Anti-LLM Discourse

By mid-2025, "anti-AI" stopped being a fringe position and became a recognizable genre with its own sub-styles. This page is a map of how the pieces in the wiki's anti-LLM cluster fit together — what argument each one is actually making, and which adjacent pieces it doesn't overlap with as much as it looks like it would.

The cluster has at least five sub-genres, and most pieces sit in exactly one of them. Reading them as if they were all making the same case loses the structure.

Economic / financial case

The bet here is that AI as currently capitalized is a bubble, and the bubble pops on numbers, not vibes. subprime-ai-crisis (Edward Zitron) is the long-form version: subsidies at every link in the chain, no profitable end-customer, predicted collapse sequence. ai-bubble-pale-horses is the running checklist of warning signs and which have fired. hold-on-to-your-hardware traces the second-order effect on consumer DRAM/NAND/HDD supply.

These pieces don't argue the technology is bad. They argue the economics don't close. The implicit prediction is that the discourse becomes moot once the money stops.

Engineering / craft case

The argument here is that LLM-assisted coding measurably hurts the work, even when it feels faster. no-silver-bullet-llms (James Bennett) applies Brooks' essential-vs-accidental complexity frame and points to DORA / CircleCI data showing rising instability. peril-of-laziness-lost (Bryan Cantrill) names the missing virtue: human time constraints used to drive simplicity, unconstrained generation produces bloat. cult-of-vibe-coding (Bram Cohen) is the workflow-level version: refusing to read AI-generated code is dogfooding gone cult.

building-syntaqlite-ai is the same case told as autobiography — vibe-coding failure, disciplined rewrite, the addiction loop.

reviewing-ai-code (Thomas Depierre) is the most falsifiable entry in this slot: it takes the standard "just review the AI's code like an intern's" defense and runs it against the empirical code-review literature, showing that review throughput (code-review-throughput-limits) caps LLM-assisted output well below the promised speedup — and that reviewers of LLM code find fewer defects while feeling more confident. He states exactly what data would change his mind, which is what makes it engage rather than refuse.

i-will-never-use-ai (Anthony Manning-Franklin) is unusual for stitching the engineering, economic, institutional, and sociological cases into one nine-reason refusal — single-author, declarative, no rhetorical exit. The skill-decay framing in particular extends skill-atrophy-supervision-paradox from the supervisor role to the individual user.

Institutional / policy case

A small number of projects have institutionalized anti-LLM stances. simonw-zig-anti-ai documents Zig's three-line ban (no LLMs for issues, PRs, or comments) and Loris Cro's contributor-poker framing — review time as investment in growing trustworthy contributors, which an LLM can't be. i-dont-want-your-prs (Dawid Ciężarkiewicz) is the post-LLM rethink of open-source contribution surfaces — prompts and forks instead of PRs, because PR review economics broke. phk-goodbye-bikesheds (Poul-Henning Kamp) is the bleakest institutional entry: he treats LLM code review as a bounded fad but argues that age verification, attested computing, and EU accountability law will end the single-maintainer BDFL model outright — a legal/liability death for the FOSS culture the bans are trying to defend.

Sociological / cultural case

These pieces argue the harm isn't to code or money — it's to people and institutions. ai-great-leap-forward is the metaphor essay: corporate AI mandates as backyard furnaces, fabricated metrics, anti-distillation, institutional knowledge loss. ceo-ai-psychosis (HandyAI on Garry Tan) names the structural sycophancy that makes executive agent-orchestration feel productive without being productive, with ai-sycophancy-loop as the underlying mechanism. agentic-coding-fatigue (0xsid) is the worker-side version — decision fatigue, gacha-loop pacing, the 4-5 hour daily ceiling.

llm-enshittification (Michał Górny / Gentoo) is the broadest piece in this slot — open-source infrastructure under DDoS-grade scraping, FLOSS community fracturing on contributor policy, copywashing, labor displacement, environmental harms. Górny's framing: "fundamentally unethical regardless of utility."

programming-still-sucks (stvn) is the worker-side narrative entry — drunk-at-a-birthday-party voice, the burning-ship metaphor, and the Sara character holding the cron job (and the company) up below decks. The argument is structural, not about AI per se: AI didn't take our jobs. Greed did. The "no more juniors" passage names what skill-atrophy-supervision-paradox mitigations all run into — the apprenticeship pipeline that produces capable supervisors is what got abolished. Pairs with ai-great-leap-forward (institutional knowledge loss) and ceo-ai-psychosis (executive-side performance vs worker-side burning ship).

Rhetorical refusal

A separate slot, occupied by very few pieces, is the explicit refusal to argue at all. i-am-an-ai-hater (Anthony Moser) is the canonical example: cite the harms, refuse the etiquette of softening, name the actual social negotiation ("you want permission"), and walk out. The Miyazaki "insult to life itself" line is the rhetorical anchor.

This slot is rare because it's hard to do without sounding like a tantrum. Moser's piece holds together because the citation paragraph rests the affect on documented claims, the "you want permission" line names what the etiquette is hiding, and the Miyazaki anchor borrows established moral authority. Without those three, refusal is just venting.

Disclosure and marking

A newer, milder slot: writers who don't argue anyone should stop using the tools, only that provenance should be declarable. no-ai-statements (James Zhan) is the declare-human case — handmade-label precedent, detector failure in both directions, and the marker as a stance rather than verifiable information (human-made-disclosure). readme-not (William Woodruff) is its mirror, declaring machine authorship (slop-marker-convention). ai-detector-arms-race is the shared premise: nobody can read provenance off the surface, and trying produces the worst outcome — writers stripping formatting and detail from their own work to get past automated detectors.

