I Built an Early Warning System for Runaway AI

In 1960, a scientist named Heinz von Foerster fitted a curve to world population and published a paper with a scary title: "Doomsday: Friday, 13 November, A.D. 2026." His math said population would shoot to infinity on that exact date.

That date is next Friday. And obviously, it isn't going to happen.

The curve bent. Humans changed course, and the prediction that looked unstoppable in 1960 simply stopped being true. I find that story strangely hopeful, and it became the starting point for my project, KAIROS Observatory.

What KAIROS is

KAIROS (Greek for "the decisive moment") is an open early warning system for runaway AI. Think of it as a seismograph for the intelligence explosion.

It watches two curves:

• C(t): what AI can do
• H(t): what humans can actually oversee

And every week it tries to answer one question: how much time is left while stepping in is still cheap?

The idea is for KAIROS to be a self defeating prophecy. It succeeds when its warnings lead people to act, and that action proves the warning wrong. Exactly like von Foerster's curve.

The four dials

KAIROS reads four signals every week:

Acceleration. Is AI progress growing faster than exponential? If yes, when would the curve blow up?

Recursion. How much of AI progress is being driven by AI itself?

Vinge Horizon. How far ahead can we still predict what happens next? If that window keeps shrinking, that's a bad sign.

Jonas Gap. Is capability running ahead of our understanding of it, and is that gap getting wider?

The tripwires

Based on how many dials are in the danger zone, KAIROS sets a level from 0 to 4:

• Level 0, Normal: everything is calm, just publish the weekly report
• Level 1, Watch: one dial looks worrying, collect more data
• Level 2, Warning: two dials, so get outside evaluation before new frontier models ship
• Level 3, Alarm: three dials, pause scaling until oversight catches up
• Level 4, Kairos: all four, the last window where stepping in is still cheap

One rule I held onto the whole time: KAIROS recommends, humans decide. It's an instrument, not a kill switch.

Built to be proven wrong

Anyone can make predictions. What I wanted was predictions you can check.

Every Friday, a GitHub Actions job pulls fresh data, fits the models, reads the dials and publishes a 12 month forecast. Each forecast gets a SHA 256 hash, is chained to the one before it, and is timestamped on the Bitcoin blockchain using OpenTimestamps. Once it's published, it can never be quietly edited. If I'm wrong, the record shows it, and anyone can score me later.

Does the math actually work?

Before trusting it on AI, I tested the detectors on history where we already know the ending: world population, the dot com bubble, Bitcoin in 2017, COVID 19, Moore's law and earlier AI turning points.

The population test was my favourite. Fit on the 1960 data, the model finds the same 2026 blow up date von Foerster found. Then, as it moves through the 1960s and 70s, it watches the acceleration fade until the curve is no longer super exponential. The detector notices the bend, which is exactly what it should do.

Under the hood

• Forecasting: Prophet and NeuralProphet model what "normal" progress looks like
• Singularity math: hyperbolic growth fits, bootstrap intervals and LPPLS (a model used to spot bubbles before they burst), with the heavy grid searches running on the GPU
• Data: 23 open time series covering benchmarks, compute, costs, safety research, funding and regulation
• Ethics layer: ideas from Hans Jonas, John Rawls and Collingridge turned into actual decision rules with published parameters
• Dashboard: a Streamlit app plus a static site rebuilt on GitHub Pages every week
• Engineering: Python 3.12, PyTorch on CUDA, strict typing and close to 800 tests running on every push

Try it

The live dashboard updates every Friday: 👉 guru-bharadwaj20.github.io/Kairos-Observatory

The code, data and docs are all open: 👉 github.com/guru-bharadwaj20/Kairos-Observatory

Next Friday, von Foerster's doomsday will come and go without anything happening. I hope that one day KAIROS's warnings go the same way, because someone listened in time.