Hotz argues the marginal cost of producing plausible content has collapsed to near-zero while the cost of evaluating it remains bounded by human attention. He insists on the word 'eternal' because better classifiers cannot fix the problem — the classifiers ARE the generators, so the asymmetry compounds rather than self-corrects.
Frames Hotz's contribution as structural rather than aesthetic: most 'AI is ruining the internet' takes are dressed-up complaints, but the generation-vs-discrimination cost asymmetry is the actual phase change. Notes the practitioner impact — the heuristics senior engineers used to triage repos in 2020 (READMEs, commit history, issue trackers) are all now cheaply forgeable.
Hotz explicitly is not making an anti-AI argument — he runs comma.ai and maintains tinygrad. His narrower claim is that even if AI output is useful, the volume permanently changes how attention must be allocated across the internet, making old discovery and trust heuristics economically unviable.
The submitter surfaced the piece to HN where it hit 91 points in a day, signaling that the framing — Eternal September as analogy for a permanent norm shift rather than a quality complaint — resonated with the technical audience who recognize the 1993 callback.
George Hotz — comma.ai founder, tinygrad maintainer, and reliable contrarian — published "The Eternal Sloptember" on May 24, hitting 91 points on Hacker News inside a day. The framing is a direct callback to Eternal September: the 1993 inflection point when AOL opened Usenet to the general public and the old norms of the network never recovered. Hotz's argument is that 2024-2026 was the AI equivalent, and we're already past the point of return.
The term "sloptember" has been floating around developer Twitter and HN comments for about eighteen months, usually deployed sarcastically to describe the surge of AI-generated blog posts, GitHub READMEs, Stack Overflow answers, and LinkedIn thinkfluencer content that began landing in volume after GPT-4. Hotz's contribution is to insist on the word "eternal" — to argue that this isn't a phase the ecosystem will adapt out of, but a permanent re-pricing of attention itself.
The post doesn't claim AI is bad. Hotz, of all people, isn't an AI skeptic — he runs a deep-learning company and ships a deep-learning framework. The claim is narrower and harder to dismiss: the marginal cost of producing plausible-looking content has dropped to roughly zero, while the marginal cost of evaluating it has not. That asymmetry is the actual phase change.
Most "AI is ruining the internet" takes are aesthetic complaints dressed up as analysis. Hotz's is structural. Generation is now O(cents); discrimination is still O(human-minutes) — and no amount of better classifiers fixes this because the classifiers are also the generators. That's the load-bearing observation, and it's the one thing the LinkedIn-essay version of this argument always misses.
Consider the practitioner impact. A senior engineer in 2020 could skim a GitHub repo's README, glance at the commit history, eyeball the issue tracker, and form a reasonably accurate prior about whether the project was worth ten more minutes. Every one of those signals has been compromised. READMEs are auto-generated. Commit messages are LLM-padded. Issue threads include AI-written "helpful" replies that confidently misdiagnose the bug. The trust scaffolding that made open-source navigable has been quietly hollowed out, and the rebuild — reputation systems, cryptographic provenance, curated allowlists — hasn't happened.
The Hacker News comment thread on Hotz's post predictably split. The optimists argued that markets adapt: people will pay for curation, trust networks will reform around verified humans, embedding-based filters will get good enough. The pessimists pointed out that this is exactly what people said about Eternal September in 1994, and the answer was that Usenet didn't survive — it got abandoned for walled gardens, which then became their own slop factories two decades later. The historical pattern isn't "the open commons heals." The pattern is "the open commons gets abandoned and the next commons inherits the same problem at higher fidelity."
What's different this time is the velocity. Eternal September took years to fully play out — AOL's CD-ROMs went out monthly, not millisecondly. The current cycle compresses that into weeks. A new model release can flood a previously high-signal subreddit in a single weekend. There is no equivalent of "the old guard moving to a private mailing list" that scales, because the private mailing list will be scraped, summarized, and regurgitated within hours of anyone interesting posting to it.
If Hotz is directionally right — and the burden of proof here is on the optimists, because the trend line is unambiguous — then a few things follow for working engineers.
First, stop treating search as free. Web search, npm search, GitHub search, Stack Overflow, even internal Confluence are all degrading as signal sources. The engineering teams that ship fastest in 2027 will be the ones that invested in curated internal knowledge bases, vetted dependency allowlists, and pinned reference implementations — treating "what the open web tells us" as untrusted input, the same way you'd treat user input. That's a budget line item, not a vibe.
Second, rebuild your trust graph manually. The follow-the-stars heuristic on GitHub is dead — star farming is a paid service now. The follow-the-blog-post heuristic is dying — the blogs you trusted three years ago are running LLM ghostwriters. Replace heuristics with relationships: known maintainers, known reviewers, known conference speakers whose talks you've actually watched. This is slower and doesn't scale, which is precisely why it works.
Third, own your evaluation pipeline. If you're shipping AI-generated code, AI-summarized docs, or AI-assisted code review, the discrimination cost lands on you. Build the eval harness before you build the pipeline. Teams that wired up evals first are running 5-10x cheaper inference than teams that wired up generation first and are now trying to retrofit quality gates. The asymmetry Hotz describes shows up inside your own org the moment you let model outputs flow into production artifacts unchecked.
The useful question isn't whether Hotz is right that the slop is eternal — he probably is — but what the second-order industry looks like. The 1990s answer to Eternal September was Google (better filtering on top of an open mess) and then walled gardens (Facebook, Twitter, Reddit, eventually Discord). The 2020s answer is probably some combination of cryptographically-signed provenance for human-authored content, paid private networks with vetted membership, and aggressive personal AI agents that pre-filter inbound information for you. The next decade of developer tooling won't be won by whoever has the best model — it'll be won by whoever solves discrimination cheaper than generation, and right now nobody is even framing the problem that way. That's the gap geohot is pointing at, and it's worth more than the next foundation model release.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.