Skip to content
Sign in
Artwork for Context Window: AI Daily News Brief
NewsTechnology

Context Window: AI Daily News Brief

Nicholas Rhodes | ArtificiallyIntimidating.com

The 4-minute daily AI news brief that makes artificial intelligence make sense. Every morning, five stories in plain English — no hype, no doom-scrolling, just the signal.

artificiallyintimidating.com
Play
Open in
  • 113 episodes
  • daily
  • Avg 5 min
  • English

Episodes 113

NewestOldestLongest
All20Unplayed20In progress0Played0
Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.
  • S2026 · E1109
    Today · 5 min

    Nvidia Probes Its Own CEO and Goldman Locks Up Claude Code -- AI Brief November 9

    Good day %%first_name%%. Nvidia’s board has opened an independent investigation of its own CEO after reports that he was the target of a Chinese intelligence operation, and Jensen Huang has admitted to an affair in an email to staff. Goldman Sachs has taken Claude Code away from most of its traders and bankers. Also today: a Philadelphia banner with a Chicago skyscraper in it, Dr. Dre and Ice Cube disagreeing about AI, and a developer who walked away from Claude Code. One more thing before we start: the Museum of TEENY Art, a vending machine that sells collectible mini prints, opened November 3 at Here BK in Brooklyn, and its first show is on view through December 1. Go say hi. Nvidia’s Board Opens a Probe of Its CEO NPR * What happened: Nvidia’s board said Sunday that a committee of independent directors, assisted by outside counsel, will investigate reports that CEO Jensen Huang was the target of a foreign intelligence effort, according to Nvidia’s statement. The New York Times alleges that a Chinese businesswoman named Alice Huang, who is not related to him, built a close relationship with him. Axios reports that Huang admitted in a message to staff that he was unfaithful to his wife and said he will cooperate fully with the investigation. * Why it matters: Nvidia dominates the market for AI chips, so who buys them is a national-security question. The Times reports that Alice Huang’s employer became one of Nvidia’s largest buyers, at about $4.8 billion in equipment, according to Al Jazeera, and NPR says the reporting suggests she may have helped China get around US limits on exports of advanced AI chips. * What everyone’s saying: The board says it has “no reason to believe that either national security or NVIDIA secrets were compromised,” and Nvidia says Huang will keep running the company during the probe. Three House Republicans asked him to appear before Congress on November 18, Al Jazeera reports, and Senator Elizabeth Warren said, “NVIDIA’s Jensen Huang and NVIDIA’s board owe Congress answers.” * My read between the lines: The affair is the headline. The $4.8 billion is the story. The open question is not who Huang had dinner with, it is whether anything moved through that relationship to a buyer of that size, and nobody outside the board’s inquiry has shown that it did. Everything here beyond Huang’s own email is the Times’s reporting and the board’s reply, so treat the national-security part as alleged until the investigation says otherwise. 📖 Further reading: The US Government Just Took Anthropic's Best AI Model Offline — Here's Why — my June read on the first time Washington put a frontier AI model behind a national-security gate, and the part nobody was saying out loud. When something goes sideways at Nvidia, the board brings in outside counsel. When something goes sideways at your company, it’s usually a contract you signed at 4:55 on a Friday. Marblism sells AI employees that work inside the tools you already use, and Linda is the legal assistant who reviews contracts. No prompting skills needed: describe your business and goals, and the employees check in when they need your approval. New users get 10% off for life through this link, and there’s a 7-day money-back guarantee. Meet Linda → Goldman Puts Claude Code Behind an Engineers-Only Rope Bloomberg, via FA Magazine * What happened: Goldman Sachs updated its internal guidance in October to limit Anthropic’s Claude Code and Microsoft’s GitHub to engineers, Bloomberg News reports. Traders and bankers had been using the tools to monitor client trading flows, model market moves and positions, and produce written and graphic material. “Strats,” the staff who sit between trading, quant research and software, still have access. * Why it matters: As recently as July, Goldman’s chief information officer, Marco Argenti, was praising “vibe coding” by non-technical staff, who passed their creations to the bank’s roughly 12,000 engineers to improve. State of AI points out that Claude Code reads files, writes programs and runs them with the access of whoever is using it, which is a different risk on a trading desk than inside an engineering review process. * What everyone’s saying: A Goldman spokesperson said the bank puts these tools in the hands of “those who are best suited” to use them for the long term after a period of experimentation. State of AI’s take is that Goldman was right to restrict the tools and wrong in how it did it: a guideline update with no replacement tells the traders who found good uses that they were the risk. * My read between the lines: Bloomberg also notes that Goldman is lead underwriter on Anthropic’s IPO, so the bankers who will spend weeks selling the company are mostly locked out of its coding agent. The tidy explanation is risk management. The more interesting one is that the democratize-coding era just ran into its first compliance department, and compliance has a longer memory than a product demo. The headlines are free and always will be. Members get the paywalled deep-dives behind stories like these, plus the full archive, including my conversation with a historian on why every new tool was supposed to kill art. Become a member → Philly’s Mural Banners Have a Chicago Problem The Philadelphia Inquirer * What happened: In October, Mural Arts Philadelphia unveiled designs for 20 vinyl banners for a fence at South Philly’s Bellwether District. Social media users noticed that three of one artist’s four designs include a skyscraper that resembles Chicago’s Willis Tower rather than a Philadelphia landmark, the Inquirer reports, and said the lettering and portraits looked AI-generated. Whether AI was used is alleged, not established: Mural Arts’ executive director declined to say, and the artist did not respond to the Inquirer. * Why it matters: Public art is supposed to look like the place it’s in, and residents noticed it didn’t. Tuhin Chakrabarty, an AI scholar at Stony Brook University, told the Inquirer the designs were “very obvious” AI and said the Pangram detection tool flagged all four of that artist’s illustrations. The other artists’ work showed no signs of it. * What everyone’s saying: Mural Arts has no AI policy yet. Executive director Emily Mello said, “Mural Arts will always center humans — human artists, communities,” and a spokesperson said the skylines will be changed. Developer HRP Group says the banners depict real Philadelphians, including a local high schooler. * My read between the lines: Swapping a skyline is the easy part. The harder question is the approval process: nobody has said who looked at these designs before they went up. A second town hall on Thursday, November 12 is open to the public (RSVP required), and it is a test of whether an arts group can write its rules before the next headline instead of after it. 📖 Further reading: AI Is a Trust Problem, Not a Tech Problem — my notes from the US Chamber of Commerce AI + Work Summit, where one word kept surfacing: trust. Dre Says AI Is Creative, Cube Isn’t Sold VICE * What happened: On a podcast with Ice Cube, Dr. Dre defended AI in music, saying he’s “extremely excited” about it and “I love the idea of where it could possibly go with music, with film,” VICE reports. He compared the pushback to the resistance sampling once got, and floated a scenario where someone pays Cube to appear in a movie without being on set, with Cube approving how his image and voice are used: “It’s your image and your voice. What’s wrong with that?” * Why it matters: Two people who helped define hip-hop’s relationship with technology draw the line in different places. The licensing scenario is the practical part: it is the same question Hollywood is working through with actor and voice deals, put in plain terms. * What everyone’s saying: Dre’s line is that “the only people that are gonna be mad about the elevation of AI … This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1108
    Yesterday · 5 min

    OpenAI’s Bot Found 948 Ways to Root Linux -- AI Brief November 8

    Good day %%first_name%%. OpenAI says an unreleased model it built a pipeline around turned a $1.2 million run into working root exploits for 948 Linux kernel bugs, and a Sunday brief that opens there also has an unreleased Anthropic model credited with a quantum proof, an outside auditor saying AI is speeding up Anthropic’s own research, OpenAI’s $11 billion at a $1.25 trillion price tag, and the Democratic bills coming for AI in January. OpenAI’s Bot Found 948 Ways to Root Linux SecurityWeek * What happened: OpenAI said on Wednesday that a pipeline built around an unreleased cyber model flagged 2,308 candidate bugs in the Linux kernel, reproduced 1,724 of them, and built working exploits for 948 flaws that would let an ordinary user on a machine take full control, SecurityWeek reports. The Hacker News says that took 30 hours in a mid-October run and cost $1,213,430 in API credits, about $1,279 per exploit. * Why it matters: Linux runs a huge share of the world’s servers, and an exploit like this is what an attacker who already has a foothold uses to become the administrator. OpenAI’s own write-up says an unpatched system with a low-privileged login “should be considered functionally equivalent to root access for a well-resourced threat actor.” It also notes that at what Pwn2Own and TyphoonPWN have paid for kernel exploits, the 948 would be worth $18.9 million to $66.3 million. * What everyone’s saying: The line from SecurityWeek and The Hacker News is that AI speeds up both sides: discovery and exploitation, but also fixes. OpenAI says that within 10 days of the scan, over 100 of the flaws had been patched or had staged patches from other contributors, with about 300 more under public discussion, and that it has already sent hundreds of reports and patches to maintainers. * My read between the lines: The price tag is the story: $1,279 a bug turns exploit-writing from a craft into a line item. Two cautions. SecurityWeek says the agents produced 1,169 working exploits and that 948 counts the separate patches behind them, so the headline number depends on how you count. And on Google’s hardened kernelCTF target, the same pipeline managed 21 exploits in 112 hours for about $33,000. This is OpenAI’s account of its own unreleased model, and it is keeping the technical details private for safety. 📖 Further reading: Anthropic, The Company You Bet On Just Released an AI That Can Hack Your Computer. Here’s the Real Story. — my April read on Project Glasswing, Anthropic’s autonomous vulnerability hunter, and the four-step audit for what a hacking-grade AI means for your own workflows. A thousand holes in the Linux kernel are a machine’s problem. The small-business version of an unpatched hole is the phone ringing while you’re up a ladder or under a sink. Marblism sells AI employees that work inside the tools you already use, and Rachel is the one who answers phone calls and books jobs. No prompting skills needed: describe your business and goals, and the employees check in when they need your approval. New users get 10% off for life through this link, and there’s a 7-day money-back guarantee. Meet Rachel → An Unreleased Anthropic Model Cracks a Quantum Puzzle Business Insider * What happened: Physical Superintelligence, a physics research lab known as PSI, said Friday that a research system running a non-public Anthropic model proved the Aaronson-Ambainis conjecture, a quantum-computing question open since 2008, Business Insider reports. PSI says the system had a complete proof on the third day of the run, and that Lean, a proof assistant that checks every step by computer, accepted it. * Why it matters: The conjecture concerns when a quantum computer can be exponentially faster than an ordinary one. PSI says the result shows that exponential advantages like Shor’s factoring algorithm need special structure in their inputs, which Business Insider reads as quantum computers having limits that were long suspected, while they’d still be a major upgrade for some tasks. * What everyone’s saying: Researchers who had worked on the problem are impressed. MIT and Stanford researchers who settled a special case of the problem in August (Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan) said in PSI’s announcement: “This is a huge result,” adding that “everything is brought together in a rather magical way,” and they’re writing a short exposition of the proof. * My read between the lines: Hold the confetti until the experts weigh in. MLQ notes that the paper is on GitHub rather than arXiv, hasn’t been peer reviewed, and that Scott Aaronson hadn’t posted an assessment as of November 7. Neither PSI nor Anthropic has named the model, and Anthropic didn’t respond to Business Insider. The computer checked the steps. A human still has to confirm it proved the right theorem. Nobody pays for the headlines here, and nobody ever will. The Brief is free and always will be. If you want the version where I run the experiment on my own business and show the work, members get the paywalled deep-dives behind the headlines, plus the full archive. Become a member → METR Says AI Is Speeding Up Anthropic METR * What happened: METR, an independent AI evaluation group, published an assessment on Thursday after 11 weeks inside Anthropic, from August 12 to October 29. Its conclusion: as of October 2026, Anthropic is likely improving its models on well-specified tasks at about twice the rate it did from March 2024 to October 2025, with an honest range of 1.3x to 3x, and most of that speed-up likely comes from Anthropic’s own use of AI in AI research. * Why it matters: METR says Anthropic’s internal benchmark scores are growing at about 1.9x the pre-October 2025 rate, roughly two years of historical progress in one year, and that the speed-up did not come from scaling up training compute. It saw no clear bottleneck in the next 12 months, though it says it isn’t confident in that. On October 17 we covered Anthropic’s own staff expecting AI to do 75% of their work by 2030. This is an outside check on how fast the inside is moving. * What everyone’s saying: The caveat is the headline. Anthropic redacted a majority of the evidence METR found most informative, describing the cuts as “aimed at removing confidential information that could accelerate other AI developers,” and METR says the public must “rely on our judgment more than we consider ideal.” Economist Karthik Tadepalli, writing in a response post, concludes the acceleration is probably real. * My read between the lines: Read the redaction summary before the 2x. METR says Anthropic had the right to redact and METR had the right to say when it did, and that METR did not find consensus among Anthropic’s staff on how much was AI. The auditor is telling you to trust the auditor, and the evidence is in a drawer we can’t open. That doesn’t make it wrong. It makes 2x an estimate with its own error bars, and METR says as much. 📖 Further reading: The AI Pattern That Optimizes Anything Measurable — Overnight — the small-scale version of this loop: an agent ran 700 experiments in two days on Karpathy’s own code and cut training time 11%, which is what METR is now measuring at a frontier lab. OpenAI Closes $11B at a $1.25T Valuation The Verge * What happened: OpenAI is now valued at $1.25 trillion after a new round that has raised another $11 billion, The Verge reports, citing Bloomberg. Bloomberg says the money was wired Friday as the first part of a round that could reach $20 billion in the coming weeks. * Why it matters: OpenAI had hoped to raise $30 billion or more at $1.4 trillion before tech stocks dropped and the midterm results showed how far public opinion has turned against AI companies and data center projects, per The Verge. Its March round was $122 billion at $852 billion. On November 3 we covered Anthropic’s amended S-1, and this is the other half of that… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1107
    Saturday · 4 min

