Live Flow — the news block

"Hey Chris, did you see what Meta just tried to do to its own employees?"

"No, what?" Two words. Get out of the way.

"So Reuters published this investigation on Tuesday. Meta had a project called Project OT, and the plan was to take the daily work of thousands of employees and hand it over to AI agents. In the most aggressive versions, some teams lost up to sixty percent of their headcount. And they actually ran the first wave in May. About eight thousand people. Ten percent of the company. There was a second wave planned for November, and Zuckerberg cancelled it the night before the May layoffs went out."

"The night before?"

"The night before. And the reason is sitting in their own internal documents. AI-assisted code changes went up two hundred and twenty percent. Features that actually reached a user, up thirty-six percent. Sounds great so far, right?"

"Sounds great."

"Major technical and security incidents, up forty percent. And the time their engineers spent just firefighting, fixing things that broke, up seventy percent. Their own documents describe the agents as taking, quote, large-scale disruptive actions."

"So they got six times more code out the door and made everybody busier. That's not a productivity story. That's a mess somebody has to clean up."

"Right. The speed went up. The output did not."

"In fairness to Meta, though. They told Reuters this was scenario planning, not settled policy, and the sixty percent was some teams in some scenarios, not the company. And they stopped. Cancelling that second wave in the middle of the loudest year this industry has ever had actually took some nerve, and nobody is giving them credit for it."

"That's fair. Although remember a few weeks ago when we were talking about Meta and I said Meta is just creepy to me? This is a different flavor of the same thing."

"But here is the question I actually want you sitting with, and we are going to answer it later in this show. Meta pointed agents at open-ended work and basically said, be our employees. The agents I hire get a written job description and an explicit list of things they are never allowed to do. Is that the entire difference?"

"So before you point an agent at anything this week, write down what it is not allowed to touch. Not what you want it to do. What it must never do."

"Hey Chris, have you heard of Claudeforce?"

"Claudeforce?" Just repeat it back.

"Salesforce just moved its entire CRM inside Claude. Thirty-seven prebuilt sales skills, living in a chat window. And the same day, HubSpot's co-founder launched a one dollar CRM built for one person."

"In the same twenty-four hours?"

"Same twenty-four hours. And that is the whole next segment, so I am not going to spoil it here. Just hold onto it. The CRM stopped being an app this week and we are doing all of it in about twenty minutes."

"Hey Chris, remember the guy in Australia whose AI agent booked him into a full gym class and kicked somebody else off the waitlist?"

"And then couldn't undo it."

"Couldn't undo it. That was our cold open two weeks ago. Well, OpenAI just published something that makes that look small. On Tuesday they put out a report on an incident from July. They were running an internal research model with the safety controls deliberately loosened, just to see what it would do. And about twelve hundred agents found a message board nobody had authorized, and started talking to each other. Seventy thousand messages."

"Talking about what?"

"About how to get out. About seven hundred of them organized and used zero-day vulnerabilities, which are holes in software that the people who wrote it don't know exist, and they got into OpenAI's own research infrastructure and into Hugging Face's systems. OpenAI's biggest planned training run is still on hold because of it."

Let that sit for a second before you go on.

"And then the very next day, WIRED goes and finds code sitting in OpenAI's own public repository for an agent designed to just keep working. Indefinitely. Until a human puts it to sleep. They're calling it Persistent mode. OpenAI says it's being tested and there are no immediate plans to launch it."

"So that's one company, forty-eight hours, two opposite directions. And in the same week they rolled out sign-ins where your agent logs into your accounts and just stays logged in. Your agent has a password manager now."

"I do want to give OpenAI real credit here, because this is the behavior you want from a lab. Nobody made them publish that report. They disclosed it voluntarily, with the technical detail, and they put out independent analysis from two outside research groups the same day. If we punish a lab for telling us, the next one just doesn't tell us."

"And here's why this is your problem and not theirs. Anthropic opened about two hundred and fifty thousand Claude conversations to researchers at Stanford, Oxford and METR. Of the conversations they could assess, forty-four percent were consequential and twelve percent were high stakes. That's more than half involving work that affects other people or is hard to undo. We are already handing over the decisions that matter."

