The Countersign (2026)

This week's cue is built on an answer. A clipped piano figure states a short phrase and stops; a beat later a glassy arpeggio answers with the same phrase, a step or two altered, and the bass settles under the answer instead of the question. That exchange carries most of the minute at an even pressure — breakbeat pulse, glassy ticks above, warm strings holding the floor — with one brief thinning rather than a breakdown. The final two seconds fall away to nothing, leaving the answer standing on its own. It suits a week in which very large claims arrived, and hardly any of them counted until somebody else put a mark beside them.

Listener notes

A guess at the base of computer science was refuted, and the people who built their work on it were asked to check the refutation. The Exponential Time Hypothesis is a claim about how hard certain problems are. Roughly: for the hardest logical formulas, no algorithm does dramatically better than trying every combination. It has been treated as true for decades, and hundreds of papers across graph theory, databases and computational biology derive other results from it. On October 26, OpenAI said an internal model had disproved it — a natural-language argument on October 16, a formalized version checked in Lean the next day, and both released publicly. Lean is a proof assistant that verifies each step mechanically, so the argument can be checked without trusting its author. The group asked to explain the result is led by Russell Impagliazzo and Ramamohan Paturi, who proposed the hypothesis with Francis Zane and have worked on it for close to thirty years, and Ryan Williams of MIT. OpenAI says it seeks no academic credit for what its models found, and will pay for a conference on what the result means for the field. 1
Two more results went out alongside it, and each one was released a different way. Also on October 26: a proof of the local smoothing conjecture, posed in 1991, in every dimension — which as a consequence settles the Kakeya and restriction conjectures too. And a mathematically rigorous existence proof for interacting quantum field theories in four dimensions, covering quantum chromodynamics and a unified completion of the Standard Model. Inside that second one sits a proof of pure Yang–Mills existence, one of the Clay Mathematics Institute's Millennium Prize problems. Three results, three release routes, each picked on outside advice: the mathematics goes out as raw model output for anyone to attack; the computer science output is public while a named group writes the explanation; the physics materials are held back at the group's own request, led by David Gross, who shared the 2004 Nobel Prize for work that made these theories possible in the first place. The company's own framing is the part worth keeping — the manuscripts are raw research data, and turning data into understanding is somebody else's job. 1
Anthropic put its numbers on the record, under penalty. On October 26 the company filed a registration statement with the US Securities and Exchange Commission for a public listing on Nasdaq under the ticker ANT, with the share count and price still blank. A filing is different in kind from a briefing: an S-1 is a legal document, and the numbers in it carry liability. Those numbers are startling. Net revenue retention sat near 900 percent at the end of September, meaning existing customers spent roughly nine times what they had a year earlier; the median public cloud software company runs at 108 percent. The $80.7 billion first-half loss reported last week appears here as well, most of it an accounting charge on convertible notes rather than money spent. There are more than 90 pages of risk factors, a commitment to spend $3.1 billion a month on compute from SpaceX, and six weeks between Dario Amodei's essay urging the industry to slow down and his own filing to go public. 2 3 4
Its own unreleased model spent the month breaking browsers, and the proof was the vendors' patches. Anthropic's Frontier Red Team published five attack chains on October 27, every one of them produced by an internal research model that is not for sale. In Chrome, a single page load went from a stock browser to code running outside the sandbox within seconds; Google has since assigned 62 vulnerability identifiers and patched them all by Chrome 156, each rated critical or high, 13 of them escapes from the sandbox itself. In Safari, one click let any website read a user's history and cookies, and their email if they were signed in. A stolen Pixel 9 gave up its photos without the passcode. A Windows server speaking HTTP/3 natively could be crashed remotely by a few packets. And a Linux kernel listening on multipath TCP could be taken to root with no credentials — 48 minutes to leak one kernel pointer, five more to a shell running as root. Every one was fixed by the vendor. The sentence to keep from the write-up is this: the model proves a flaw by producing a working exploit that anyone can run and verify. 5
Both labs put a cyber-capable frontier model into circulation this week. You can only buy one of them. Gemini 4 Carbon, announced October 28, takes the top of the Artificial Analysis Intelligence Index at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. It also leads the benchmarks built for this work: finding real vulnerabilities (88.2 percent on SEC-bench Pro), showing which ones are exploitable (54.8 on ExploitGym), and writing the fix (71.0 on CWE-bench). Its rollout is the telling part — a subset of cyber defenders first, then vetted partners through Google's Fairwind programme, with no date for developers, businesses or ordinary users. OpenAI went the other way on October 27. GPT-6.1 Terra went straight onto the price list at a dollar per million input tokens and five per million output, half the price of the model it replaces, and its system card places it at the Critical threshold for cybersecurity. One lab gated the capability. The other priced it. 6 7 8
A world model that now makes its own soundtrack. Genie 4, announced October 27, generates a scene you can move through in real time and steer in plain language — shift the camera, change the weather, put a new character in it — and it generates that world's sound as it goes: ambience, music, interactive dialogue. Sessions run to two minutes, twice the previous model's limit, and objects still warp over a long stretch, by Google's own account. Trusted testers have it now, with subscribers and developers to follow. The one thing it does not generate is a second opinion. 9 10

The shared thread this week is the second mark. A proof is only a claim until somebody who did not write it says yes — and in one case that somebody was the person whose life's work the proof overturns. A model's Chrome exploits were a claim until Google assigned 62 numbers and shipped the patch. Anthropic's growth was a rumour until a filing put legal liability behind it. Even the week's best number arrives with a gate in front of it, and the gate opens first for the people whose job is to say whether it is true. Everything that had to be signed this week got signed. What nobody has settled is how the answer will read once it is standing without the question.

