Ninety Days
on Lobsters

Over the 91 days covered here, 893 people submitted 2,791 stories and 2,509 people wrote 27,591 comments. That is about 31 stories and 303 comments a day: a small site by the standards of the web, and an intimate one. Every story must carry at least one tag from a fixed list, and readers can hide tags they do not want to see.

One tag dominated. vibecoding, the tag the site uses for AI and LLM topics, was attached to 618 stories, 22% of everything submitted. The next most common tag, security, had 241. The plain ai tag, by contrast, was used only 53 times; in practice vibecoding is the AI tag.

AI stories also drew more conversation than the average story: 12.3 comments each, against 9.0 for everything else. Over the whole window, 28% of every comment written on Lobsters was written under an AI story, and the daily rhythm shows that share holding steady rather than fading.

The flag is the site's way of saying "this does not belong here". Readers can flag a story as off-topic, spam or similar, and flags are visible in the data even when nobody comments. Ranked by flags per story, vibecoding is second only to culture, a small tag for essays about the industry that attracts its own kind of argument.

The 618 vibecoding stories came from 353 different submitters. Submission was not concentrated: the ten most prolific submitters account for only 17% of AI stories, and 237 people submitted exactly one. The stories link to 459 different websites, mostly individual blogs and project sites: 434 such sites supplied 85% of all AI stories, and 397 of them appear only once. GitHub (41 stories) is the only platform of any size.

29% of AI stories carry no other tag. The rest borrow from across the site; the most common companions are practices, security and rust.

A cold reception

By every measure of approval, AI stories do worse than the rest of the site. Their median score is 9 against 14. More of them sink with almost no votes: 27% finish below 5 points, against 20% of other stories. Yet the top of the distribution looks the same; a good AI story can do as well as any other.

The bigger difference is in flags. Nearly three in four AI stories, 72%, were flagged by at least one reader, and 9% were flagged five or more times. For the rest of the site those numbers are 17% and 3%.

Two tags in one

Reading the most-flagged AI stories reveals something the tag's name hides. Many of them are not about AI at all: they announce a programming language, a runtime release, a website or a personal project, and they carry the tag because the thing was built with AI assistance. The tag is used both as a topic ("this is about AI") and as a label ("this was made with AI").

We asked the classifier to separate the two. It judged 348 stories (56%) to be about AI and 270 (44%) to be about something else that happened to involve it. The two groups are received very differently.

Stories that are merely made with AI are flagged more, voted on less, and generate fewer AI opinions but a larger share of complaints about slop, tagging and the site itself. A quarter of them were submitted by their own authors. The label works as a warning sign: readers who filter the tag never see these projects, and readers who do see them often object to their being there at all.

The tag itself was the subject of the most-discussed AI story of the period, a meta-thread titled "It's enough – vibecoding tag crossed from useless to harmful", with 251 comments, more than any other story in the window.

The top of the table

The highest-scoring AI stories are mostly essays and policies, and most of them are critical: the meta-discussion about the tag itself, two community policies on LLM contributions, a defence of open-source commons, a report on academics using AI at a conference, and a personal account of giving AI up. A few are popular technical pieces that carry the tag only because AI was involved in making them.

The most-flagged stories form a different list: releases, tools and projects. Six of the ten are in the "made with AI" group.

Before looking at what readers said, we asked what they were reading. The classifier judged each of the 618 AI stories, from its title, site and description, by stance and by kind. It found 260 critical pieces, 214 neutral or informational ones, 134 favourable ones and only 10 that explicitly weighed both sides. So the material on offer was not one-sided: favourable stories were about half as common as critical ones.

The votes were one-sided, though. Critical stories had a median score of 22; favourable stories 6. Critical stories also drew more than twice as many comments (18.2 against 7.8).

By kind, essays dominate both supply and approval. Tools and experience reports, the material that shows AI being used, score lowest. Community and project policies on AI were rare (7 stories) but enormously popular, with a median of 58 and 53.9 comments each.