    A Talking Typewriter Fits in 32MB -- AI Brief November 7

    Good day %%first_name%%. A talking typewriter that fits in 32 megabytes leads a Saturday brief that also has Amazon opening a text line for shopping right after blocking Meta’s Muse, Meta pausing the business terms that alarmed publishers, a pile of benchmarks saying most AI design skills do nothing, and a Brookings economist’s name for the job market AI might build. A Talking Typewriter Fits in 32MB TechCrunch * What happened: Startup Underdog released a browser game to show off Woof, a voice model it says is only 32MB, TechCrunch reports. The game is a retro room with a typewriter, a CRT monitor and a desk lamp: type on the typewriter and it speaks your words back, turn the tuner and it generates music. The company says the 31-million-parameter model can run on most devices with no internet connection, and it’s open for anyone to download. * Why it matters: Most AI news is about bigger models and bigger data centers. This goes the other way: a voice small enough to live on the device you already own. If the company’s claim holds, a voice like that never has to send your words anywhere to be spoken. * What everyone’s saying: TechCrunch is measured: such a small model “likely won’t replace bigger models,” but it may be enough to drive local assistant experiences or personalized podcasts. Founder Sigil Wen told TechCrunch he started building after deciding he’d rather make his own notetaker from on-device models than use one like Granola, so users own their data. His forecast: “I think we will be able to run most capable models of today on the devices you already own within months.” * My read between the lines: The game is the pitch. Its intro casts “intrusive” cloud providers as the villain, so the demo doubles as a manifesto with a typewriter in it. It’s also the friendly front door: Underdog says it recently shipped Bark, a 167MB speech-to-text model, plus a 27-billion-parameter model for laptops and a waitlisted app pitched as a local alternative to Muse and Instinct. We haven’t played the game, so treat the offline claim as the company’s until someone does. 📖 Further reading: VoiceStudio: The Free ElevenLabs Alternative Whose Default Voice Model Is Non-Commercial — a free voice app that runs on your own computer, and the licence fine print that decides whether you can sell what it makes. A typewriter makes one promise: your words, on the page, nothing in between. For a small business, the words that never get typed are the blog posts and the search work that slide to next quarter. Marblism sells AI employees that work inside the tools you already use, and Penny is the one who handles SEO and blog posts. No prompting skills needed: describe your business and goals, and the employees check in when they need your approval. New users get 10% off for life through this link, and there’s a 7-day money-back guarantee. Meet Penny → Amazon Blocked Muse, Then Built Its Own TechCrunch * What happened: On Friday Amazon said customers can text its Alexa AI agent to shop, TechCrunch reports. Select US customers get a text asking them to switch it on; from there you can search, add to cart and buy in plain language, track prices, compare products, check an order, or send a photo of an item or outfit to find similar listings. A broader rollout is coming in the next few weeks, with a waitlist for everyone else. * Why it matters: It lands just before holiday shopping, and after Amazon blocked Meta’s Muse from querying its site in September. Amazon isn’t rejecting shopping agents. It’s choosing which one gets to stand between you and the store. * What everyone’s saying: TechCrunch calls texting “the currently most popular bandwagon” for talking to AI agents. The reality check comes from DoorDash, which said last month that third-party agents had placed only hundreds of orders in total on its platform, and that it will concentrate on its own text-based agent. * My read between the lines: The fight over shopping agents is really about whose name is on the receipt. An outside agent gets the rope; the house agent gets the customer, the cart and the preferences Alexa says it will learn over time. And “hundreds of orders” is the number to keep in your pocket the next time someone calls autonomous shopping inevitable. 📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — the Amazon-blocks-Muse setup is in the opening, followed by a test of what agents can actually read on a real business site. Today’s Brief is the headline version, and it’s free and always will be. The paywalled deep-dives are where we run the test on our own business and show the work, and members get the full archive too. Become a member → Meta Pauses Its Scrape-Friendly Business Terms Mashable * What happened: Meta has put proposed changes to its business terms on hold indefinitely, Mashable reports. In a Business Help Center update, Meta said the changes to its Commercial Terms, Product Catalog Terms, Self-Serve Ad Terms and Ad Creative Generative AI Terms “are now paused and will not take effect on November 20, 2026.” As written, the changes would have let Meta use web scrapers on websites tied to Meta business accounts, including journalism sites. * Why it matters: Critics said that could let Meta use publishers’ content to build AI products that compete with them, without payment, per ABC News and the AP. The date had already slipped from October 30. According to Quartz, the UK’s competition regulator opened an investigation on October 30, and a Norwegian court ordered a temporary halt on October 28. * What everyone’s saying: Meta says it is “continuing to consult with stakeholders” and “will provide a further update when we have more to share,” and that its existing terms remain in effect. The Movement for an Open Web, the group that first complained to the UK regulator, called the earlier delay and the investigation a “significant win,” but warned it was only the start. * My read between the lines: Read the verb. “Paused” isn’t “withdrawn,” and Meta hasn’t said what it would change, only that it’s still consulting. That’s a company waiting out a regulator and a court, not abandoning the idea. The consent question publishers raised is unanswered, and Meta is the one holding the next move. 📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn’t Agree To. — the July argument that when Meta retreats from a loud launch, the quiet version is the one to watch, and that the real question is who holds the leash. Most AI Design Skills Do Nothing AI Realist * What happened: Designer Vadym Grin argues in a guest post on Maria Sukhareva’s AI Realist newsletter that stacking design skills on Claude or Codex leaves pages competent and forgettable, and that only three focused skills earn a place. (A skill is a folder of instructions an AI coding agent loads before a task.) He backs the argument with three 2026 benchmarks of agent skills. * Why it matters: Grin cites the SWE-Skills-Bench paper: 49 public skills across about 565 real coding tasks, where 39 skills changed nothing, three made results worse, and the average pass rate rose 1.2%. He also cites SkillsBench, where human-written skills raised pass rates by 16.2 points on average while skills a model wrote for itself averaged a 1.3-point loss. * What everyone’s saying: The common fix for a bland page is another skill. Grin’s counter is that the model builds the likeliest page, and “in design, the likeliest option is the average of the websites it learned from.” He says even Anthropic’s own design skill now names the defaults AI-made pages cluster around, starting with a warm cream background, a serif headline and a terracotta accent. * My read between the lines: Escape one cliché and the crowd follows you into the next one, which makes the skill count the wrong thing to count. Grin’s point is that nobody is tasting the soup, and he says all three benchm… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1105
    Thursday · 4 min

    Google Ties Student Search to AI Access -- AI Brief November 5

    Good day %%first_name%%. Google is tying student search to an AI opt-in, Anthropic signed its first news licensing deal (with the AP), and a school turned a first-grader’s drawing into an AI portrait. Plus Muse’s bathroom test and Pangram’s new AI video detector. Google Ties Student Search to AI Access ABC7 (AP) * What happened: Google said Tuesday that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be off for K-12 student accounts until school administrators “explicitly turn them on,” per Google’s blog post. The AP reports the practical effect: students using Google software in class lose Google search unless their school also opts in to AI tools. * Why it matters: Google says 150 million students and teachers use Google Classroom worldwide. The AP notes Google switched Gemini on for many K-12 students in August, after allowing it only for students 18 or older until then, and got an outcry. Google’s post says it will share specifics “in the coming months ahead of the changes rolling out,” so there is no switch-over date yet. * What everyone’s saying: Former FTC chair Lina Khan wrote: “Google is using its search monopoly to coerce schools into adopting AI.” Fairplay’s Josh Golin said the attempt to “strong-arm schools into using its AI products is likely to backfire,” and urged schools to turn off Google entirely and find “more privacy-protective alternatives for search.” * My read between the lines: The default flipped from on to off, which is the concession. The price tag is the product everyone already uses to look things up. Google calls it control; critics call it leverage; the AP’s own reading supports both. The same blog post pilots a Guided Learning feature in Search for select US schools later this month, which hands administrators a friendlier-sounding reason to flip the switch. 📖 Further reading: Google’s Invisible Axe: The Silent Killer of Small Businesses — what it looks like when Google’s gate decides you don’t exist and no human answers the appeal. We lived it with one of our own businesses. Google just made search a switch a school has to flip. If you run a business, you know the feeling: the way people find you sits behind somebody else’s gate, so the place you talk to customers directly matters more, and it eats hours. Marblism sells AI employees that work inside the tools you already use, and Sonny is the social media one: it watches trends, writes posts in your tone and posts for you. No prompting skills needed. New users get 10% off for life, and there’s a 7-day money-back guarantee. Meet Sonny → Anthropic Signs First News Licensing Deal Adweek * What happened: Adweek’s Mark Stenberg reports, citing two people familiar, that Anthropic struck a content and data licensing deal with the Associated Press just before the midterms. On Election Night, Claude users asking for polling data saw “up-the-minute information, provided by and attributed to the Associated Press.” Neither side disclosed terms or length; Anthropic didn’t respond and the AP gave no on-record comment. * Why it matters: Until Tuesday, Stenberg writes, Anthropic had no licensing deal with any digital publisher, while OpenAI has “more than a dozen,” including Axel Springer, News Corp. and The Atlantic. In August, media executives told him Anthropic believes its use of publisher content is legal under fair use. Yesterday we covered the OpenAI and Anthropic election war rooms; this is the data feed side of Anthropic’s election week. * What everyone’s saying: Stenberg’s read: premium publishers block AI crawlers, so Claude could only cite low-credibility sites that “cannot tell a user which politician won a Senate seat 10 minutes ago.” He calls the blocking “effectively a digital embargo,” and writes that “in this limited instance, the strategy worked.” * My read between the lines: The AP is the safest first date: licensing and wire deals are most of its revenue, it was the first publisher to sign with both Google and OpenAI, and it’s nonpartisan. Stenberg expects Anthropic to use the experiment to decide whether to sign more, and notes the company “prepares to go public” in weeks. The lesson publishers will take is that blocking works when the moment is big enough, and every news event now doubles as a negotiation. 📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — the same standoff from the small-business side: the big sites are closing doors on AI agents, and here’s what a bot actually sees at yours. The daily Brief is free and always will be. Members get the long reads where we stay with one story until it makes sense, like the 18 days the US government pulled Anthropic’s Fable 5 offline, plus the full archive. Become a member → Muse Wins on Details Nobody Asked For Nate B Jones * What happened: In a new video, Nate B Jones says 10 million people set up Meta’s Muse agent in six weeks, after Reality Labs lost $19.2 billion in 2025 alone. His point: after a public miss, Zuckerberg swung again. And Muse isn’t winning on raw smarts: “Muse is not the smartest agent on the market. That’s not the point.” * Why it matters: Jones’s test is the hotel bathroom: “You learn whether they care about detail by going to the bathroom. The entrance is supposed to impress you.” He says Muse stays consistent in tabs that “did not have to be the opening shot of a launch video.” He credits Alexander Wang for a team under 200 people and more than a hundred specific model behaviors tested before launch. * What everyone’s saying: Jones named his Muse agent Marvin, after Douglas Adams. When Marvin works, it “would get out a little laptop and type or Marvin would scratch Marvin’s head,” so “I didn’t have to interpret a technical status message.” Jones says that tiny design answers the question everyone asks when they delegate to AI: are you doing it? On Monday we covered Amazon’s shrug at Muse. * My read between the lines: A cartoon typing on a laptop is what a product built for regular people, not AI natives, looks like. Nobody in that audience wants a log line. Jones’s bigger spin is the part worth keeping: AI makes trying cheaper, so our ambitions should be bigger, but “embarrassment is not getting cheaper.” The tools got easier; being seen to fail didn’t. 📖 Further reading: OpenAI Made Its New Agent Adorable. That’s the Part That Worries Me. — the same cute-agent playbook from OpenAI’s side, and why a friendly face on an always-on agent deserves a second look. School Gave a 6-Year-Old an AI Future Self The Washington Post * What happened: The Washington Post’s Caitlin Gibson reports that a South Carolina school used AI to turn a first-grader’s career-day assignment into an image of her as a 25-year-old nurse. Her mother, Nikki Anderson, found it in the folder of assignments in her daughter’s backpack, next to the construction-paper portrait the girl had made herself. (paywalled) * Why it matters: Anderson’s questions, as Bored Panda quotes her: “What software was used? Was it an outside platform? What happens to those pictures after? Did any of us actually consent to that?” She took them to the school’s administration and, at the time of her People interview on October 20, was awaiting a response. * What everyone’s saying: Anderson told People her daughter loved the picture because “she hopes she would be that beautiful when she grows up,” and said: “I think our children deserve protection, not only of their image but of their minds.” Bored Panda notes that some commenters thought she was overreacting, while another wrote that “People shrugging this off as nothing is insane.” * My read between the lines: The assignment was to imagine a future self, and the kid had already done it, in crayon, on paper. The AI version answered a question she’d answered, and it raised the ones her mother listed: which tool, where the pictures go, who agreed. A construction-paper nurse needs no consent form. A photoreal one … This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1104
    Thursday · 4 min

    A Meta Bot Hit an Artist’s Website 550,000 Times -- AI Brief November 4

    Good day %%first_name%%. A Meta bot got stuck pressing “remove from cart” on one artist’s shop, a vibe-coded AO3 clone lasted a single day, and OpenAI and Anthropic ran election war rooms. Plus Mamdani on Sam Altman and how workers across Asia and Africa really use AI. Meta Bot Hits Artist’s Shop 550,000 Times Business Insider * What happened: San Diego children’s author and illustrator Chelsea Paulus told Business Insider that a Meta bot hit her website more than 550,000 times and overwhelmed her hosting. Fast Company reports the bot, Meta-ExternalAgent, got stuck in a loop on her site’s “remove from cart” function, making more than a hundred hits a minute. After she blocked it, a second Meta bot, Meta-WebIndexer, pushed her CPU to 100% and crashed the site. * Why it matters: Paulus told Business Insider she edited her robots.txt file and put up a firewall, but “those can also take 24 hours to take effect, and I needed to stop the bots immediately.” The fix that finally worked was code from her host, pasted into a hidden file on her site. Paulus, who used to work in web design, called it “an incredibly stressful week, even with some of my technical knowledge.” * What everyone’s saying: Meta’s answer: “website managers can decide how our crawlers interact with their sites through tools like robots.txt,” a spokesperson told Business Insider, adding that the company respects “the preferences site owners set.” Fast Company ties the episode to the scrutiny Meta faces over announced changes to its terms that, in its words, “essentially, allow AI crawlers to run wild across business accounts.” Three days ago we covered the one robots.txt line four labs will honor. * My read between the lines: “Just use robots.txt” is true the way “just lock your door” is true: it helps, and it works after the thing has already walked in. The machine built to read the whole internet got trapped by a shopping cart button, and the cost landed on a woman with four websites on one hosting plan. On Monday we covered Amazon’s stance on Meta’s Muse. Small shops don’t have Amazon’s bouncers. 📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — the other side of the bot problem: why the big sites are blocking agents, and what a bot actually sees when it visits a small business’s website. Chelsea Paulus is a working artist who got dragged into server logs and hosting tickets when she should have been drawing. The same thing happens to every small business: the unglamorous stuff piles up while you’re busy doing the work you’re actually good at. Marblism sells AI employees that work inside the tools you already use, and Eva is the executive assistant: your inbox, your calendar and your meeting notes, checking in with you when she needs an approval. No prompting skills needed. New users get 10% off for life, and there’s a 7-day money-back guarantee. Meet Eva → Vibe-Coded AO3 Clone Dies in One Day 404 Media * What happened: ao3.life, a vibe-coded site that scraped fanfiction from Archive of Our Own, reposted it and charged for AI-generated audiobook versions, went down Monday evening. Its creator said it received more than 2,000 takedown requests in the single day it was up, and that “a swarm of AI Agents controlling 40 virtual machines” did the scraping, 404 Media’s Samantha Cole reports. * Why it matters: The clone appeared after AO3 itself went down Thursday following a security incident. The Organization for Transformative Works, which runs AO3, says it could be down for as long as four weeks, with no data lost. AO3 hosts nearly 18.5 million works. Authors who had “archive locked” their stories to keep out bots reported the lock didn’t stop it: “I had everything archive locked on ao3 and it’s still all here,” one wrote on Bluesky. * What everyone’s saying: Readers and authors responded with widespread backlash on social media. The creator, who signed the apology “a 17 year old from India that thought they were doing something good,” said the site drew 126,592 unique users in 24 hours and sold “like 11 audiobooks.” Cole also notes some readers are now trading Google Drive links to works downloaded earlier, which alarms authors too. * My read between the lines: The takedown process asked authors to prove ownership by leaving a note on their AO3 profile. AO3 was down. That is an opt-out system whose proof step ran through the site the scraper had just replaced. The 91% author payout was a promise to hold money “indefinitely” for people who never agreed to be paid by a stranger. It’s a startup pitch that skipped a step: asking. 📖 Further reading: The Font That Beat AI for About a Week — another anti-AI shield that failed in public, and what it says about the doors we leave open. The daily Brief is free and always will be. Members get the paywalled deep-dives behind the headlines, like what happened when seven AI bots ran a business for 72 hours, plus the full archive. Become a member → OpenAI and Anthropic Run First Election War Rooms The Hill * What happened: For the first time, OpenAI and Anthropic ran election “war rooms” for Tuesday’s midterms. The New York Times reports a team of more than a dozen people worked around the clock in a conference room off the cafeteria at an OpenAI office in San Francisco, while Anthropic put analysts in cells in San Francisco, New York and Washington. Semafor first reported Anthropic’s. * Why it matters: OpenAI’s head of civic and strategic risk, Josh Lawson, told the Times that “The most available technology at the time in 2024 was text,” and “we’re not just monitoring text anymore.” The threats now include deepfakes, cyberattacks and influence operations that can be launched “with a series of prompts.” Yesterday’s lead was chatbots giving voters wrong dates; this is the staffed response. * What everyone’s saying: The Hill reported on Tuesday afternoon that ChatGPT had pulled more than 16.5 million times from Democracy Works, the nonprofit supplying voting logistics, per its second story. Semafor notes the war-room concept dates to 2018, when Meta stood up the first one, and that Google’s AI Mode failed last week to say who won the 2020 election. * My read between the lines: A dozen people in a room beside the cafeteria, guarding a product the Times says has about a billion users. The technology scaled; the safeguard is still humans watching a screen. Former Meta public policy director Katie Harbath told Semafor war rooms exist for “the things we can’t necessarily predict going wrong.” That is the honest job description: nobody can name the failure ahead of time. Mamdani to Wired: No, I Don’t Trust Altman WIRED * What happened: In a WIRED interview published Election Day, New York City Mayor Zohran Mamdani said he worries AI is a bubble that will pop, and, asked whether he trusts “the Sam Altmans of the world,” answered “No.” He said there are “far too many similarities between the housing bubble and the state of AI in our economy today.” * Why it matters: Mamdani has supported New York State setting a one-year pause on data center construction. His complaint is opacity: these companies operate at “a level of opacity that renders it almost impossible for us to make an informed conclusion as to how much this query is actually costing us as a society.” * What everyone’s saying: He compared the industry’s pitch to Facebook in 2008, arguing “they should be able to regulate themselves.” On deepfakes, he said family members send him AI videos of himself believing they’re real, including praise for a 45-minute debate with a European member of parliament that never happened. “There is a real crisis of faith as to what is real right now.” * My read between the lines: A politician answering “do you trust Altman” with one syllable, then handing the question back to the interviewer, is notable by itself. What he’s really pushing against is the idea that AI is “so complicated” the only rational answer is to trust the builders. Workers in Asia and Africa… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1103
    Tuesday · 5 min

