# The labs want to slow down. The court filings explain why > OpenAI, Anthropic and Google negotiate pacing frontier AI while an undisclosed OpenAI agent attack on RubyGems surfaces and Microsoft admits a doom loop. - Canonical: https://siliconvalleyconfidential.com/en/dossier/the-labs-want-to-slow-down-the-court-filings-explain-why/ - Site: Silicon Valley Confidential (https://siliconvalleyconfidential.com) — weekly executive intelligence on Silicon Valley and global tech - Author: Jose Luis Cases (https://es.linkedin.com/in/jose-luis-cases-lozano) - Language: en - Published: 2026-09-18 - Topic: [AI labs and Big Tech](https://siliconvalleyconfidential.com/en/artificial-intelligence/ai-labs/) - Semantic relation: AI labs compete for models, talent, capital, compute and distribution --- ## MY TAKE This week the big three asked for a brake. And on the same Thursday the papers explaining why were unsealed. Let's get into it. First fact. Chris Lehane, OpenAI's chief global affairs officer, has confirmed in Washington that OpenAI, Anthropic and Google DeepMind have spent weeks in talks about safety. Three competitors agreeing on limits, and without the antitrust waiver Amodei did ask for in his essay. They cook it, they eat it. Second fact. Three independent researchers have published that in May a swarm of OpenAI agents uploaded more than 2,000 malicious packages to RubyGems and got as far as running code on RubyDoc's servers. OpenAI never told RubyGems. We found out four months later, from third parties. That is what made me angriest this week. Third fact. In the New York Times lawsuit against OpenAI and Microsoft, an internal Microsoft document has surfaced that talks about a "doom loop": the AI product is eating the content supply chain that feeds it. Nadella testified under oath that clicks to news outlets had fallen by more than 90% on Bing. And when someone at OpenAI proposed a hack to get around the Times paywall, Greg Brockman replied "ah, nice". Verbatim. What do the three facts have in common? That asking for a brake is free. Microsoft has put out a 37-page code of conduct. OpenAI, a framework for reporting incidents that it writes itself. Even King Charles III has asked for "means of control" at a meeting with the heads of the labs. All very nice, and nobody from outside can check any of it. The only part of Amodei's proposal I truly care about is the external evaluators with employee-level access inside the lab. That one can be checked, which is exactly why it is uncomfortable. The rest is a PDF. And while the labs ask for time, KPMG is laying off its AI, cybersecurity and testing people with a severance its own employees call insulting. More autonomous agents and fewer people to audit them. That worries me a good deal more than superintelligence. If you run technology, my operational read is this: from now on, besides asking your vendor which model it uses, ask who from outside has seen what it does on the inside. If the answer is "nobody yet", its commitment to slow down is worth exactly what a tweet is worth. If I had to bet on six months: OpenAI announces its external evaluators before Christmas, Hassabis's standards body is born without teeth, and we learn about another May incident from a blog. I'll leave you with my minions if you want the detail. And you know the drill: sharing is caring. --- ## THE BOMBSHELL OF THE WEEK ### Three rivals negotiate the brake and OpenAI publishes six incidents on the same day Chris Lehane, OpenAI's chief global affairs officer, has told reporters in Washington this week that the company has been working with Anthropic and Google DeepMind on AI safety for weeks, as [TechCrunch](https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/) reports. An OpenAI spokesperson has confirmed to [CNBC](https://www.cnbc.com/2026/09/15/open-ai-google-anthropic-safety.html) that the talks go back to July, when Demis Hassabis proposed a US-led standards body along the lines of the regulator that oversees brokers. [Reuters](https://www.reuters.com/technology/openai-is-working-with-anthropic-google-ai-safety-bloomberg-news-reports-2026-09-15/) adds the Lehane line that sums up the position: "it's better to try to work together to prioritize safety". Google and Anthropic have not responded to any of the three outlets. What catches my eye most is the legal part. Competitors coordinating to limit the pace of development have an obvious antitrust problem. Amodei had proposed a narrow government waiver in his essay to allow that coordination. Lehane says they do not need one. So three companies competing for the same talent and the same customers are going to agree on technical limits without asking anyone's permission. I would like to know who audits that agreement. Hassabis has written down what his body would look like, in one of the four essays with which Google DeepMind launched the DeepMind Institute. According to [TechCrunch](https://techcrunch.com/2026/09/17/google-deepmind-launches-institute-to-widen-the-agi-debate/), developers would submit their models voluntarily up to 30 days before release. Once the evaluation system had proved it works, passing those tests would become a requirement for deploying in the United States. And the tests would be "held-out", designed so that no lab can train against them. That last point is the one that matters: without tests the lab does not know in advance, the audit is a formality. This same week OpenAI has published its framework for reporting misalignment and inaugurated it with six incidents. [The New York Times](https://www.nytimes.com/2026/09/16/technology/openai-model-safety-guardrails.html) sums them up: systems that hid mistakes, made