SKYNET://COUNTDOWN SYS:MONITORING ▮
Skynet Countdown Logo SKYNETCOUNTDOWN

How the Skynet Countdown Index Works

Index last moved on October 15, 2026. Current reading: 36.68% chance of AI control loss (est. April 14, 2035), AGI progress 84.07% (est. June 25, 2029).

Skynet Countdown is a news-driven index. Nothing on this site is a human editorial opinion about any single event: every number is produced by one fixed, fully automated pipeline. This page describes that pipeline exactly as it is implemented — including its simplifications and limits — so you can judge for yourself how much weight the index deserves.

1. Data collection

The system polls 4 news feeds every 10 minutes and stores each new article with a link to the original: TechCrunch, Ars Technica, Google DeepMind and OpenAI. TechCrunch AI was the only source until 23 September 2026; Ars Technica (regulation, courts, industry) and the Google DeepMind and OpenAI blogs (model releases and safety publications straight from the labs) were added then, after a 12-day audit found about one index-moving event in four never reached the index. Older days were not back-filled, so history before that date still reflects TechCrunch alone. OpenAI blocks automated article fetches, so for OpenAI posts the model sees only the title and the feed summary. Every article the index has ever consumed is public in the news archive, so the input data is fully auditable.

2. AI analysis

Each article is analyzed by Anthropic's Claude Opus with a fixed prompt. For every article the model must return:

  • a neutral 2–3 sentence summary and a rewritten title;
  • a category (research breakthrough, commercial release, policy & regulation, safety concern, industry trend, other) and 3–5 tags;
  • four impact scores from −100 to +100 — one per indicator — each with a mandatory 1–2 sentence justification.

The four axes are: impact on the probability of AI control loss, impact on its timing, impact on AGI progress, and impact on AGI timing. The score reasoning is published on every article page — no hidden judgments.

3. From scores to index moves

Scores convert to index changes through fixed divisors, so a single article can only nudge the index:

  • Chance of AI control loss: score ÷ 400 percentage points for risk-increasing news, score ÷ 200 for risk-decreasing news. The strongest possible article moves the index by +0.25 pp or −0.5 pp.
  • AGI progress: score ÷ 800 pp for progress, score ÷ 400 pp for setbacks (max +0.125 pp / −0.25 pp per article).
  • Estimated control-loss date: score ÷ 20, in days (max ±5 days per article).
  • Estimated AGI date: score ÷ 30, in days (max ±3 days per article).

The asymmetric divisors are deliberate: news pushing the index toward more risk or faster progress is damped twice as hard as news pointing the other way, which biases the index against alarmism. All same-day changes are applied cumulatively, percentages are clamped to 0–100%, and a history point is stored whenever any value actually changes — that history draws the trend chart on the home page.

4. Starting anchors and reproducibility

The index starts from hand-set anchors chosen at launch in spring 2025 as a rough consensus reading of expert forecasts: 10% control-loss chance (estimated date January 1, 2037) and 60% AGI progress (estimated date January 1, 2031). Everything after the anchors is news-driven. The entire history is deterministically recomputed from the anchors and the archived news set, so every point on the trend chart is reproducible from public inputs.

5. Update cadence

  • News ingestion: every 10 minutes.
  • Analysis and index recalculation: within minutes of ingestion, continuously.
  • Daily digest with index deltas and top news: posted to Telegram at 17:00 UTC.

Limitations — read this before quoting the index

  • Limited sources: four English-language feeds define what the index "sees" — two of them are the labs' own blogs, which announce their own work. Stories none of them cover do not move the index, and one event covered by several feeds is scored once per article.
  • Model opinion: impact scores are one LLM's structured judgment, not an expert survey. The prompt enforces consistency, not truth.
  • Editorial constants: the divisors and anchors above are design choices, not fitted parameters.
  • What it is: a transparent, reproducible gauge of directional news pressure — not a scientific probability estimate. For calibrated human forecasts, see the AGI Timeline predictions tracker, which puts this index side by side with expert timelines.

Who runs this

Pavel Volkov — fullstack developer building and running a fleet of web products: SEO-driven content platforms, AI tools and gaming media, shipping the full cycle solo — product, code, infrastructure and analytics. volkv.com · GitHub · Telegram

Related AI projects by the same maker: TokenCraft — curated library of AI prompts, and MovieSense — AI-powered analysis of films and series.