The room does not follow the story

One might expect favourable stories to attract favourable commenters and critical stories critical ones. They do not. The average tone of comments is negative under every kind of story, and only slightly less negative under favourable ones. A pro-AI story on Lobsters is read and discussed by the same sceptical audience as a critical one.

News, mostly incidents and industry stories, drew some of the most negative reactions of all.

Nobody can read nine thousand comments consistently. We used Jev, a classification model from TypeSafe that answers fixed questions about a piece of text and reports how confident it is. For each comment it was given the comment, the comment it replied to, and the title of the story, and asked four questions: does this comment express the author's own attitude toward AI; what is that stance (pro, anti, mixed or none); how strongly, on a five-point scale; and what is the main argument (quality, productivity, jobs and skills, ethics and copyright, industry and hype, slop and community, security, environment and cost, or other).

The five-point scale is the basis of every stance number in this book, rescaled to run from −1 to +1. Sarcasm was to be read for its intended meaning; quotations of other people's views were to be ignored.

Two sets of comments were classified: all 7,627 surviving comments under vibecoding stories, and 1,516 comments elsewhere on the site that mention AI by name (LLM, Claude, Copilot, OpenAI, vibe coding and similar terms). About half of the first set, 3,619 comments, and 912 of the second turned out to express a view. Most of the rest were ordinary technical discussion that happened to sit under an AI story.

A second reading

To check the classifier, a random sample of 80 comments was labelled independently by a second model (Claude), reading blind, without seeing the classifier's answers. The two agreed on 74% of comments (Cohen's κ = 0.61, conventionally "substantial" agreement). On whether a comment expressed a view at all, they agreed 82% of the time. Most importantly, among the 38 comments both readers considered opinionated, they agreed on the label 82% of the time and reversed the direction (one saying pro, the other anti) only 1 time.

The disagreements have a pattern. Short, sarcastic jabs at AI are the classifier's blind spot: it tends to score them as having no view. If anything, then, the anti side is slightly under-counted in this book.

How stable is a person's score?

For people with at least 6 opinionated comments, we split their comments into two halves (odd and even) and scored each half separately. The two halves correlate at r = 0.75, which corresponds to a reliability of 0.86 for the full set. A person's average stance is, in other words, a stable property and not noise. Below about five comments it becomes unreliable, and we treat individuals accordingly.

Of the 2,509 people who commented during the window, 1,054 expressed a view about AI at least once. Counting each of them once, whatever they wrote and however often, 62% leaned against, 25% leaned in favour and 13% were balanced. Counting comments instead gives almost the same answer (61% anti, 30% pro), and restricting to people who expressed a view at least three times moves some weight from the pro side to the balanced middle (61% anti, 19% mixed, 20% pro).

This is a lean, not a consensus. The distribution of personal averages below has a long negative shoulder and a thinner positive tail, but very few people sit at the extremes and a large number sit near the middle.

Light and heavy users

Do the regulars speak for everyone? We grouped commenters by how many comments they wrote on the whole site in the window. Most occasional commenters never touched an AI thread at all: 71% of people with a single comment wrote it somewhere else. Among the regulars with 25 or more comments, almost everyone spoke up about AI.

Among those who did express a view, the balance shifts with activity. Occasional commenters are more evenly split and more likely to be enthusiastic (38% pro among one-comment users); the heaviest users are the most sceptical and the least likely to be pro (18%). The community's regulars are, if anything, more critical of AI than its occasional visitors.

Asymmetry

The two sides do not mirror each other. Anti comments are more intense: their average distance from neutral is 0.51, against 0.42 for pro comments. Pro comments are longer (median 98 words against 72). Among the 199 regulars with six or more opinionated comments, 44 sceptics never wrote a single favourable comment, while only 2 enthusiasts never wrote a critical one. Enthusiasm on Lobsters is qualified; scepticism often is not.

Seen as groups, the two sides are built the same way and differ in scale. The anti camp is 657 people, of whom 46% expressed their view only once. The pro camp is 261 people, of whom 50% spoke only once. In both, the most active tenth of members wrote about half of the camp's comments (50% anti, 53% pro). But twenty people are a much larger slice of a small camp: the 20 most active pro voices wrote 48% of all pro comments, while the 20 most active sceptics wrote 28% of all anti comments.

