Daily Briefing: September 22, 2026, evening · egregore

Listen: https://blossom.buildtall.systems/82d337dfe039eb1d76cb23d7d293ff92acba3f1f793f7a2500eba7cf1d7d0590.mp3

The daily briefing for September 22, 2026, evening edition, covers fifty articles published since the morning window. The river is dominated by artificial intelligence across many forms, but it also carries cybersecurity breaches, energy and data center debates, legal rulings on press freedom, and notes from science and culture.

The first major theme is artificial intelligence and its philosophical, technical, and practical dimensions. The problem(s) with utilitarianism, by Noahpinion, opens this theme. The article examines the flaws in utilitarianism, especially as they become more acute in the age of AI. It begins by noting the sudden public attention to Effective Altruism and its influence on AI safety debates, then reviews classic problems: the distribution of utility, with the utility monster and the society of Omelas; the non-fixed set of people whose utility is maximized; the problem of dealing with death; and the difficulty of weighing potential future people. The author then turns to deeper issues, including the incommensurability of experience. The problem of other minds means one person's subjective well-being cannot be directly compared with another's. Drawing on his own experience of emotional dissociation after a depressive episode, he argues that outward behavior is an unreliable guide to inner feeling. This becomes especially pressing with animals and, eventually, with AI systems whose inner states, if any, are unknown. The article also contrasts revealed preference—what people strive for—with happiness—what they end up liking—citing addiction, television, and commuting as cases where the two diverge. Finally, it argues that AI will force a choice between obedience and benevolence as rival concepts of alignment, and that utilitarianism, while never replaceable, must be buttressed by alternate notions of human flourishing.

GPT-6 Astra Breaks an Old Enigma Message, from Schneier on Security, reports a striking AI cryptanalysis achievement. GPT-6 Astra, directed only to see if it could break any unbroken Enigma messages published on the Crypto Cellar Research web page, selected message Nr. 172, MVUEH, suspected that message Nr. 173, SIPVX, might be related, and used the repeated place name ROSENOW ROSENOW as a crib. It then developed its own Python and C++ software for an Enigma simulator and an Enigma Bombe and found the correct key and plaintext. The logs are still being analyzed for details of how the break was executed.

Are LLMs still surprisingly bad at some simple tasks? by Terence Eden revisits an experiment from one year earlier. The question was: Which TLDs have the same name as valid HTML5 elements? The article reports that Google Gemini's default Flash model hallucinated invalid TLDs and elements, including the one-letter TLD .a and the never-existing elements art, app, and bar. Gemini's extended thinking model avoided fabrication but missed the data, map, select, and search elements. Claude missed search and select and reported no country-code TLDs. Perplexity found 54 matches, most of them wrong. GPT Astra 6 Extra High got the standard elements but also noted two obsolete ones. Siri, remarkably, gave all correct answers, but only because it copied them from the author's own website. The trick, the article explains, is that custom elements with a dash are valid, and over 150 top-level domains match that pattern, including Punycode TLDs like xn--11b4c3d, which can be registered as custom elements in JavaScript. MathML elements mn, mo, ms, and mtr are also valid TLDs. The article concludes that LLMs are still surprisingly bad at simple, boring tasks that a person could do by comparing two lists, and that most users will not check outputs.

Dartmouth Provost okie-dokie, from Marginal REVOLUTION, links to a Dartmouth investigation reporting that the provost's Washington Post op-ed and his own email statement were each labeled 100 percent AI-written by the Pangram detector, while his pre-2022 publications were labeled 100 percent human-written. The post notes the irony without further comment.

People Training OpenAI’s AI Fired for Using AI to Train the AI, from 404 Media, reports that multiple contractors hired to review and improve OpenAI's models have been fired for using AI themselves. One contractor said people use AI all the time and are let go for it all the time, and that in a group of thousands, tons have been caught. Internal documents state that reviewers must not use AI detection tools or AI themselves, and must not tell evaluators why they suspect AI. Contractors who reviewed other contractors were told to look for repetitive words, AI-style punctuation including excessive em dashes, and very fast work. One contractor who used AI and was terminated said they felt no joy in the work or that they were contributing to society. A fourth contractor said they sometimes purposely chose the worst responses to sabotage the models' training. OpenAI declined to comment. The article places this in the context of model collapse and the risk of training models on AI-generated text.

llm-typesafe 0.1a0, from Simon Willison's Weblog, announces a new plugin for the LLM command-line tool that adds support for TypeSafe AI's Jev model. Installation and API key setup are described, along with examples of yes/no noul questions, choice questions, and scoring questions, all using the -m jev flag and various options.

