Good day . Anthropic polled its own staff and found they expect AI to do 75% of their work by 2030, nearly triple what US workers expect. Plus ChatGPT gets a look at your open Chrome tabs, Linus Torvalds calls the kernel’s AI disclosure rule “utterly stupid,” one developer’s fleet of single-job bots, and Pew on who is actually using chatbots. Let’s get into it.
Anthropic Staff Bet AI Takes 75% of Work Anthropic
What happened: Anthropic published two reports on Friday. In an internal poll of 1,334 staff on August 9 to 12, employees expected AI to be able to do 75% of their current work by 2030. US workers, surveyed by Morning Consult, said 28%. More than half of Anthropic staff said 80% or more, against 5% of US workers.
Why it matters: The two groups agree on what AI can do today and split on how fast it improves: 91% of staff expect AI to make a Nobel-level discovery by 2030, against 39% of US adults. A companion report says AI agent hours at Anthropic outnumbered employees’ 6 to 1 by August, and that the company has hired for a smaller share of junior technical roles. Last Sunday we covered Ethan Mollick on the vanishing intern. This is one lab’s own data on it.
What everyone’s saying: The framing making the rounds is the gap itself: the people building the thing expect it sooner than the people who will live with it. The report adds that both groups reason alike. At the same level of expected automation of their own job, staff and US workers forecast about the same 2030 unemployment rate.
My read between the lines: This is an unweighted internal poll of Anthropic’s own staff, which the report says plainly, and it’s a number that never hurts the case for the product. The better line is in the companion report, where one engineer describes his job as “an army of Claudes, with every one asking for my attention.” The prediction is 75%. The lived experience is a queue.
📖 Further reading: The Tools That Just Replaced 40% of Block's Workforce Are Free in Your Browser — my March piece on Block cutting 40% of its staff, the early, rougher preview of the argument Anthropic’s staff are now making about 2030.
One Anthropic engineer says his job is now “an army of Claudes, with every one asking for my attention.” Your version is probably a Monday report. Viktor is an AI agent that lives in your team’s Slack and connects to 3,000+ tools, then does the work: reports, dashboards, code and campaigns. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor →
ChatGPT Can Now Read Your Chrome Tabs The Verge
What happened: OpenAI updated its ChatGPT extension for Chrome so the chatbot can use the tabs you already have open. Tag Chrome in ChatGPT’s web or desktop app and it can look across them. The demo has someone asking it to compare three tabs of sofas. The sidebar now carries your full ChatGPT history, and ChatGPT Work can “control” tabs for tasks like “sending messages, filling out forms, or downloading files.”
Why it matters: An AI agent used to mean a separate app, or a separate browser you had to switch to. This one lives in the browser you already use and can see the pages you’re already logged into. OpenAI’s own Atlas browser is gone, so this is how OpenAI gets into the browser instead.
What everyone’s saying: Coverage frames it as OpenAI borrowing the browser most people already use rather than asking them to switch to its own. It also works much like Google’s Gemini-powered “auto browse,” so Google now has a rival inside its own browser.
My read between the lines: Your open tabs are a diary: banking, medical searches, the sofa you’re embarrassed about. Making it opt-in per chat by tagging Chrome is the right instinct. The open question is what happens when one of those tabs is a page written to give the agent orders. The sofa comparison is the demo. The form-filling is the product.
📖 Further reading: An AI That Can Use Your Computer Better Than You Can. I'm Not Sure How to Feel About That. — my April piece on a benchmark where an AI navigated a real desktop, which I read as a direct challenge to my own Monday-morning workflow. This is that capability arriving as a browser extension.