What separates this slot from the rest of the map is that it's compatible with heavy AI use. Zhan's argument would be satisfied by a world where everyone uses models and says so. That makes it the only part of the cluster with a plausible path to being adopted by people who disagree with the rest of it.

What the genre as a whole isn't

The cluster is not a unified ideology. The economic-case writers don't necessarily share the political-case writers' frame. Bennett and Cantrill are LLM-skeptical engineers who care about software quality; Moser and Górny make political-ethical cases; Zitron is a financial journalist. Reading them as one position erases the structure.

It's also not — yet — a movement with shared institutions. Zig's ban, the rare exception, is the thing several of these pieces point to. Most of the cluster is individual writers describing what they see, not a coordinated rejection.

The reconciliation slot

A slot that sits between refusal and endorsement: writers who accept the critics' premises in full and use LLMs anyway, trying to describe a disciplined middle. llm-critics-are-right-use-anyway (Jeremy Theocharis) is the canonical example — he grants the copyright, environmental, bubble, OSS-trust-collapse, junior-pipeline, and geopolitical-cutoff arguments, then argues that LLMs amplify what you already have rather than supplying what you don't, so the value depends entirely on the human being able to evaluate the output. Its load-bearing idea, credibility-as-slop-test, is the positive-frame contribution the "what the cluster is missing" section below asks for: a test for slop that doesn't depend on surface signals. simonw-vibe-coding-agentic (Willison conceding the vibe/engineering line collapsed in his own practice) is the adjacent honest-admission piece.

This slot is distinct from the adapt-or-die cluster: Theocharis isn't offering career strategy, he's describing a working practice for someone who finds the critics correct and the tools useful at the same time.

Since mid-2026 it's the fastest-growing slot in the cluster, and it now splits into variants by who's writing and how much guilt they carry:

• The applied-tooling version. tao-apps-coding-agents (Terence Tao) is the calm end — he grants that agent-written code carries bugs and uses it anyway, but only for non-load-bearing visual aids where he can eyeball correctness. A second Fields medalist (after Gowers in chatgpt-5-5-pro-mathematical-research) reporting scoped, checkable use.
• The harm-reduction version. know-thine-enemy (Amy J. Ko) is the same accept-the-critics-in-full move under a stronger political-economy frame — a three-month expert experiment that documents the cost to focus, joy, money, and values and keeps using the tools as survival rather than endorsement. The bleakest member of the slot.
• The no-guilt version. control-the-ideas-not-the-code (antirez) drops the dissonance entirely: line-by-line review of LLM output was never the high-value activity, so stop doing it and own the design instead. not-understanding-your-codebase (Goedecke) is the adjacent descriptive claim — partial understanding is the honest baseline at scale, and Naur's "theory of the program" is one engineering value traded against speed and turnover, not a sacred one.

The through-line is credibility-as-slop-test: every writer in this slot keeps the tool only where they can still tell good output from bad, and differs mainly on how wide that zone is and how they feel about it.

The adapt-or-die counter-cluster

A separate cluster — not anti-LLM, but worth pinning next to this map — argues the opposite: don't refuse the tools, restructure your career around them. elena-verna-job-2027 is the canonical example so far, from a growth/PM vantage rather than an engineering one. The piece is striking because it doesn't deny the harms the anti-LLM cluster names — its prescription against ai-sycophancy-loop, its caveat against eager Tier-6 agent use, and its "AI just helps you do the wrong thing faster" line all read like concessions to the engineering-skeptic critique — but routes around them via ai-native-tiers discipline and career optionality.

Putting Verna next to this cluster is more useful than putting her against it: the engineering-skeptics describe what gets lost; she describes what to do when you can't make the loss not happen.

who-manages-the-agents (off-policy.com) sits in the same counter-cluster from a build-side / organizational vantage. It accepts the concentration and skill-gap worries the anti-LLM writers raise — its 100x-superuser-vs-unmoved-median framing leans on the same METR data as no-silver-bullet-llms, and its accountability refrain answers agent-principal-agent-problem — but routes around them by pushing agent management down to every worker rather than restructuring careers around scarcity. Where Verna prescribes individual career optionality, this prescribes an organizational design (and, in the final paragraph, discloses the product that sells it).

What the cluster is missing

A few things noticeably absent:

• A serious technical analysis of what would have to be true for LLM-assisted dev to work well. Partly filled since: programming-differently-difficult (Jeremy Osborn, CACM) is peer-reviewed, argues from the empirical literature on programming cognition rather than workflow anecdote, and cites the evidence against its own optimism (Shihab's comprehension gap, Alanazi's unstable learning gains). It refuses both boosterism and refusal — its claim is that difficulty relocated from recall to judgment (recall-to-judgment) rather than falling. llm-critics-are-right-use-anyway is the closest the cluster itself has — the /grill-me, Pitch, and adversarial-subagent patterns are a partial constructive answer, though workflow-level rather than a full technical account. Note that chatgpt-5-5-pro-mathematical-research (Gowers) provides this for mathematical research — a documented case where the LLM contributed an "original and clever" idea verified by the framework's original author. That's not in the engineering domain but it is the missing positive case from a technically authoritative voice.
• An equivalent of contributor-poker for non-code domains — a positive frame for what gets lost when AI mediates a relationship.
• A clean economic case from inside the AI labs. The financial-case writers are all outsiders.

Cross-references

The pieces above; also ai-sycophancy-loop, tokenmaxxing, llm-as-average-democratizer, cult-of-vibe-coding, building-syntaqlite-ai for adjacent framings.