We found an undocumented bug in the Apollo 11 guidance computer code

Original Article ↗ Hacker News Discussion ↗

An article claiming to uncover a previously undocumented bug in the Apollo 11 guidance computer, allegedly found using an AI-driven specification tool, triggered debate over both the technical claim and the way it was presented. Commenters scrutinize whether the defect is genuinely new, how it would manifest in flight, and an Apollo software expert ultimately explains it was a known issue discovered and fixed during the original program, with some technical details in the article overstated or wrong. A large part of the exchange centers on whether the piece itself was written or heavily polished by an LLM, the reliability of AI-detection tools, and broader frustration with “AI slop” and undisclosed machine-generated content in technical writing.

Bug validity and historical context

  • Several commenters ask whether the described AGC bug is real or an AI hallucination.
  • An AGC specialist involved in reconstruction work states it is a real, documented defect, previously logged as anomaly L‑1D‑02 and fixed between Apollo 14 and 15.
  • According to that account, the article’s proposed two‑instruction fix is incomplete; the historical patch also restructured code and woke pending jobs.
  • The bug’s impact is described as less dramatic than the article suggests: mode changes zero the flag that leaks the lock, and realistic manifestations would trigger a 31202 alarm (Apollo 12+ analogue of 1202), not a silent failure.

Use of AI tools and Allium

  • The article claims the bug was found by extracting a behavioral specification of the IMU subsystem using an “AI‑native” language (Allium) and then using Claude to explore all paths.
  • Some readers find that role of AI clear; others say the description is vague and similar checks could be done by non‑AI static analysis.
  • Linked Allium docs show a natural‑language‑ish rules language enforced by an LLM, positioned between informal plans and formal specs.

Debate over AI-generated prose

  • Large subthread argues whether the article is LLM‑written:
    • Accusers cite “LLM tells” (short punchy sentences, “it’s not X, it’s Y” rhythm, repetition, “Claude‑isms”, marketing tone, high output volume).
    • Defenders note these are common human styles too, tools like Pangram misclassify, and blind tests show humans are bad at detection.
  • Ethical disagreement: some call using AI with a human byline “cheating”; others see AI‑assisted drafting or polishing as acceptable if content is solid.
  • There’s concern that constant witch‑hunt‑style accusations are corrosive and violate HN guidelines against shallow dismissals, while others argue guidelines should evolve for an AI era.

Quality of article and repo

  • Many praise the piece as fascinating, gripping, and well‑written; a few say the dramatized “Collins alone” scenario is overblown or weakly supported.
  • A technically detailed critic calls parts “garbage,” pointing to:
    • Sloppy AGC hardware numbers (RAM/ROM units, clock vs instruction timing).
    • Mischaracterizations of 1202‑alarm causes, task priorities, and rendezvous procedures.
    • Overly dramatic failure narrative given orbital mechanics and ground support.
  • The public “bug reproduction” repo is criticized: one commenter shows the final “deadlock” phase was initially just print statements of imagined behavior rather than an actual emulator run; they submit a fix to make it real.

Technical clarifications about AGC and related claims

  • AGC expert clarifies:
    • For Apollo 11 programs, we only have printouts, not full rope dumps; other programs are reconstructed from mixed sources.
    • The guidance software wasn’t designed to drop low‑priority jobs under load; landing guidance itself was low priority.
    • The rendezvous‑radar issue involved power‑supply details; early testing suggests voltage differences, not just phase drift, were key to the interrupt load.
  • Another commenter flags a Rust claim in the article (“ownership makes lock leaks compile‑time errors”) as wrong: Rust can still leak resources or deadlock, though some patterns are harder.

Meta-discussion about HN & AI content

  • Some want fewer “AI slop” complaints, arguing:
    • They’re repetitive, tangential, and often evidence‑free.
    • Downvotes/flags should handle low‑quality links instead.
  • Others say early warnings about AI‑written or AI‑padded posts are valuable, given time wasted on low‑effort content and even AI‑generated plagiarized projects.
  • There’s worry that the mere possibility of AI involvement devalues all text and turns reading into constant authenticity policing.

Related side topics

  • Recommendations for in‑depth AGC material, especially a YouTube restoration series, and discussion of the 1202 alarm’s real complexity.
  • Explanation of a famous AGC code snippet (“TEMPORARY, I HOPE HOPE HOPE”) and AGC assembly details like CADR.
  • Broader reflections on how many bugs can hide in tiny, mission‑critical codebases, and on the high‑risk nature of early aerospace and test‑pilot eras.