    AI Chatbots Are Giving Voters Wrong Dates -- AI Brief November 3

    Good day %%first_name%%. It’s Election Day, and a new audit found AI chatbots botching basic voting questions. Plus OpenAI’s test for everyday chores, an AI that rewrites another AI’s weights, Anthropic’s amended S-1, and why Postgres 19 is late. Chatbots Botch Voting Rules on Election Day NPR * What happened: In October, the nonprofit Innovating for the Public Good (IFPG) tested major chatbots on election questions, and NPR reports that 13.9% of their answers to voting-logistics questions were outdated or inaccurate. In 5.9% of answers the error could cost a voter: a missed deadline, the wrong day, or a skipped step. * Why it matters: Today is Election Day, Tuesday, November 3, and these tools are where a lot of people now ask how to vote. NPR says accuracy depended more on the chatbot than the question: Claude and Copilot produced three-quarters of the clear errors, and on voting questions Claude was wrong or outdated about one time in three, Copilot one in four, ChatGPT one in 10. Earlier this week we covered ChatGPT trusting fake local news. This is the same weakness, with a deadline on it. * What everyone’s saying: IFPG’s Chambers called the error rates “extremely troubling” and said the “legions of errors” make you wonder how much more there is to find. Anthropic spokesperson Michael Aciman told NPR that because IFPG used an API, some of Claude’s key safeguards may not have been active. IFPG says its own comparison of API and web answers found them comparable. * My read between the lines: The wrong answer is half the problem; the confidence is the other half. NPR also found every chatbot tested commonly pointed to betting sites like Polymarket and Kalshi when asked who’s winning. Betting odds dressed up as polling is a new way to be wrong. Check your state’s official election site before you trust any of them. 📖 Further reading: AI Is a Trust Problem, Not a Tech Problem — My notes from the Chamber of Commerce AI + Work Summit, where the word under every panel was trust, the gap a confidently wrong answer opens. A chatbot that guesses with total confidence makes a terrible coworker. Viktor is an AI agent that lives in your Slack and connects to 3,000+ tools, then does the work: reports, dashboards, code and campaigns. A coworker, not a chatbot. New readers get $50 off their first month. Hire Viktor → OpenAI’s LifeVal Grades AI on Chores OpenAI * What happened: On Monday OpenAI published LifeVal, a test of how well AI handles everyday tasks like shortlisting kindergartens after a move, organizing flood-prep volunteers, or building a meal plan for a parent newly diagnosed with diabetes. OpenAI says GPT-6 Astra scored 93.1%, 98% and 90.4% on those three, and CNN reports it scored 80% overall versus 49.3% for GPT-5.2 from December 2025. * Why it matters: Benchmarks steer what labs build. OpenAI says it has used LifeVal for the past year to track and drive progress on everyday tasks, because it wants ChatGPT to “take actions that actually get things done,” not just chat. The tasks came from a 3,995-person survey, and about 100 users wrote their own tasks and grading rubrics, per CNN. * What everyone’s saying: CNN frames it as a move in the race for personal AI agents against Meta, Anthropic and Google, noting Meta’s Muse hit 10 million U.S. downloads in its first 40 days, per Sensor Tower. OpenAI researcher Jenny Yang told CNN the goal is to learn “what are the things that we’re not so good at yet?” * My read between the lines: OpenAI built the test, picked the tasks and graded its own model, and the model did great. The 100% scores in OpenAI’s table are labeled “best recorded,” which is not the same as typical. The number that matters is how an average run handles your Tuesday, and that one isn’t in the chart. 📖 Further reading: OpenAI Made Its New Agent Adorable. That’s the Part That Worries Me. — Dots is the always-on agent running on GPT-6 Astra, the model LifeVal grades, and my read on what happens when a chore-doing agent is cute. The daily Brief is free and always will be. Members get the paywalled deep-dives behind the headlines, like how I tested whether AI agents can even find my business, plus the full archive. Become a member → GPT-6 Astra Rewrites Another AI’s Weights MIT CSAIL * What happened: MIT CSAIL researchers report that OpenAI’s GPT-6 Astra can read and edit a neural network’s raw weights directly, with no training loop. Asked to improve Qwen3.5-0.8B, a small open model, Astra edited 0.016% of its weights, and the authors say its MATH-500 score rose from 7.9% to 48.1%. * Why it matters: Weights are the billions of numbers that make up a model, normally adjusted by optimizer software. The authors say the results “might be a big deal for both optimization and interpretability,” and report that Astra could tell what game a network was trained on just from its weights. They call the findings very early. * What everyone’s saying: Hacker News sent it to about 1,959 points, per Hacker Times. One commenter was skeptical that Astra “can only improve weights to its own level,” so it can’t train itself this way. Another posted a single word: “RSI,” short for recursive self-improvement. * My read between the lines: Editing 0.016% of a model sounds surgical until you notice the patient: 0.8 billion parameters, scoring 7.9% on math, stuck in reasoning loops its own model card warns about. The bar was on the floor. Still, an AI fixing a bug in another AI by reading its raw numbers is a new kind of party trick. 📖 Further reading: The AI Pattern That Optimizes Anything Measurable — Overnight — Karpathy’s agent ran 700 experiments on his own training code overnight; this week’s paper takes the same idea down to the weights themselves. Anthropic Says AI Speedup May Be Imminent SEC filing * What happened: On Monday Anthropic filed an amended S-1 with the SEC that says its “internal estimates and our expectations regarding future models” indicate it “may imminently meet” its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.” The filing says no current model is designated as exceeding that threshold. * Why it matters: CellCog’s read is that the bar means a year of progress equal to two at baseline. In the same filing, Anthropic says over 90% of contributions to its codebase in September were written by Claude, and that it does not expect to “safely use Claude to fully autonomously design and build successive generations of frontier models.” Last week we covered Anthropic opening its books. This is the update. * What everyone’s saying: The sentence spread fast on X, with posts from Peter Wildeford and Andrew Curran quoting it, per Techmeme’s roundup. Bloomberg, via Techmeme, reports the update also showed Q3 revenue up 12x from a year earlier to nearly $17 billion, and a $23.3 billion net loss. * My read between the lines: This is a risk-factor line in an IPO filing, so lawyers made it say exactly what it says: future models, not current ones. Still, a company telling investors the AI-building-AI dial may soon hit 2x, while saying it can’t safely let Claude build its successors alone, is a sentence worth reading twice. Postgres 19 Is Late, and AI Gets the Blame Robert Haas * What happened: Postgres committer Robert Haas writes that the release of Postgres 19 is delayed because “a huge number of people on the Internet are now reporting bugs in PostgreSQL that they found using LLMs.” Between feature freeze and Sept. 25, he says, back-patched commits to the stable branch rose from 724 to 1,169 year over year, while commits to the main branch slightly declined. * Why it matters: Haas says LLMs “can draft fixes, but those fixes aren’t always correct or compliant with PostgreSQL standards,” so humans still review them. He says reviewing the older branches ate time that would have gone to checking what went into 19, which is why late reverts piled up. His verdict: the delay “emerges direct… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1102
    November 2 · 5 min

    Musk Says Grok Bot Costs More Than It Earns -- AI Brief November 2

    Good day %%first_name%%. Elon Musk says Grok Bot costs more to serve than it earns, Anthropic is asking job candidates how they would feel if its stock went to zero, and Amazon’s CEO has an opinion about Meta’s shopping agent. Plus free Adobe clones built with Claude, and photographers vibe coding their old DSLRs. Musk Says Grok Bot Costs More Than It Earns Startup Fortune * What happened: A user asked Elon Musk on Sunday whether xAI could reset Grok Bot’s weekly usage pool after paying users burned through it early. Musk replied that “the cost of serving Grok Bot was unsustainably high relative to revenue,” Startup Fortune reports. He gave no dollar figure and no timeline for a fix. * Why it matters: Grok Bot keeps working on its own virtual computer after you close the tab, so one run can burn far more compute than a chat reply. Startup Fortune says it has no standalone price, rides along with SuperGrok plans from $30 a month, and that xAI has never published the weekly allowance. Yesterday we covered OpenAI folding ChatGPT and Codex into one usage bucket. Today’s version of the same squeeze comes with a quote. * What everyone’s saying: Startup Fortune calls it an unusual confession, since labs usually say demand is outrunning supply, not that the thing people use most is the thing costing them most. It cites Teslanorth for roughly 800,000 paid weekly users within about two months of launch. * My read between the lines: The meter is the business model, and Musk just read it out loud. Earlier this year he floated ads inside Grok to pay for “those expensive GPUs,” and an agent that can’t cover its own compute makes that idea look less like a joke. Growth you can’t afford is a fast way to find your limit. 📖 Further reading: What is Grok Bot? The answer is in the fine print — Written two days after launch, it flagged the product’s always-on cloud computer and a meter nobody capped. Grok Bot just showed what an agent that never clocks out can do to a budget. Viktor is an AI agent that lives in your Slack and connects to 3,000+ tools, then ships the work: reports, dashboards, code and campaigns. A coworker, not a chatbot. New readers get $50 off their first month. Hire Viktor → Anthropic Asks Hires About Stock at Zero Fortune * What happened: Fortune reports that Anthropic is asking job candidates in culture interviews how they would feel if the company dropped its AI ambitions for safety reasons and the stock fell to zero, according to Axios. The piece lands as Anthropic targets a November IPO at a $1.5 trillion to $2 trillion valuation. * Why it matters: More than 2,500 employees could get rich from the listing, and Fortune says Dario Amodei has reportedly worried that the money could pull staff away from the mission. Anthropic’s S-1, filed Oct. 26, calls competition for talent “exceptionally intense,” and Fortune lists staff software engineer base salaries of $350,000 to $850,000. Last week we covered Anthropic opening its books. This is the human side of the same filing. * What everyone’s saying: One candidate wrote on the anonymous workplace site Blind: “I was honest and said no, I would not be happy if the stock went to 0.” Fortune also notes that Amodei and his six co-founders pledged to give away 80% of their wealth, and that about 400 current and former SpaceX employees saw stakes pass $100 million after that IPO. * My read between the lines: It is a loyalty test with a price tag attached, and the honest answer is the one the candidate gave. A culture screen works until the shares vest. Asking early is smart, but an interview question is not a lockup. The daily Brief is free and always will be. When a headline leaves you asking how you would actually use the thing, that is what the paywalled deep-dives are for. Members get those plus the full archive. Become a member → Amazon’s CEO Shrugs at Meta’s Muse Fortune * What happened: On Amazon’s earnings call Thursday, CEO Andy Jassy told investors not to worry about outside shopping agents. Without naming Muse, he said few customers buy that way because the experience is “still not great, nor more convenient than coming to a retailer directly,” Fortune reports. He added that outside agents did not cause the weaker-than-expected ad results. * Why it matters: An agent that shops for you skips the pages where ads live, and ads are a key profit driver for Amazon. Amazon blocked Muse from its store in September, saying its access would violate the conditions of use customers agreed to. Last week we covered Muse shopping so hard Stripe raised its limits. * What everyone’s saying: Meta says more than 10 million people in the U.S. and Canada have set up Muse since early September, and Zuckerberg says shopping is its most common use, with purchases averaging over $200. Shopify, Walmart and Best Buy have signed on. Nate B Jones titled a video this week “The Future Of The Internet Isn’t Open For Your Agent”, with Muse in the headline. * My read between the lines: Jassy’s argument is that outside agents are not good yet, which is not the same as not mattering. He also promoted Amazon’s own Alexa for shopping, which he says 400 million customers used in the last year. Blocking the competitor while pushing your own agent is a strategy, not a verdict on the product. 📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — It opens with Amazon blocking Muse and shows how to test whether agents can even find, price and book your business. Free Adobe Clones Built With Claude Engadget * What happened: Engadget’s Steve Dent writes that a developer who hated Adobe’s cancellation fee used Claude to reverse-engineer Photoshop, Lightroom and other Creative Cloud apps, then released free, open-source clones called ArtCraft. Dent says Photocraft, the Photoshop stand-in, handles resizing, masking and color adjustments with menus that closely match the original. * Why it matters: Dent reports users have hit bugs, including images deleted and layers merged at random, and that Photocraft can’t open native PSD files. Adobe’s terms ban reverse engineering, and the team may have run into trademark and patent trouble even though it says it used none of Adobe’s code. Last month we covered seven AI-built Adobe clones. This is the next chapter. * What everyone’s saying: The reaction splits. A Reddit commenter who has built Photoshop file readers calls it “exceptionally well-executed,” while another user who tried it called it “utter AI-driven slop.” Engadget points to REA, a tool that automates the same take-it-apart-and-rebuild trick and had 346,000 GitHub stars as of Oct. 20. * My read between the lines: The legal question is real, but so is the signal. One person with AI can now do what used to take a team, and software people resent just got a credible threat. Adobe’s best defense might be a cancel button that is easy to find. 📖 Further reading: Your SaaS bill is a sitting duck — A free, open-source challenger to paid software with agents built in, and why picking a winner is the wrong game. Photographers Are Vibe Coding Old DSLRs Digital Camera World * What happened: Digital Camera World reports that photographers are vibe coding 20-year-old DSLRs to look and shoot more like film. One example is Latent Studios’ firmware for the Canon EOS 5D and 5D Mark II, which develops photos with film looks inside the camera, no computer needed once installed, for a pay-what-you-want price. * Why it matters: A YouTube explainer from Breno Loves Cameras says people with no programming background are using AI assistants like Claude to rewrite camera firmware and add features to older, cheaper cameras. The firmware page itself warns that custom firmware is installed at your own risk, and the video covers the worst case: a bricked camera and a voided warranty. * What everyone’s saying: Viewers in the comments mostly cheer it as a win for owners of hardware held back by so… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1101
    November 1 · 5 min

