Cuauhtémoc Pacheco claims kcc — a formal C17 compiler in ARM64 assembly set down as a ~1000-page literate program co-authored with Claude without agentic tooling on about $100/month — compiles itself, runs Lua, SQLite, and DOOM, and boots a Linux kernel; independent builds and conformance runs resolve whether plain LLM-assisted development can produce substantial systems software without agents.
state: seedheat: lowuncertainty: highcontradictsscott: highllm-assisted-development coding-capability compilersCuauhtémoc Pacheco
What is this?
The supplied search results never mention Cuauhtémoc Pacheco or kcc — they instead document a rival, high-visibility project with nearly identical verification targets: Anthropic's 'Claude's C Compiler' (CCC), built in February 2026 by safeguards researcher Nicholas Carlini using 16 parallel Claude Opus 4.6 agents (~2,000 sessions over two weeks, ~$20,000 in API costs, 100,000 lines of Rust), which compiles Linux 6.9, passes ~99% of the GCC torture tests, and builds QEMU, FFmpeg, SQLite, PostgreSQL, Redis, Lua, and DOOM. kcc, as described only in the case's own evidence, is the mirror-image methodology claim — one human co-authoring a C17 compiler in ARM64 assembly as a ~1000-page literate program, using Claude without agentic tooling for ~$100/month — and its demo list (self-compilation, Linux kernel boot, SQLite, Lua, DOOM) tracks CCC's almost item for item, suggesting deliberate positioning against the agent-team result, but nothing in the snippets independently corroborates kcc's existence or its claims. The Anthropic coverage also stresses that the real engineering was the test harness and GCC-oracle feedback loop rather than the model, which sharpens what an independent kcc build would actually test: whether harness-free, chat-only LLM collaboration can clear the same systems-software bar.
Why it matters to Scott
Directly stresses his load-bearing 'agents don't hallucinate less, they loop better' thesis (Agentic Coding Plain and Spicy, Three Ingredients Framework): kcc claims the same verification targets as the ~$20k agent-built CCC — self-hosting, SQLite, DOOM, Linux boot — via chat-only collaboration at ~$100/month, with its ~1000-page literate program doubling as a live instance of his spec-as-asset claim (retained source, generated assembly as disposable IR). Since CCC's coverage credits the harness/oracle as the real engineering, kcc-without-harness isolates exactly the variable his canon says is decisive — if independent builds confirm it, the loop thesis is bounded; either way it's a controlled pair that feeds his agent-vs-chat ebooks and his own all-agents dev stack.
ip:source.agentic-coding-plain-and-spicy-ebookip:framework.three-ingredients-frameworkip:source.think-in-whole-stories-why-ai-coding-agents-write-better-code-when-they-see-the-complete-picture-ebookip:concept.spec-as-assetip:concept.characterisation-testingip:concept.token-economicsradar:trip-plain-c-transformer-stackradar:codex-assisted-apple-gpu-driverradar:concept.inference-economics
queries asked of Scott's wikis
- literate programming AI-authored codebases
- agents vs plain chat LLM development workflow
- self-hosting compiler built with LLM
- test oracle harness LLM-generated systems code
- LLM development token cost economics
- formal C17 conformance verification
Measured heat
now 0 pts/hpeak 8 pts/hcomments 0/hpeers p14momentum: steady1 platformsage 236h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
How the heat travelled
pace: p60 vs 1188 stories at the 168h mark (now 236h old) — ahead of ling-spark-mtp-throughput (1.0x), behind local-kv-cache-pressure-probe (1.0x)
Evidence (3) — ⭐ canonical anchor
| source | object | author | score | comments |
| 🟧 hn ⭐ | Kcc – a C17 compiler built solo with an LLM on $100/month, no agentsRetrieved article excerptOpen article · Retrieved 2026-10-01T20:35:58.452815+00:00 # kcc — A Formal C17 Compiler in ARM64 Assembly
**kcc** is a formal C17 compiler written entirely in ARM64 assembly, and set
down as a single **literate program** using [weft](https://github.com/temo-pacheco/weft),
a metalanguage built to be authored and read by human and AI collaborators
alike. Every instruction is explained in prose before it is written in code.
It checks C against the Standard, optimises it, allocates registers, and emits
native machine code; it compiles itself to a byte-identical binary; and it
**cross-compiles to x86-64** (Linux and macOS). It builds and runs Lua, SQLite,
and DOOM — and boots a Linux kernel.
The whole book **is** the compiler (≈1000 pages): read
**[kcc.pdf](https://github.com/LiterateDrivenDevelopment/kcc/blob/main/kcc.pdf)** in this repository, also published on
[Releases](https://gitlab.polinesso.org/literate-programming/kcc/-/releases),
or build it yourself (see below). Learn more about Literate Driven Development
at <https://literatedriven.com>.
## The `.weft` files are the only source of truth
Everything the machine builds — the assembly modules, the verification scripts,
and the `Makefile` itself — is **generated** by tangling the literate sources.
You never edit generated files; you edit the book.
```
kcc.weft master document (LaTeX preamble + @i includes)
literate/ all the chapters (prose + code fragments)
assets/ cover art (SVG → the PDF covers)
kcc.pdf the typeset book (committed; rebuilt with `make pdf`)
README.md this file
LICENSE MIT
```
The typeset book, `kcc.pdf`, **is** committed, so the repository always carries
a readable copy of the current text. Rebuild it with `make pdf`.
## Build
There is a one-time bootstrap: because the `Makefile` is itself tangled from the
book, you generate it first, then use it.
