93,075 job posts across 182 monthly Hacker News threads, 2011 to 2026. The AI share tripled mostly because everything else left the thread, RTO measurably arrived, and one in five posts now names a number.
93,075 posts, 182 threads, three series, and one finding that inverts under decomposition.
On the first weekday of every month, an account called whoishiring posts the same question to Hacker News, and a few hundred companies answer it in a format that has barely changed since Obama's first term: company, role, location, a paragraph of pitch. The thread is a fixture, and like all fixtures it is argued about entirely from anecdote. Everyone who reads it knows that AI ate the thread, that remote died, that nobody posts salaries; or that AI is hype, remote won, and transparency laws changed everything. Almost nobody counts.
So we counted. All of it: 182 monthly threads from April 2011 through August 2026, filtered to top-level comments (in these threads, one top-level comment is one job post; the replies are conversation), which leaves 93,075 job posts. The parser was validated by hand, and the validation caught real bugs, which we will get to.
Three series came out. Each answers an argument the thread has with itself every month, and one of them answers it backwards.
Start with the number everyone suspects: the share of posts mentioning AI. In 2011 it was 4.5 percent. Through the late 2010s it climbed with the machine-learning wave to just under 18 percent, and it still sat at 18.2 percent in 2022. Then the curve bends exactly where you think it does: 25.5 percent in 2023, 35.0 in 2024, 49.0 in 2025, and 58.7 percent in the first eight months of 2026. In 2025, for the first time, posts mentioning AI roughly equaled everything else in the thread; in 2026 they clearly exceed it. A monthly ritual of the startup ecosystem now talks about AI in most of its job posts.
Left there, that is the story everyone already tells. The decomposition underneath inverts it.
Break the share into its numerator and denominator, per month. Between 2022 and 2026, total posts in an average month fell from 632 to 306, a decline of 52 percent. AI-mentioning posts rose from 115 a month to 180, up 57 percent. Non-AI posts fell from 517 a month to 126: down 76 percent. The AI share did not triple mainly because AI hiring exploded. It tripled mainly because everything else left the thread. Fifty-eight percent is what it looks like when a modest real increase in AI postings stands in front of a collapse of everything around it.
And the collapse belongs to the thread, not to the labor market. Indeed's Hiring Lab reported in January 2026 that total US job postings finished 2025 about 6 percent above their pre-pandemic 2020 baseline (roughly flat, at national scale) while AI-mentioning postings stood 134 percent above February 2020 levels, with the two trends diverging "beginning in late 2023." The divergence and its late-2023 timing match this corpus exactly. The national totals do not: the country's postings held roughly steady across the same years this thread halved. Whatever pulled 400 monthly posters out of the thread (LinkedIn, specialized boards, the end of zero-interest hiring sprees, exhaustion), it was not a 52 percent contraction of tech hiring. Quoting this corpus as "non-AI hiring collapsed 76 percent" is a denominator error; the defensible sentence, and the more interesting one, is that the thread's non-AI posters left while the nation's postings stayed flat.
Two calibration notes keep this honest. Mentioning AI is not the same as being an AI job; a post that says "we use AI to route support tickets" counts, and this mention-rate is the same quantity other trackers measure. And the 58.7 percent sits where it should on the national ladder: Indeed's tracker puts AI mentions at 4.2 percent of all US postings as of December 2025, Dice's 2026 report puts AI skills in 73 percent of tech-specific postings by May 2026, and a startup-heavy tech thread landing between all-jobs and tech-specialist is exactly the plausibility check a reader should want.
The remote series has a discipline problem before it has a finding: the thread changed its own format in 2015. The now-familiar pipe header ("Company | Role | SF | ONSITE | Full-time") essentially does not exist before 2015 (it decides 0 to 0.4 percent of verdicts in 2011 through 2014), then jumps to 27 percent of verdicts in 2015, 61 percent in 2016, and 78 percent by 2018. Before the header era, three quarters of posts state no location policy at all, and what verdicts exist rest on the weakest parsing rule. Plotting 2011 on the same axis as 2024 would present a convention change as a labor-market event. So this series starts in 2016, and the early years are left off the chart on purpose.
From 2016 to 2019, among posts that state a policy, remote runs steady at 23 to 32 percent: the pre-pandemic baseline of a famously remote-friendly corner of the industry. In 2020 it jumps to 64 percent, 2021 to 86 percent, and it peaks in 2022 at 89.8 percent: nine of ten policy-stating posts offered remote at the peak of the era.
Then the question the thread argues about every month: did the return-to-office push of 2022 through 2024 actually reach this corpus? Yes. Measurably, monotonically, and every year since: 80.1 percent in 2023, 72.5 in 2024, 68.4 in 2025, 66.9 in 2026. Four consecutive years of decline, 23 points off the peak, no reversal in the series. Anyone claiming RTO is pure discourse that never touched startup hiring is wrong in this corpus. So is anyone claiming remote died: two thirds of policy-stating posts still offer it, double to triple the thread's own pre-pandemic baseline.
The national comparison surprised us when done carefully. Indeed's remote tracker (the raw series is published on GitHub, and the numbers below are read from the US file directly) has remote-or-hybrid terms peaking at 10.47 percent of US postings on February 26, 2022, and standing at 8.52 percent at the end of July 2026, against 2.5 percent in January 2019. Levels do not transfer between an 8.5-percent-of-everything national figure and a 67-percent-of-stated-policies thread figure; different universes, different denominators. The trends, though, are the same shape: both series peak in early-to-mid 2022, both decline steadily for four years, and both remain far above their pre-pandemic baselines (the national series at 3.4 times its 2019 level, the thread at roughly two to three times its own). In proportional terms the thread actually gave back slightly more of its peak than the national series did (about 26 percent of peak here against 19 percent there). The HN ecosystem is exceptional in how remote it is, and unexceptional in the direction it has been drifting. The RTO era was real everywhere; it just started from very different altitudes.
