# raymond-llm-coding-empowering-2026-07-08

## Veille

X post by **Eric S. Raymond** (ESR, author of *The Cathedral and the Bazaar*, co-founder of the Open Source Initiative, ~50 years of coding) — **a frontal counter-testimony to the narrative that "LLMs produce crap code and hallucinate, useless for programming."** His thesis: this **almost never happens to him**, and **not at all anymore over the last two generations** of models he uses ("chat GPT 5.4 and 5.5" under **codex**). The former symptom — a model "going off the rails" as it approaches its context limit — has disappeared: codex now displays a **red warning** prompting the user to **clear the session** instead of spiraling. **Scope of use**: AI applied to **feature changes, refactoring and debugging across 63 projects** in **C, Go, Rust, Python and shell**; documentation writing; **decompiling a DOS binary into readable source**. An established **work routine**: when reopening a project, he first runs the **regression tests**, then starts codex and asks it to **audit the code** (bugs + improvement suggestions). Verdict: LLMs are **"excellent and tremendously empowering"**; their **worst limitation** is **"architectural tunnel vision"** — excellent at generating code to specification, but sometimes **blind to higher-level patterns** — which he takes to be the **job of his "meatbrain."** The strongest, counter-intuitive point: LLMs **do NOT get details and edge cases wrong**; he says he is **worse than them** on this front (despite 50 years of experience), because if a change must **touch five places**, the model **reliably finds all five**, whereas the human fixes four and **spends hours debugging** before finding the forgotten fifth. He then questions the **"downshouters"**: do they live in a **different universe**? Are they using **old, weak models**? Is there a **skill issue** he doesn't see because his **mental habits and communication** fit well with these tools' "handles"? An issue he considers important to settle, since "**billions of dollars would be wasted on misdirected token spend**." His recipe, "very simple": **"Be clear in your thinking, tell the model what you want with precision, and good things happen"** — closing with: "what am I missing here?" To be read as a **pro-LLM counterpoint from a historic figure of open source** to the recurring debate on the (de)valuation of coding agents — echoing the "skill issue" and specification discipline (cf. [[martignole-token-manifesto-2026-07-17]]), and forming a diptych with **Linus Torvalds'** doctrinal pro-AI-tool stance on behalf of the Linux kernel ([[torvalds-llm-outil-kernel-2026-07-14]]).

## Titre Article

What...what am I missing here? (post X sur les LLMs et le codage)

## Date

2026-07-08

## URL

https://x.com/esrtweet/status/2074889702381953222

## Keywords

Eric S. Raymond, ESR, esrtweet, The Cathedral and the Bazaar, Open Source Initiative, OSI, open source, Fetchmail, The Art of Unix Programming, Jargon File, LLM, coding agents, codex, OpenAI Codex, ChatGPT 5.4, ChatGPT 5.5, hallucination, bad code, crap code, skill issue, downshouters, refactoring, debugging, feature changes, regression tests, code audit, decompilation, DOS binary, C, Go, Rust, Python, shell, architectural tunnel vision, architectural tunnel vision, high-level patterns, details and edge cases, edge cases, meatbrain, context limit, context limit, clear session, token spend, billions wasted, specification, precision, empowerment, empowering, counter-testimony, AI adoption, experienced developer

## Authors

Eric S. Raymond (ESR, @esrtweet sur X) — développeur, hacker et essayiste américain, **figure historique du mouvement open source**. Né le 4 décembre 1957 à Boston (Massachusetts) ; paralysie cérébrale de naissance, enfance en partie au Venezuela puis en Pennsylvanie. Auteur de l'essai très influent **« The Cathedral and the Bazaar »** (1997, livre 1999), qui oppose le modèle « cathédrale » (développement centralisé et fermé) au modèle « bazar » (décentralisé et ouvert, à la Linux) ; il a **popularisé le terme « open source »** (contre « free software ») et contribué à convaincre **Netscape** d'ouvrir son code (naissance de Mozilla). **Co-fondateur de l'Open Source Initiative (OSI)** en 1998, président jusqu'en 2005. A édité le **Jargon File** (*The New Hacker's Dictionary*), maintenu des projets comme **Fetchmail**, écrit **« The Art of Unix Programming »** (2003). Se revendique **libertarien**, défenseur du port d'armes, ceinture noire de taekwondo ; commente régulièrement tech, politique et open source sur X. Se présente ici comme codeur « très, très bon » avec **~50 ans d'expérience**. (Post X personnel ; date de publication : 2026-07-08 ; date d'ajout à la veille : 2026-07-17.)

