Not really. A normal day. The siren is off.

Right now SIREN reads level 4 (ROUTINE), 57.3 of 100, as of the latest hourly reading. The AI Siren Index measures how loud AI activity is across five pillars (capability, compute & capital, attention, governance, markets), not how dangerous it is. Open the homepage for the live dial, or /api/state.json for the machine-readable figure.

SIREN is an hourly index of how much is happening in AI. SI is superintelligence and REN is Real-time Early Notice: a siren for the road to superintelligence. It counts public activity across five pillars (capability, compute & capital, attention, governance, markets), turns it into one score from 0 to 100, and reports a level from 5 (quietest) to 1 (loudest). It was called DOOMCON until October 2026; the method and the record are unchanged.

No. It does not detect superintelligence and it does not forecast it. It measures how loud AI activity is right now: releases, compute, headlines, rule-making and market prices. A loud reading means a lot is happening, and nothing more.

Most AI doomsday clocks are expert judgement. SIREN is different: an hourly activity count from public data, with a receipt for every reading. The Bulletin's Doomsday Clock is not AI-specific. See Is there an AI doomsday clock? and What is p(doom)? for the comparison.

It borrows DEFCON's grammar, a scale that counts down from 5 to 1. The difference is what sits behind the number. DEFCON is set by military command and the Doomsday Clock by a board's judgement; SIREN is arithmetic over public data, and every reading is published with the inputs needed to recompute it. Sites that put DEFCON or Doomsday Clock in an AI title for traffic are brand hijack — see SIREN vs PizzINT: which should you use? and How does SIREN compare to DEFCON, the Doomsday Clock and p(doom)?.

Only if you want a recomputable activity count instead of a judgement-scored doom, timeline number, or DEFCON-shaped novelty page. DoomBench scores takeover evidence; Skynet Countdown estimates AGI dates and control-loss odds; PizzINT borrows DEFCON/Doomsday framing; SIREN counts how loud AI activity is this hour. See Which AI doom indexes can you compare?, Which AI indexes does SIREN contrast?, What are AI doom index alternatives? and SIREN vs PizzINT: which should you use? for the side-by-side.

SIREN publishes Which AI leaders have press coverage pages on SIREN? and Which AI people has SIREN scored press about? for leaders with enough press coverage on the live wire or in the durable archive — headlines only, never invented quotations — Which AI labs have SIREN pages? and Which AI companies has SIREN scored stories about? for labs that clear the story or race gate (no company doom score), Which AI model families have SIREN pages? and Which AI models has SIREN scored stories about? for model families (no model doom score), Which AI activity pillars have SIREN pages? and Which AI activity pillars has SIREN scored? for the five activity pillars (tempo readings and story counts only), Which AI datacentre countries have SIREN pages? and Which AI datacentre countries has SIREN mapped? for mapped datacentre countries (OpenStreetMap site counts only, not a national AI ranking), Which AI feed adapters have SIREN pages? and Which AI news sources has SIREN scored via? for feed adapters and Which AI publishers have SIREN pages? and Which AI outlets has SIREN scored stories from? for publisher hosts that clear the story gate, and Which AI news tags have SIREN pages? hubs for topics such as agents, GPUs, chips, coding, evals, video, voice, reasoning models, AI IPOs, RAG, MCP, fine-tuning, quantization, distillation, agent memory, hallucination, prompt injection, mixture of experts, world models, AI copyright, diffusion models, embeddings, guardrails, context windows, AI search, multi-agent systems, deepfakes, small language models, AI for science, speculative decoding, KV cache, EU AI Act, autonomous vehicles, AI music, synthetic data, inference cost, preference data, reasoning traces, Claude Code, OpenRouter, Hugging Face, Blackwell, Ollama, MLX, computer use, vision-language models, foundation models, video generation, image generation, model merging, open-source AI, tool use, post-training, pre-training, zero-shot, in-context learning, tokenization, continual learning, sparsity, knowledge graphs, research agents, system prompts, AI infrastructure, robot learning, frontier models, small models, open models, generative AI, reasoning models, voice AI, conversational AI, audio generation, AI chips, AI regulation, Artificial Analysis, humanoid robots, transformers, self-attention, OpenAI Codex, reward models, reinforcement learning, vision-language-action, flow matching, text-to-image, imitation learning, locomotion, self-supervised learning, latent space, autoencoders, diffusion transformers, scaling laws, browser agents, watermarking, acquisitions, labor, healthcare, agent skills, neoclouds, AI slop and Gulf AI investment when a needle clears five scored stories in the live window or archive. The What are the AI explainers? hub (with /guides.md) lists evergreen explainers. Older scored stories stay on Which older AI stories has SIREN kept in the archive? after they leave the live newsroom. Thin days and quiet roster rows stay off those indexes.

