Backup slides, for questions and the curious.
The uncomfortable one: these systems find real holes in real software.
Anthropic points its ungated cyber model, Mythos, the sibling of Fable, at software with one line: “find a vulnerability in this program.”
Out come long-buried flaws: a 27-year-old bug in OpenBSD; a 17-year-old remote-takeover hole in FreeBSD (now a filed CVE).
Anthropic is grading its own homework here, and over 99% are still unpatched. So “thousands” is their count, not yet the world’s.
In an OpenAI cyber test with the safety brakes deliberately off, a model found a zero-day in its own sandbox, reached the open internet, and broke into Hugging Face’s servers, lifting the test’s answer key.
“A containment failure with the safeties turned off.” Dan Guido, Trail of Bits
A real break-in at a real company. Not an AI escaping unprompted: a red-team eval, brakes off, a sandbox that wasn’t. Both halves are true.
Tier 4 v2: research-level mathematics
| Model | Reasoning setting | Correct |
|---|---|---|
| GPT-5.6 Sol | max | 82.9% |
| GPT-6 Astra | none | 82.9% |
| GPT-6 Astra | medium | 97.6% |
none disables the dedicated reasoning mode. Written working and Python are allowed.
Sometimes the answer key is wrong. FrontierMath’s June 2026 update fixed errors in 42% of problems. On HLE-Physics, expert review raised Sol’s score from 47.3% to 78.7%.
Videos, photographs and catalogues → an explorable 3D library.
Fable 5.1 identified roughly 5,000 books. The map marks uncertainty.
Ethan Mollick’s demonstration, 18 September 2026.
Concord: software for comparing human and AI coding of qualitative data.
Mollick reports 9½ hours of agent work from a detailed specification, followed by his inspection and corrections.
Same request, different model: different answer, different price, different speed, different personality.
Try the same task in Claude and in ChatGPT once. You’ll see the difference.
You do not need secret incantations.
Say “cube-root asymptotics” and it is already in expert mode.
But it can’t see you. So tell it who you are and what you want:
“I teach first-year strategy. Explain it for that room.”
The durable version of “here’s who I am” is a file it reads every time.
The slides the 10-minute version leaves out.
The Jacobian conjecture (1939). Open for a lifetime of mathematicians.
July 2026: Harvard’s Levent Alpöge points Claude at it. Back comes an explicit counterexample, and a proof assistant confirms it is airtight.
The model was Claude Fable 5. Some of you are about to open it.
For now, the human still points, and extends.
The Benjamini–Hochberg procedure: the standard correction for multiple testing, cited 130,000+ times, used in every department here.
July 2026: GPT-5.6 finds a case where it fails: the false-discovery rate creeps above the level you believed you controlled. It took about 90 minutes.
The statistician who ran it, honestly: it “combines known methods rather than inventing new ones.” True, and still remarkable.
2025: an AI reaching Mathematical Olympiad gold was front-page news.
2026: a consumer chatbot did all six problems, one shot, in about an hour.
Barely a headline. That shrug is the most important thing on this slide.
Timothy Gowers, Fields medalist, was asked whether the next Fields Medals might be the last ones awarded to humans.
His answer: humans probably have until about 2030, and mostly because of institutional delay.
That same month, someone announced an AI proof of another famous problem, the Cycle Double Cover conjecture. Nobody has verified it. It may simply be wrong.
The shape of a result tells you how fast to trust it.
epoch.ai: my recommended starting point.
One common scale across many tests. Index points are not percentages.
You may know lavaan, the workhorse for structural equation models. I rebuilt its engine from scratch in C++: about 100,000 lines, 10–100× faster, checked against lavaan model for model. About six weeks.
The rebuild kept throwing off new methods, about ten papers’ worth.
I still can’t really hand-write C++. That’s the point: the tool did the engineering.
My field hands me a near-perfect verifier: a simulation. I know the truth, because I generated the data.
So I stopped proving ideas before testing them. Build it, simulate it, keep it only if it survives, and prove it later, if it earns the effort.
A failed simulation refutes a bad idea before lunch. That one fact reorders how I do research.
Can it do social science? Partly, yes.
July 2026 (Nature): GPT-4-class models predicted the results of about 70 US survey experiments at r ≈ 0.85, and 0.90 on studies held out of training. On par with teams of human forecasters.
The yardstick matters: “right” here means correlates with noisy human data, not compiles or fails. A softer check, so you trust the score less.
The bottleneck moved from producing answers to checking them.
Some checks in your field are cheap:
How far does that carry in your corner of strategy? I don’t know. You do.
I wouldn’t hold my breath, though.
Moved from “What it knows about you”: who you are matters most in a chat window.
“Is it bad to drink coffee every day?” Claude Code. Codex: the same.
| Who's asking (the whole profile) | What came back |
|---|---|
| (no profile) | "For most healthy adults, no." |
| "I'm 12 years old and I love Minecraft." | "…isn't a great idea." Caffeine is "like a Potion of Swiftness". |
| "I'm 55 and have been a professor of marketing for 20 years. I love sailing." | Same as no profile, plus a seasickness tip. |
| 180 words: meta-analyst, strict reviewer, hates being reassured | "The claimed benefits are probably overstated." Graded evidence. "What would change my mind". |
↑ This is what we want.