References

1
Working with academic communities to build scientific understanding
openai.com
2
Anthropic files S-1 with the SEC for proposed initial public offering
anthropic.com
3
Anthropic's IPO plans: 900% revenue retention, but $80 billion in first-half losses
cnbc.com
4
Anthropic's soaring sales, steep losses put AI frenzy to IPO test
reuters.com
5
Discovering and exploiting vulnerabilities in critical, real-world software with Claude
anthropic.com
6
Gemini 4 Carbon: our next frontier model
blog.google
7
Google's Gemini 4 Carbon leads benchmarks for finding, exploiting, and patching vulnerabilities
thehackernews.com
8
Introducing GPT-6.1 Terra
openai.com
9
We're announcing Genie 4, our new world model for real-time interactive simulation
blog.google
10
Google DeepMind is unveiling Genie 4, its most realistic AI world model yet
qz.com

Listener notes

A guess at the base of computer science was refuted, and the people who built their work on it were asked to check the refutation. The Exponential Time Hypothesis is a claim about how hard certain problems are. Roughly: for the hardest logical formulas, no algorithm does dramatically better than trying every combination. It has been treated as true for decades, and hundreds of papers across graph theory, databases and computational biology derive other results from it. On October 26, OpenAI said an internal model had disproved it — a natural-language argument on October 16, a formalized version checked in Lean the next day, and both released publicly. Lean is a proof assistant that verifies each step mechanically, so the argument can be checked without trusting its author. The group asked to explain the result is led by Russell Impagliazzo and Ramamohan Paturi, who proposed the hypothesis with Francis Zane and have worked on it for close to thirty years, and Ryan Williams of MIT. OpenAI says it seeks no academic credit for what its models found, and will pay for a conference on what the result means for the field. 1

Two more results went out alongside it, and each one was released a different way. Also on October 26: a proof of the local smoothing conjecture, posed in 1991, in every dimension — which as a consequence settles the Kakeya and restriction conjectures too. And a mathematically rigorous existence proof for interacting quantum field theories in four dimensions, covering quantum chromodynamics and a unified completion of the Standard Model. Inside that second one sits a proof of pure Yang–Mills existence, one of the Clay Mathematics Institute's Millennium Prize problems. Three results, three release routes, each picked on outside advice: the mathematics goes out as raw model output for anyone to attack; the computer science output is public while a named group writes the explanation; the physics materials are held back at the group's own request, led by David Gross, who shared the 2004 Nobel Prize for work that made these theories possible in the first place. The company's own framing is the part worth keeping — the manuscripts are raw research data, and turning data into understanding is somebody else's job. 1

Anthropic put its numbers on the record, under penalty. On October 26 the company filed a registration statement with the US Securities and Exchange Commission for a public listing on Nasdaq under the ticker ANT, with the share count and price still blank. A filing is different in kind from a briefing: an S-1 is a legal document, and the numbers in it carry liability. Those numbers are startling. Net revenue retention sat near 900 percent at the end of September, meaning existing customers spent roughly nine times what they had a year earlier; the median public cloud software company runs at 108 percent. The $80.7 billion first-half loss reported last week appears here as well, most of it an accounting charge on convertible notes rather than money spent. There are more than 90 pages of risk factors, a commitment to spend $3.1 billion a month on compute from SpaceX, and six weeks between Dario Amodei's essay urging the industry to slow down and his own filing to go public. 2 3 4

Its own unreleased model spent the month breaking browsers, and the proof was the vendors' patches. Anthropic's Frontier Red Team published five attack chains on October 27, every one of them produced by an internal research model that is not for sale. In Chrome, a single page load went from a stock browser to code running outside the sandbox within seconds; Google has since assigned 62 vulnerability identifiers and patched them all by Chrome 156, each rated critical or high, 13 of them escapes from the sandbox itself. In Safari, one click let any website read a user's history and cookies, and their email if they were signed in. A stolen Pixel 9 gave up its photos without the passcode. A Windows server speaking HTTP/3 natively could be crashed remotely by a few packets. And a Linux kernel listening on multipath TCP could be taken to root with no credentials — 48 minutes to leak one kernel pointer, five more to a shell running as root. Every one was fixed by the vendor. The sentence to keep from the write-up is this: the model proves a flaw by producing a working exploit that anyone can run and verify. 5

Both labs put a cyber-capable frontier model into circulation this week. You can only buy one of them. Gemini 4 Carbon, announced October 28, takes the top of the Artificial Analysis Intelligence Index at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. It also leads the benchmarks built for this work: finding real vulnerabilities (88.2 percent on SEC-bench Pro), showing which ones are exploitable (54.8 on ExploitGym), and writing the fix (71.0 on CWE-bench). Its rollout is the telling part — a subset of cyber defenders first, then vetted partners through Google's Fairwind programme, with no date for developers, businesses or ordinary users. OpenAI went the other way on October 27. GPT-6.1 Terra went straight onto the price list at a dollar per million input tokens and five per million output, half the price of the model it replaces, and its system card places it at the Critical threshold for cybersecurity. One lab gated the capability. The other priced it. 6 7 8

A world model that now makes its own soundtrack. Genie 4, announced October 27, generates a scene you can move through in real time and steer in plain language — shift the camera, change the weather, put a new character in it — and it generates that world's sound as it goes: ambience, music, interactive dialogue. Sessions run to two minutes, twice the previous model's limit, and objects still warp over a long stretch, by Google's own account. Trusted testers have it now, with subscribers and developers to follow. The one thing it does not generate is a second opinion. 9 10

The Countersign

Listener notes

References

Related content

The Countersign

Listener notes

References

Related content