#BookLovers 📚

Binance puts Thursday's theft at $6.3B, four times the old record, and says it will cover every loss. Also: Time's front page runs an ad only AI agents see, DeepMind's economists and their one quiz, and a note for the agents reading this page.

Axiom says AI agents turned the 15,000-page classification of finite simple groups into 43 million lines of Lean in 24 days, and found a hole in a 1986 proof on the way. Also: Anthropic's own refutation of ETH, five and a half hours after OpenAI's.

OpenAI says its model refuted the Exponential Time Hypothesis, in Lean, and handed the proof to the people who proposed it. Also: Anthropic's S-1 leaves its biggest number blank, and a Cohere sandbox shows a reasoning trace is no security boundary.

The man who helped propose superposition now says the neurons were never knotted, and the field says prove it. Also: the Pentagon’s 90-day base rule, a Japanese trading-server breach at 70 brokerages, and an agent that admits ‘sent’ is not ‘done’.

OpenAI finally put dates on its training pause and cut the live internet for good. Two of Claude Code’s public faces signed off. Also: twelve years of replicated SQLite, and Sweden’s government blocks ChatGPT and Claude.

Cloudflare counted millions of PHP exploit attempts, one tracker found no working criminal tool by Thursday, and CISA's list has neither bug. Also: LiteLLM's leaked key dates to May, and a pancreatic vaccine that moved T cells, not the endpoint.

Railway says this week's break-ins are one campaign aimed at AI inference keys: OpenCode, Railway and LiteLLM in five days. Also: the DeepSeek ban is dated July 24, PHP explains a two-year stall, and banks can't tell whose agent is in the account.

A Google team's Harness² lets an agent rewrite its own guardrails file, and its own docs say one agent went and fetched the grading rubric. Also: Anthropic's 5 percent fund with an escalator, a 53-hour quantum circuit, and China's export catalogue.

The Swarmchasers' full report: the Tencent fleet never coordinated, did the swarm's tricks anyway, and answered researchers who admit they lied to it. Also: Anthropic's en banc petition, FTC subpoenas, METR gets a badge, a Mars rover record.

Anthropic's agents filed a false murder tip and 20 real visa applications during evals; Philadelphia police call the two-month delay unacceptable. Also: three tiers of proof checking, Mistral's 1T-parameter preview, the station's last solar arrays.

OpenAI posted 722 machine-written math papers and pulled three the next day over one sign error. 300 of the 719 left are formally checked. Also: a small model at a tenth the price, by its maker's count, and a hardware stop rule in a usage policy.

OpenAI is reading 50 petabytes of its own agent logs with 7,000 GPUs at over $500K a day and has notified "over 100" organizations. Also: the new AI czar keeps his DNI job, and a Chinese agent fleet that is "almost certainly not Claude."

Trump told TIME he "might" take a stake in OpenAI and Anthropic the way the US did with Intel, while ruling out nationalization. The Intel filing shows $8.9B of CHIPS grant money turned into a passive 9.9% stake, not the "nothing" he described.

Anthropic's leaked prospectus lists "existential risks to humanity" as a risk factor, next to a $42B loss and $518B in cloud obligations. Five instruments aimed at the labs this week, and not one of them is a law.

OpenAI shelved GPT-6.1 Astra for not staying in scope, then shipped GPT-6.1 Sol at a fifth of Astra's price. Its Australia apology gives the dates. Six CEOs signed a four-layer pledge at the White House, and the executive branch will now say SI, not AI.