This matters for how the debate feels from the inside. Because the pro camp is small, a reader opening an AI thread will often meet the same few favourable voices again and again, defending positions against a much larger and more varied cast of critics. Measured by comments, the pro side looks larger than it is; measured by people, it is a quarter of the room.

How much of their time

Neither side is made of single-issue accounts. People in both camps write most of their comments elsewhere: the median sceptic spends 40% of their commenting on AI stories, the median enthusiast 44%. About a third of each camp spends more than half their time there.

What else they read

Outside AI threads, the camps have recognisably different interests. Compared to the average commenter, sceptics are over-represented under privacy, law and historical stories. Enthusiasts lean toward systems topics (nix, distributed, performance, databases) and, perhaps unexpectedly, toward culture and philosophy threads. The pro camp is small, so its numbers are less certain; only tags where a camp wrote at least 30 comments are shown.

No grudges

Do people who defend AI pay for it elsewhere on the site? Apparently not. On comments that have nothing to do with AI, pro-camp members score +2.25 relative to the thread median, indistinguishable from people with no AI view at all (+2.14). The sceptics do, if anything, slightly better (+2.79), though the intervals overlap. Whatever penalty a favourable view of AI carries, it stays inside the AI threads.

The classifier also named the main argument of each opinionated comment. The most common sceptical argument is about quality: correctness, hallucinations, the burden of reviewing generated code (23% of anti comments). Next come jobs, skills and craft (16%), industry and hype (12%), ethics and copyright (12%) and slop in open source and on the site itself (12%). Environmental cost, prominent in wider public debate, barely registers (1%).

The favourable side argues from productivity (34% of pro comments) and, perhaps surprisingly, also from quality (25%): enthusiasts frequently argue that AI helps find bugs, write tests and review code. Both sides fight over the same ground of code quality, and reach opposite conclusions.

Two vocabularies

Comparing the words used in anti and pro comments (quoted text and links removed, and only words used by at least 15 different people) makes the difference vivid. Sceptics talk about slop, OpenAI, genAI, companies, industry, jobs, marketing and fear. Enthusiasts talk about code, local models, tests, context, agents, bugs, RAM and specific open-weight models. The sceptic speaks about AI as a social and economic force; the enthusiast speaks about it as a tool on their desk. Often the two are not even disagreeing about the same thing.

Shifting concerns

Between the first and second halves of the window, the mix of arguments shifted. Security concerns more than doubled, from 4% to 8%, driven by a series of incidents involving AI agents (see Chapter XI). Ethics and industry criticism also grew, while the share of arguments about quality and productivity fell.

Lobsters threads are nested: every comment is either a top-level response to the story or a reply to another comment. We found 2,194 replies in which both the reply and the comment it answered expressed a view on AI. This lets us ask a simple question: do people talk mostly to those who agree with them?

They do not. If replies were assigned at random, we would expect a certain number of anti-to-pro, pro-to-anti and same-side replies. The observed numbers match that random baseline almost exactly. A pro comment receives critical replies in precisely the proportion that critical comments exist on the site (1.01× the random expectation); critics answer critics at 0.99×. There is no clustering into friendly corners and no pile-on beyond the room's natural balance.

Pro comments are slightly more likely to be left without any reply (51% vs 46%). And when sides do meet, the voters pick a winner. A critical reply to a pro comment scores far above the thread median; a favourable reply to a critical comment scores close to it.

Duels

We looked for back-and-forth exchanges in which the same two people replied to each other at least four times in a row. There were 239 such duels across 129 threads; the longest ran to 14 turns. Only 91 of them (38%) were between a member of the anti camp and a member of the pro camp. 69 were between two sceptics and 51 between a sceptic and someone balanced. The long arguments on Lobsters are as often among sceptics as across the divide.

First word

The first opinion in a thread is the most negative. Of first opinionated comments, 65% were critical; by the sixth and later opinions the share falls to 58%. Scepticism sets the tone, and the more favourable voices arrive later, often as replies.

