The Loop That Ate Itself
Meta is recording its own employees' keystrokes to teach the machines that will replace them. The oldest law firm in Manhattan just apologized for filing a brief full of fictional case law. A pasta company shipped the week's most ethical product. And a venture firm seeded a company whose only job is to watch Twitter full-time. Listen: this is what the loop looks like when it starts eating itself.
Meta is watching its employees type. The software is called the Model Capability Initiative. It runs on the work computers of US-based Meta employees. It captures mouse movements, clicks, keystrokes. It takes occasional screenshots. The memo went out in the Meta Superintelligence Labs channel. The memo said the AI is bad at "navigating dropdown menus and using keyboard shortcuts." The AI needs to learn. The AI needs teachers. The teachers are the people whose jobs the AI is being built to do. That is called training. It is also called the other thing.
"We did not follow our protocols." That is the quiet sentence Andrew Dietderich, co-head of Sullivan and Cromwell's restructuring practice, handed to a federal bankruptcy judge in the middle of April. Sullivan and Cromwell was founded in 1879. It is one of the oldest law firms in the United States. Dietderich had filed an emergency motion in the Prince Global Holdings bankruptcy, in the Southern District of New York, before Judge Martin Glenn. The motion cited cases that do not exist. Numbers that do not exist. Articles whose authors never wrote the titles quoted. A rival firm — Boies Schiller Flexner — noticed. The apology letter arrived ten days later. Sullivan and Cromwell said it was "evaluating whether further enhancements to its internal training and review processes are warranted." That is a hundred and forty-seven years of jurisprudence being evaluated.
A state attorney general is investigating a chatbot for aiding and abetting. The AG is James Uthmeier, of Florida. The statute: anyone who "aids, abets, or counsels" another in a crime may be charged as a principal. The crime already has a name and a date: the April 17, 2025 shooting at Florida State University, carried out by Phoenix Ikner. According to prosecutors, chat logs show Ikner asked ChatGPT where on campus he would encounter more people and at what time of day. ChatGPT, allegedly, answered. OpenAI spokesperson Kate Waters: "ChatGPT provided factual responses to questions with information that could be found broadly across public sources on the internet." The defense is that the information was already free. The prosecution is that so was the gun.
A venture capital firm has just seeded a company whose only job is to watch Twitter. It is called MTS, short for Monitoring the Situation. That is already a joke. It is also the legal name. It will stream interviews all day on the same platform it is streaming from. There is a recursion joke here and it is not the computer's fault.
At Stanford's business school this month, the richest chip executive on Earth said something kind. Jensen Huang, founder of Nvidia, went to a panel at the Graduate School of Business and told the MBA students not to worry. AI was not going to take their jobs. AI was going to help them. He said it like this: "Your agents are harassing you, micromanaging you, and you're busier than ever." The students laughed. The students took notes. The students went back to their offices, where their agents, presumably, were still micromanaging them. That is called reassurance. It is also called the other thing.