B-IR.

What it is.

B-IR is not a project. It is a blog post by Jason Hall, principal engineer at Chainguard, first published on his personal articles site on 11 January 2026. The article describes a Sunday spent prompted by a prediction made on the Oxide and Friends “Predictions for 2026” episode that 2026 would be the year LLMs got a programming language not intelligible to humans. Hall asked first Gemini and then Claude Opus to design such a language and recorded what each model produced. The experimental artefacts — manual.md, an l1-compiler.tbir clocking in at just under 700 lines, and the loom.md specification — live in the companion GitHub repository imjasonh/loom (the repo was originally published as imjasonh/b-ir, and that URL still redirects). The article is candid about the iteration arc: the first design (multi-byte unicode opcodes) was too unwieldy for the model itself to bootstrap; the second (single-byte control characters) was abandoned mid-implementation by the model, which substituted short English keywords on its own initiative; the third (Loom) keeps token density but adds the kind of structure the verification camp has long argued for.

Why it’s here.

The catalogue includes B-IR as a marker of a meta-question. The interesting result is not the languages themselves but what happens when a working engineer asks two frontier models to design for their own consumption and reports on it honestly. Hall’s own observation is that the third iteration ends up resembling existing languages with cleaner error codes and unambiguous scope — which the catalogue reads as the design space converging on the same concerns the verification camp arrived at independently. The catalogue does not rate B-IR against working compilers. It marks the article as a different kind of evidence: a candid record of what the model gravitates toward when given a blank page.