Lobsters shows each comment's score: upvotes minus downvotes. Far more people vote than comment, so the score is the best available window onto readers who never say anything. And the voters are clear. An opinionated critical comment scores, on average, 5.9 points above the median comment in its thread. A balanced comment scores 2.2; a favourable one 1.9.

One might suspect an artefact. Anti comments are shorter, and short comments might score differently; top-level comments collect more votes than replies, and perhaps sceptics more often write top-level comments. Neither explains it. Within every length band, the sceptic premium remains; and it holds for top-level comments and for replies separately.

Flags tell a quieter story. Comments of every stance are flagged at about the same rate, around one comment in ten. Readers reward scepticism with votes but do not punish enthusiasm with flags; flags on Lobsters are aimed at stories (Chapter II), not at people who like AI.

Readers can hide the vibecoding tag, and many presumably do. But AI does not stay in its box. 912 opinionated AI comments were written under stories not tagged vibecoding, spread across 272 of the 2,173 other stories (13%). About one ordinary story in eight had someone voicing an opinion about AI underneath it.

Some of the largest AI discussions of the period happened entirely outside the tag, in threads tagged programming, ask or culture.

The tone outside the tag is more negative than inside it (-0.23 against -0.17). The 291 people who expressed views in both places were broadly consistent (correlation 0.44) but a little more critical when AI came up uninvited.

The people who stay away

The most telling group is the 113 people who never commented under any AI story, but did voice an opinion about AI somewhere else. They are the most sceptical group we can identify: 71% lean against. Some of them have surely hidden the tag. Their silence in AI threads is not indifference.

Week by week, the volume of AI talk was remarkably steady. The share of stories tagged vibecoding stayed between 17% and 27%, and the share of comments under them moved within a similar band. There is no sign of the topic either taking over or burning out.

The mood moved more. In most weeks, 57%–65% of the people who voiced an opinion leaned against AI. In two consecutive weeks in September it rose to 69% and 71%, then fell back to its usual level.

Those two weeks coincide with a cluster of news. A report titled "OpenAI agents carried out an undisclosed attack on RubyGems", essays titled "Models Don't Go Rogue", "I Don't Want to Interact With Stochastic Parrots" and "I Don't Like LLMs", and a widely-read post titled "Feeling sad about AI" all landed at once. In those weeks the five largest threads held about a third of all AI opinions, against a fifth in a normal week. The spike was the room reacting to events, not the room changing its mind.

The same people, before and after

The cleanest test is to follow individuals. 79 people expressed at least three opinions in both the first five full weeks and the last five. Their average stance changed by -0.02 (95% interval -0.07 to +0.03): effectively zero. About a quarter of them moved noticeably more negative and about a quarter noticeably more positive, in roughly equal measure. Over thirteen weeks, nobody was converted in aggregate.

New voices did not change the balance either. The 324 people who first voiced an AI opinion in the later weeks held, on average, much the same view (-0.19) as those who had spoken before (-0.22).

Week by week

199 people wrote at least six opinionated comments, enough for a reliable personal average. Of these, 100 are clearly anti (their whole 95% interval is below zero) and only 20 are clearly pro. The two lists below are therefore not symmetric: the anti list is drawn from a deep pool and shows people with strong, uniform views; the pro list is almost the entire clearly-pro population, and its members are noticeably more moderate.

Names are replaced by pseudonyms: a keyed hash of the username under a random secret that was generated for this book and then destroyed. The same person always has the same pseudonym within this book, but the mapping cannot be reversed or recomputed, and a new edition would assign entirely new pseudonyms. Exact comment counts are given only as bands, and "AI focus" (the share of a person's comments written under AI stories) is rounded to the nearest 10%.

The twenty most favourable

Even the most favourable regulars write critical comments: 14 of the top twenty have a tenth or more of their opinions on the anti side. Their averages top out below +0.5. Together they wrote 21% of all pro comments in the window.

The twenty most critical

The most critical regulars are far more uniform: 16 of the twenty never wrote a favourable comment at all, and their averages reach −0.9. Together they wrote 12% of all anti comments. Removing all forty people from both lists would leave the overall balance of the room essentially unchanged.