The full arithmetic — the percentile normalisation, the NowCast smoothing and the six
anti-flap layers that stop the level oscillating — is in the
How the SIREN AI activity index is computed · methodology, and every observation is
written to a Where are the SIREN index moves? carrying the inputs
it was computed from.
Three source states exist, and they stay distinct, because collapsing them is how an index
starts lying. Live: the source answered and is scored. Dark: the fetch failed, so
the source is excluded — never imputed, never carried forward, never counted as zero — and
while any pillar is dark the level cannot change at all. Awaiting baseline: the source
answered perfectly well, but there is no frozen history to score it against yet, so it is
collected and published and left out of the composite. Reporting that third state as an outage
would claim a failure that is not happening.
What questions do people actually ask?
What is the history of AI risk?
A sourced timeline of the argument and of the people who tried to measure it: Good's intelligence explosion, Dartmouth, Asilomar, Bostrom's paperclip maximizer, the 2015 and 2023 open letters, Hinton's resignation, Bletchley and Seoul, and the EU AI Act — each entry dated with a primary or near-primary link. SIREN counts what is happening now; this page is the record that makes the count mean something.
What is the intelligence explosion?
The argument that a machine able to design machines better than humans can would design a successor better than itself, and that the process would repeat. I. J. Good set it out in 1965 in "Speculations Concerning the First Ultraintelligent Machine", calling such a machine the last invention humans need ever make. Vernor Vinge renamed the idea the singularity in 1993.
Who came up with the paperclip maximizer?
Nick Bostrom, in a 2003 paper titled "Ethical Issues in Advanced Artificial Intelligence", expanded in his 2014 book Superintelligence. The example is a superintelligence whose top goal is manufacturing paperclips and which therefore converts available matter into paperclip production. It is an argument about goal specification and instrumental convergence, not a claim about stationery.
What did the 2023 AI pause letter actually ask for?
An immediate six-month halt to the training of AI systems more powerful than GPT-4, and the use of that time to develop shared safety protocols. The Future of Life Institute published it on 22 March 2023 and tens of thousands of people signed it. No pause took place.
Why did Geoffrey Hinton leave Google?
Hinton resigned in May 2023 and told the New York Times he wanted to speak about the risks of AI without those comments reflecting on his employer. He shared the 2018 Turing Award and, in October 2024, the Nobel Prize in Physics.
What was the Asilomar conference, and why does AI keep citing it?
A February 1975 meeting of about 140 scientists at Pacific Grove, California, which agreed containment guidelines for recombinant DNA research after a self-imposed moratorium in July 1974. It is cited as the precedent for a research field regulating itself. The 2017 Asilomar AI Principles were deliberately produced at the same venue.
When did the EU AI Act take effect?
It entered into force on 1 August 2024. Prohibited practices applied from 2 February 2025, obligations on general-purpose AI models from 2 August 2025, and general application from 2 August 2026, with further high-risk phases dated in the regulation.
When does the idea of AI risk arrive before the machines?
Every argument being had in 2026 was written down before anyone had the hardware to test it. Two of these three documents are funding proposals.
Which precedent can nobody in AI stop citing?
In 1974 a field called a halt on itself, and in 1975 it met in a conference centre on the California coast to write the rules. Every AI governance proposal since has been measured against it, usually by people invoking the name and skipping the details.
How does the AI-risk argument acquire a shape?
Between the second AI winter and the deep learning era, the worry got precise enough to be criticised — which is the point at which an intuition becomes a position.
Where did the capability jumps land?
Five dates on which something that could not be done became something that had been done. Note how ordinary the delivery mechanism is: a chess match, a contest entry, a preprint.
When does the field sign its name on AI risk?
The distinguishing feature of this round of AI worry is that the people building the systems are the ones signing the letters. That is a dated, countable event, whatever you make of the content.
When does AI governance acquire dates?
The stretch where the governance pillar of this index became countable. Declarations are cheap and statutes are not, and both leave machine-readable records with numbers on them.
What test did Turing propose instead of a definition?
Alan Turing opens "Computing Machinery and Intelligence" by rejecting the question "can machines think?" as too ill-defined to answer, and substitutes an operational test that can actually be run. The move that starts the field is a refusal to argue about definitions and an insistence on something observable. This…
Where was the name artificial intelligence coined?
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon submit a proposal to the Rockefeller Foundation for a two-month, ten-man study at Dartmouth College the following summer. The phrase "artificial intelligence" is invented in that document, largely to sidestep the baggage of "cybernetics". The…
How did Good describe the intelligence explosion?
I. J. Good, who had worked with Turing at Bletchley Park, sets out the recursive argument in "Speculations Concerning the First Ultraintelligent Machine": a machine that designs machines better than people do would design a better machine than itself, and the process would not obviously stop. Good called such a…
When did a field call a moratorium on itself?
Paul Berg and ten co-signatories — among them David Baltimore, Stanley Cohen, Herbert Boyer and James Watson — publish a letter in Science asking laboratories worldwide to defer two named classes of recombinant DNA experiment until the hazards had been assessed. It is the one occasion on record where a research…
What rules did Asilomar write?