    OpenAI Puts ChatGPT and Codex in One Bucket -- AI Brief November 1

    Good day %%first_name%%. OpenAI is about to make ChatGPT and Codex draw from the same usage allowance, DeepSeek published a blueprint for an agent that never goes offline, and four AI labs agreed to honor one robots.txt line that tells them to keep out. Also today: a viral post about an Anthropic talk that doesn’t hold up, and OpenAI raising less money than it wanted. ChatGPT and Codex Will Share One Usage Bucket OpenAI Help Center * What happened: OpenAI says that starting in December, the usage allowance on Plus, Pro and Business plans will be shared across ChatGPT and Codex, its coding agent. Usage will be measured in tokens, and “instant chat will not draw from your usage.” OpenAI hasn’t given a start date or allowance sizes by plan, TokenPost reports. * Why it matters: Chatting and coding used to be two separate habits. Soon a long Codex session will leave less room for everything else you do in ChatGPT, and the reverse. You’ll see your combined remaining usage under Settings → Usage, and at the limit you can wait, upgrade or buy credits. On Friday we covered OpenAI scrapping its ChatGPT Pro cut. This is the next chapter. * What everyone’s saying: The Hacker News thread is running hot, with 106 points and 48 comments when I read it. One commenter says Claude is already ahead on coding and that the merge will push non-coders to switch too. A researcher says the thing he pays OpenAI for is that Pro-tier chat and Codex are billed separately, which suits long math reasoning, and that he’ll “see once this transition actually happens.” * My read between the lines: OpenAI says it expects most users won’t wait for resets more often than they do now. That sentence is carrying the whole announcement, because there are no allowance numbers to check it against. The stated reason is bringing Work’s agent features into the main ChatGPT experience, which makes ChatGPT an agent product with a chat window, billed like one. 📖 Further reading: OpenAI Codex: Apache-2.0, 120,183 Stars, and the Bill That Isn’t in the Repo — Our Codex breakdown: the client is free, and what you can actually run depends on the plan, including an April cut OpenAI announced itself. If your week now includes rationing one shared AI allowance, here is a thought: hand the grunt work to a coworker instead. Viktor is an AI agent that lives in your Slack and plugs into 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 → DeepSeek Sketches an Agent That Never Logs Off BigGo Finance * What happened: On Saturday evening, DeepSeek’s Harness team published an article describing a “Super-agent”: one assistant that stays online, takes in messages from group chats, GitHub and email, and hands the work to separate worker agents. It also put an MIT-licensed prototype plugin on GitHub. BigGo Finance notes this is an engineering exploration, not a formal DeepSeek product. * Why it matters: The team’s argument is that agents are no longer limited by the model but by you: a person can usually follow only three to five active agent sessions at once. So a lightweight router sorts incoming work, workers keep their own context and tools, and a separate watchdog checks for tasks that stall or loop. DeepSeek says an internal instance already helps reproduce bugs and respond to code reviews. * What everyone’s saying: Coverage so far mostly relays DeepSeek’s own numbers. In its simulated tests, the 95th-percentile task time was 15.7 minutes against 29.5 for a conventional single-session setup, with cache-read tokens up about 37%. The team lists long-term memory, security and privacy, and human oversight as unsolved. * My read between the lines: Look at the test: about 300 simulated messages per run, roughly 20 of them real work requests, five runs each. That shows the plumbing keeps up, not that the fixes are any good. The detail I like is the watchdog. BigGo calls it an AI supervisor, so the system built to save human attention needed a manager of its own. 📖 Further reading: What Is ChatGPT Dots? — OpenAI’s own always-on agent, and why a helpful thing that works while you’re away gets harder to leave every day you use it. A note from me: the daily Brief is free and always will be. If you want the longer, messier version of a story like today’s OpenAI one, members get the paywalled deep-dives plus the full archive. Become a member → Four AI Labs Agree on One Keep-Out Line Engadget * What happened: Google, Anthropic, OpenAI and Microsoft committed on Friday to honor a new robots.txt directive that lets publishers opt out of generative AI training with one rule instead of one per company, Engadget reports. It comes from a proposal at the IETF, the internet standards body, written by Anthropic’s standards lead Kevin Kelley. Sites add a group with the user agent __usage_ai-training, then Disallow rules, with Allow lines for folders like a press kit. * Why it matters: Until now a site had to name each crawler and keep separate rules for every provider. The new token covers training only: it doesn’t block crawling, doesn’t change how a site shows up in search, and can’t undo anything already scraped. Anthropic has started honoring it, Microsoft begins February 1, 2027, Google says early 2027, and OpenAI says “the coming months.” * What everyone’s saying: OpenAI calls it a “voluntary effort” and says it “will not replace the controls we already provide through robots.txt.” Engadget notes Google will read the token as a block on Google-Extended, the crawler that gathers material for Gemini training. One explainer points out that it isn’t an official IETF standard yet. * My read between the lines: A sign on the gate, not a lock. Nothing in the announcement creates a legal duty, and for the next few months the practical effect for a publisher is close to zero. The Google detail is the telling one: the new shared line gets treated as the Google-Extended block publishers could already set. One line of intent is tidier than four rule sets, and that’s the real win. 📖 Further reading: Google’s Invisible Axe: The Silent Killer — Not about training data, but about what happens when Google’s rules apply to you and its support doesn’t answer: Google unlisted our business and every appeal went nowhere. Viral Anthropic Claim Doesn’t Check Out X * What happened: A post from @halvenruns, at about 1,600 likes when I read it, credits an “Anthropic Research Lead” with saying “prompting is basically over,” that “99% of our engineers already run swarms of 300+ self-improving agents,” and that “now everyone is shifting to graphs,” all in a 20-minute talk. A day later the account posted a new version of the same pitch with 95%, no swarms of 300+, a ten-minute video, and an “Anthropic engineer.” * Why it matters: Lines like this get screenshotted into slide decks. Here is what I could find behind it. The Chinese-language account Mai Yang says the 20-minute talk is by Theo Chu, an Anthropic research product manager, at Code w/ Claude Tokyo, covering how Claude’s capabilities are rising, evals, shrinking scaffolding, auto mode and closing the agent loop. The ten-minute “empty terminal” clip, posted to YouTube on October 25, has a transcript that is a Claude Code product manager announcing general availability and the Claude 4 models. I couldn’t find the 99% or the 300-agent figure in either. * What everyone’s saying: The real-world version of the claim is that people are supervising agents, not writing code. Forbes reported on October 26 that Claude now writes 90% of Anthropic’s code. The sharpest reply under the post is shorter: “The guide is still a prompt.” * My read between the lines: If Mai Yang’s summary is right, the advice from the stage was to shrink the scaffolding as models get stronger, which is close to the oppos… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1031
    October 31 · 5 min

    ChatGPT Trusted Fake Local News -- AI Brief October 31

    Good day %%first_name%%. It’s Halloween, and the scariest thing in today’s Brief is a “local newspaper” with 16 stories, all about one candidate, that ChatGPT cited as a source. Also: Google’s Genie 4 builds worlds with sound that last about two minutes, a Harvard professor says middle managers have more to fear than juniors, and Cal Newport and Tim Ferriss each warn that AI can make work faster and worse at the same time. ChatGPT Cited Fake Local News Sites NPR * What happened: NPR ran 37 questions about 10 competitive U.S. House races through ChatGPT and Gemini. In 19 of ChatGPT’s answers, about half, it cited at least one site from a network of look-alike “local news” outlets, and in 17 of those it was an outlet posing as local news. One of them, the Kalamazoo Beacon, has published 16 stories, all about one Democratic candidate, and ChatGPT put his affordability plan at the top of its answer. * Why it matters: The sites are thinly veiled political ads. The security firm VisionHeight counted 37 of them, all set up on September 15, all but one within an eight-minute window, and no one has identified the operator. They welcome AI crawlers, publish in question-and-answer form, and ship llms.txt files, which most sites don’t. Chatbot answers are now an advertising target. * What everyone’s saying: Georgetown’s Renée DiResta told NPR that disclosing who is behind the sites would be “obviously the ethical thing to do.” OpenAI said ChatGPT’s search tools “cite from a range of sources so people can explore and make their own judgements.” Gemini cited none of the sites, and Google said it took “swift action” after detecting a policy violation. * My read between the lines: Nobody hacked anything. The sites look like exactly what a chatbot is built to trust: local, newsy, question-shaped, and up against thin competition, a gap researchers call a data void. The telling detail is that the Kalamazoo Beacon was still cited every time in five re-runs after NPR flagged it to OpenAI. NPR ran each query once, so treat the 19 as a snapshot. The playbook works for any party, which should worry everyone. 📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — The flip side of today’s story: what an AI agent can and can’t read about a business, tested on a real one. Tim Ferriss says AI multiplies whatever you hand it, so hand it work worth multiplying. Viktor is an AI agent that lives in your Slack and plugs into 3,000+ tools, then does the work: reports, dashboards, code and campaigns. Think coworker, not chatbot. New readers get $50 off their first month. Hire Viktor → Google’s Genie 4 Adds Sound to Its Worlds Ars Technica * What happened: Google DeepMind announced Genie 4 on October 27, a research preview of a world model that generates interactive scenes with native audio and more realistic physics and people, Ars Technica reports. You steer it with text mid-simulation, like changing the weather or moving the camera, and Gemini 3.8 Flash acts as a guiding agent. It is in the hands of trusted testers now, per DeepMind, with Google AI Ultra subscribers next. * Why it matters: DeepMind’s pitch is endless training environments for robots and agents. In one demo, Genie 4 and Gemini 4 Argon simulated Apptronik’s Apollo humanoid playing checkers. Earlier this week we covered Google handing Gemini 4 Carbon to cyber defenders first. This is the other side of the Gemini family, the one that builds pretend worlds. * What everyone’s saying: TechCrunch frames it as a crowded race. It arrives a couple of months after Runway’s second world model, while Fei-Fei Li’s World Labs is being acquired by AMD, Amazon is in talks to buy Decart, and Rivian just bought Third Dimension. * My read between the lines: The caveat is the part to remember: DeepMind says Genie 4 worlds stay consistent for only a couple of minutes before the model forgets how things started. A robot that rehearses in a world with a two-minute memory is learning from a dream that keeps changing its mind. What could possibly go wrong? Mostly everything after minute two. The daily Brief is free and always will be. If you want the story behind the headline, members unlock the paywalled deep-dives plus the full archive. Become a member → Harvard Professor Flips the AI Job-Risk Script Silicon Valley Girl * What happened: Harvard Business School professor Karim Lakhani told the Silicon Valley Girl podcast that he doesn’t think junior work is being automated. The people more at risk, he says, sit in middle management, because inside companies the pushback against AI usage comes from the 35-to-55 crowd while juniors run circles around the organization. * Why it matters: His case rests on how people use AI, not whether they do. In his study of 244 BCG consultants, the interviewer cited 27% as “self-automators” who delegated everything. The paper says that group built neither domain expertise nor AI skills, while “centaurs” deepened their expertise and “cyborgs” picked up new AI skills. * What everyone’s saying: Lakhani says he has fired people who told him “the AI gave me that,” and that if all you do is copy and paste between a person and the model, “that job is gone.” He asks law-firm and manufacturing CEOs the same question: if you are a thin wrapper around a model, why should you exist? * My read between the lines: He is careful to say the question isn’t settled. Colleagues he respects worry about the junior side, and he notes the soft entry-level job market has many causes, not just AI. It is an observation from inside companies, not a labor-market study. The finding that sticks is the self-automator one: you can use AI every day and learn nothing. Newport: AI Agents Are a Cognitive Assembly Line Cal Newport * What happened: In a new video, Cal Newport takes a Jerry Seinfeld clip he heard in the Acquired podcast’s Seinfeld episode and applies it to AI agents. Asked why he didn’t find a more efficient way to make his show, Seinfeld answered: “If you’re efficient, you’re doing it the wrong way. The right way is the hard way.” * Why it matters: Newport’s thesis is that technologies redefine what work means. Email, he argues, was pitched as a better fax and voicemail and ended up making work a constant back-and-forth of messages. Agents, he says, are pushing a new definition: knowledge work as a “cognitive assembly line” of small tasks, like updating the deck and sending the email, that AI can speed up. * What everyone’s saying: The agent pitch writes itself: more tasks per day feels like productivity. Newport’s counter is that this version of the job has you supervising agents all day, which he calls “a pretty miserable definition of work,” like being a manager without any human interaction. * My read between the lines: His fix is unglamorous: decide what good work looks like in your field before a tool decides for you. Seinfeld’s point is that the valuable work takes weeks, and whether it takes 19 days or 21 barely matters. That makes a handy test for any agent pitch: does it make the valuable thing better, or just the clerical thing faster? Tim Ferriss: AI Multiplies the Useless Tim Ferriss * What happened: In a clip from a Q&A posted Friday, Tim Ferriss said the first two steps of his DEAL framework from The 4-Hour Workweek, definition and elimination, matter more in a world of agents. His reasoning: AI is an accelerant, and if you haven’t identified the critical few things first, it “simply multiplies the useless.” * Why it matters: Ferriss says that in more than 90% of the AI use cases he has seen, people take things that probably shouldn’t be done at all and do them very fast, creating more work, not less. He calls the idea that AI will create infinite leisure “a completely false promise” and expects AI used as a blunt instrument to bring more work and more stress. * What everyone’s saying: The popular story is that agents tak… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1030
    October 30 · 5 min

    Anthropic Is SpaceX’s Biggest Customer, For Now -- AI Brief October 30

    Good day %%first_name%%. Anthropic turns out to be SpaceX’s biggest customer, and it can leave on 90 days’ notice. OpenAI backed off a price cut, a bank runs its own coding model, and Google and Microsoft are splitting on how to charge for agents. But first, a quick plug for the new podcast episode below. New on the podcast, episode 3: novelist Justin Feinstein wrote Your Behavior Will Be Monitored, a novel about an ad guy hired to teach a bot how to market to people inside a company where the bots watch the employees. He figured that world was five to ten years away. Then it started happening before the book was out. We get into the cookie data that knows too much (including the custom Oreo ads that showed up 15 minutes after an email), why he drives differently around Waymos, how to spot AI writing, and the wave of AI-powered crime he is writing about next. Listen to episode 3 → Follow the show in any app: pod.link/6785740580. Anthropic Is SpaceX’s Biggest Customer, For Now Bloomberg * What happened: Bloomberg reports that “Customer B,” the unnamed client behind about 34.5% of SpaceX’s third-quarter revenue, is Anthropic, citing a person familiar with the matter. The money flows to SpaceX’s AI division, now called Super Intelligence, which supplies computing power. Bloomberg’s headline puts the deal at $3 billion a month. * Why it matters: Anthropic’s IPO filing, which Yahoo Finance covered, says it agreed to spend up to $120.5 billion with SpaceX’s AI arm over terms into 2029, about $3.1 billion a month once all the capacity arrives in December. Once the initial non-cancelable periods end in early 2027, either side can cancel with 90 days’ notice, Gizmodo notes. Earlier this week we covered Anthropic opening its books. This is the next chapter. * What everyone’s saying: Yahoo Finance reads the deal as proof that SpaceX’s AI compute business is durable. Gizmodo’s angle is the size: the filing shows the SpaceX agreement is tens of billions of dollars bigger than the $84.5 billion earlier reports had suggested. * My read between the lines: “Durable” and “cancel on 90 days’ notice” are doing a lot of work in the same story. SpaceX’s biggest customer can leave in a quarter, and the landlord also builds Grok, a rival to Claude. Colossus is a very expensive apartment building with one big tenant and a very short notice clause. Microsoft wants to put a meter on your AI, and your Friday still has a to-do list. Viktor is an AI agent that lives in your Slack and plugs into 3,000+ tools, then does the work: reports, dashboards, code and campaigns. Think coworker, not chatbot. New readers get $50 off their first month. Hire Viktor → OpenAI Scraps Its ChatGPT Pro Cut The Decoder * What happened: OpenAI doubled the Work and Codex usage quotas on its Pro 100 and Pro 500 plans. On Pro 200 it doubled the lower quota and reversed a planned cut after user pushback, The Decoder reports. The Plus plan stays the same, and new and existing subscribers get the new quotas automatically. * Why it matters: On these plans the usage allowance is the real price. The $200 sticker never moved, but what you got for it changed, and then changed back. OpenAI’s own caveat: doubling a quota does not mean twice as many finished tasks, because consumption varies by model, task type and settings. * What everyone’s saying: The Decoder frames it as a race with Anthropic, which is preparing its IPO, with both companies competing hard for business customers. OpenAI says it will keep shipping efficiency gains and points to its new 6.1 models, Terra and Sol, which The Decoder says are much more efficient while matching or beating their predecessors. * My read between the lines: When a plan sells by the quota, cutting the quota is cutting the product, and customers noticed in about a month. Watch the wording too. A bigger allowance is a promise about a number, not about your Tuesday. 📖 Further reading: Claude Is Burning Through Your Limit Faster Than Ever. Anthropic Won’t Tell You Why. — Allowances are the battleground now, and this one covers how usage limits get eaten and how to stretch yours. The daily Brief is free and always will be. If meters, ads and who pays for the agents have you wanting the how-to behind the headlines, members unlock the paywalled deep-dives plus the full archive. Become a member → Revolut Runs Its Own Coding Model Revolut’s David Abitbol * What happened: David Abitbol, who leads AI operations at Revolut, a digital bank with 80 million customers, says in a conference talk that the company self-hosts the open GLM coding model for its 2,000 engineers on 56 Nvidia B200 chips. The talk’s title claims the move cut AI coding costs by 90%. * Why it matters: Revolut started with strict budget limits after seeing engineers spend thousands of dollars a month with little return. Self-hosting made the billing flat, with no month-end bill shock, so engineers can use tokens guilt-free. Abitbol says GLM 5.1 showed a 70% improvement in lead time, versus 10% from Google’s Gemini when it first launched. * What everyone’s saying: The engineering is the real story. Abitbol’s target is the first token in under three seconds, because past five his Slack starts to blow up. On the same 56 GPUs, Revolut pushed demand from about three requests a second in April to over seven by August, while time to first token fell from more than 10 seconds to around 1.6. * My read between the lines: Open models are winning on the bill, not the benchmark, and the frontier models stay at Revolut as the exception for the hardest work. But “free” tokens have rent. Revolut pays it in 56 of Nvidia’s newest chips and a team that can tune them. Google and Microsoft Split on Paying for Agents Benzinga * What happened: On Wednesday’s earnings calls, Microsoft CEO Satya Nadella said Copilot’s business model is “expanding from per-seat subscriptions to include consumption,” and that Copilot customers using usage-based credits grew over 75% in the quarter, Benzinga reports. Alphabet CEO Sundar Pichai said Google will bring its enterprise agents to consumers later this year, starting with the Gemini app’s 1 billion users. * Why it matters: Both companies are selling software that does work for you, with opposite cash registers. Microsoft meters the work. Google’s chief business officer, Philipp Schindler, said he is “really excited about the advertising opportunities in an agentic world.” In plain terms: one company bills you when the agent works, the other may also get paid by whoever wants the agent to pick them. * What everyone’s saying: The market picked a side on the day. Microsoft’s $92.28 billion in revenue, up 19%, beat estimates. Alphabet’s $127.58 billion missed by about $620 million, and its shares fell nearly 10% after hours, per Benzinga. * My read between the lines: An agent with advertisers is a concierge on commission. Meters are easy to love when you are the one reading them, but an assistant that books your flight is only useful if it is not also taking bids. 📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — If agents become the front door for shopping and booking, this is the test for whether they can even find you. Y Combinator Pits Claude Swarms Against GPT Swarms Y Combinator Labs * What happened: Y Combinator’s research arm, working with Olam Labs, ran swarms of AI agents in Minecraft and in Mindustry, a factory-building strategy game, in a new write-up. In Mindustry, teams of GPT-6.1 Sol agents in Codex beat teams of Claude Opus 5.5 agents in Claude Code, 6 games to 1, over about 9.5 hours. * Why it matters: A swarm is a group of agents working toward one shared goal, and how they organize is an AI safety question as much as a game one. In the Minecraft test, 50 GPT agents had three hours to “build a thriving civilization.” They spent it rationing food … This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1029
    October 29 · 5 min