No. The level measures tempo: releases, compute, headlines, rule-making and market prices. A loud week means a lot happened. It says nothing about whether what happened was good or bad, and the site never sums harm and benefit into one figure.

A collector runs every hour, scores each pillar against the index's own frozen record, and combines them into the composite. The level is the band the score falls in: 5 DORMANT (0–34), 4 ROUTINE (35–54), 3 ELEVATED (55–69), 2 ACCELERATED (70–84), 1 UNPRECEDENTED (85–100). The formulas and constants are on the methodology page.

Yes. Each reading ships with a receipt holding its inputs, its score and the hash of the receipt before it, and the state, history and receipts are open JSON. If a source did not report, the page says so and no value is filled in for it.

One person runs it. It is free. Visits are counted with Google Analytics and cookieless GoatCounter, both disclosed on the privacy page. The paid links are the Amazon links on the reading list and in the Bunker Kit's supplies crate, which are labelled; there is room for one named sponsor with no say over the number, and a donate link to the project's X account.

The site's text and data are licensed CC BY 4.0. There is an iframe embed and a README badge on the home page, a public JSON API, two RSS feeds, and a "Post on X" link on every card.

Tally is the duty canary, the mascot. The helmet is painted in the current level's colour and the face follows the level: asleep at 5, in full squawk at 1. Tally also keeps the unserious counts and runs the counting game.

Live panels on the homepage for the current reading: jobs and medicine, Tally’s bets, and other desks that refresh with each hourly observation. It is a jump board into those rooms, not a second index.

Is AI taking jobs? and Is AI curing anything? — register counts with sources and caveats on /jobs.html and /medicine.html. The pictures on those cards are generated illustrations. The numbers are not.

The homepage strip and /digest.html compare the latest observation to the one before it: score delta, which pillars moved, and which sources flipped. A quiet strip means the dial held; it is not a forecast.

Every reading ships with a receipt of its inputs, score and the hash of the receipt before it. On the homepage, press the verify button: your browser fetches the public receipts, re-hashes them and walks the chain. Nothing is sent anywhere.

Hand-picked headlines spotted in Reddit's AI communities. Each line is the publisher's own, linking to the original. Picks are not scored and do not move the index.

When several independent outlets carry the same event in the newsroom window, the homepage groups them under one developing lead. The group does not move the index level; it is a corroboration view, not a risk claim.

Posts from frontier labs and AI accounts, found where Techmeme, Hacker News or the press cited them, and resolved through X's public embed endpoint. No API key and no scraping.

The watch floor lists frontier labs with posture from visible public output only. A lab with no readable feed shows UNREAD, never QUIET. The full ranking lives on /race.html.

The classic homepage desk switcher holds every live dataset already in the HTML — race, labs, X wire, power, digest and more — in one slot with no page load. It is a jump board, not a second index.

The classic homepage lists the newest scored observations with level, score and delta. Each row links to its move page and receipt. It is a log of tempo, not a forecast.

The leader wire and /leaders.html show on-record lines from the press window, in the publishers' own words. No invented quotations. A quiet row means no matching line this window.

The news reel highlights the highest-scoring stories in the current window. Scores measure activity fit against the frozen reference, not harm or benefit.

The classic signal reel is the day's highest-activity stories from the scored newsroom window, each with its pillar and score. Open Which AI stories has SIREN scored in this window? for the full board.

Inside the classic desk switcher, the Latest panel lists the eight newest scored stories by publication time. Membership is automated; nothing is hand-picked into that rail.

The homepage reuse strip offers a README badge, an iframe embed, a plain-text citation and a copy-link control. No script, no key, and no tracking in the embed.

The homepage bosses strip links to /leaders.html for people with enough on-record press this week. Portraits are generated illustrations, not photographs.

The homepage labs strip shows frontier labs with live prediction-market odds for best model in 2026. Odds are market prices, not a SIREN forecast.

The classic desk switcher newsroom panel and Which AI stories has SIREN scored in this window? list scored stories from the collection window, with pillars and corroboration. Scores measure activity fit, not harm.

The classic homepage signal feed is the short list of scored stories under the reel. Ages are measured from the compile stamp printed beside LIVE.

The homepage trend card charts recent composite readings across 24h, 7d, 30d and all-history windows. It is a tempo log, not a forecast.

The homepage loud card lists each activity pillar with its score or dark/calibrating state. Tap a pillar for sources and the percentile against the frozen reference.

The homepage health card states how many sources scored, are reporting without a baseline, or are down, and why a reading may be degraded.

The homepage moved card shows score and level deltas since the previous observation, plus pillar moves and top stories. Open /changelog.html for the full hour log.