About 140 people — molecular biologists, plus lawyers, physicians and the press — meet at the Asilomar Conference Grounds in Pacific Grove, California, and agree a tiered containment scheme matching physical and biological safeguards to assessed risk. The summary statement was published in PNAS that June; the NIH…
Who named the singularity and dated it?
Vernor Vinge presents "The Coming Technological Singularity" at the VISION-21 symposium held at NASA Lewis Research Center. He takes Good's recursive argument, gives it the name that stuck, and — unusually for the genre — commits to a dated window. This index notices that detail: a dated claim can be scored…
Who built the paperclip maximizer argument?
In "Ethical Issues in Advanced Artificial Intelligence", Nick Bostrom introduces the thought experiment that has outlived every other piece of vocabulary in this field: a superintelligence whose top goal is the manufacture of paperclips, which therefore converts first the Earth and then whatever else it can reach…
When did Deep Blue take the rematch?
IBM's Deep Blue beats Garry Kasparov 3½–2½ in New York, winning the deciding game in nineteen moves. Chess had been the standing proxy for machine intelligence since the 1950s. It stopped being one the moment it fell, which is the pattern every benchmark on this list follows.
How did AlexNet end the argument about deep learning?
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton enter a deep convolutional network trained on two consumer GPUs into the ImageNet challenge and record a 15.3% top-5 error rate against 26.2% for the next best entry. The gap was too large to argue with, and the field reorganised around it inside about eighteen…
When did AlphaGo beat Lee Sedol 4–1?
Go had been the standing counter-example: too large a search space, too dependent on intuition. DeepMind's AlphaGo took the five-game match in Seoul. Lee Sedol won game four, and that game remains the last one a top human won against the system.
When was the transformer paper posted?
Eight authors at Google publish an architecture that drops recurrence and convolution entirely in favour of attention. Almost every system this index counts today is a descendant of that paper. It is the clearest case in the record of a step change arriving as an ordinary preprint on an ordinary Monday — which is…
When did ChatGPT ship as a research preview?
The underlying model was not new and the interface was a text box. What changed was who had access. This is the date the general public acquired an opinion about AI, and the date the attention pillar of this index stopped being quiet.
What was the research-priorities letter, and the money behind it?
The Future of Life Institute circulates "Research Priorities for Robust and Beneficial Artificial Intelligence" to attendees of its Puerto Rico conference (2–5 January 2015) and publishes it on 12 January. Stephen Hawking, Elon Musk and Steve Wozniak sign alongside several hundred working researchers, and Musk…
How was OpenAI announced as a non-profit?
Sam Altman and Elon Musk announce OpenAI as a non-profit research company with a stated billion dollars in committed funding, framed explicitly as a safety argument: that the safer route is broad distribution rather than concentration. Musk left the board in 2018 and later founded xAI. Almost every sentence written…
What are Concrete Problems in AI Safety?
Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman and Dan Mané publish a preprint recasting the worry as five ordinary engineering problems: avoiding negative side effects, avoiding reward hacking, scalable oversight, safe exploration, and robustness to distributional shift. It is the…
Why do the Asilomar AI Principles borrow the venue on purpose?
The Beneficial AI conference, held at Asilomar on 5–8 January 2017, produces 23 principles across research questions, ethics and values, and longer-term issues. The choice of venue is the entire rhetorical move: chapter II, invoked by geography. Whether an intangible general-purpose technology admits the same…
What is the pause letter?
Eight days after GPT-4 shipped, the Future of Life Institute publishes an open letter calling for an immediate six-month halt to the training of systems more powerful than GPT-4, gathering tens of thousands of signatures including senior researchers and industry figures. No pause occurred. It is the clearest marker…
Why did Hinton leave Google to speak freely?
Geoffrey Hinton — whose students built AlexNet, who shared the 2018 Turing Award for the work underpinning modern deep learning, and who shared the 2024 Nobel Prize in Physics — resigns from Google and tells the New York Times he wants to discuss the risks without it reflecting on an employer. A field's most…
What one sentence was signed by the people building it?
The Center for AI Safety publishes a single-sentence statement placing extinction risk from AI alongside pandemics and nuclear war as a global priority. Its signatories include the chief executives of the leading labs. The document is one sentence by design: it was written so that nobody could decline to sign it on…
How did the United States regulate AI by executive order?
Executive Order 14110 sets reporting duties for training runs above a compute threshold and directs a long list of agencies to act. Whatever its later fate, it is the moment AI capability acquired a number in United States federal law, and the moment the governance pillar of this index acquired a signal worth counting.
What was Bletchley Park, and twenty-eight signatures?
The first international AI Safety Summit produces the Bletchley Declaration, signed by twenty-eight countries and the European Union, including both the United States and China. The venue was chosen for the symbolism. The substance was an agreement that the thing is worth meeting about — which, on a list where most…
How did Seoul turn a declaration into commitments?
The AI Seoul Summit produces the Seoul Declaration and, alongside it, the Frontier AI Safety Commitments, under which frontier developers agreed to publish safety frameworks including the thresholds at which they would judge risks intolerable. Bletchley established that the summits exist. Seoul is where they…
When did the EU AI Act enter into force, on a staircase?