The Brief is free and always will be. If today’s 75% number left you wondering what to do about your own job, that’s the kind of question the paywalled deep-dives are for, with the full archive behind them. Become a member →
Torvalds: AI Disclosure Rules Are “Utterly Stupid” The Next Web
What happened: Linux kernel maintainers are dropping the “Assisted-by” tag, the rule since January asking contributors who used AI coding tools to say so in the commit message. Kernel developer Sasha Levin wrote that the Maintainers Summit reached consensus to deprecate it. On October 13, Linus Torvalds asked a critic whether he could not see how “UTTERLY STUPID” such rules would be, and said he expects people to understand what they submit, whatever tools they used.
Why it matters: The kernel is among the most scrutinized codebases there is, so its rules get copied. The Register notes AI agents only recently started producing code good enough for it. Torvalds has gone from skeptic to fan: “Now I’m at the point where I actually really like using AI,” he said at Open Source Summit Europe. He also said it is “stressing maintainers out to the point that it’s a problem.”
What everyone’s saying: The maintainers split. Jeff Layton of Red Hat argued the tag “basically serves as free advertising for proprietary LLM companies,” and that not everyone follows it anyway. Laurent Pinchart objected that a rule sets expectations even if some people cheat, and wants another disclosure mechanism before this one goes.
My read between the lines: Torvalds’ test is understanding, not provenance, which is hard to argue with and impossible to measure. The Register also points out that Claude adds provenance messages by default and Codex does not, so the disclosure the kernel just dropped was already uneven by vendor. The tag never measured the cost he named himself: “bots talking to bots, and it does actually happen.”
One Developer, a Fleet of Single-Job Bots Geoffrey Huntley
What happened: Developer Geoffrey Huntley wrote up the fleet of small AI bots he now runs on Grok Bot, xAI’s persistent-agent product. One sends a 7am brief from his calendar, email and site traffic. One reports blog signups and churn. One finds newsletters worth unsubscribing from and lists them for him to click himself. His rule: one bot per function, like Unix commands you can pipe together.
Why it matters: The interesting part is why he picked a managed service. He tried OpenClaw and dropped it because the “lethal trifecta” (an agent with private data, exposure to untrusted content, and a way to send things out) is still unsolved. His answer: “I want them losing sleep over the lethal trifecta, not me.” Last Sunday we covered Grok Bot’s new inbox. This is what people do with it.
What everyone’s saying: The security discussion keeps coming back to the trifecta. TechRadar has a rundown of OpenClaw’s security risks, and Huntley is pointing at the same problem from the user’s side.
My read between the lines: Managed moves the risk to a vendor’s name. It doesn’t remove it. A bot with your email and your finances is still the trifecta, now with a security team on the invoice. Worth noting that Huntley discloses the credits behind his switch came from a nerd-snipe on X, and says he wrote the post of his own free will.
📖 Further reading: What is Grok Bot? The answer is in the fine print — my August explainer, two days after launch, on one cloud computer holding every login you hand to every bot you create.
Women Caught Up to Men on Chatbot Use Pew Research Center
What happened: Pew published a short read from its June study of 5,119 US adults, surveyed February 17 to 23. About 47% of women and 50% of men say they ever use AI chatbots. In 2024 men were 11 points ahead. Men still use them daily more often, 27% to 20%.
Why it matters: Same adoption, different habits. Men were more likely to say chatbots help their productivity, 35% to 25%, and to use Copilot, 22% to 13%, while ChatGPT use was identical. Women were slightly more likely to use chatbots for emotional support or advice, 11% to 8%. Yesterday we covered the Harvard survey on who bluffs about AI. Two surveys, one theme: same tool, different relationship with it.
What everyone’s saying: The framing, as eMarketer puts it, is parity in use and a split in daily habit, which matches Pew’s own headline numbers.
My read between the lines: “Parity” here means having ever used one, a low bar, and the 2024 question was worded differently and asked only of people who had heard of chatbots, so part of that jump is the ruler. This is also an October repackaging of data collected in February. The number worth sitting with is productivity: 35% to 25% on the same tools.
That’s your AI Brief for Saturday.
—Nicholas from Artificially Intimidating