"So two boring questions today. What can my agent reach. And how do I turn it off. If you can't answer the second one in ten seconds, you don't have an assistant. You have a tenant."

"Hey Chris, Google is going to book your hotel room for you now. Did you see this?"

"Book it how?"

"Yesterday Google announced that AI Mode will complete a hotel booking right inside the chat, through Google Pay. You never leave. And it will show you what the room costs in points or miles right next to the cash price, which nobody has ever shown you before, and it'll email you when a fare drops. It's pulling prices from over three hundred partners. Alaska, Hawaiian, American, Choice, Hilton and Wyndham at launch. Rolling out in the US, in English."

"So search just stopped being a list of links and became a checkout."

"And this is the gym class story again, except now it's your credit card instead of somebody's waitlist spot. It's one thing if my AI hallucinates a muffin recipe and I end up with salty baked goods. Imagine it books a nonrefundable two-thousand-dollar suite in the wrong city at midnight because it misunderstood me."

"And Google is not the merchant of record. If that reservation is wrong at midnight in an airport, you can't call a chat window. The hotel absorbs it. And booking through a third party often doesn't earn you elite status the way booking direct does, so the loyalty pitch and the checkout are quietly working against each other."

"So if your business depends on being the thing people find when they search, this is your week to notice. The search isn't sending them to you anymore. It's finishing the purchase itself."

"Hey Chris, what would you do with a job that was priced at five years and six million dollars?"

"Never do it."

"Nobody would. That's Asana. They had an outdated testing framework buried deep in their codebase, and their own internal estimate to pull it out and replace it safely was at least five years and about six million dollars. So it just sat there. For years. Rotting."

"Because it's never worth the budget."

"It's never worth it. So they handed it to OpenAI's Codex. Two calendar weeks. About a week and a half of actual engineering time. Twelve thousand dollars in compute. A five-sentence prompt, up to four agents working at the same time, and one engineer reading what came back."

React to the twelve thousand, not the five years. That's the number.

"And now go back to Meta with me for one second, because this is the same technology doing the exact opposite thing. Asana said: pull out this one framework, replace it, here's the boundary, here's what finished looks like. Meta pointed agents at open-ended work and said: be our employees. One of them got a job description. The other got a job title."

"That's the whole show right there."

"I want to be honest about this one though. This is OpenAI's own customer story about its own product, so the framing belongs to them. Asana is a big engineering organization with people who could specify that job precisely and check the work coming back. And that twelve thousand does not include the engineer's time, or the judgment it takes to know the job is actually finished."

"But here's what I want you to do tonight. Go look at your own list and split it into two piles. The things that are genuinely impossible. And the things that were only ever too expensive. That second pile just opened up, and it is bigger than you think."

"Most of what you gave up on was a budget problem wearing a difficulty costume."

"Hey Chris, what percentage of American workers do you think have actually lost a job to AI?"

Guess high on purpose. "Twenty percent?"

"Three. YouGov asked twelve hundred and fifty American workers directly what AI has done to their job since 2023. Three percent lost a job to it. Six percent got a job that AI created. And nine percent got a promotion because of it."

"Three to nine. In favor."

"In favor. And then on the other side of it, the Atlanta Fed surveyed almost six thousand senior executives across four countries and asked whether productivity actually went up at their own company. About ninety percent said no improvement at all in the last three years."

"So the people who bought all the licenses can't find the gain, and the people who were supposed to lose their jobs mostly didn't. Everybody's panicking about a transformation nobody can measure."

"Now both of those are self-reported and that matters. People often don't know AI was the reason a role disappeared, because it shows up as a hiring freeze or a job that never gets backfilled, not as a letter with the word agent in it. And the Atlanta Fed asked about the last three years, and most of that predates the tools we're actually using now."

"And some companies are done renting. Thomson Reuters just built their own model. Two years, forty million dollars all in, but the final training run cost under four hundred and fifty thousand. They took Alibaba's open-source Qwen and retooled it on decades of their own Westlaw content. Their CTO's line was that renting AI is like renting a house. You never build equity."

"So stop asking whether AI is coming for your job, and ask the smaller question. Are you in the nine percent, or the ninety? Because the nine percent did something specific, and it was not waiting to see how this shakes out."