new method for orchestrating and substantially increasing the amount of test-time compute applied to solve a difficult task.
Anthropic utilizes Claude 'routines' to autonomously maintain the codebases for its own products, including the command-line interface, iOS app, Android app, and desktop application.
The Opus 5 model is capable of running continuously on a task for extended periods, such as days, weeks, or months, without needing to be restarted.