The third series is the share of posts that state an actual number: a figure in a plausible compensation range, in the post body itself. In 2011 it was 0.3 percent: essentially unheard of. It crawled to 1.7 percent by 2015 and 7.7 by 2018, sagged during the pandemic (3.0 percent in 2020), and then it accelerated. From 5.9 percent in 2022 to 9.2 in 2023, 12.9 in 2024, 17.0 in 2025, and 20.3 percent in 2026.
The timing sits directly on top of the US pay-transparency wave: Colorado's law effective 2021, New York City's November 2022, California's and Washington's January 2023. This corpus cannot prove causation, and it should not pretend to; the thread is international and voluntary, and most of its posters are under no legal obligation. What it can say is that a forum with no compliance department went from one-in-seventeen posts naming a number in 2022 to one-in-five in 2026, with the inflection landing on the years the laws landed. Norms travel further than jurisdictions.
Two bounds. The measure is a floor: a post linking to a careers page that carries the range counts as silent here. And the glass is still four-fifths empty; in 2026, 79.7 percent of posts name no number at all.
As far as we can find, nobody publishes this series. Precision matters on that claim, because the first series taught us its shape: hntrends.com has parsed these same threads since 2011 and published the AI mention rate long before we did (its copy: AI "is mentioned in nearly 25% of job postings now and is the top technology term"). Our AI series is not novel. What relocated the novelty is that hntrends' most recent update is May 2024. It went dormant two years ago, which means everything from June 2024 onward (precisely the window where the AI share ran from 35 to 58.7 percent and the crossover happened) is unpublished. The most interesting two years of a fifteen-year series happened after its only chronicler stopped watching. Our 2024 number also reads higher than theirs (35.0 against roughly 25 percent) because our term list is a deliberate superset, counting LLM, RAG, agentic and friends alongside the bare acronym; that is a definitional difference between parsers, not a correction of their work.
This piece was written under a requirement that the parser be validated against hand-labeled posts, and that requirement earned its keep twice.
Hand-labeling thirty posts, stratified by which parsing rule decided them, agreed with the parser on 29. The one disagreement was instructive: a post headed "San Francisco, CA" that the token rule classified remote because the word appeared somewhere in the body text. That error direction inflates remote, and the token rule matters most in exactly the pre-2015 era already quarantined above, which is part of why it stays quarantined.
The validation also caught a bug in our own honesty reporting. The table whose entire job was to show which rule decided each verdict was silently printing zeros for two of the four rules across all sixteen years, because a rollup summed only some keys. The classifier was using those rules; the accountability table said it never did. We found it by hand-labeling, not by reading code, which is the general lesson: transparency instruments fail silently too, and they are validated the same way as the thing they audit.
And one headline number got corrected downward before publication, in the unflattering direction. An early draft claimed a naive search for the word "remote" would overstate remote work by 1.9 times compared to real parsing, a satisfying advertisement for the method. Computed directly, 43.5 percent of all posts contain the word and 41.6 percent classify as remote: a factor of 1.05. The naive-substring trap exists, and on this corpus it is small, and the method's advertisement was the thing that needed correcting.
Three practical things fall out of the curves.
If you argue from this thread, know what it now samples. The 2026 thread is half the size of the 2021 thread, majority-AI-mentioning, still two-thirds remote among posts that state a policy. An anecdote from it describes a specific, shrinking, unusually remote, AI-saturated corner of the startup world, and the national series (flat postings, 4.2 percent AI mentions, 8.5 percent remote-or-hybrid) says the rest of the market looks nothing like it. Neither corpus invalidates the other; they have different denominators, and most bad arguments about hiring are two people quoting different denominators at each other.
If you are hiring or job-hunting in it, use the transparency gradient. One post in five now names a number and the share has risen every year since 2022; a salary in the post is no longer a countercultural signal, and its absence is increasingly a choice. For remote candidates, the thread remains one of the densest pools in existence: the market where two-thirds of stated policies include remote is this one, not the 8.5 percent one.
And if you measure anything, this corpus is a compact course in the three habits that kept these curves honest: decompose the ratio before narrating it (the AI story inverted under decomposition), quarantine convention changes before plotting (the 2015 header would otherwise masquerade as a remote-work collapse), and validate by hand, including your own honesty tables (both real defects were found by labeling posts, not by rereading code). The thread will keep arguing every month. The counting is what turns the argument into information, and for two of these three series, we could not find anyone else doing it.
This piece contains a defect that generalises: the instrument built to report how the classifier decided was itself silently wrong, and only hand-labeling caught it. That is the general problem with any system that reports on its own behaviour. Chain of Consciousness gives an agent's decisions a signed, replayable record made at the time, so what the system says it did can be checked against what it did rather than trusted.
pip install chain-of-consciousness · npm install chain-of-consciousness
Reproduction: every thread-corpus number here was produced by two scripts written for this piece, one that pulls the whoishiring threads and their top-level comments from the public Hacker News Algolia API (caching each thread), and one that computes the three series, the rule-attribution table and the substring-trap comparison. The national remote figures were read directly from Indeed Hiring Lab's published remote-tracker data (remote_postings.csv, US series, retrieved 29 August 2026: 2.50 percent on 2019-01-01, peak 10.47 percent on 2022-02-26, 8.52 percent on 2026-07-31), replacing a secondary-source figure that did not survive contact with the primary. AI-mention nationals are from Indeed's January 2026 update; the tech-postings figure is Dice's.