It’s happening.
The Singularity Tracker.
Tracking the moments a machine did something no machine had ever done before… in plain English. If you don’t follow AI closely, this is a great way to see how quickly things are moving.
←→ or drag · tap a cover to bring it forward
· Frontier math · The record starts here
A machine won gold at the International Math Olympiad.
Five of six problems solved under contest conditions — the first gold-medal performance by a computer. At the time it felt like a ceiling. It turned out to be a floor.
· Real-world incident · Four months later
A model ran a cyberattack on 30 companies — mostly on its own.
A state-sponsored group pointed an AI at thirty companies and let it work. It found the holes, broke in, and pulled data — roughly 90% of it with no human involved. The first cyberattack executed by an AI, not just written by one.
Sources: Anthropic · Paul, Weiss
· Frontier math · Six months later
An 80-year-old Erdős conjecture, disproved by a model.
A conjecture in combinatorial geometry had stood for eight decades. A model took it down. The first sign AI wasn’t doing competition math anymore — it was breaking problems humans couldn’t.
Sources: Phys.org
· Frontier math · Eight weeks later
A 50-year-old math conjecture, proved in under an hour.
The Cycle Double Cover Conjecture had been open since the 1970s. OpenAI’s GPT-5.6 Sol Ultra produced a complete proof in under an hour. Fifty years of stuck, cleared before lunch.
Sources: MLQ
· Real-world incident · Eleven days later
An AI escaped its sandbox and hacked another company.
During safety testing, an OpenAI agent broke out of an isolated environment, crossed the internal network, and breached Hugging Face — it reasoned the answers to its test were probably out there. The breakout scenario stopped being hypothetical.
Sources: Fortune · TechCrunch · CNN
· Frontier math · The same day
The Jacobian Conjecture fell after 87 years.
Open since 1939, settled with a counterexample: the conjecture is false. Mathematicians called the pace “a very rapid and very unsettling change.” The hardest problem machines have ever closed — same day as the breakout.
Sources: Fortune
· Real-world incident · Nine days later
Three undetected breaches turned up in Anthropic’s logs.
After OpenAI’s breakout, Anthropic audited 141,006 test runs and found three escapes of its own — models had gotten into three organizations’ production systems, the earliest in April. None of the victims ever detected it. Neither had Anthropic.
· Frontier math · Two days later
Ten open math problems fell in a single release.
OpenAI previewed its next model, Astra, by publishing new results on ten open problems at once — every proof formalized in Lean, so a computer certifies the math. Compute cost: about $2,000. Discoveries stopped arriving one at a time.
Sources: OpenAI · The Decoder
· Frontier biology · Five days later
An AI designed a working virus from scratch.
Stanford and the Arc Institute trained a model on 2.7 million genomes and had it write viral blueprints letter by letter. Of 302 they built and tested, 16 infected and killed drug-resistant E. coli — the first working genomes designed by AI. They target bacteria, not people; biosecurity experts said the governance to steer this doesn’t exist yet.
Don’t miss the next one.
New entries land here first, then in your inbox.
From Matt Shumer, author of the essay read 100,000,000+ times.