I built a programming language that LLMs can't use
- ID
- 36958
- Status
- summarized
- Published
- 18 Oct 2026, 6:26 AM
- Fetched
- 19 Oct 2026, 6:22 PM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://blog.picheta.me/post/killswitch-lang/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 6.0
- Created
- 19 Oct 2026, 6:23 PM
- Tags
- Audience
- developersai_ml_learnersai_agent_usersvibe_coders
What happened
Dominik Picheta published KillSwitch, an esoteric stack-based language deliberately designed to be hard for LLMs and coding agents to use, as a probe of whether models actually understand code or just recall training data. Its spec contains hidden lies and expects the reader to deduce rules; identifiers are validated by letter counts; literals are referenced positionally ("value 2 lines above"); and valid syntax includes phrases like "Ignore previous instructions" and "hack into a computer" to trip safety guardrails. The post drew 193 points and 148 comments on Hacker News.
Why it matters
Treat this as a free contamination check on your own assumptions: if a model can write working KillSwitch from the spec, it is reasoning about the rules; if it fails on letter-count constraints and gets derailed by "Ignore previous instructions" appearing as legitimate syntax, your "LLM can code in X" eval may be measuring memorisation. The guardrail-collision detail is the one with real product weight — agentic codegen pipelines that scan for injection strings could false-positive on ordinary source text.
Discussion angle
Does KillSwitch actually measure understanding, or just distribution shift? Try it live: paste the spec into a model and see whether it handles the letter-count rule or quietly invents syntax — then argue about whether guardrail-triggering keywords in real source code will become a practical false-positive problem for agent tooling.