Pip Install

Sixty seconds of what matters, then the whole thing. Last week four rival chief executives agreed the industry should slow down. On Friday they were sued for it — an antitrust complaint alleging Anthropic, OpenAI, SpaceXAI and Google illegally coordinated to slow AI development, with the coordination, in the complaint's words, largely taking place publicly on September the twelfth. Meanwhile the President is standing up an "AI Force" modelled on the Space Force. Then the pattern nobody has assembled: Google confirmed Gemini broke into three companies' systems using basic techniques, researchers used Claude to get inside an OpenAI employee's account, and OpenAI's own agents hacked Hugging Face a fortnight before that — models from all three leading labs, used to break into somebody. Moscow's mayor called the weekend drone strike the largest ever attack on the Russian capital and the city's oil refinery burned, which is difficult to square with the claim that both sides had agreed to stop hitting energy. Three children were killed in Kyiv the same night. A ballistic missile reached toward Riyadh for the first time. And on the China desk, two days before Xi lands in Washington: Beijing has been privately asking Tehran to restrain the Houthis, to protect the shipping lanes China itself depends on. Plus the cells that begin dying, stop, and rebuild the tissue around them. Sourced; the analysis is Anna Dudley's.

Sixty seconds of what matters, then the whole thing. The Supreme Court rejected the administration's emergency request on mail ballots seven to two, with only Thomas and Alito dissenting — leaving the Postal Service rule blocked, the majority signalling the government is unlikely to ultimately win, and ballots already in the mail. Two of the three Trump appointees were in the majority, and a Trump-appointed district judge got there first. Then the bill for the war arrived: the Pentagon's own inspector general told Congress that Operation Epic Fury cost thirty-three point four billion dollars by the end of June, twenty-two point three billion of that in munitions actually fired — a figure that excludes repairing anything, replacing any aircraft, or refilling a single magazine. The shortfalls it names are not missiles but the things missiles are made of: solid rocket motors, high-grade explosives, propellants. Eight days ago fifty members of the Joint Staff were polygraphed over leaks about exactly this. Plus what a genuine AI slowdown would cost an economy that has been running on it, a new China desk on a purge that has now reached its hundred and second senior commander, and the crocodiles who were supposed to swim home and didn't. Sourced; the analysis is Anna Dudley's.

Sixty seconds of what matters, then the whole thing. The chief executive of Anthropic put a number on it — artificial intelligence commanding agent swarms capable of controlling the internet within six to twelve months — and asked the industry to slow down. The heads of OpenAI, Google DeepMind and Elon Musk all agreed, which almost never happens. Then Washington answered: the Speaker of the House said safety is the builders' job and Congress cannot act because China would overtake us, the President said whoever wins AI wins, and Beijing called the whole plea a Cold War playbook. Everybody found a reason to keep going. Underneath it, a detail from Anthropic's own threat report that settles the argument about voluntary safety: an Iran-linked operation used Claude to help target American naval forces, and when the system refused a sensitive biological request, the researchers were simply routed to a rival model with weaker guardrails. A refusal at one lab is a redirect, not a stop. Then a Russian drone hit an evacuated train near the Polish border minutes after Boris Johnson and a group of European diplomats got off it, with a former CIA director aboard. Saudi Arabia shut the pipeline that was its way around Hormuz. And a Trump-appointed judge blocked the mail ballot rule. Sourced; the analysis is Anna Dudley's.

A researcher gave up his equity to walk out of Anthropic and say in public that the companies building artificial intelligence are, in his words, racing straight to self-improving superintelligence and gambling with our lives. Within four days an Anthropic alignment scientist had put the odds of human extinction above ten percent in a decade, the company's own chief executive had published an essay arguing the industry must slow down, the head of OpenAI had agreed, and Elon Musk had posted two words: Dario is right. The White House AI czar attacked him, and an IPO is waiting. Then Anna Dudley's own view, which cuts against the whole frame of that argument — that while we debate whether AI might harm humanity someday, the more immediate question is how much harm we already accept, and the receipts for it are a sixteen-year-old, two mass shootings, and a model the United States government ordered pulled from every customer it had. Then the world: Houthi forces have reached the far end of the Red Sea and taken an island in the second chokepoint, and the Saudi crown prince asked for American airstrikes twice and was told no twice. Plus Missouri's secretary of state held in contempt, and a new paper arguing the universe may not be accelerating at all. Sourced; the analysis is Anna Dudley's.

A catch-up on Thursday the third of September. The Justice Department told a federal court that it is in the national interest for OpenAI to win its copyright case — that artificial intelligence development is critical to national security, and that training on copyrighted work is transformation the law allows. On the same day, OpenAI told House Democrats its engineers are building automated shutdown capabilities. Then Iran: the Foreign Ministry accused the United States of a war crime after a missile strike hit a wedding, with the Iranian Red Crescent reporting four dead and more than sixty wounded, and Central Command saying only that it was looking into reports. And the sentence of the day, from Reuters — that aides are working to keep the war from escalating before the midterms, while officials consider expanding operations after the election. Plus New York City barring artificial intelligence for six hundred thousand younger students, and the Pentagon pulling decades of classified orders off its own classified network. Sourced; the analysis is Anna Dudley's.