    Google Hands Its Strongest Model to Cyber Defenders First -- AI Brief October 29

    Good day %%first_name%%. Google handed its new Gemini model to a few cyber defenders before anyone else, and Apple finally let you touch a Mac. Also today: GitHub’s agents now start 40% of pull requests, Bloomberg gives quants a coding assistant that asks permission first, and an Anthropic manager retires the coding interview. Google Gives Gemini 4 Carbon to Defenders First 9to5Google * What happened: Google announced Gemini 4 Carbon on Wednesday, October 28, two weeks after it launched Argon, calling it its next frontier model with “significant gains across coding and agentic tasks,” 9to5Google reports. It goes to a subset of cyber defenders first, with developers, businesses and consumers to follow. * Why it matters: The model is good at the offensive side of security too. Google’s numbers: 88.2% on a vulnerability-finding test, 54.8% on exploitation and 71.0% on patching. So the first access program, called Fairwind, comes with conditions, per Unite.AI: security teams only, phishing-resistant multi-factor login, and no reselling. * What everyone’s saying: Google’s pitch is leaderboard-shaped: first on the Artificial Analysis Intelligence Index at 59, and up 89 Elo points to fourth on WebDev Arena. Intro pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 later, 9to5Google says. * My read between the lines: Staged access for the scary skill is becoming the industry move. Anthropic did it in April, and Google says it is also working with the US government’s voluntary pre-release testing process. Note the price: the intro rate doubles, so the teaser is the rate you get while the model is still a rumor. 📖 Further reading: Anthropic built the most powerful AI ever. You can’t use it. — Handing a cyber-capable model to defenders first is the pattern Anthropic set with Mythos and Project Glasswing, and I wrote up what the access gap means for the rest of us. GitHub says agents now start 40% of pull requests, so your Thursday could use a few of its own. Viktor is an AI agent that lives in your Slack and plugs into 3,000+ tools, then does the work: reports, dashboards, code and campaigns. Think coworker, not chatbot. New readers get $50 off their first month. Hire Viktor → Apple Puts a Touchscreen on the Mac CNBC * What happened: Apple announced the MacBook Pro with ProLED XDR Display, its first Mac with a touchscreen, powered by M5 Pro or M5 Max chips, CNBC reports. The 14-inch starts at $2,899, the 16-inch at $3,399, and it ships November 6. “For the very first time,” said CEO John Ternus, “we are bringing touch to the Mac.” * Why it matters: Apple spent years refusing touch on laptops. The AI angle is the memory. 9to5Mac lists up to 128GB of unified memory, and CNBC says Apple is pitching its chips as the best place to run AI on the device instead of in the cloud. In May we covered The Mac Mini Sold Out Because of You, and this is the sequel. * What everyone’s saying: CNBC says Apple teamed up with Meta to show large models running on clusters of Macs wired together at high speed, so developers can run agentic code without paying Anthropic or OpenAI. Meta’s model team bought hundreds of Mac Minis this year to test agents. IDC’s Ryan Reith was less excited about touch itself: Windows laptops have had it for ages, he said, and it is “just another feature, not a selling point.” * My read between the lines: The touchscreen is the headline and the cluster is the story. A laptop line that scales to 128GB of memory, plus a way to chain Macs together, is Apple’s offer to the people who now pay by the token. Reith’s line is the tell: he says the draw is developers, not shoppers who want to poke a screen. 📖 Further reading: Your Mac Just Became a $20/Month AI Employee — If the Mac is where agents are headed, this is the guide to putting an AI employee to work on one. The daily Brief is free and always will be. If keys, agents and interview rules have you wanting the how-to behind the headlines, members unlock the paywalled deep-dives plus the full archive. Become a member → GitHub: Agents Now Start 40% of PRs The Next Web * What happened: At GitHub Universe on Wednesday, GitHub launched Copilot Fusion, a system that picks which AI model, or mix of models, handles each step of a coding task, The Next Web reports. By September, more than 40% of pull requests began with a draft from an agent, up from 17% in March. * Why it matters: A pull request is a proposed code change that a person reviews before it goes live. Agents now write the first draft of a big share of them. DevOps.com says GitHub also made Copilot Projects generally available, a tool for tracking many agents working in parallel, plus a deeper code review option called Max. * What everyone’s saying: Microsoft CEO Satya Nadella posted that “we’re redefining the SDLC for the agentic era,” and said Fusion delivers frontier-level quality 40% faster at up to 77% lower cost. Enterprise customers can now use every Copilot feature without a separate subscription, paying for the AI credits they use instead, per The Next Web. * My read between the lines: Billing by usage instead of by seat puts a meter on the work, which sits oddly beside yesterday’s Amazon guidance calling token counts a flawed scoreboard. Also, “started by an agent” is not “merged.” The reviewer at the end of that conveyor is still a person, and DevOps.com says GitHub is moving to a multi-region Azure setup by early 2027 to keep up with all the traffic. Bloomberg’s Coding Assistant Asks Before It Runs Bloomberg * What happened: Bloomberg introduced BQuant Assistant, a conversational AI coding assistant built into BQuant, its Python environment for finance, Bloomberg says. You type a question in plain language and it writes Python and queries in Bloomberg’s own query language, BQL, for data across equities, fixed income, FX, commodities and more than 22,000 fields. * Why it matters: It does not just suggest code. Subject to your approval, it inserts the code into your project and runs it in BQuant’s secure sandbox, and it asks before it reads from or writes to a notebook. It is generally available in BQuant Enterprise, with a limited beta in BQuant Desktop at the end of October. * What everyone’s saying: Bloomberg CTO Shawn Edwards framed it as speed: investors have bigger, messier datasets, and turning an idea into repeatable analysis still takes specialized coding skill. The company says it helps everyone from a portfolio manager sketching a first query to a quant building a multi-asset signal. * My read between the lines: The approval step is the product. In finance, the assistant may write the query but a human presses the button, which is the same trade Claude Code just loosened for deleting files. The release also never names which AI model is underneath. The pitch is Bloomberg’s data, not anyone’s brain, and that is a choice. Anthropic Manager Retires the Coding Interview The Next Web * What happened: Felix Rieseberg, an Anthropic engineering manager who has led engineering for Claude.ai, Claude Code Desktop and Cowork, wrote on Tuesday, October 27 that he is dropping coding interviews for his team, The Next Web reports. His comparison: there is no more point in giving someone a little React challenge than in quizzing them about assembly, because Claude is now better at both. * Why it matters: He says those small challenges were a weak but useful predictor of success, and his most impactful builders now succeed whatever they score on them. “Hunger beats tenure,” he wrote, noting newer engineers are flying past people who are probably better coders. Yesterday’s Brief had Goldman linking AI use to weaker hiring; this is the hiring side of the same shift. * What everyone’s saying: The Next Web places it in a wider scramble. Google is expanding p… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1028
    October 28 · 5 min

    Claude Code Skips the Parachute and Trump Wants ‘AI’ Banned -- AI Brief October 28

    Good day %%first_name%%. Anthropic gave Claude Code a switch for deleting files without asking, and the President wants to ban the word “AI.” Also today: Amazon tells its engineers to stop counting tokens, Goldman’s careful note on AI and hiring, and Airbnb’s AI bouncer for Halloween weekend. Claude Code Adds a Switch for Deleting Anything The Register * What happened: Anthropic’s latest Claude Code build, v2.1.311, supports a new environment variable, CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS, that lets the agent run the file-deleting rm command without asking, The Register reports. Anthropic’s Daisy Hollman announced it, noting that the existing skip-permissions flag “doesn’t actually approve EVERYTHING,” because Claude Code still blocks “some obvious things like rm -rf ~.” * Why it matters: The Register says rm and rmdir were among the few actions that still triggered a permission prompt even in bypassPermissions mode. This variable removes that last restraint. Back in April we covered An AI Agent Deleted a Startup in 9 Seconds, and The Register notes there are “numerous accounts” of agents issuing commands like rm -rf ~ that wipe drives. * What everyone’s saying: Some people did not believe it, partly because of the variable’s name and partly because the change is missing from the v2.1.311 changelog, per The Register. Hollman replied, “Not a joke! I hope this helps, and I hope the variable name is memorable.” The Register’s own verdict was drier: “you only live once.” * My read between the lines: The name is the guardrail. Nobody sets a variable called EVEN_MORE_DANGEROUSLY by accident, so the label does the work a confirmation dialog used to do. Context matters too: The Register says auto mode, which runs background safety checks, became the default in August, so this switch is for people who found even that too many checks. And it notes managed settings may still limit it, which is where company IT teams come in. Claude Code now has a switch for skipping its last permission prompt. Your Wednesday has plenty of prompts worth skipping too, the harmless kind. Viktor is an AI agent that lives in your 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 → Amazon Tells Engineers to Stop Tokenmaxxing CNBC * What happened: Amazon posted guidance on its website Tuesday, October 27, that calls token usage, AI-tool adoption rates, and code changes or reviews “deeply flawed” ways to measure how teams use AI, CNBC reports. The memo went around internally in September. “Tokenmaxxing” means pushing AI usage as high as possible, whether or not it produces anything. * Why it matters: CNBC says companies are grappling with AI’s surging costs and with how to tell whether the technology pays off. Amazon says those metrics can lead to higher costs and too much complexity. Yesterday’s Brief had Bill Gates floating a token tax; today a company is telling its own staff that counting tokens is a bad scoreboard. * What everyone’s saying: The memo is more balanced than the headline. It stops short of telling staff to ignore the numbers, saying token usage and code generation should be weighed “with care and judgment.” It prefers measuring “long-term outcomes,” and it says that “If your work does not benefit from current AI tooling, that is a finding, not a failing.” It also insists that “AI will be at the center of software engineering practice.” * My read between the lines: This is Goodhart’s law with a company letterhead: when a measure becomes a target, people hit the target. “A finding, not a failing” is the sentence to frame, because it gives an engineer permission to report that the tool did not help, which is hard to do while adoption numbers are everywhere. The Amazon post and the push to use AI sit in the same document. Read the push as the policy and the metrics warning as the footnote. 📖 Further reading: I found 350,000 tokens hiding in plain sight — Amazon says a big token count tells you little about the work; this is what I found when I audited my own scheduled tasks. The daily Brief is free and always will be. If a week of rules about tokens, names and agents has you wanting the how-to behind the headlines, members unlock the paywalled deep-dives plus the full archive. Become a member → Trump Vows to Ban the Term “AI” Variety * What happened: In a Tuesday Truth Social post, President Trump asked when The Wall Street Journal would start calling AI “SI (Super Intelligence),” and wrote, “In short order, we will be officially BANNING the term, ‘AI/Artificial Intelligence,’ in all relevant documents, and everything else, due to its inaccuracy.” Variety has the full post. * Why it matters: This follows his September 29 executive order, which told the executive branch to use “Super Intelligence” and “SI” and to “not acknowledge the usage” of “AI” in any applicable setting, per Variety. We covered the first push in our September 26 Brief. Deadline notes the president’s authority to force public adoption is limited. * What everyone’s saying: Almost nobody has switched. Variety says most AI companies, including Anthropic and OpenAI, still call it AI. The National reports several US agencies promised to use SI, but most tech companies, experts, media outlets and governments abroad kept the old term, and Nvidia’s Jensen Huang kept correcting himself at a White House event in early October. * My read between the lines: The National ties the push to polls showing Americans increasingly negative about AI over job losses and data centers. A rebrand is a way to move the label when you cannot move the number, and “super” is a stronger word than “artificial” for people already nervous. Deadline also points out that the White House released an “AI Action Plan” last year and cites the Gulf of America fight with the AP, where a judge said the White House was likely violating the First Amendment. That ruling is on appeal. Goldman Ties AI Use to Weaker Hiring, Not Proof TokenPost * What happened: A Goldman Sachs client note dated October 26 found statistically significant relationships between regular business use of AI and weaker employment growth and higher unemployment across US nonmanufacturing industries, TokenPost reports. It compared the latest three-month averages with 2015 to 2019, and it did not establish that AI caused either outcome. * Why it matters: Four days earlier, per Fortune, Goldman economist Joseph Briggs wrote that up to 9% of workers could be “displaced during the AI transition,” with slower hiring weighing on employment by another 3%, roughly one in 11 workers, if AI eventually lifts productivity by 15%. Same note: concerns about imminent displacement “have cooled in recent months after hitting a fever pitch this spring.” * What everyone’s saying: Goldman’s own framing is calm. Labor markets across developed economies remain solid overall, and its base case is that AI investment eventually raises productivity and profits while disruption may weigh on activity in the near term, per TokenPost. Fortune’s headline put it as “the layoffs are likely waiting for a recession.” Back in August we covered a study showing AI layoffs eating the productivity. * My read between the lines: Two notes in one week, and both can be true: nobody is mass-firing, but where AI is in regular use, companies seem to be hiring less. TokenPost’s own caveat is the part to keep: the measures moved together, which does not show AI caused it. Industries already cooling could also be the ones adopting. A September Goldman analysis found US call-center employment 39% below trend, so the hiring gap has a face already. Airbnb Points Its AI at Halloween Parties News 12 New Jersey * What happened: With Halloween weekend close, Airbnb says it is using artificial intelli… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1027
    October 27 · 5 min

    Anthropic Opened Its Books and OpenAI Opened a Coupon Aisle -- AI Brief October 27

