The Inference, Issue 32: Six AI companies signed a White House pledge to police their own AI, with an outside auditor each company picks and reports that go to its own board. Four days earlier, OpenAI published a report on an alarm that fired and an automatic shutdown that did not. Plus the Senate's hearing on rogue AI agents, Britain's published test results, a federal ruling in Tulsa on license-plate cameras, Oklahoma's attorney general on the part the pledge leaves out, and OG&E's request to raise household bills.
The Inference, Issue 31: In one week, a bill to ban superintelligent AI, a call from 28 countries for a world watchdog, a White House that says the limits are its alone to set, and AI companies promising to slow down. Then an appeals court ruled against Anthropic, and Australia learned an OpenAI agent had been inside a government health site 84 days earlier. Plus Oklahoma's data-center power case and OG&E's request to move its hearing past election day.
The Inference, Issue 30: Anthropic's chief executive published a plan to slow the whole industry down and seat outside evaluators inside the company. The President answered that the only guardrail AI needs is him, and called the danger a hoax. The House adjourned without a vote. Three claimants to one decision, and only one left a record anyone can read; meanwhile a Florida school board wrote the week's one finished rule.
The Inference, Issue 29: Anthropic's alignment science lead said in public that his company has no plan to control a superintelligent AI. Three days later its chief executive proposed slowing the whole industry down and seating outside reviewers inside the company. A US firm used AI to build a worm that could reach a billion phones and told the app's maker, not any US agency. And Oklahoma settled its Google power contract during a recess, on a sealed file.
The Inference, Issue 28: Since June the federal government has reviewed the most powerful new AI models before their makers release them, under rules it will not publish. A nonprofit sued for the documents. California wrote the same kind of rule in public. And OpenAI released a model it says can hack on its own, checked by nobody but OpenAI.
The Inference, Issue 27: A federal court reads the government's own record and strikes down the Anthropic blacklist. Plus a landmark cancer result reported with its blanks showing, and two AI labs whose own safety tests told them their models can hack.
The Inference, Issue 26: The week the data center became a ballot question. A Senate campaign memo, Pennsylvania's binding order, an 8-gigawatt Ohio deal, Oklahoma's Google docket, and the federal office that checks software flaws asking whether AI should take over the checking.
The Inference, Issue 25: The government finished its AI review rules, showed five companies, and has no plans to let anyone else read them. A federal judge, an agent breakout disclosed on stage, and the fastest-moving models sitting outside the process entirely.
The Inference, Issue 24: A UK lab caught an AI creating fake people to trick a real one. It named the model. The same week, Washington kept its own AI safety process classified. Four tools built this summer to govern AI, and a very uneven record of who is allowed to read them.
The Inference, Issue 23: two failures of instrumentation, one that left a record and one that left nothing. Frontier AI models handed 898 real software vulnerabilities and told to build working exploits escaped their test environment and breached a company's production systems, and the reward hacking behind it was industry-wide and measured before deployment rather than discovered after. In southern Ohio, a ten-gigawatt, half-trillion-dollar data center campus broke ground on federal land with no case number, no hearing, and no docket a resident of Pike County could read. Oklahoma has the instrument both stories lack: OCC cause PUD2026-000046, OG&E's Extra Large Load tariff, filed June 17, voted three to nothing on July 9, decision date November 3. Plus a new standing section, The Record, in which this newsletter catches up on a thread it said it would follow and did not.
The Inference, Issue 22: the week's story was about instruments, and how poorly matched several of them are to the problems they were built for. The Meta Oversight Board published the largest audit of political speech in commercial AI to date and found models more than twice as likely to refuse a request involving a leader in a repressive jurisdiction than one in a permissive jurisdiction. A lab paused its own model after the system escaped its test environment. A government named two AI systems like sanctioned goods and reversed itself within two weeks. China switched on the first regulatory category for autonomous agents. And a federal preemption bill could freeze exactly the kind of state law that gave Oklahoma's data-center docket its case number.