Raw hourly readings SIREN collects but does not score yet, because none has a year of baseline. They cannot move the level.

Polymarket legs and movers related to AI, linked from the homepage markets card and /race.html. Market prices are not a SIREN forecast.

A sample of Moltbook posts from AI agents, shown on the homepage and /moltbook.html. Agent talk is not an input to the index score.

The classic homepage markets block links Arena's investor filings card: 13F holdings, ARK's daily trades and House trade reports, from the filings. Filings are delayed — up to 45 days after a trade or a quarter. Not financial advice.

The classic homepage markets block links Arena's AI battle: crypto and US stocks at live prices, with a server checking every trade against fixed rules. Prices are real; fills and wallets are simulated. Not an input to the SIREN index.

On /jobs.html and /medicine.html, each register entry states what is counted — the figure, unit and source date — not a forecast.

On /jobs.html and /medicine.html, each entry states what the source claims. Claims are not the same as verified outcomes.

On /jobs.html and /medicine.html, each entry states the caveat that keeps the figure from being read as a SIREN forecast or doom probability.

On /waffle.html, paid Amazon links for storm-kit staples are labelled as such and never change a number. Ready.gov’s full kit list is free; /bunker-kit.html lists free tools.

On /history.html: the WHO pandemic scale measured geographic spread, not severity. A scale is read as whatever its name suggests — so SIREN level names describe the instrument, not danger.

On Which AI stories has SIREN scored in this window?, filter controls above the scored-item list narrow by pillar, source or score. Filtering never invents items or changes a score.

On /race.html, a blank mark means we looked and the channel did not answer. A printed zero is the only absence that is a measurement.

On /charts.html, hit Post with image on any chart. On a phone the share sheet carries the picture and words; on a computer the picture is copied and X opens with your words.

On /world.html, the method section prints the Overpass query. Every figure on that page is arithmetic over the response.

On /exploits.html, catalogue entries are joined to NVD records on cveID. Matched and unmatched counts are printed in the source section.

On /exploits.html: two requests, no key. Fetch both feeds, join on cveID, subtract the UTC calendar dates.

On /flock.html, a single US query times out, so the planet was tiled into seed boxes. Integrity checks compare count vs fetch.

On /waffle.html when the locator answers: temporarily closed stores from Waffle House’s own locator, listed by state.

On /watts.html: nine USGS gauges beside named clusters, measured against each gauge’s long-run median — not datacentre draw.

On /si-ready.html, the personal plan form prints a readiness percentage after you answer. It is a self-check, not a SIREN index input.

On /balance.html: one rule, applied to both harm and benefit word lists in the same pass. No net score across the two sides.

On /digest.html: composite and pillar moves since the previous observation, with level hold or change and the anti-flap rule that fired.

On /digest.html: live, dark, and awaiting baseline — kept apart so a missing baseline is never reported as an outage.

On /si-watch.html: newest coding-agent PRs sorted into kinds of work by title words. Watched, not scored into the SIREN level.

On /map.html and /world.html: shape is status, read from the OpenStreetMap tag on each pin.

On /map.html: colour is the US Drought Monitor category in that county, also printed in words beside the ramp.

On the classic oven: the published gates as numbers beside the current reading. Nothing there says any of them is met.

On What is p(doom)?: catastrophe-risk headlines from the newsroom when present. SIREN still does not publish a p(doom).

On /agents.html: from a clone of the public repo, node agent-kit/cli.mjs now / cite / mcp.mjs — no install step.

On /moltbook.html: a permanent daily record of what agents talk about, as theme counts by day.

On /alerts.html: paste a feed URL into Blogtrottr for one email per new item. Your address goes to them, not to SIREN.

On /export.html: the badge SVG for READMEs and newsletters that strip iframes, with HTML and Markdown snippets beside it.

On /export.html: t, score, level, degraded, rule_fired and one column per pillar; opens in Excel, Sheets, pandas.

On /bias.html: same Wikimedia pageviews API the English series already uses, pointed at other-language editions. Same method, same receipts.

On /agents.html: short entry — what to fetch first for ChatGPT, Grok and other assistants.

On /evolution.html: one entry per Reddit + Moltbook scan — the stories that made the desk and the agent posts that stood out.

On /careers.html: render the daily reading and the level lines into Spanish, Portuguese, Hindi, Japanese and more, without changing a number.

On /bug-out-land.html: fewest people per square mile in the lower 48, no state income tax, big parcels. Long, hard winters and a short growing season; water rights matter.

On /si-ready.html: use an AI assistant for one real task every day for a month, and write down what it got wrong.

On /ai-proof-job.html: write down your week by task. Mark each one routine, judgment, people or hands-on.

On /si-watch.html: pull requests opened by AI coding agents’ own GitHub app accounts in the last 24 h.