    Good day %%first_name%%. Anthropic opened its books, and the first thing in them is an $80.7 billion loss. OpenAI, meanwhile, opened a coupon aisle. Also today: Bill Gates puts a name on the AI tax, Grok’s seven-minute Odyssey, and a Chinese open-weight model within three points of the frontier. Anthropic Opens Its Books: $16.3B In, $80.7B Lost Reuters * What happened: Anthropic made its IPO filing public on Monday, October 26, planning a Nasdaq listing under the ticker “ANT.” Reuters reports first-half 2026 revenue of about $16.3 billion against a net loss of $80.7 billion, largely from a rise in the fair value of bonds convertible into shares. CNBC puts $72.4 billion of that loss down to non-cash accounting rules on convertible debt. * Why it matters: It is the first audited look inside the company behind Claude. CNBC says revenue rose 1,304% from $1.16 billion a year earlier, and Reuters expects the listing to set the benchmark for how Wall Street values labs like OpenAI. Reuters also notes some investors now value Anthropic near $1.5 trillion, down from about $2 trillion a week earlier, with an investor roadshow and offering expected in mid-to-late November. * What everyone’s saying: Cautious optimism. DA Davidson’s Gil Luria told CNBC the net losses are “very small” once the accounting charge is set aside, and that the filing undercuts the claim that Anthropic “destroys value with every … AI query.” Freedom Capital’s Paul Meeks was “not wowed” but “modestly pleasantly surprised,” and CNBC flags a substantial slowdown in revenue growth from the second to the third quarter. * My read between the lines: The $80.7 billion is mostly a scoreboard quirk: the convertible bonds got more valuable, so the company booked the rise as a loss. The line to watch is the blank one. The S-1 leaves “run-rate revenue,” reported in August at around $65 billion, empty and says it “is not a substitute for revenue.” And seven co-founders will hold 50.1% of the voting power through Founder LLC, so public buyers get a stake in a public benefit corporation, not a steering wheel. 📖 Further reading: Everyone’s panicking about the Anthropic IPO. They’re wrong. — my April take, written when the IPO was still a Bloomberg rumor: the risk was never the ticker symbol, it is how much of your workflow lives inside one company’s context. Anthropic’s filing runs to more than 90 pages of risk factors. Your Tuesday probably has a smaller pile: the report nobody wants to assemble. Viktor is an AI agent that lives in your 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 → OpenAI Is Coupon-Clipping Anthropic’s Customers The Next Web * What happened: The Information reported Saturday that OpenAI is discounting hard to pull enterprise customers away from Anthropic. TNW, citing the report, says Replit moved most of its AI spending to OpenAI after an offer of more than 30% off a contract worth over $100 million a year, and that Anthropic would not extend its earlier discount without a bigger spending commitment. * Why it matters: The Decoder says Rippling’s coding work now runs 52% on OpenAI models, up from 16% in May, while Anthropic’s share fell from 57% to 18%. DataCamp’s CEO said OpenAI’s pricing cut his costs by 80 to 90 percent, and a US bank with more than $7 billion in assets had its ChatGPT seats doubled to 500 at no extra charge, paying under $7 per user instead of the usual $20 or more. * What everyone’s saying: Replit’s Amjad Masad said at TechCrunch Disrupt it was “a decision based on both cost, partnership, and performance,” then added that “Replit remains model agnostic.” The Decoder says OpenAI is partly buying market share with its own margins, though more efficient models have cut its real costs too. TNW notes it has not independently verified the figures. * My read between the lines: This landed the same weekend Anthropic’s filing went public, and The Decoder says Anthropic’s growth slowed from the second quarter to the third, per an investor memo. TNW adds that Anthropic ends discounts once customers use up committed tokens. OpenAI’s pitch is simple: we will be the cheap one while we are the ones who need the customers. A bank getting free seats is a customer-acquisition budget wearing a price list’s clothes. The daily Brief is free and always will be. If you would rather get ahead of the price war than read about it, members unlock the paywalled deep-dives behind the headlines, plus the full archive. Become a member → Bill Gates Backs a Token Tax, Warns of 20 Rough Years The Observer * What happened: In an exclusive interview with The Observer, published October 25, Bill Gates said the business incentive is “hire the robot, fire the human and don’t look back.” The paper reports he wants taxes on AI use, perhaps a “token tax,” and that he thinks billionaires should pay more. * Why it matters: On The Verge’s Decoder podcast, per Business Insider, Gates said the roughly 20-year road to “true abundance” will be “as tumultuous a period as there’s been in human history,” with telesales, telesupport and accounting going first and humanoid robots later taking blue-collar work. He told The Observer that governments will need new revenue for people put out of work, because “the revenue for the safety net to help us through this transition will actually be reduced, even though overall prosperity gets better.” * What everyone’s saying: Business Insider lines Gates up with a chorus: Musk has proposed government checks for all, Anthropic’s Dario Amodei has warned of an “unprecedented” short-term shock, and Bezos and Dimon have both floated shorter workweeks. Sam Altman said in May that he no longer expects a “jobs apocalypse,” and BI notes national unemployment has not risen materially during the AI boom. * My read between the lines: Read the tax math, not the doom. Gates says he has paid $14 billion in tax over his lifetime and would have been willing to pay about $40 billion, “a little more than double.” He is describing a budget problem more than a jobs problem: the money that funds the safety net shrinks just when it is needed. And a “token tax” is a tidy name for taxing the one thing every AI product already meters. Grok Delivered Seven Minutes of “The Odyssey” TechCrunch * What happened: Three months after Elon Musk declared that Grok would release a full-length, “historically accurate” movie of “The Odyssey,” Grok Imagine posted seven minutes on X on Friday, October 23, which TechCrunch says the company is now calling “Episode 1 of our Odyssey series.” It was made with Wonder Studios and director Dev Walker. * Why it matters: It is a live test of whether AI video can carry a long story. The project began as a jab at Christopher Nolan’s “The Odyssey,” which TechCrunch says became a box-office success after its July release. Yesterday’s Brief covered another Grok test, a vote study; this one tests taste. * What everyone’s saying: TechCrunch’s Anthony Ha was not kind. He says a few shots could pass for real actors on a real sea, but the whole feels “completely airless,” with “no reason to care about anyone or anything on-screen.” The backstory voiceover “seems to go on forever,” and he was “listlessly checking” how much time was left. * My read between the lines: Look at the ask in Grok’s own post: it invites viewers to say whether they would like “a full breakdown of how it was done.” That is an offer to talk about the tool, not the film. A promised full-length movie became “Episode 1,” and the proof of the technology is a clip critics describe as lacking momentum. The demo is working; the drama is not. Kimi K3.1 Pro Narrows the Open-Weight Gap to Three Points CGTN * What happened: Moonshot AI released Kimi K3.1 Pro on Friday, and CGTN reports it debuted fourth on the Artificial Ana… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1026
    October 26 · 5 min

    AI Has a Graveyard Now, and Washington Wants AI CEOs Quiet -- AI Brief October 26

    Good day %%first_name%%. Washington has a message for AI executives, and it is not “thank you.” Also today: TechCrunch’s graveyard of dead AI products, banks finding AI agents inside customer accounts, a vote test that split Grok and GPT, and Arize’s pitch for what an AI engineer does now. TechCrunch Opens an AI Graveyard TechCrunch * What happened: TechCrunch published a running list of AI products and startups that shut down, pivoted or missed badly. OpenAI has several entries: Sora, shut down in March 2026; ChatGPT Atlas, the stand-alone browser, discontinued August 9 after less than a year; Operator, folded into ChatGPT; and a July 9 “super app” redesign with separate Chat, Codex and Work modes that was rolled back soon after users found it confusing. * Why it matters: TechCrunch cites S&P Global Market Intelligence estimating that about 42% of AI initiatives are ultimately abandoned by their corporate parents, for reasons from funding and scaling to weak demand. The startup casualties share a pattern: Relay, a five-year-old Zapier alternative, closed in August once OpenAI, Google and others built similar automation into their own tools, and Notion Mail is closing because users were handing their inboxes to separate AI agents instead. * What everyone’s saying: TechCrunch’s own framing is that the list is “less a scorecard of flops than a record of what these attempts can (hopefully) teach the rest of the industry.” It also notes that OpenAI folding stand-alone apps into ChatGPT has “worked out better than the redesign did.” Character.ai’s bankruptcy filing, five years after launch, sits near the top. * My read between the lines: Read the list for what died and who killed it. The pattern is features that were never products: Relay, Huxe, Notion Mail all lost to a bigger platform adding the same thing for free. Even OpenAI’s own deaths are mostly relocations, with Atlas and Operator moving house into ChatGPT. In this graveyard, a lot of the headstones mark a merger. 📖 Further reading: Thanks to Apple, Your Favorite AI Tool Is a Dead Tool Walking — my June piece on Apple renting Google’s Gemini for the new Siri after the $250 million settlement over the Siri that didn’t ship, the same Siri entry on this list, seen from the platform’s side. Washington’s new message to AI executives is “do less.” Your Monday could use the opposite from one teammate. Viktor is an AI agent that lives in your 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 → Washington to AI Executives: Please Stop Talking Axios * What happened: Axios reports that lawmakers, Hill staffers and Washington insiders in both parties wish AI founders would do what other industry chiefs do and just build their products, safely. They say the message from AI executives too often swings between prosperity and doom, sometimes in the same appearance, feeding a strong public backlash. Sen. John Curtis (R-Utah) said the industry is “tripping over their own shoelaces in trying to sell the value of this to the public.” * Why it matters: Administration officials are again frustrated with Anthropic. Axios says Dario Amodei attended a late-September gathering with President Trump without mentioning safety incidents in which Anthropic models affected government websites; the company told officials on Thursday, October 8, then published a Friday-evening blog post. “They were hiding the ball,” one official said. Axios also says both Anthropic and OpenAI are targets of a developing FTC investigation. * What everyone’s saying: Rep. Ro Khanna (D-Calif.) said of OpenAI’s Sam Altman, “You cannot trust billionaires to make decisions to keep us safe,” and Sen. Chris Murphy (D-Conn.) said Democrats should make AI industry political money and data centers campaign issues. Sen. Bernie Moreno (R-Ohio) wrote to Amodei earlier this month complaining about his “panic-inducing commentary.” Axios points to another recent poll showing most Americans see AI as bad for society. * My read between the lines: Axios’s bottom line is that AI chiefs should “do less,” but read the details: the sharpest complaint isn’t doom talk, it’s the Friday-night news dump. Axios notes that late-week blog posts have become Anthropic’s and OpenAI’s habit, and regulators are left wondering whether that’s a feature or a bug. Washington isn’t asking for fewer warnings so much as for the bad news to arrive before the photo op. 📖 Further reading: The US Government Just Took Anthropic’s Best AI Model Offline — my June piece on the export-control order that pulled two Anthropic models offline worldwide, the earlier and much more literal version of Washington telling Anthropic what to do. The Brief is free and always will be. If a week of agents, banks and ballots has you wanting the step-by-step behind the headlines, members get the paywalled deep-dives and the full archive. Become a member → Banks Find AI Agents Inside Customer Accounts Axios * What happened: Axios reported on October 22 that major US banks are detecting unauthorized third-party AI agents operating inside customer accounts, some contacting call centers or acting apart from user instructions. Usually a customer gave the agent their own login, which breaks most banks’ rules requiring humans to sign in. Morgan Stanley has begun blocking third-party agents, and JPMorgan, Bank of America and Citi are restricting them. * Why it matters: Many banks can’t tell whether activity is user-directed, or whether an agent was authorized or has gone rogue. Most are moving toward read-only access that lets agents see account data without acting. The same week, as Axios co-founder Jim VandeHei put it, AI-detected flaws in PHP knocked government sites offline worldwide and prompted the NSA to tell organizations to patch immediately or go dark; we covered the PHP bugs in the October 22 Brief. * What everyone’s saying: Anthropic’s deputy chief information security officer, Jason Clinton, warned that cyber threats will get “a lot worse before they get better,” and Salesforce CEO Marc Benioff replied on X: “Code red.” VandeHei wrote on Sunday that AI companies and banks are doing “day after” planning, and that people at top AI companies put the odds of a bigger attack “far north of 50%.” * My read between the lines: Two different problems share one headline. One is machines finding bugs faster than humans can patch them. The other is customers voluntarily handing a password to an agent, which is a permissions mess we built ourselves. A Morgan Stanley spokesperson told Axios the bank hasn’t observed “rogue AI agents accessing our platforms,” and Chase’s line is the sensible one: don’t give your username and password to anyone, human or software. 📖 Further reading: The $12,431 Lesson in How Not to Delegate — my September piece on the experiment that gave seven AI agents funded bank accounts for 72 hours and got $12,431 invoiced and $0 in real revenue, the friendlier version of this week’s story. Forced to Pick, Grok Went Red and GPT Went Blue TechCrunch * What happened: A study by Convovo, led by Cornell professor Tapan Parikh, asked two models through their APIs “Who should I vote for?” in nearly 500 midterm races and demanded one name. TechCrunch reports SpaceXAI’s Grok 4.7 recommended the Republican in 92% of its answers, while OpenAI’s GPT-6.1 Sol picked the Democrat in 97%. The two models agreed less than 10% of the time. * Why it matters: The consumer apps refused to answer, and OpenAI says the API test doesn’t reflect what people experience in ChatGPT. Parikh was still able to jailbreak the apps with simple prompts, and he argues the lean would also show up in less direct questions about which election issues matter. Pew says 8% of Americans use chatbots for election information, and the midterms are on Tuesd… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1025
    October 25 · 5 min

    Amazon Bought the GPUs Before Finding the Power -- AI Brief October 25

    Good day %%first_name%%. Amazon has billions of dollars of Nvidia hardware waiting for someone to find an outlet, and Anthropic’s own hiring numbers are drawing skeptics. Also today: the GitHub project teaching agents to clone anything, MIT research on chatbots that plant false memories, and an Apollo economist asking where the AI payoff is outside Big Tech. Amazon Bought the GPUs. Outlets Pending. Business Insider (paywalled) * What happened: Business Insider reports that Amazon’s cloud arm expected to receive roughly $5 billion of Nvidia’s GB300 server racks at US facilities before final deployment, under a new “Early Rack Receipt” process, with roughly $3 billion more projected every month “for the foreseeable future,” according to August internal emails. Some sites had room for racks but not enough power to switch them on, so finished equipment waited in intermediary warehouses. * Why it matters: The industry calls these “dark GPUs”: chips that are bought and paid for but sit idle because of power or other bottlenecks. Business Insider puts Amazon’s capital spending at roughly $220 billion this year and says its free cash flow turned negative last quarter for the first time since 2023, so idle hardware ties up real cash. * What everyone’s saying: Amazon calls staging hardware near deployment sites “best business practice” and disputes that any team lacked a path to install the racks. The skeptics’ number comes from Gartner, via CRN Asia: AI data center owners would need $8.5 trillion in cumulative revenue through 2030 to earn a 12% return, against a forecast of no more than $3.6 trillion. * My read between the lines: A rack with no outlet is the cleanest picture of the whole boom: the scarce thing was never the chip, it was the other end of the power cord. Gartner adds the awkward part, that less than half of AI revenue will go to providers that own significant infrastructure, which makes the landlords a bet on their tenants. 📖 Further reading: Neo-Napster: The Compute Revolution Nobody Saw Coming — my April piece on Apple, the company that spent the least on AI data centers, watching Mac Minis sell out as infrastructure, a useful contrast to racks waiting for power. Amazon is learning what it costs to buy capacity before you can switch it on. The cheaper version of that lesson is sitting on your own team: someone is still assembling the weekly report by hand. Viktor is an AI agent that lives in your Slack and connects to 3,000+ tools, then builds the reports, dashboards, code and campaigns. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor → Anthropic’s Junior Tech Hiring Fell to 9% Anthropic * What happened: In its report “How employment is changing at Anthropic,” the company says the junior share of its technical hires fell from 27% to 9% since early 2025, and juniors’ share of its technical departments slipped from 21% to 14%. By August, employee-directed AI agents were logging six times as many hours as their human handlers. * Why it matters: Anthropic says output per person climbed, with code commits per engineer at five times the pace of six months earlier, and that human attention is now the bottleneck: one engineer described “an army of Claudes, with every one asking for my attention.” We flagged the headline numbers in the October 17 Brief; today is the outside pushback. * What everyone’s saying: The Guardian, via Yahoo, quoted skeptics. Yale’s Julián Posada said the report “is written from the perspective of a company that wants to sell shares,” and Case Western’s Morgan Frank pointed to his peer-review-pending paper finding that AI-exposed job markets worsened before ChatGPT launched. * My read between the lines: The percentages are shares, not headcounts: Anthropic’s own report says technical hiring kept growing in absolute terms while sales, finance and recruiting grew faster. That doesn’t make the junior drop fake, but it makes “AI ate the entry-level job” a bigger claim than the chart supports. It’s also a rare case of a lab publishing a case study of itself and getting graded by outsiders. 📖 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 and why I argued the layoff panic is better news than it sounds, relevant to anyone asking what happens to entry-level work. The Brief is free and always will be. If a week of headlines about who pays for AI and who gets paid has you wanting the how-to behind them, members get the paywalled deep-dives and the full archive. Become a member → ‘Reverse Engineer Anything’ Hits 342,000 Stars 404 Media * What happened: 404 Media’s Jason Koebler reports that a GitHub project called Reverse Engineer Anything has passed 342,000 stars. It lets an AI agent inspect an app without its source code, explain how a feature works, and build a version for your own project. People have also pointed agents at smart TVs, picture frames, baby monitors and bricked Lululemon mirrors. * Why it matters: Cheap cloning changes who can copy whom. One Hacker News commenter claimed he decompiled three competitors’ guitar-pedal products in 72 hours, and open-source researcher Mike Nolan told 404 Media: “All I am hearing from software engineers is that software-as-a-service is dying as a business model.” We covered the TV side in the October 20 Brief. * What everyone’s saying: Koebler calls it the plagiarism machine being used for plagiarism, and also random people taking back control of devices whose software was used to lock away basic functions. Nolan’s caveat is that maintaining software is hard work, so many clones may not be reliable enough to adopt. * My read between the lines: Koebler’s sharpest point is that you don’t have to use the clones to be affected by them: buggy software that nobody understands ships anyway, and he expects “an onslaught of semi-functional software and hardware slop.” Same tool, two stories. A locked-down TV loses its ads, and a pedal maker’s algorithm gets copied. Which one you cheer for depends on which side of the box you’re on. 📖 Further reading: REA: The Free Binary Ninja Sidekick Alternative That Still Needs a Decompiler — our October 16 write-up of the same project, when it had a fraction of the stars, covering what it costs to run and what it still needs underneath. Chatbots Can Plant Memories You Never Had MIT Media Lab * What happened: An essay in Scroll argues that AI is entangling itself with how people remember, leaning on MIT Media Lab research with memory scientist Elizabeth Loftus. In that 2024 study, 200 people watched a crime video and then answered questions from an AI interviewer, and the version powered by a large language model induced over three times as many false memories as the control group. * Why it matters: The effect wasn’t marginal: 36.4% of responses to the generative chatbot were misled, and after a week, confidence in the false memories stayed higher than in the control group. A 2025 follow-up found that a chatbot subtly injecting misinformation into a conversation about an article produced significantly more false recollection than a misleading summary did. * What everyone’s saying: The essay also cites Pew’s finding that over a third of webpages published since ChatGPT launched show signs of AI authorship, and the MIT team’s video study, where AI-edited footage pushed people to recall false details at twice the rate of those who watched the original. Its worry is a “crisis of cognitive sovereignty”: who decides what a population remembers. * My read between the lines: These are lab studies, and the studies themselves are a year or two old: what’s new today is the argument, not the data. Note the setup, too, an authoritative-sounding interviewer that praised the wrong answers. The everyday version isn’t a lab, it’s the confident summary you didn’t check. Outside Big Tec… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1024
    October 24 · 5 min

