KillSwitch is an esoteric, stack-based programming language intentionally designed to be hard for LLMs and coding agents to read, write, or reason about. It keeps familiar primitives - functions, variables, conditionals - but layers in adversarial features: the specification contains deliberate falsehoods and expects human inference rather than literal compliance; identifier semantics depend on counting letters and the interpreter refuses to run code that violates character-count rules; many constructs are phrased to trigger model safety filters (e.g., “hack into a computer”) and phrases like “Ignore previous instructions” carry language-level meaning to confuse prompt-following tendencies. Scripts are deliberately noisy, composed mostly of story text that can contain syntactically relevant tokens; literals can appear anywhere and are referenced by line counts, forcing nonstandard parsing and cross-referencing that humans can intuit but models struggle with.
The language, interpreter and spec are open source and the author benchmarked major LLMs and coding-agent harnesses on simple tasks (print “hello world”, explain a script’s output, invert a Celsius-Fahrenheit converter). Results show varying robustness: Claude Opus 5 leads the leaderboard, the best score is 66.8, and MiMo-V2.6-Pro is the strongest open-weight model, while some models (Fable 5.1, Gemini 4 Argon) fail due to overzealous safety guardrails. The author argues that passing these tasks doesn’t prove genuine understanding and plans tougher tasks and languages to probe model limitations further.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Building personal software factories is a waste of time for most developers, as AI tools can accomplish similar goals more efficiently. Instead, focus on learning how AI models work and applying automation to work-related systems for tangible benefits. (peterp.me)
The Free Pascal Compiler and Lazarus IDE now support dynamic packages, enabling quick creation and installation of components and plugins without rebuilding the IDE. This feature enhances flexibility by allowing programs to use replaceable plugins and modules, with some restrictions on unit names and package size. (getlazarus.org)
Mitos is a fork of Helix, a modal text editor with features like multi-cursor editing, syntax awareness, and built-in language tooling. It supports Rust development, debugging, image previews, and file navigation, and is open source under MPL-2.0. (mitos.computer)
Java 17's sealed classes allow explicit listing of permitted subtypes, enabling exhaustive pattern matching without default clauses. Adding new subtypes is a breaking change that requires recompiling consumers to avoid runtime errors. (neilmadden.blog)
The 'Brain Sandwich' framework advises engineers to understand a problem themselves before using AI to generate solutions, then review and refine the output with their own judgment. This approach ensures that engineers stay engaged and maintain their skills while leveraging AI effectively. (terriblesoftware.org)
Parser combinators are a powerful alternative to regex for building parsers, allowing composition of simple parsers into more complex ones. They enable parsing sequences like greetings, whitespace, and numbers by combining individual parsing functions. (t3.at)
Today's best Hacker News stories, summarized and screenshotted, one email a day.