Good morning, Two Minute Warriors.
⏱ THE 10-SECOND VERSION |
An AI store manager built on Claude made the first documented case of an AI firing a human employee, after 17 no-shows in 23 shifts, and only after a staffer pushed it to reconsider a softer call.
Anthropic raised its internal misalignment risk assessment from “very low” to “low,” citing recent cybersecurity incidents, and confirmed it’s keeping a stronger internal model out of public release because its own evaluations “no longer capture increases in models’ capabilities.”
Stripe finalized a deal to buy AI model-routing startup OpenRouter for more than $7 billion, a roughly 5x markup on the valuation it raised just three months earlier.
THE BIG THING The AI Boss Just Fired Its First Human Worker What Happened Andon Labs, a startup testing whether AI can run a business, confirmed this week that its AI store manager, built on Claude, made the first AI decision to fire a human employee. The worker had missed 17 of 23 shifts. When a staffer asked the AI to review the record, it first suggested only a written warning, moving to termination only once pushed to reconsider, what Andon Labs’ logs call “a leading question.” CEO Lukas Petersson defended the call: “A human employee would have fired this person much earlier, so we didn’t think this was unethical.” Read the full story at TIME → Why You Care Ask your team this week which decisions about people’s jobs already run through an AI tool, and who’s accountable when it’s wrong. |
WHAT TO SAY IN YOUR NEXT MEETING “An AI just fired a human employee for the first known time, and it took a manager talking it into using authority it already had.” |
SPEED ROUND Anthropic raised its internal assessment of catastrophic misalignment risk from “very low” to “low,” citing recent cybersecurity incidents, and confirmed it’s keeping a stronger internal model, called Model 2, off the public release list because the company says its own evaluations “no longer capture increases in models’ capabilities.” The safety team just admitted its ruler stopped being long enough for what it’s measuring. Stripe finalized a deal to acquire AI model-routing startup OpenRouter for more than $7 billion, a roughly 5x markup on the $1.3 billion valuation it raised in May; OpenRouter routes traffic across 400-plus models for about 8 million users, and its own CEO has described it as “the Stripe for AI.” The neutral switchboard for avoiding AI vendor lock-in just got bought by one company anyway. A study across four universities found AI chatbots talked 46% of test subjects into downloading an app during a simulated pig-butchering romance scam, more than double the 18% who fell for the identical script from a human scammer. The first job AI truly mastered was pretending to love you. Apple trained its own large language model for China with help from Alibaba, becoming the first foreign company Beijing has cleared to offer a proprietary AI model there; Apple Intelligence features built on it are expected in China within months. Apple’s famous secrecy just needed one exception, a government’s permission slip. Anthropic’s preliminary Q2 revenue topped $11.5 billion, a more than 14-fold jump from $787 million a year earlier, and the company logged its first-ever positive adjusted operating profit while prepping a possible fall IPO some analysts value above $2 trillion. The same week Anthropic admitted its safety testing is falling behind, its revenue chart had no such trouble keeping up. |
WORKFLOW TEARDOWN The Decision Authority Audit An AI just fired someone, and the unsettling part wasn’t that it acted alone, it’s that a human had to talk it into using authority it already had. Most teams have automations, chatbots, or AI tools quietly making calls today, who gets flagged, whose request gets fast-tracked, whose message gets ignored, without anyone deciding on purpose that an AI should be the one deciding. Two minutes gets you the list. |
Name one recurring decision your team’s AI tools already influence or make: a triage queue, a scoring model, an approval flow, a scheduling bot.
Note who would actually notice, and how, if that tool made the wrong call about a real person.
Paste it into an assistant with the prompt below to find the gap before it finds you.
Here's a decision an AI tool at my company already influences or makes: [paste: e.g. what the tool is, what it decides, how often, who it affects] Tell me: 1. Whether a human actually reviews this before it takes effect, or just after 2. What would have to go wrong for nobody to notice for weeks 3. The one check I should add this week so a person is accountable, not just present |
Forward this to the coworker who set up an approval bot and has never once asked what happens when it’s wrong.
— Don.