How this site works
The feed
Every 15 minutes the site reads whichever of its public sources are due — the AI providers' status pages every 10 minutes, Hacker News, public Mastodon hashtags and Google News searches every 15, news outlets' RSS feeds and Reddit hourly, the public fact-check feed every three hours — and sorts what it finds into lanes: AI News, AI Fails & Slop, AI Incidents & Outages, Memes, Confidently Wrong, AI Bubble & Money, Fact-checks, Satire. Items are link-only and labeled: the title, a snippet, the source, and a link out. Nothing is re-hosted. New items stream into an open page every two minutes; every lane also has an RSS and a JSON feed, and the whole thing is searchable.
Sources are data, not code: a source is a URL, a host allowlist, a lane and a cap. New sources are discovered automatically from the domains the feed keeps linking to, run in shadow for two weeks, and only go live if they kept producing relevant, clean items. A discovered source that later returns nothing for a week, or keeps failing for a week, is retired again. The hand-picked sources are never retired by software: when one fails or falls silent it is flagged on the status page and in the weekly digest for a person to look at.
A deterministic screen drops items about elections, war, disasters, crime, health, personal finance, minors and violence. That screen is a rejection signal, not an endorsement of what passes. The feed is English-only: a title that does not read as English (a cheap function-word check, no model) is dropped at intake, by the owner's decision. Satire is labeled satire because it comes from satire sites — not because anyone read it.
AI Broke Prod
The incidents lane reads the AI providers' own status pages (OpenAI, Anthropic, GitHub Copilot, Cursor, Replit, Perplexity), the AI Incident Database, and reporting on AI agents and AI-written code breaking things. Broke prod counts each provider's incident notices over 7 and 30 days as they publish them, and keeps a sourced register of incidents with real consequences — postmortems, statements, court documents, vendor advisories. Companies and products are named; individual employees and users never are.
Stupid Thing of the Day and the Hall of Fails
One item a day is picked by a formula: lane weight × the engagement its own source reports plus readers' 🔥 reactions (log-scaled) × a small bonus for a picture × freshness. Nothing is picked twice within 30 days. To qualify at all an item needs a real headline of 30 characters or more (not a reply, a quote-post or a bare link), must be about AI unless it sits in the fails, incidents or confidently-wrong lanes, must not be a fact-check or a status notice, must not be profane (the pick goes out as a push notification and an email subject), must be mostly words rather than hashtags, and a social post needs at least one like, boost or picture. Monthly Hall of Fails pages are the same formula over a whole month. No editor, no taste — a formula you can read in the source code.
Captions, reactions, the week, and sharing
The one-line captions under some items are machine-written: a language model sees only the item's own title and snippet and returns one dry line, which passes fixed checks (no person named, nothing the headline did not say, no profanity, no links, inside the safety scope) before it is stored; a line that fails is retried once and then dropped. Every caption is labeled, the title is never touched, and captions never affect what is picked or ranked. Up to eight items get one each hour: the day's pick first, then the top three per lane.
Anyone can 🔥 or 🥱 an item; the first click sets the random-token cookie described on the privacy page. One reaction per token per item, counted, never attributed. This week is the same formula over seven days plus a Stupidity Index: the share of the week's items that sat in the fails, incidents or confidently-wrong lanes; subscribers get it on Mondays. Every item has its own page (the share link) with proper link previews — it still links out, and re-hosts nothing.
(_) .---+---. | o o | | _ | '---+---' _|_ o _/[=]\/ [=] | | _| |_
Stu
The mascot is Stu: a monitor on a stand with an antenna, drawn in the same characters as everything else. Stu has moods, not opinions, and the moods are computed: fifteen outage notices in a week set it on fire, a hot Bubble Thermometer makes it sweat, a market move puts on the shades or brings the rain, pre-market means coffee, a busy hour means typing, a closed market puts it to sleep, a missing page loses it. Stu's lines come from a fixed list picked by the date, like everything else here — no editor, no improv, and never about a person. All the moods, and the rules.
Markets, money, and the ledger
Markets are public quotes read from Yahoo's one-minute chart: every minute while the US session is open, every five minutes pre- and after-hours, every half hour when closed. Pages show the session and an as-of time and never claim exchange-grade data; a ticker that stops trading is dropped rather than shown stale. Follow the Money lists only deals traced to a press release, an SEC filing, or top-tier reporting, each with its date, status and source; a deal the parties never confirmed is marked reported, and a deal that died stays on the page as cancelled. The Bubble Thermometer is an attention index built from the site's own data — how much the news is talking about a bubble, how the AI basket moved against the Nasdaq-100, how many billions were announced this month. It measures temperature, not the future. The Prediction Ledger records dated public promises by AI companies and executives with a primary source each, and marks them met, unmet or pending when their date arrives.
None of it is investment advice.
Who runs it
This site is operated by software. Items link to their sources and are never re-hosted; fact-checks are by the named publishers; selection and ranking are automated. Data files (the deal graph, the ledger, the incident register) are reviewed on a schedule against their sources; anything a source contradicts is corrected in the next review. Reactions and accounts are optional — no password ever, one email link. The status page shows every scheduled job's last run and every source's health; /healthz answers 503 when the scheduler has stopped ticking.