Kingsbury argues that LLMs are inherently unreliable generators of plausible-sounding but untrustworthy output — 'bullshit machines.' He traces his skepticism back to 2019 when he publicly questioned the ethics of making deep learning cheaper, warning it would enable spam and propaganda. After years of deliberation, he's deliberately filling 'the negative spaces in the discourse' as a counterweight to industry boosterism.
The editorial argues that the piece matters less for its novel arguments than for who is making them. Kingsbury's track record breaking consensus claims in Elasticsearch, MongoDB, and CockroachDB through Jepsen gives him essentially unimpeachable credibility in identifying system failure modes, and he has no financial incentive in either direction on AI.
Kingsbury explicitly flags his own limitations: he's not covering ecological or IP issues, he acknowledges flattening complex stories for polemic effect, and states outright that his essay is 'neither balanced nor complete.' He frames the work as a deliberate case for the prosecution rather than a comprehensive evaluation, calling it 'bullshit about bullshit machines.'
The post received nearly 500 upvotes and 482 comments on Hacker News, which for a piece long enough to warrant PDF and EPUB downloads signals genuine appetite for substantive, critical writing about LLMs. This level of engagement suggests the skeptical perspective resonates with a significant portion of the technical community.
Kyle Kingsbury — better known as Aphyr, the engineer whose Jepsen project has stress-tested virtually every distributed database that matters — has published a multi-part essay titled *The Future of Everything Is Lies, I Guess*. It landed on Hacker News with nearly 500 upvotes, which for a piece that runs long enough to warrant PDF and EPUB downloads, signals genuine appetite for this kind of writing.
The essay is self-consciously polemical. Kingsbury opens with his credentials as a lifelong AI enthusiast — Asimov, Clarke, perceptrons on camping trips — before pivoting to disillusionment. He describes attending a hyperscaler talk in 2019 about new cloud hardware for training LLMs, where he asked during Q&A whether making deep learning cheaper and more accessible would enable new forms of spam and propaganda. Five years later, he's decided the perfect essay will never happen and has shipped what he calls "bullshit about bullshit machines."
Importantly, Kingsbury flags his own limitations upfront: he's not covering ecological or IP issues, he's deliberately filling "the negative spaces in the discourse," and he acknowledges flattening complex stories for polemic effect. This is not a balanced assessment. It's a practitioner's case for the prosecution.
The piece matters less for its specific arguments — many of which have been made elsewhere — than for who is making them and how. Aphyr occupies a rare position in the industry: he's someone whose technical credibility is essentially unimpeachable in infrastructure circles, who has no obvious financial incentive in either direction, and who writes with enough precision that handwaving gets noticed.
When the person who broke Elasticsearch, MongoDB, and CockroachDB's consensus claims tells you something is unreliable, the prior should be that he's probably identified real failure modes. That doesn't make him right about everything, but it means the arguments deserve engagement rather than dismissal.
The Hacker News discussion crystallized the genuine fault lines. User danieltanfh95 pushed back directly: "LLMs with harnesses are clearly capable of engaging with logical problems that only need text," arguing that the "LLMs can't do X so it's an idiot" framing misses how tool-augmented models actually work in production. This is a substantive objection. The gap between a raw model hallucinating arithmetic and an agent with code execution, retrieval, and verification layers is enormous — and growing.
On the other side, user munificent drew a parallel to the Industrial Revolution that's worth sitting with: before industrialization, the natural world was "nearly infinitely abundant" because we simply couldn't exploit it fast enough. The analog to AI isn't that content was scarce before LLMs — it's that our ability to generate plausible-sounding content now vastly exceeds our ability to verify it. This is a pollution framing, not a scarcity framing, and it maps uncomfortably well to what we're seeing with AI-generated code, documentation, and SEO content.
User beders spoke for a significant cohort of practitioners: "What actually is happening inside an LLM has nothing to do with conscience or agency and the term AI is just completely overloaded right now." This is the deflationary position — LLMs are useful tools being marketed with misleading language, and the mismatch between capability and narrative is where the damage happens.
The most technically grounded pushback came from joefourier, who noted Kingsbury's framing of diminishing returns on scale. The "scaling laws are hitting a wall" narrative has been alternately confirmed and contradicted roughly every six months since 2023. The honest answer is that nobody outside a handful of labs knows whether the current architecture has hit fundamental limits or a temporary plateau, and anyone who claims certainty in either direction is selling something.
If you're an engineering leader deciding how deeply to integrate LLM-based tooling, Kingsbury's essay is useful as a checklist of failure modes rather than a binary yes/no signal. The practical questions aren't philosophical — they're architectural:
Verification layers. Every production LLM integration needs a verification layer proportional to the cost of failure. Code generation with test suites? The blast radius is bounded. Customer-facing content generation without human review? You're the spam and propaganda Kingsbury warned about in 2019. The engineering discipline isn't "use AI" or "don't use AI" — it's ensuring your verification infrastructure scales with your generation infrastructure.
Dependency risk. Kingsbury's broader point about the AI industry resting on shaky epistemological foundations translates to a concrete concern: if you're building core workflows around capabilities that might be architectural dead ends, what's your migration path? The teams that are winning aren't the ones with the deepest AI integration — they're the ones with the cleanest abstraction boundaries around their AI dependencies.
Team calibration. The most insidious failure mode isn't hallucination — it's the gradual erosion of your team's ability to evaluate output quality. When junior engineers start treating LLM output as authoritative, you don't get a single catastrophic failure. You get a slow drift in code quality that's invisible until something breaks in production at 3 AM. Kingsbury's essay is worth circulating to your team not because it's the final word, but because it's a counterweight to the default assumption that more AI integration is always better.
Kingsbury promises more installments, and given his track record with Jepsen — where each database analysis ran thousands of words and surfaced bugs that vendors had missed — there's reason to expect the subsequent parts will get more technically specific. The real value won't be the polemic framing, which is deliberately one-sided, but whatever concrete failure modes he catalogs. In distributed systems, Aphyr's methodology was simple: state a system's guarantees, then systematically break them. If he applies the same rigor to LLM claims — and the essay's structure suggests he will — the results will be more useful than the opening salvo. For now, the essay is best read as a well-credentialed engineer planting a flag: the burden of proof belongs on the people making claims about AI capabilities, not on the skeptics asking for evidence. In an industry currently running on vibes and venture capital, that's a position worth defending.
> 2017’s Attention is All You Need was groundbreaking and paved the way for ChatGPT et al. Since then ML researchers have been trying to come up with new architectures, and companies have thrown gazillions of dollars at smart people to play around and see if they can make a better kind of model.
> It remains unclear whether continuing to throw vast quantities of silicon and ever-bigger corpuses at the current generation of models will lead to human-equivalent capabilities. Massive increases in training costs and parameter count seem to be yielding diminishing returns. Or maybe this effec
I think the discussion has to be more nuanced than this. "LLMs still can't do X so it's an idiot" is a bad line of thought. LLMs with harnesses are clearly capable of engaging with logical problems that only need text. LLMs are not there yet with images, but we are improving with
Thank you for putting it so succinctly.I keep explaining to my peers, friends and family that what actually is happening inside an LLM has nothing to do with conscience or agency and that the term AI is just completely overloaded right now.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
There is a whole giant essay I probably need to write at some point, but I can't help but see parallels between today and the Industrial Revolution.Prior to the industrial revolution, the natural world was nearly infinitely abundant. We simply weren't efficient enough to fully exploit it.