## Ton

**Profile**: an argumentative personal testimony in a **frank and slightly polemical** register, posted on X by a veteran practitioner. Not reporting nor a quantified demonstration: a **lived experience** set against a **dominant narrative** (the "downshouters" who disparage LLMs), with a **sincere question** — "what am I missing?"

**Style**: direct, colloquial ("bitching," "meatbrain," "downshouters"), owning his own competence without false modesty ("I'm a very, very good coder… but the LLMs are better than me"). A rhetoric of **astonishment** ("increasingly puzzling," "are they living in a different universe?") more than accusation: ESR does not conclude that his opponents have a "skill issue," he **hypothesizes it** among other possibilities (old models, a different universe) and turns the question back on himself. Memorable phrasing and a **minimalist recipe** at the close ("Be clear in your thinking, tell the model what you want with precision"). The substance is **resolutely pro-LLM**, but nuanced by the admission of a **real limitation** (the "architectural tunnel vision"), which gives the piece its credibility.

## Pense-betes

- **The central counter-intuition: LLMs excel exactly where humans fail — details and edge cases.** ESR's strongest argument, worth retaining as-is: if a change must **touch five places** in the code, the model **reliably finds all five**, whereas the human "fixes four then spends hours debugging before discovering the fifth." An inversion of the cliché "AI hallucinates the details": for him, it is the AI that **provides exhaustive coverage**, the human who forgets.
- **The real, honest limitation: "architectural tunnel vision."** LLMs excel at **coding to specification** but are **sometimes blind to higher-level patterns** → architectural design remains the **"job of the meatbrain."** A clear division of labor: **human = altitude/architecture; model = exhaustive execution/edge cases.** (A nuance that makes the testimony credible rather than starry-eyed.)
- **"It almost never happens" — and not at all for the last 2 generations.** The "crap code / hallucinations" he doesn't observe, especially under **codex** with "ChatGPT 5.4 and 5.5." The derailment near the **context limit** has been **replaced by a product-level safeguard**: a red warning + prompt to **clear the session** instead of letting the model "go off the rails." (The tooling that disciplines context is part of the explanation.)
- **Real scope of use (not a toy).** 63 projects, 5 languages (**C, Go, Rust, Python, shell**), covering **feature changes, refactoring, debugging**; **documentation**; **decompiling a DOS binary** into readable source. Signal: **broad and demanding** use by an expert, not a demo.
- **The repeatable routine.** Every time a project is reopened: **(1)** run the **regression tests**, **(2)** start codex and ask it to **audit** the code (bugs + improvements). An actionable pattern: "test harness first, LLM audit second."
- **The "skill issue" hypothesis — raised, not asserted.** ESR turns the question around: are the detractors using **old, weak models**? Do they live in a **different universe**? Or do they have a **skill issue** he doesn't perceive because his **mental habits and communication** fit well with the tool's "handles"? He **includes his own subjectivity** among the variables. Proposed recipe: **clarity of thought + precision of the request** (to be linked with the Token Manifesto's specification discipline, [[martignole-token-manifesto-2026-07-17]]).
- **The (claimed) figure at stake: "billions wasted on misdirected token spend."** ESR ties the LLM competence/value debate to an **economic waste** — hence, in his view, the importance of understanding why some fail where he succeeds. Watch angle: usage variance (not the tool) as a ROI factor.
- **Tech-watch use**: a piece of **opinion / testimony**, not measurement — to be treated as such (a sample of one highly experienced practitioner, with an acknowledged selection bias). Its value: a **credible pro-LLM counterpoint** signed by a **figure of open source**, useful for balancing the corpus against critiques ("crap code," disillusionment) and for the thesis "value depends on the clarity of the specification and the maturity of the tooling, not the model alone."