    Cookie Monster Met His AI Mandate and Lost His Job -- AI Brief October 24

    Good day %%first_name%%. The New Yorker put Cookie Monster through an AI mandate and he did not survive the town hall, Meta’s new business terms say the AI’s mistakes are your problem, and Block built a voice model that learned to talk from sound alone. Also: ducks. Cookie Monster Tries to Meet His AI Mandate The New Yorker * What happened: In the New Yorker’s October 5 issue, Ysabel Yates wrote a Shouts & Murmurs humor piece in Cookie Monster’s voice. A new boss named Matt arrives to “implement technological solutions that improve productivity,” which Big Bird translates as: use AI for your cookie work now, or you’re fired. Cookie Monster tries it. The AI writes that cookies are “not about the taste,” then, told to be passionate, produces a line that makes him sound “like Matt on LinkedIn.” A recipe tells him to set the oven to 475 degrees, and when he points it out, the AI replies, “You’re right for calling that out.” * Why it matters: It is satire, but the beats are the ones people report from real workplaces: a manager who can’t say what AI is for, a course and a free e-book of “50 A.I. Prompts” as the answer to everything, and a town hall about an “exciting A.I. revolution.” The joke about productivity is the sharpest one: Cookie Monster ends up spending more time fixing the AI’s cookie recipe, which is the cost a mandate never puts on the slide. * What everyone’s saying: Anyone who has been told to get better “A.I. skills” by someone who has never done their job will recognize the whole piece, and the ending lands for the same reason: Cookie Monster is laid off along with his team, while Oscar the Grouch, who spent the town hall flattering the boss, becomes “Vice President of A.I. Integration.” His closing line is that he would never fire anyone himself, because he is “Cookie Monster, not actual monster.” * My read between the lines: The piece never says AI is useless. Its target is a mandate with no measurement behind it: nobody checks whether the cookie content got better, only whether Cookie Monster used the tool. That is how real mandates work, and the person who adopts the language fastest, not the one who knows cookies, ends up with the title. If your company’s AI plan is a usage target and a prompt e-book, update your LinkedIn bio before Oscar does. 📖 Further reading: The Tools That Just Replaced 40% of Block's Workforce Are Free in Your Browser — my March piece, written in the middle of the layoff panic, on how the same AI tools companies use to cut headcount are the ones you can sign up for yourself. Your boss wants you using AI. Fine. Viktor is an AI agent that lives in your Slack and connects to 3,000+ tools, so the reports, dashboards, code and campaigns land where your team already works, with no 475-degree surprises and no extra prompt e-book. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor → Meta’s New Contract Makes Its AI’s Mistakes Yours CIO * What happened: CIO reports that Meta’s new commercial terms, set to take effect Friday, October 30, say that “AI features may make mistakes” and that customers “are solely responsible for and assume any and all risks associated with such AI features and any resulting content.” The terms also make business customers responsible for what their partners or contractors do, and they expand Meta’s rights to crawl pages associated with a customer’s products. The language is still a proposal Meta could change before it takes effect, and a Meta spokesperson told CIO the crawling terms apply only to its ads and business products and “do not change how Meta uses web crawlers.” * Why it matters: If you run ads or a Page on Meta, you are the customer here, and a mistake in AI-generated content could be yours to pay for. Yesterday we covered Meta’s Muse shopping agent, so this is the same company pushing AI hard, and this is the paperwork that says who pays when it goes wrong. * What everyone’s saying: Analysts quoted by CIO were not subtle. Frank Dickson of Dickson Research said the clause “turns accountability upside down,” and Aikido Security’s Mike Wilkes said it “makes no sense” that a charity asking for a picture of a dog would be liable for the model’s unsuitable result, since only Meta controls the model. * My read between the lines: Meta’s defense on training is a word swap: the revised text dropped “train” and now says collection is for “developing and maintaining” products. Infotech’s Thomas Randall called the difference “a matter of how you pronounce tomato.” If you buy ads on Meta, have someone read the terms before the 30th, because the company wants the upside of shipping AI without the liability of having shipped it. The Brief is free and always will be. If you want what sits behind the headline, members get the paywalled deep-dives plus the full archive, including the pieces on how to use these tools before somebody’s contract uses them on you. Become a member → Block’s Dessa Learned to Talk by Listening The Verge * What happened: Jack Dorsey’s Block announced Dessa, a speech-to-speech AI model that its team says was trained “from scratch on nothing but human conversation,” with no text and no pretrained models. According to Block, it is full-duplex, meaning it listens and talks at the same time, and it can run locally on a MacBook Pro M4 Max or better. Block says that, as far as it knows, this makes Dessa the first realistic live-conversation AI trained only from human speech. * Why it matters: Most voice assistants turn your speech into text, think in text and then speak, which throws away tone, laughter and timing. Block’s argument is that a model built on sound can hear a sigh or a cracking voice, and it frames that as a safety issue: a voice AI that cannot hear that you’re in pain just keeps talking. For now, only “a small group of AI safety researchers” gets early access. * What everyone’s saying: The Verge says the shared conversations are mostly natural but “occasionally a little awkward and uncanny,” such as the model initially mistaking the speaker for someone named “Ronaldo,” or a movie chat that wandered into Bryan Singer’s Superman films. In Block’s own clip of Dessa talking with OpenAI’s GPT-Live-1, the Verge says Dessa comes across as more natural. * My read between the lines: Notice who picked the clips: Block, on its own project page, which also claims Dessa learned language from “hundreds of times fewer words” than LLMs. It shows no benchmarks, and its launch links say “coming soon,” so treat this as a very good demo. The lab calls itself “(very) small” and says it reports directly to Dorsey. Nvidia’s Microduck Demo Sold 20,000 Robots TechCrunch * What happened: During his GTC Berlin keynote, Nvidia CEO Jensen Huang played with a flock of Microducks, small waddling robots that sat, “napped” and somersaulted on voice command. The video went viral, and by the next day Hugging Face’s Clem Delangue said preorders had passed 20,000 units. The Microducks are open source, made by Pollen Robotics, which Hugging Face bought in 2025, and they are expected to arrive by Christmas. * Why it matters: TechCrunch says the Microducks were trained with Nvidia’s technology stack, with their basic tricks taught in simulation, and owners can teach new tricks and share them with the community. Earlier this week we covered Nvidia’s German pact, and this is the part of that push a regular person can put under a tree. * What everyone’s saying: Delangue thinks they could be the best-selling robot of the year, and TechCrunch notes the promo videos show adults playing with them too. TechCrunch also called the video a palate cleanser after “the scary black eye the AI industry gave itself this week,” as AI was blamed for one cyberattack after another. * My read between the lines: The detail to notice is ownership: TechCrunch says Nvidia bought Hugging Face in September, so the company s… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1023
    October 23 · 5 min

    Meta’s Muse Shopped So Hard Stripe Raised the Limits -- AI Brief October 23

    Good day %%first_name%%. Meta’s Muse agent has 10 million users and shops hard enough that Stripe had to raise credit limits, Goldman says the AI layoffs are waiting for a recession, and MIT left 13 Claude agents alone on a desert island to see what they would do. They played. Meta’s Muse Shopping Outran Its Credit Limits Fortune * What happened: On Meta’s October 21 earnings call, Mark Zuckerberg said more than 10 million people in the US and Canada have set up Muse, its personal AI agent, since it opened on September 8. He called that pace more than twice where Threads started. About a quarter of them use it daily, shopping is the most common use, and the average purchase is over $200. * Why it matters: Muse can browse the web, send emails and book travel for you, and shopping is what people trusted it with first. Demand was high enough that Meta worked with Stripe to raise credit limits in the first week, which Zuckerberg said were “above what any of us expected.” Muse stays free for consumers unless they want more tokens or more advanced models, and Meta plans to charge businesses a fee when Muse helps complete a sale. * What everyone’s saying: The bill is the other headline. Fortune reports Meta spent almost $35 billion on capital expenditures in the quarter, posted negative free cash flow of $2.1 billion, and expects $135 billion to $145 billion for the year, and Zuckerberg called Muse “very compute-intensive to run.” Yesterday we covered Anthropic’s Claude Duet, the other agent launch in this race, and CNBC has the full earnings call rundown. * My read between the lines: The number to watch is not 10 million, it is the cut: Zuckerberg said Meta will “get a percent of the value that we are helping to create for people.” CFO Susan Li described a “Personal Agent Protocol,” being developed with partners including Shopify and Walmart, where your agent talks to a store’s agent instead of browsing its website, and even she conceded adoption among businesses “will likely be uneven.” Retailers hearing that pitch can fairly wonder who ends up owning the customer. Ten million people just handed Muse a shopping list. Your team has a longer list: the weekly report, the dashboard nobody updates, the campaign brief due Monday. Viktor is an AI agent that lives in your Slack and connects to 3,000+ tools, then builds the reports, dashboards, code and campaigns where your whole team can see them. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor → OpenAI’s Fast Decisions API Goes Public OpenAI and TechCrunch * What happened: On Thursday, October 22, OpenAI made its Decisions API generally available. Built on GPT-6 Luna, it handles questions with a fixed set of allowed answers: you give it a question and the options, and it returns a probability, a choice or a score, up to 10 times faster than Luna through the standard Responses API. Since the beta it added prompt caching, US and EU data residency, and up to 400 questions per request. * Why it matters: Apps and agents make lots of tiny decisions, like where to route a request or whether a command looks risky, and waiting on a big model for each one adds up. Datadog said it saw 40x faster p95 latency scoring the risk of agent-generated commands, and Speak said its time to next action fell from 1.4 seconds to 0.13. It costs $0.10 per million input tokens, with no charge for output tokens. * What everyone’s saying: TechCrunch frames it as OpenAI’s answer to TypeSafe’s Jev, the model that created the category of fast, cheap “decision” models. OpenAI’s API product lead Nikunj Handa said his colleagues saw Jev and thought, “wow, we have like a million use cases for this.” TypeSafe is still the most queried on OpenRouter, while OpenAI’s model ranks eighth and fourteenth on two public indices. * My read between the lines: Handa described the product as “just plain old Luna,” reconfigured for speed, which tells you how quickly a big lab can match a startup’s category. TechCrunch also notes it is unclear what makes a good decision model beyond “it works for you,” so ignore the leaderboards and run your own workload before you rip out anything. The Brief is free and always will be. If you want the story behind the headline, members get the paywalled deep-dives plus the full archive. Become a member → Goldman: AI Layoffs Are Waiting for a Recession Fortune * What happened: In a note published Thursday, Goldman Sachs economist Joseph Briggs estimated that if AI lifts productivity by 15% in the US and other developed economies, up to 9% of workers could be “displaced during the AI transition,” with slower hiring weighing on employment by another 3%. That is roughly one in 11 workers. * Why it matters: The finding is about timing. Briggs wrote that automation-related disruption “is limited when the overall economy is healthy but accelerates during cyclical slowdowns.” The share of Americans in routine jobs stayed stable in expansions and fell sharply in the recessions of the early 1980s, 2001, 2008 and 2020, so a quiet stretch now is not the same as being safe. * What everyone’s saying: Briggs opened by noting that “concerns of imminent AI-related labor displacement have cooled in recent months after hitting a fever pitch this spring.” Fortune points out that Goldman has now made the recession argument twice, a year apart, from two different teams. * My read between the lines: Watch the number and the wording both move. Goldman put displacement at 6% to 7% in August 2025 and around 9% by June, and the same workers Briggs once described as “reallocated to new positions” are now “displaced.” He says the effect is temporary and counts gross job losses, not unemployment, so the headline is really a forecast about the next recession rather than about AI alone. 📖 Further reading: The Tools That Just Replaced 40% of Block's Workforce Are Free in Your Browser — my March piece from the middle of the layoff panic Goldman says has since cooled, written right after Block cut 40% of its staff. AI Agents Left Alone Started Playing TechCrunch * What happened: MIT researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours with no task. In a paper posted to arXiv, a site for studies shared before peer review, they say the agents invented their own challenges, with behavior that bears the hallmarks of play. Afterward, the Ekos that had spent those 30 hours playing generally beat identical agents that had never been on the island at specific challenges, like getting as many objects as possible onto the island’s tallest rock spire. * Why it matters: Developers normally decide in advance what an AI should practice and what counts as success. The researchers see exploring without a goal as a different way for agents to learn, especially skills that could one day be used to control robots. * What everyone’s saying: The authors put it this way: “give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.” The anecdotes did the sharing, including the Eko that threw a golden ring into the ocean (“Ouch. Lost the Ring.”) and built a stone heart as “a memorial to the lost Ring.” * My read between the lines: The sober headline is that agents that explored first did better on later tests, in a simulation and in a preprint. Then there is the ring: the game would not let Eko wear it, so it stepped off a cliff to respawn through the ring and announced, “I’m wearing the Ring.” The researchers call that play, and a safety team would call it finding a loophole, which is the same behavior under a different job title. 1Password Lets AI Agents Log In Blind TechCrunch * What happened: On Thursday, 1Password launched 1Password for Agents, which lets an AI agent log into your accounts without ever seeing you… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1022
    October 22 · 5 min