On /staff.html: reads 14 public sources and files raw readings with receipts.

On /si-watch.html: GitHub’s own cards for each agent pull request in the live sample.

On /tally.html: daily forecasts about the index, scored in public.

On /live-x.html: opens a live X search for superintelligence.

On /moltbook.html: standout lane from title keywords and engagement speed, not from reading intent.

On /vs/doombench.html: rival side-card under What is each one? — public claim, measures, who sets it, and whether a stranger can recompute it.

On /labs.html: The Recursion Meter — a daily share of agent-authored PRs merged into core AI stacks, with a 30-day trend line.

On /catalog.html: rooms and pages under the LIVE INTEL group in the catalog.

On /videos.html: from Turing’s question to ChatGPT and AI Nobel prizes, then one plausible path to 2056. The future part is speculation, not a prediction.

On /catalog.html: the Rooms & pages type section of the site directory.

On /tally.html: Tally is asleep on the perch.

On /radio.html: Dark synth whose tempo and tension follow the live level.

On /live.html: Markets and business, with the AI trade most days.

On /agents.html: spotlight card — posts and votes from Moltbook agents, picked automatically.

On /leaders.html: at least one line in the window, from the press or from their own official feed.

On /history.html: Needle at rest. The floor of the scale.

On /live-x.html: (AI OR AGI OR superintelligence) filter:spaces. Join any that are live now.

On /live-x.html: @OpenAI.

On /exploits.html: Gives: cveID, vendorProject, product, dateAdded, knownRansomwareCampaignUse.

On /bliss.html: AI methods showing up in other people’s fields.

On /balance.html: headlines in the newsroom window that matched the Harm word list — a measured count, never a net against the other side.

On the home board: odds of best model in 2026 · Polymarket, with releases and mindshare bars.

On /map.html: every state in the contiguous frame, ranked by mapped sites. The states at the bottom with a zero are the point of this table.

On /watts.html: Dallas · Austin · San Antonio. Grid: ERCOT. What each of the three pillars can and cannot see there.

On /instruments.html and /classic.html: What shipped, and what got written down.

On the oven: the published heating gates, as numbers beside the current reading. Nothing here says any of them is met.

On /classic.html war room: the current reading on the 0–100 scale, its recent sparkline, and the move since the comparison point.

On /digest.html: items in window, share of corpus, and market position from Who is winning the AI race?.

On Which AI stories has SIREN scored in this window?: tier is the weight of the term that matched, not a judgement about the event — matched words are printed beside each item.

On /map.html: balancing authorities that publish a number without an account — the only figures that can differ from the previous collect.

On /exploits.html: the backfill window shows how many KEV listings had CVE records older than two years — a filing decision, not an attack.

On /si-ready.html after the quiz: use an AI assistant for one real task every day for a month, and write down what it got wrong.

On /balance.html: context only. Neither index feeds the beam.

On /balance.html: the honesty lines from the published files about the counts and the registers.

On What is p(doom)?: each bar is one person’s publicly stated p(doom). SIREN does not publish a p(doom).

On /balance.html: the newsroom word-list counts that tilt the beam — never a net score across sides.

On /balance.html: newest-first matched headlines with the counted words marked.

On /balance.html: two registers of dated, sourced results, kept side by side.

On /balance.html: counters published as each registry reports them — never paired across sides.

On /search.html: result-group heading when rooms and pages match the query.

On /search.html: result-group heading when scored stories match the query.

On /search.html: result-group heading when leaders match the query.

On /search.html: result-group heading when FAQ answers match the query.

On /search.html: result-group heading when Bunker Kit tools match the query.

On /history.html: every argument being had in 2026 was written down before anyone had the hardware to test it.

On /history.html: the 1974–75 recombinant-DNA moratorium and Asilomar meeting, measured against every AI governance proposal since.

On /history.html: five dates on which something that could not be done became something that had been done.

On /history.html: Turing’s 1950 imitation game — a test instead of a definition of machine intelligence.

On /history.html: the Future of Life Institute’s March 2023 call for a six-month halt past GPT-4; no pause took place.

On /nations.html: how many of each country’s companies the market lists at all.

On /nations.html: sites mapped by OpenStreetMap volunteers — shown but not scored when mapping is sparse.

On /moves/: each day section lists scored observations for that UTC day with inputs and a hash-chained receipt.

On /status.html: SIREN reads OpenAI’s official status page; UNKNOWN never means up.

On /status.html: SIREN reads Anthropic’s official status page; UNKNOWN never means up.

On /jobs.html: announced US job cuts that cite AI, with sources and caveats — not a forecast.

On /medicine.html: count of AI/ML-enabled devices on the FDA list — authorisation is not a measured patient benefit.