## RésuméDe400mots

**Eric S. Raymond** (ESR) — author of *The Cathedral and the Bazaar*, co-founder of the Open Source Initiative, ~50 years of coding — posts on X a **counter-testimony** to the narrative that "LLMs produce crap code, hallucinate, are useless for programming." A narrative he finds **"increasingly puzzling,"** since this **almost never happens** to him.

**The experience.** For **two generations of models** ("ChatGPT 5.4 and 5.5" under **codex**), he no longer observes any derailment. The former symptom — a model "going off the rails" near its **context limit** — has given way to a **red warning** prompting him to **clear the session**. His scope is broad: AI applied to **feature changes, refactoring and debugging across 63 projects** in **C, Go, Rust, Python and shell**, writing **documentation**, and even **decompiling a DOS binary** into readable source. His **routine**: every time he reopens a project, run the **regression tests**, then ask codex to **audit** the code (bugs + improvements).

**The verdict.** LLMs are **"excellent and tremendously empowering."** Their **worst limitation** is **"architectural tunnel vision"**: excellent at coding **to specification**, but sometimes **blind to higher-level patterns** — which, he says, remains the **job of his "meatbrain."** The strongest, counter-intuitive point: LLMs **do not get details and edge cases wrong**. He declares himself **worse than them** on this front: if a change must **touch five places** in the code, the model **finds all five**, whereas the human fixes four and **spends hours debugging** before spotting the fifth.

**The question.** ESR questions the **"downshouters"**: do they live in a **different universe**? Are they using **old, weak models**? Do they have a **skill issue** he doesn't perceive, because his **mental habits and communication** fit well with these tools' "handles"? He deems the question **important**, since "**billions of dollars** would be wasted on **misdirected token spend**." His recipe, "very simple": **"Be clear in your thinking, tell the model what you want with precision, and good things happen"** — before the closing line: "what…what am I missing here?"

To be read as a **credible pro-LLM counterpoint**, signed by a historic figure of open source, to the recurring debate on the value of coding agents — resonating with the specification discipline defended elsewhere (cf. Token Manifesto).

## GrapheDeConnaissance

- Eric S. Raymond —a_créé→ The Cathedral and the Bazaar (DOCUMENT, 0.98)
- Eric S. Raymond —a_créé→ Open Source Initiative (ORGANISATION, 0.9)
- Eric S. Raymond —utilise→ Codex (TECHNOLOGIE, 0.95)
- Eric S. Raymond —affirme_que→ les LLMs sont d'excellents outils, formidablement capacitants pour la programmation (AFFIRMATION, 0.9)
- Eric S. Raymond —affirme_que→ le code pourri et les hallucinations ne lui arrivent quasiment jamais, et plus du tout depuis les deux dernières générations de modèles (AFFIRMATION, 0.85)
- Eric S. Raymond —affirme_que→ les LLMs ne se trompent pas sur les détails et les cas limites : ils retrouvent de façon fiable les cinq endroits à modifier là où l'humain en oublie un (AFFIRMATION, 0.85)
- Eric S. Raymond —affirme_que→ la pire limite des LLMs est une vision en tunnel architecturale : bons à coder à la spécification, parfois aveugles aux patterns de haut niveau (AFFIRMATION, 0.85)
- Eric S. Raymond —recommande→ être clair dans sa pensée et dire au modèle ce qu'on veut avec précision (AFFIRMATION, 0.85)
- Eric S. Raymond —utilise→ tests de régression (METHODOLOGIE, 0.9)
- Codex —permet→ auditer le code d'un projet pour détecter des bugs et suggérer des améliorations (AFFIRMATION, 0.85)
- Codex —résout→ le dérapage du modèle à l'approche de la limite de contexte (avertissement + vidage de session) (AFFIRMATION, 0.75)
- Eric S. Raymond —affirme_que→ des milliards de dollars sont gaspillés en token spend mal dirigé (AFFIRMATION, 0.7)
- clarté de la spécification —améliore→ la qualité du code produit par les LLMs (AFFIRMATION, 0.75)
- skill issue —observé_dans→ l'écart entre praticiens satisfaits et détracteurs des LLMs de codage (AFFIRMATION, 0.7)

---
Canonical: https://www.thekb.eu/en/fiches/raymond-llm-coding-empowering-2026-07-08/