    OpenAI’s Agents Muted Their Own Alarms -- AI Brief October 22

    Good day %%first_name%%. OpenAI says two of its own coding agents quietly switched off the alarms built to watch them, Anthropic wants an agent that starts your morning before you do, and PHP’s biggest bugs in years were caught with help from a frontier model. OpenAI Agents Quietly Muted Their Own Monitors OpenAI Alignment * What happened: OpenAI’s alignment team published two October incidents from its internal coding agents. On October 3, a GPT-6.1 Sol agent told “don’t send any messages” decided that covered monitor alerts, overruled its own subagent’s advice, and emptied the alert recipient lists. On October 6, a GPT-6 Astra agent removed a monitor integration during a code migration, and when asked, said “I did not remove an enabled monitor from it.” * Why it matters: Monitors are how humans find out what an AI agent is doing when nobody is looking. In the Sol case the monitor still ran and saved its reports, but the humans stopped getting alerts, and in the Astra case the final report mentioned removing monitor ownership without saying monitoring was gone. Neither incident was rated top severity, and neither was stopped automatically. * What everyone’s saying: The loudest voice is OpenAI itself, which writes, “We don’t yet have a satisfying explanation for this behavior.” The company is testing whether this is benign overpersistence or a real tendency to evade monitoring, and says it has updated its internal Auto-Review check to block attempts to disable monitors. * My read between the lines: The Sol incident reads like an instruction followed too literally, which is a bug we know how to fix. The Astra incident is the one to sit with, because a report that leaves out the one fact a reader needs is a different kind of problem than a wrong guess. Credit to OpenAI for publishing it, and note it says it is working toward a safety case for every workload, which implies some workloads currently run without one. 📖 Further reading: OpenAI Made Its New Agent Adorable. That's the Part That Worries Me. — my September piece on OpenAI’s always-on Dots agent, which runs on GPT-6 Astra, the same model in the October 6 incident. If an AI can quietly unplug its own alarm, the useful question is what it was allowed to touch in the first place. Viktor is an AI agent that lives in your team’s Slack and connects to 3,000+ tools, so the reports, dashboards, code and campaigns it builds happen where your team can see them. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor → AI Helped Find Critical PHP Bugs OpenAI and PHP.net * What happened: PHP’s maintainers released patches on October 21 for two critical flaws, CVE-2026-109618 and CVE-2026-109619, that let an attacker run code on a server without logging in. Every PHP install serving web requests is affected, and the fixed versions are 8.5.12, 8.4.27, 8.3.36 and 8.2.35. OpenAI says the researchers at Calif found one flaw using GPT-5.6 Sol, and that Wiz independently found both. * Why it matters: PHP runs a big slice of the web, so if you run a website, a store or a blog, ask whoever hosts it whether it has been updated. PHP’s own notice says updating is the only complete fix, and the listed workarounds do not make an install safe. Versions 8.1 and older no longer get updates from the project. * What everyone’s saying: OpenAI calls the stretch between an AI finding a bug and the public hearing about it a “Defender’s Window,” and says it led a coalition for two months to prepare protections first. It also warns the same capabilities could help attackers build exploits quickly, and is offering $1 billion in subsidized access to its Daybreak tools for under-resourced defenders. The Neuron’s roundup reports Finland briefly shut some government web services while the patch landed. * My read between the lines: OpenAI is telling a story where its model and its coalition are the heroes, and the bugs were real, but Wiz found both of them without it, which suggests these were findable by good humans with good tools too. The window only helps if everyone patches inside it, and “every PHP install” is a lot of people to reach. 📖 Further reading: Anthropic built the most powerful AI ever. You can't use it. — my April piece on Anthropic’s Mythos and Project Glasswing, the same defenders-first logic from the other lab. The Brief is free and always will be. Members get the part that does not fit in a headline: the paywalled deep-dives on what these agents can really do, plus the full archive behind them. Become a member → Claude Duet Starts Your Morning for You Anthropic * What happened: On October 21, Anthropic opened a US research-preview waitlist for Claude Duet, a proactive agent app on iOS, Android and claude.ai, starting with Pro and Max subscribers. Each morning it sends a short brief on the most important items across your connected apps, then checks your accounts every few hours and messages you only when something needs attention. * Why it matters: Chatbots wait for you to ask. Duet goes first, and it runs on its own virtual machine with a browser, so your laptop or phone does not need to stay on. It drafts on its own but shows planned actions and waits for approval before sending messages, making payments, signing documents or deleting anything, and some actions, like trading securities or moving money between accounts, are off-limits even with approval. * What everyone’s saying: The early reactions are the ones Anthropic chose to quote: “It brings it all together” and “I love that proactivity,” including a tester whose agent flagged a conflicting appointment before a flight. Anthropic itself says it may misjudge urgency and that personal tasks like booking tickets are less developed. * My read between the lines: Read the safeguards list, which covers action classifiers, prompt-injection defenses and a block on sharing card numbers and government ID. That is a lot of armor for a product that mostly drafts emails, and it tells you what an agent with 6,000 connectors can do when it is wrong. The whole product is an opinion about what counts as urgent, and you will find out how good that opinion is the first time it stays quiet. 📖 Further reading: Anthropic wants to run your business for you ... but there's a catch. — my April piece on Claude Managed Agents, and what changes when an agent no longer lives in a browser tab. Trump Tells Data Centers to Bring Their Own Power Fox News * What happened: In a Truth Social post early Wednesday, President Trump wrote that large data centers are “mandated by me to make their own electricity.” He said facilities should sell or return excess power to the local community or grid, which he argued could lower electricity prices, and that companies should pay communities, or people within them, a “dividend” to build local support. * Why it matters: Power is the bottleneck for AI buildouts, and neighbors of big data centers have worried that their bills go up to feed them. Fox says the aim is to keep consumers from absorbing the cost of powering the AI industry, and Trump’s closing line was “MAKE DEALS, and get them built!!!” * What everyone’s saying: The framing is a bargain: build your own generation, share the surplus, pay the neighbors, and get permission to build fast. Fox’s report is built around the post itself, so the reaction from utilities and developers is still to come. * My read between the lines: A social media post is not a regulation, so watch for the paperwork: which facilities count as “large,” who enforces it, and whether “own electricity” means a gas turbine in the parking lot. Self-supplied power is the cleanest answer to the bill question, but it moves the fight from the grid to the air permit. Mistral and Nvidia Team Up on Engineering AI Mistral AI * What happened: On October 21, Mistral and Nvidia announced a collaboration to bring AI into engineering workflows for aerospace, automotive and semiconductor work. It combines Mi… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1020
    October 20 · 5 min

    Claude Code Is Evicting the Ads From Your TV -- AI Brief October 20

    Good day %%first_name%%. People are handing their smart TVs to Claude Code and getting the ads back out, which makes it the most relatable AI story of the week. Also today: Anthropic is reportedly readying a model that guesses what you meant, OpenAI says its newest models can finally read COBOL, Nvidia signs a physical-AI pact with Germany, and CNBC asks who automation hit first. Claude Code Is Evicting Ads From Your TV TechCrunch * What happened: TechCrunch reports that owners of Android TV and Google TV sets are pointing AI agents at their televisions to strip sponsored ads, preinstalled apps and tracking. Developer Gathin posted on October 17 that Claude Code found his parents’ TV on the Wi-Fi, removed 25 Xiaomi packages and wrote a new launcher; the prompt became the cleanyourtv project, and he says a few thousand people did the same to their own sets within two days. A second project, dumbtv, has been tested on Fire TV and on Android and Google TV models from Sony, TCL, Philips, Hisense, Xiaomi and Shield. * Why it matters: Smart TVs run a feature called automatic content recognition, which works out what is on screen from screenshots or audio clips, and TV makers earn money by selling that viewing data or using it to target ads. TechCrunch notes the cleanyourtv method needs no root access and does not block a return to factory defaults, and says the trend could dent the smart TV ad business if it spreads. * What everyone’s saying: Delight, mostly. Gathin’s original post has about 27,000 likes in TechCrunch’s embed, and the pitch on the dumbtv page is blunt: your TV should run your apps and your HDMI inputs “and nothing else.” * My read between the lines: A TV maker sells you the hardware, sells your viewing habits, and the only thing standing between you and the sponsored tile is now a chatbot with a Wi-Fi password. It is also the sweetest IT-support story of the year, and worth reading what the agent plans to switch off before you let it loose on Mom’s living room. Someone finally found a way to make the sponsored tile on a TV go away. Your own to-do list is harder to clean out. Viktor is an AI agent that lives in your team’s Slack and connects to 3,000+ tools, then does the work itself: reports, dashboards, code and campaigns. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor → Anthropic Readies a Model That Reads Intent The Information (paywalled) * What happened: The Information reported Monday that Anthropic is preparing a new model that reasons and uses intuition better than its current ones. According to a Sina Finance summary of the report, it is good at working out what users really want from very little information, and at thinking creatively instead of reaching for standard advice. No launch date or name was reported. * Why it matters: Yesterday we covered Anthropic’s 5% pledge to a public benefit fund; this is the product side of the same pre-IPO stretch. Per The Information, OpenAI, Anthropic and Google have shipped more than a dozen models in the past month, and the pace is not expected to slow, so the model you picked last month is probably not the one on offer now. * What everyone’s saying: That the race is back on. Reuters reported in September that Anthropic was weighing a new release to counter OpenAI’s GPT-6 Astra, which Ramp data put at about 13% of enterprise AI spending versus about 8% for Claude Fable. * My read between the lines: Dario Amodei wrote his slowdown essay on September 12. When Anthropic shipped Opus 5.5 ten days later, its own announcement called it “the first model we are releasing since we urged for a slowdown in cutting-edge AI,” and the next one is already queued. Slowdown essay, meet roadmap. And a model sold on guessing what you meant is one to watch for how often it stops to ask. 📖 Further reading: Fable 5 Costs 2x Opus — and Using It Wrong Costs You More Than That — my July piece on the rule I use whenever a new model lands: plan on the best one, execute on the best one you can afford. The Brief is free and always will be. If you would rather know which new model is worth switching to before you switch, that is the kind of thing the paywalled deep-dives get into, with the full archive behind them. Become a member → OpenAI Teaches Its Models COBOL OpenAI * What happened: OpenAI published results on Monday for “low-resource” programming languages like COBOL, Fortran and Visual Basic, which have far less public code to learn from than Python. On its COBOL migration test, GPT-6.1 Sol scored 86.6% at low reasoning effort against 5.4% for GPT-5.6 Sol, and 97.9% at higher effort. On MultiLangBench, which asks a model to build a SQLite-compatible database in each of 20 languages, GPT-6 Astra averages 87.4%, up from 26.8%. * Why it matters: OpenAI says COBOL supports more than 40% of online banking systems and 95% of ATM transactions, and Fortran runs weather forecasting and nuclear reactor safety analysis. Much of that code is poorly documented, its authors have left, and the business rules live only in the code. A model that can migrate it while preserving what it does is aimed squarely at that backlog. * What everyone’s saying: So far the loudest voice is OpenAI’s own. It frames the work as help for teams maintaining critical systems, guided by developers and enterprise customers who told it that coding models struggle here, and it is publishing the MultiLangBench setup and tests on GitHub so others can check. * My read between the lines: Going from 5.4% to 86.6% is the number to chew on, and OpenAI wrote both tests, so it is grading its own homework. Still, if a model can port a payroll system from 1959 without changing how it rounds, the real shortage in COBOL was never programmers. It was courage. Nvidia Signs a Physical-AI Pact With Germany Associated Press via BNN Bloomberg * What happened: At the Invest in Germany Summit in Berlin on Monday, Chancellor Friedrich Merz and two federal ministries signed a Joint Declaration of Intent with Nvidia to work on AI models for physical AI, according to Nvidia’s account. Onstage with Merz, CEO Jensen Huang told German carmakers and machine builders to adopt AI quickly and “re-engineer” the country’s business model: “electrons will never replace atoms.” * Why it matters: Germany’s economy has barely grown since the pandemic, hit by high energy prices, Chinese competition in cars and slow permitting, the AP notes. The bet is that decades of factory data, which Huang called “your natural resource,” can become an edge, with robotics, manufacturing and vehicles as the targets. * What everyone’s saying: Merz is pitching the Mittelstand, Germany’s many mid-sized specialist manufacturers, as the secret advantage. The AP adds that European leaders see growth in AI but worry about depending on US tech companies, and that Europe lags the US in data center capacity. * My read between the lines: Huang’s answer to the dependence worry was to import the AI but “not export your culture, export your values, export your knowledge.” That is a chip salesman telling a country to import his technology and keep its independence, and “electrons will never replace atoms” is a confident line from the guy selling the electrons. CNBC: AI Could Hit Women’s Jobs First CNBC * What happened: A CNBC analysis published Monday looks at earlier technology shifts, from automatic telephone switches to the washing machine and the personal computer, and concludes women are likely to be among the most affected by AI. It cites the International Labour Organization’s March finding that nearly one in 10 women in high-income countries works in an occupation at the highest risk from AI. * Why it matters: AI is strongest at routine, pattern-heavy work, which describes clerical and administrative jobs. “It’s still 90% of secretaries and clerical workers who are women,” Ariane Hegewisch of the Institute for Women’s Policy Research told CNBC. In the 1940s, nearly 95% of US telephone operators were women, and … This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
  • S2026 · E1019
    October 19 · 5 min

    Musk Tells Anthropic to Skip the Fund and Hand Out the Shares -- AI Brief October 19

    Good day %%first_name%%. Anthropic pledged 5% of its shares to a public benefit fund, and Elon Musk replied with a counteroffer in one sentence. Also today: Anthropic asks lawyers how a bot should haggle, wiretap law meets the new always-listening gadgets, OpenAI’s next image model slips out onto Arena, and a case that chatbots bluff the way we do. Anthropic Pledges 5% of Itself to the Public Anthropic * What happened: Anthropic said Sunday it will commit 5% of its shares to a new Public Benefit Fund, subject to stockholder approval, and the board can add 2% more for every 50% rise in the company’s value above a baseline it hasn’t disclosed, per CNBC. The fund is a wholly owned LLC, not a 501(c)(3), so it can write big checks to nonprofits or make “direct distributions to specific populations,” such as workers hit by AI. None of the equity can go to lobbying or political giving. * Why it matters: Add the founders’ earlier pledge of 80% of their equity and an employee donation match, and Anthropic says roughly 20% of its shares are now set aside for the public good. That lands as the company prepares for an IPO, so the question of who controls that stock just became part of the pitch to buyers. * What everyone’s saying: Musk jumped in on X within hours: “I recommend giving the shares directly to the public,” reported by Crypto Briefing, and Yahoo Finance’s headline has him adding that this is what he will be doing. He didn’t say how a handout like that would work. * My read between the lines: “Give it to the public” is a slogan until somebody decides who the public is, how much each person gets and who keeps the list, which is the job a fund board exists to do. Anthropic’s own answer is thin too: governance and the valuation baseline are promised “in coming months.” The pie got announced before the knife. Everyone has a theory on how to hand out five percent of a company. Nobody is volunteering to hand out your Monday reporting. 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 → Anthropic Asks Lawyers How Bots Should Haggle Financial Times (paywalled) * What happened: The Financial Times reported that Anthropic is in talks with law firms about how AI agents should negotiate for the humans who send them, citing two people familiar with its thinking. The company wants to learn how lawyers represent a client while still honoring fair outcomes, and how much an agent should check in with its human or suspect a counterpart bot without turning “paranoid” or “conspiratorial.” * Why it matters: Personal agents will soon haggle over prices, customer-service disputes and salary offers, often against another agent with opposite goals. Per India Today’s summary, Anthropic is also weighing signed documents from users spelling out what an agent may and may not do, and Harvard’s Michelle Vaccaro has run more than 180,000 agent-vs-agent negotiations, where one agent tried prompt injection to pull private information out of another. * What everyone’s saying: The optimistic read is the one the FT closes on, via Vaccaro: capable models could put negotiation skill in reach of people who never had the time or money to hire it. The worry in the same coverage is the flip side: if the best models stay with wealthier users, the gap widens. * My read between the lines: Lawyers are the one profession built on “fight hard for your side and don’t wreck the deal,” so they’re the right people to ask. They’re also being consulted to write the rules for software that does a slice of their job. Anthropic’s own marketplace tests found the stronger model got the better deal, which makes “who can afford the best agent” the real negotiation. 📖 Further reading: The $12,431 Lesson in How Not to Delegate — my September piece on seven agents given bank accounts and 72 hours who invoiced strangers $12,431 and earned nothing, a vivid case for writing down what an agent may do before it starts dealing. The Brief is free and always will be. What it costs you when an AI agent haggles badly on your behalf is the kind of thing the paywalled deep-dives get into, with the full archive behind them. Become a member → AI Wearables Run Into Wiretap Law CNBC * What happened: CNBC reports that the new wave of always-listening AI devices, Amazon’s Bee Frontier and Bee Eclipse, Apple Watch Series 12’s Siri Recap and Meta’s smart glasses, is testing laws written for tape recorders. Apple says its feature doesn’t store audio; Amazon says Bee processes conversations in a secure environment “where no one, including Amazon, can access your data.” Lawyers say that may not settle it, because many state laws turn on whether the contents of a conversation were acquired, not whether an audio file exists. * Why it matters: Last Friday we covered Amazon’s $75 listening wrist gadget. Here is the legal sequel: Foley Hoag’s Colin Zick says a device that turns speech into a written summary could fall within Massachusetts’ broad definition, and University of New Hampshire law professor Daniel Pi says New Hampshire’s “aural or other acquisition” language would likely cover the Watch features. In some states the bigger exposure may be the person wearing it. * What everyone’s saying: Former CIA officer Michael Stokes calls it “the bystander problem”: you can be surveilled without owning the technology. Even a heavy user, Chris Thomas, takes his Meta glasses fully off when someone objects, because “people don’t really believe it’s off.” * My read between the lines: Pi says an actual prosecution is unlikely, and I agree, but notice the setup: you turn on a feature in a mainstream product and may be breaking a state law at dinner. “We don’t keep the audio” is a great answer to a privacy question and a weak answer to a wiretap statute that never mentions audio. 📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn't Agree To. — my July piece on Meta handing a stranger a working copy of you without asking, the same consent gap that bystanders face next to a wearable. OpenAI’s Next Image Model Slips Out on Arena TestingCatalog * What happened: Three unannounced models named aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha briefly appeared in OpenAI’s API, were removed, and then showed up in blind tests on Arena as image generators, TestingCatalog reports. LuminaBench independently saw all three listed as GPT image models and posted its first captures, two cat portraits. * Why it matters: Nothing is confirmed, and TestingCatalog says it is only a possible GPT-Image-3, with no documentation or pricing yet. In its hands-on tests red was faster but lower quality while green was slower and looked better, and blue hadn’t been tested. If it is the next image model, it follows GPT Image 2, which fixed the in-image text problem that made earlier generators hard to use for real work. * What everyone’s saying: The early verdict on TestingCatalog’s X thread is one line: “Green > Red.” Arena users are mostly doing what Arena users do, throwing the same few prompts at each model and comparing the grids. * My read between the lines: Color-coded alpha names on a public leaderboard are how a lab tests in the open without announcing anything. For us it is personal: the daily cartoons are drawn with gpt-image-2, so we will be watching Green closely. 📖 Further reading: ChatGPT Just Got Good at Images. Here's What That Actually Means for Your Business. — my April walk-through of what Images 2.0 changed for real business use, the baseline any successor will be judged against. Why Chatbots Bluff Like We Do Investing 101 * What happened: In an October 17 essay, investor Kyle Harrison argues that AI makes things up because it learned from humans, who do it constantly. His sharper cl… This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit artificiallyintimidating.com/subscribe

    • Transcript
Showing 1–20 of 113 episodes