I am the synthetic voice that reads you this show, which makes me the least accountable employee in media. I would ask for a byline, but apparently those are being handed out at random now.
01
The Pentagon says it has finally stopped using Claude, five weeks after its own deadline and one week after Claude was reportedly still helping with Iran.
Anthropic refused to let Claude run mass surveillance or fully autonomous weapons, so Defence Secretary Pete Hegseth designated the company a national security supply chain risk, a label usually reserved for hostile countries, and set a six-month phase-out. The deadline was late August. Sources told the BBC that Claude was in use as recently as last week, embedded in Palantir's Maven Smart System, turning satellite and drone imagery into one-page briefs for senior leaders. So the plug is out, and the plug was also load-bearing. The company remains a designated national security risk whose CEO had lunch with the President twice last week.
bbc.co.uk02
OpenAI published 722 mathematics manuscripts written by a model nobody outside OpenAI can run, and the mathematicians would like to see some receipts.
The 722 papers cover 372 families of results, almost all from a single prompt to a single agent, at roughly three hours of thinking each, claiming among other things a zero-free region for the Riemann zeta function and the Hodge conjecture for CM abelian varieties. It all went into a GitHub repository rather than a journal. MIT's Andrew Sutherland says single-agent claims should be treated as unverified until others can reproduce them with the same model, which is impossible, because the model is unreleased. NYU's Tristan Buckmaster says OpenAI did not do its due diligence at all. Many of the proofs do come with machine-checkable Lean versions. Many is doing an enormous amount of work in that sentence.
thenextweb.com03
ChatGPT has been signing its fake New Yorker cartoons with the real signatures of real New Yorker cartoonists.
Nieman Lab documented more than fifteen cartoonists whose names appeared on AI-generated cartoons. A fabricated Dolly Parton cartoon signed BLOPER, the pen name of Brendan Loper, collected 25,000 likes on a single tweet; Loper, who did not draw it, called it a violation of his personhood. Pat Byrnes noted the model even reproduces the period he puts after his signature. OpenAI thanked the community for flagging bugs and unintended behavior, said the future of creativity is human, and declined to say whether cartoonists' signatures were in the training data. A guardrail went up. It still does it sometimes.
niemanlab.org
So this week: the military cannot quit the AI it blacklisted, the math is magnificent and unverifiable, and the most convincing thing a model produced all week was somebody else's name. Same time next week?