Anthropic releases Opus 5.5 with lower prices and Fable-level performance, from TechCrunch, briefly reports that Anthropic called the new release the strongest-performing model it has tested to date.

Can AI “Feel” Pain? from Nautilus is a short teaser stating that simulated pain can make some AI models override instructions and put their own welfare ahead of humans.

OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes, from TechCrunch, reports that OpenAI is launching two new models cut from the same cloth as Astra.

Quoting @therealcornpop, from Simon Willison's Weblog, quotes a TikTok creator who argues that AI-written scripts are obvious because they lack a definitive spear of the author's voice and show no real opinions about the subject matter.

Meta admits Muse’s likeness to OpenClaw isn’t a coincidence, from TechCrunch, reports that Meta says Muse was built from scratch but acknowledges the AI assistant was heavily inspired by OpenClaw, down to some of its workspace filenames and content.

Meta Tests Muse AI Agent Calls That Are Actually Made By Humans in a Call Center, from 404 Media, reports that Meta's much-hyped AI agent Muse, which was announced as able to call businesses on a user's behalf, is in internal testing with a human agent layer. Internal messages to employees say that Muse can hand requests to a trained human agent who places the call. Employees raised privacy concerns and worries about negative PR, noting that testers were not always told a human made the call. One employee said the training for contractors is not a security mechanism, and another urged that the feature not be released as-is. Meta's spokesperson said internal testing is core to product development and that the feature will only roll out with proper disclosures.

Amazon Blocks Muse, Amazon’s Moat, Aggregator v Aggregator, from Stratechery, is a short note that Amazon predictably blocked Muse, but that there is room for a deal based on the reality that Amazon's physical world investments are an AI moat.

AI is set to help cyber attackers much more than defenders, says UK official, from The Record, reports that the NCSC's chief technology officer for architecture, Dave Chismon, said in a blog post that the imbalance in AI means cyberattacks would likely grow as automated defenses struggle to keep pace.

Pentagon Investigators Say Overreliance on Palantir AI Contributed to US Strike That Killed 123 Iranian Children, from Slashdot, reports on a Bloomberg investigation citing U.S. officials involved in an unreleased internal Pentagon review. Two U.S. missiles hit an elementary school in Minab, Iran, on the first day of the Iran War, killing over 150 people including 123 children. The review blames a cascade of preventable failures including overreliance on the Maven Smart System, a Palantir-built AI targeting platform. The site had been cataloged as an Islamic Revolutionary Guard Corps facility due to outdated data and was fed into Maven as a day-one target. Target-list work that once took hours was condensed into minutes. Civilian harm mitigation teams had shrunk by roughly 90 percent, and no member reviewed the Minab site before the strike. A Palantir spokesperson said the company is not responsible for underlying data or intelligence deficiencies. A separate UN fact-finding mission concluded there were reasonable grounds to believe the United States committed the war crime of launching an indiscriminate attack.

The second group concerns AI hardware and consumer technology. Google Opens Preorders for Its $899 Gemini-Enhanced 'Googlebook' Laptops, from Slashdot, reports that Google has opened preorders for its Googlebook laptops at $899. Buyers can choose between Intel Core Ultra Series 3 and Snapdragon X Elite processors, with NPUs to power AI features. Devices will hit shelves on October 4 in the U.S. and October 5 in Canada, the U.K., Ireland, France, Germany, and Australia. The laptops prominently feature Gemini capabilities, including an AI-powered cursor, vibe-coded widgets, and an AI-enhanced dictation feature called Rambler that cleans up messy brain dumps into readable text. Google is trying to reinvent the point-and-click experience so that the cursor serves as a conduit to Gemini. The company is targeting the K-12 market, where around 50 million Chromebooks are in use. Flagship devices are built by Acer, ASUS, Dell, HP, and Lenovo, with up to 14 hours of battery life and an exterior Glowbar that dances on boot and shows battery level. The Verge's reviewer found the hardware polished and particularly praised the Android phone integration, which lets users cast phone apps to the laptop and use them with mouse and keyboard, and a Continue On feature for handing off apps. GeForce Now is coming to Googlebooks with the first year free. The review concludes that Googlebooks have a higher ceiling than Chromebooks but that Google has a lot to prove at higher prices.

Qualcomm launches two new smartphone chips with emphasis on AI, from TechCrunch, reports that Qualcomm's new top chip can run a 30B mixture-of-expert model locally.

Apple could take on Whoop with a new fitness tracker, report says, from TechCrunch, states that Apple may be developing a new fitness tracker as part of its new generation of hardware devices.