The first comprehensive statutory AI regime anywhere takes effect: risk-tiered, with some uses banned outright. The obligations arrive on a published staircase — prohibited practices from 2 February 2025, general-purpose model obligations from 2 August 2025, general application from 2 August 2026, and further…
How did the Nobel committee credit the machinery?
John Hopfield and Geoffrey Hinton receive the Nobel Prize in Physics for foundational discoveries and inventions enabling machine learning with artificial neural networks. One of the two laureates had resigned from his job seventeen months earlier in order to talk about where the work was going. Both facts are in…
How did Paris change the subject, and which two governments declined to sign?
The AI Action Summit in Paris gathers participants from over a hundred countries and produces a non-binding Statement on Inclusive and Sustainable Artificial Intelligence with 61 signatories. The United States and the United Kingdom did not sign it, and the framing moved from safety toward investment and adoption.…
What is the second International AI Safety Report?
The International AI Safety Report, chaired by Yoshua Bengio and written by over a hundred independent experts with backing from more than thirty countries and intergovernmental organisations, publishes its second edition; the first appeared in January 2025. It is the closest thing this field has to an IPCC-style…
What is the paperclip maximizer, properly attributed?
Bostrom's 2003 argument is not that a machine would hate anyone. It is that capability and goals are independent — a system can be arbitrarily competent in the service of an arbitrarily trivial objective — and that almost any objective implies the same instrumental sub-goals. The paperclips are deliberately absurd.
What is the SIREN scale, level by level?
Five levels, counting down toward louder, borrowing DEFCON's grammar. Each is a statement about where the composite sits in this index's own frozen reference distribution. Not one of them is a statement about danger.
What does SIREN 5 · DORMANT mean?
Needle at rest. The floor of the scale. Papers, releases, filings and market odds are all moving slower than this index calls normal. That is not safety — it is a quiet hour on the instruments, and we post it exactly as loudly as we post the rest.
What does SIREN 4 · ROUTINE mean?
Needle breathing. Business as usual for a field whose usual is already loud. The scale is centred here by construction — 50 is the middle of the frozen reference distribution — so a reading in this band means nothing across the five pillars is behaving out of character.
What does SIREN 3 · ELEVATED mean?
Off the rest stop. Something is carrying the average. One or more pillars have pulled clear of their own reference distribution and the composite has followed them up. It says louder than usual. It does not say why, and the pillar breakdown is where you go to find out.
What does SIREN 2 · ACCELERATED mean?
Pinned high. The top decile of the record, with several pillars high at once. That is what it looks like when releases, capital and paperwork land in the same window — and it is also what it looks like when one source is having a strange week. The receipt tells them apart.
What does SIREN 1 · UNPRECEDENTED mean?
Off the top of the chart. Past the top of the reference distribution: the frozen backfill holds nothing like this. The name is a statement about our own record running out, not about the world ending. Those are two different sentences and this index refuses to merge them.
What does this index not claim?
It is not a probability of doom, a forecast, or a statement about danger. A scale is read as whatever its name suggests; these level names describe the instrument rather than the danger.
What is the cautionary tale, in full?
The WHO six-phase pandemic scale measured geographic spread and nothing else. When Phase 6 was declared for H1N1 in 2009, the public read it as a statement about severity. The lesson: a scale is read as whatever its name suggests. So SIREN level names describe the instrument rather than the danger, every surface that prints a level also prints the disclaimer, and the copy generator is barred from the future tense by a unit test.
What does this index actually count?
Five pillars, fixed and never reordered. Each is a mean over its live public sources, normalised against a frozen reference distribution built once from backfill and never touched again.
What does the Capability pillar count?
Preprints, model releases, release notes, repository activity. The field leaves a paper trail every time it moves. This pillar counts the trail; it does not read it.
What does the Compute & Capital pillar count?
GPU spot prices, rentable supply, disclosed capital expenditure, and what the market pays for the companies selling the shovels. Weights can be kept secret. The power bill cannot.
What does the Attention pillar count?
News volume, search interest, forum and repository chatter. The softest pillar, and the first to spike at nothing — which is precisely why it is one of five and not one of one.
What does the Governance pillar count?
Bills, consultations, statutory instruments, enforcement notices. The slowest pillar, and the one that leaves the cleanest dates: a regulation has a number, a journal and a day.
What does the Markets pillar count?
Polymarket, Kalshi, Manifold. The only pillar whose sources pay for being wrong, which makes it the only one where an opinion carries a price instead of a byline.
Is SIREN a p(doom)?
No. The SIREN composite is a percentile of observed public activity across five pillars, rescaled to 0-100. It carries no units of risk, contains no forecast, and is not scored by any model or any person. It measures how much is happening, not how bad it is.
Does SIREN publish a history-page doom score?
No. Activity tempo, not probability of harm. This page is a sourced timeline of the AI-risk argument — not a probability of harm.
Search-shaped sibling:
How the SIREN AI activity index is computed · methodology
· How does SIREN compare to DEFCON, the Doomsday Clock and p(doom)?.