```
# 1. Tangle sources, scripts, and the Makefile into gen/ (and ./Makefile).
weft -p gen/ --mkdirs kcc.weft
# 2. Build the compiler (gen/kcc).
make
# 3. Compile something with it.
./gen/kcc gen/hello.c -o gen/hello && ./gen/hello
```
`-p gen/` prepends `gen/` to every tangle output; `--mkdirs` creates the
directories. The generated tree lives entirely under `gen/` (plus the root
`Makefile`); `make clean` removes it and the command above regenerates it.
### Common targets
| Target | What it does |
| --- | --- |
| `make` / `make build` | assemble + link `gen/kcc` |
| `make difftest` | compile a corpus with kcc **and** clang, compare |
| `make diagtest` / `make warntest` | invalid programs are rejected; `-Wall` warns |
| `make test-codegen` / `make e2e` / `make interop` | run generated programs |
| `make selfhost-*` | recompile a module in C with kcc, relink, verify |
| `make meta-c` | full self-host fixpoint on Linux (needs Docker + qemu) |
| `make ci` | the complete verification suite |
| `make pdf` | weave the book and typeset `kcc.pdf` |
The verification scripts locate their own base directory, so they work under any
`-p` prefix.
## The book
`make pdf` weaves the `.weft` sources to `kcc.tex` and typesets `kcc.pdf` with
LuaLaTeX. The covers are regenerated from `assets/*.svg` when `rsvg-convert` is
available. The resulting `kcc.pdf` is committed alongside the sources, so the
book can be read straight from the repository; tagged versions are also
published as a **Release** asset on GitLab.
## Prerequisites
- **weft** ≥2.0.0 (the literate tool).
- A C toolchain's `as` and `ld` (Xcode command-line tools on macOS).
- For the book: a TeX distribution with LuaLaTeX and `latexmk` — a full
[TeX Live](https://tug.org/texlive/), or the lightweight
[TinyTeX](https://yihui.org/tinytex/) (a minimal TeX Live) — and, optionally,
`rsvg-convert` for the covers.
- For `make meta-c` and the kernel-boot reproduction: Docker and qemu-user.
## License
MIT — see [LICENSE](https://github.com/LiterateDrivenDevelopment/kcc/blob/main/LICENSE). Copyright © 2026 Cuauhtémoc Pacheco, co-authored
with an AI collaborator (Claude). | temo_pacheco | 19 | 25 |
| 🟧 hn | Maylang – A self-hosted systems language compiled with LLMs | mirged | 2 | 1 |
| 🟧 hn | Show HN: Kcc, a C compiler built solo with an LLM on $100/mo boot Linux kernel | temo_pacheco | 2 | 0 |
Interpretation history
2026-10-06T19:28:42Z
The new attach is temo_pacheco re-submitting kcc as a second Show HN that flopped (2/0) — same first-party source, not a new evidence line, so no material change; the original thread's growth to 18/25 is engagement with questions and distillation skepticism, still no independent build, conformance run, or coverage after ~6 days. The meaning is unchanged but sharpened: a stalled first-party claim whose author is re-promoting, gated entirely on a third-party make ci/meta-c run; low heat stands — the 59th-percentile reading is again a dead cohort artifact against 0.33 pts/h at age 120h.
2026-10-06T19:27:01Z
evidence attached: hn.story.49981314 — shared external link with case evidence
2026-10-02T05:20:05Z
The Maylang attach does not hold as corroboration: its submitter states it was built 'with LLM agent harnesses', placing it beside CCC on the agents side of the controlled pair — which leaves kcc the sole claimed chat-only exemplar and still a single first-party evidence line. Thread traction is tail-tier warmth (9 pts / 5 skeptical comments, one platform; the 72nd-percentile cohort speed reflects a mostly-dead peer set, not periphery expansion), so low heat stands pending an independent `make ci` run.
2026-10-02T04:27:22Z
evidence attached: hn.story.49929136 — Second independent LLM-authored self-hosting compiler, corroborating the pattern that plain LLM-assisted development produces substantial systems software.
2026-10-01T21:35:30Z
grounded: contradicts/high — Directly stresses his load-bearing 'agents don't hallucinate less, they loop better' thesis (Agentic Coding Plain and Spicy, Three Ingredients Framework): kcc c
2026-10-01T21:27:12Z
case created — First-party artifact with runnable verification targets making a distinctive no-agents claim that complements rather than duplicates the Bez episode.
Decision trace
- 10-07 06:28repriceThe new attach is temo_pacheco re-submitting kcc as a second Show HN that flopped (2/0) — same first-party source, not a new evidence line, so no material change; the original thread's growth to
- 10-07 06:27attachshared external link with case evidence
- 10-07 06:25propose_attachshared external link with case evidence
- 10-04 03:57review_screenThe only change is a new third-party comment (pjmlp) offering opinion on the significance of the ARM64-assembly aspect; it adds no new fact, independent build/conformance result, contradiction, or fir
- 10-02 15:20repriceThe Maylang attach does not hold as corroboration: its submitter states it was built 'with LLM agent harnesses', placing it beside CCC on the agents side of the controlled pair — which leave
- 10-02 14:27attachSecond independent LLM-authored self-hosting compiler, corroborating the pattern that plain LLM-assisted development produces substantial systems software.
- 10-02 14:23propose_attachSecond independent LLM-authored self-hosting compiler, corroborating the pattern that plain LLM-assisted development produces substantial systems software.
- 10-02 11:21sensor_dirtycomment_update
- 10-02 07:35groundDirectly stresses his load-bearing 'agents don't hallucinate less, they loop better' thesis (Agentic Coding Plain and Spicy, Three Ingredients Framework): kcc claims the same verificati
- 10-02 07:27createFirst-party artifact with runnable verification targets making a distinctive no-agents claim that complements rather than duplicates the Bez episode.