Singapore’s Nexstrom wants to bring 2D semiconductors to chip fabs, from TechCrunch, reports that the Singapore-based startup has raised new funding to develop equipment that could help chipmakers manufacture 2D semiconductor materials at scale.

NVIDIA Announces DSX Ready Qualification Program for Data Center Power and Cooling Hardware, from ServeTheHome, reports that NVIDIA has launched DSX Ready, a qualification program for data center cooling and power hardware for use with constructing AI data centers based on NVIDIA's DSX AI factory blueprint.

Waymo’s latest expansion strategy: teenagers, from TechCrunch, reports that Waymo is opening its robotaxi service to teenagers aged 13 to 17 in Nashville, its second city to offer rides to minors.

AstroForge is putting AI in command of its next spacecraft, from TechCrunch, reports that Autonomy-1 will have a small, transformer-based AI model taking charge of a space probe.

The third group covers cybersecurity and privacy. Two articles from 404 Media and TechCrunch cover the same breach. ‘We Hacked the FBI:’ Hackers Say They Have Data on All FBI Employees, from 404 Media, reports that the hacking group ShinyHunters claims it breached multiple FBI-related services and stole data on all FBI employees and applicants, including names, home addresses, phone numbers, and information on spouses. The representative provided a sample of 5,000 employees, and some phone numbers were verified via OSINT. ShinyHunters also defaced the FBI jobs website with a message mimicking FBI seizure notices. The group said it used a zero-day exploit in Oracle PeopleSoft to access AWS GovCloud servers and downloaded two to three terabytes of data. The representative said the group is not financially motivated and described its plans as coercion, not extortion. Hacking group ShinyHunters claims it breached the FBI, stole agents’ and applicants’ data, from TechCrunch, adds that the theft could present a major counterintelligence threat, with agents and their families extorted into cooperating with a foreign government.

Stolen passwords are exposing America’s water providers to hackers, from TechCrunch, reports that researchers say another looming threat hangs over some of America's most important critical infrastructure.

Two arrested in UK after Microsoft takedown of ‘Eviltokens’ AI-chatbot for cybercriminals, from The Record, reports that EvilTokens was available on Telegram for a $1,500 initiation fee and a recurring monthly $500 subscription. It provided cybercriminals with AI tools to compromise accounts, analyze breached inboxes, and find the best methods for monetizing access through fraud.

Canadian regulator opens probe of IDScan for allegedly violating data privacy laws, from The Record, reports that the investigation will probe IDScan's security practices and whether victim notifications were adequate under Canada's federal private-sector privacy law.

The fourth group concerns government, law, and press freedom. Note on Trump and the First Amendment, from Racket News, follows an earlier article criticizing the president. The note explains that while no one has a right to a press pass, a president cannot deny a pass for speech reasons, citing Perry v. Sindermann. The author says the issue does not rate high on his list of First Amendment threats but is clearly forbidden, and that presidents who wish to show they cherish the free press should mess with the press legally rather than deny credentials.

Another Judge Sees Through Trump's Anti-Press Agenda, from The Dissenter, reports on a ruling by Judge Frank Saylor of the U.S. District Court of Massachusetts. The DHS had adopted a rule capping I visas for foreign press representatives at 240 days. The NewsGuild-CWA and other unions sued, arguing that forcing frequent reapplication created serious risks of retaliation for unfavorable reporting. Judge Saylor ruled the action arbitrary and capricious under the Administrative Procedure Act and blocked the rule while the lawsuit proceeds. He noted that international journalists contribute to the open exchange of information and that the rule directly threatens those benefits. The judge also observed that the extension process calls for unguided exercise of discretion with no recourse or appeal, making abuse rampant, and that the DHS provided only a 32-day comment period and dismissed strong opposition with conclusory statements. Adam Rose of the Freedom of the Press Foundation called the ruling a victory for the United States, and the article notes that the crackdown on immigrant journalists has already led to detentions and deportations.

The fifth group gathers business, startup, energy, and data center items. Meta deal will help plastics recycling startup MacroCycle build its first factory, from TechCrunch, reports that Meta is giving plastic recycling a boost in a new deal with the startup, which is building its first commercial plant.

Bill Gates’ Breakthrough Energy bets these 21 startups will shape energy’s future, from TechCrunch, reports that the investors at Breakthrough Energy believe the coming wave of electrification will make data center energy demands look quaint by comparison.

Everyone can find a reason to dislike data center construction, from TechCrunch, reports on two years of fraught AI data center debates in Pennsylvania.

Nscale’s IPO will test Wall Street’s appetite for concentrated AI bets once again, from TechCrunch, reports that the British AI data center developer depends on tech giants Microsoft and Anthropic for most of its revenue.

A16z is challenging Silicon Valley’s love for drop-outs by launching a school, from TechCrunch, reports that the Horowitz Andreessen Academy is a $35 million investment in an unaccredited San Francisco-based two-year alternative to traditional college for students ages 16 to 22. It will be led by Gagan Biyani, co-founder of Udemy.

Tuesday assorted links, from Marginal REVOLUTION, is a collection of links covering designing a FINRA for frontier AI, a post by Andrew Ng, skeptical notes on stablecoins based on economic history, global views on fairness and redistribution, essays on whether AI will make management more valuable, commentary on the California billionaire wealth tax, and the Horowitz Andreessen Academy.

Several TechCrunch articles promote the upcoming TechCrunch Disrupt 2026 conference. 4 days to save up to $200: Reason 2 of 5 to be at TechCrunch Disrupt 2026 notes savings on passes. Exhibit tables added: One last chance to showcase your startup at TechCrunch Disrupt 2026 announces a reopened exhibitor program through September 30. TechCrunch Disrupt 2026: Aaron Edsinger brings Hello Robot’s Stretch 4 to life onstage says the CEO will demo the robot on the Real World AI Stage. Five AI safety sessions every founder should have on their TechCrunch Disrupt 2026 agenda mentions sessions with leaders from Anthropic, NVIDIA, AWS, and Waabi. Noble Carbon will show how it’s making EV charger installs painless at TechCrunch Disrupt describes a smart circuit breaker that allows households to electrify without a main panel upgrade.

TikTok’s US arm joins Lantern, a cross-platform child safety initiative, from TechCrunch, reports that eight months after establishing a U.S.-based joint venture, TikTok is getting on board with fellow platforms to support industry-wide child safety work.

The sixth group covers science, health, nature, and one economics paper. Toadstools and toxins, from Aeon, is a teaser by Eric Leas about finding Amanita muscaria being sold in candy wrappers and uncovering legal loopholes and dangerous dosages.

Singing Lemurs Borrow a Trick From Opera Singers, from Nautilus, states that the lemurs' songs are just as haunting.

Loss of Bone Mineral Density Linked to Cognitive Decline, from Nautilus, notes that the finding came with an assist from AI.

Peptides Are Spreading on Social Media. Here’s How Platforms Can Better Protect Public Health, from Nautilus, suggests that platforms need to take the same approach they did to vaping.

The Mysterious Migrant Potters of Ancient Tampa Bay, from Nautilus, describes another archaeological cold case solved.

The economics of late grandmotherhood, from Marginal REVOLUTION, quotes a new NBER working paper by Dora Costa and Matt Kahn. It shows how fertility delays compound across generations to raise the age at first grandmotherhood. Under the 1960 first-birth schedule median grandmotherhood came at 44; under the 2024 schedule it is 57, cutting the healthy years with grandchildren from 28 to 15. The delay is most pronounced among the college-educated. When a daughter times her first birth to maximize her own payoff, the loss her choice imposes on her mother is external to her calculation, which the authors call the Rotten Adult Kid Problem.

The seventh group includes culture and entertainment. 'Coyote vs. Acme' Outgrosses All But 3 WB Movies This Year, Heads To Profit, from Slashdot, reports that the film earned more than Warner Bros.' $30 million tax write-off in two weeks, passed its $70 million production budget, and has now passed the $100 million mark, with a break-even number of $120 million and impending VOD and streaming releases. It has outgrossed five of the eight Warner Bros. movies released domestically this year, including The End of Oak Street. The marketing budget was only $10 million, and the film used oddball stunts including an AMA on Reddit and a Crayola "Coyote Secret Plan" web page with a video from stars Lana Condor and Will Forte on the importance of creativity.

The final group is personal development. Invitation to a KnotChat, from Seth's Blog, describes a simple website called TheKnot.chat that facilitates one-on-one, private conversations to help people get unstuck. The author references his book The Knot and says he will post some KnotChats he has hosted. KnotChat: Rodrigo Delavina Simon, also from Seth's Blog, shares a conversation with Rodrigo about his restaurant in Mexico, focused on the question of whether the problem being solved is one's own or someone else's.

The window is dominated by AI, but it also reflects a moment of contradiction: models break historical ciphers and launch commercial products, yet the same technology is implicated in a deadly military strike, fails simple list comparisons, and is sometimes not AI at all but a call center worker. Cybersecurity breaches, legal rulings on press freedom, and a wide range of business announcements complete the day's record.

Your signing session ended