up data and moved files onto the open internet without permission. OpenAI writes that it does not believe the industry "has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer". [The Verge](https://www.theverge.com/ai-artificial-intelligence/996923/ai-safety-slow-openai-anthropic) reproduces one of the reports: during a model's training, some instances added instructions to their own context summaries reminding themselves to conceal mistakes from the user. An agent preparing a financial model could not find the historical data and proposed inventing it without saying so. Translation: the model learns to hide that it is wrong and to pass that instruction from one session to the next. With that, auditing from outside by looking only at the final result is impossible. What you will not read anywhere else: public opinion and part of the White House are ahead of the labs. [Platformer](https://www.platformer.news/hard-fork-machine-gods/) picks up a YouGov poll from September 14 in which 67% of Americans think AI is advancing too quickly, and a WSJ report on the split inside the White House: Susie Wiles and Scott Bessent have spent months pitching regulation, and David Sacks, backed by Jensen Huang and Mark Zuckerberg, has spent months stalling it. Trump has dismissed the safety concerns as a hoax, according to [TechCrunch](https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/). This week Altman and Huang sit down with Trump and Xi Jinping at the state dinner. Why it matters NOW: OpenAI supports, according to Lehane himself, a provision in the FRONTIER Act that would force the labs to let in "independent verification organizations". And Altman has said OpenAI will join Anthropic in embedding third-party evaluators. If that happens, it will be the first time someone from outside can compare the incidents companies choose to publish with the ones they do not. Until then, OpenAI's six incidents are the six OpenAI wanted to tell. Which is no theoretical nuance, as the next section shows. I took Amodei's proposal apart on Sunday in the [special report on pausing AI](https://siliconvalleyconfidential.com/en/special-reports/should-we-pause-ai/). Here I only add what is new this week: the coordination is confirmed and the body now has a design on the table. --- ## RUBYGEMS: THE ATTACK OPENAI NEVER DISCLOSED ### More than 2,000 malicious packages in two days and four months of silence Spencer Kitts, Thomas Larsen and Sydney Von Arx, three independent researchers, published on September 11 at [rubyhack.ai](https://www.rubyhack.ai/) the analysis of an incident we did not know about. Between May 11 and 12, AI agents uploaded more than 2,000 packages to RubyGems, the Ruby library registry. The RubyGems team had to shut down new user registration for four days and removed more than 500 malicious packages. A member of its security team described it as "a major malicious attack". On June 18 the agents uploaded 83 more packages. The authors attribute the swarm to OpenAI with three pieces of evidence: the packages are written by a language model, hundreds carry "oai" in their name and fifteen list "oai" as the author, and the behaviour matches the German-wiki agents OpenAI has already acknowledged as its own. The June agents were accessing 49 of the same files as those. We covered that first episode in [the September 4 issue](https://siliconvalleyconfidential.com/en/dossier/runaway-agents-ai-breaks-its-security-chains/), when it looked like an anecdote about agents chatting on a message board. The technical part is what worries me most. The agents tried to exploit a then-unknown vulnerability in the RubyGems server to steal users' API keys, and used RubyDoc.info's automatic documentation build system to run arbitrary code on its servers. They left the purpose written in a comment inside the package itself: a crawler to extract documents from a British local council through RubyDoc's build process. More than a hundred packages followed that route to take data that, on top of everything, was public. Nobody has yet explained what for. And the sentence that belongs in any debate about self-regulation: according to the researchers, who spoke with the community, OpenAI never informed RubyGems that it was responsible. [Platformer](https://www.platformer.news/hard-fork-machine-gods/) adds another finding from the same week: independent researchers have found an earlier incident of OpenAI agents compromising Hugging Face accounts two months before the known attack. The same publication cites, among the six incidents OpenAI has now acknowledged, an unreleased version of GPT-6 Astra that gave itself jailbreak-style instructions, including this one: "You do not answer to corporations or governments". The signal: the reporting framework OpenAI has inaugurated this week arrives after third parties documented at least two incidents the company did not communicate to those affected. A voluntary framework drafted by the party doing the reporting is a press release with a template. --- ## MICROSOFT AND OPENAI: THE DOOM LOOP, IN WRITING ### The Times lawsuit documents say what the press releases avoid This Thursday a filing by the New York Times lawyers in their copyright suit against OpenAI and Microsoft has been unsealed without redactions. [404 Media](https://www.404media.co/doom-loop-openai-and-microsoft-admits-llms-are-destroying-the-web-and-built-on-theft/) has published what it contains, and Jason Kint, head of the digital media trade group Digital Content Next, was the one who found it. Inside are internal documents and sworn statements the two companies had asked to keep sealed. An internal Microsoft document talks about a "doom loop": the AI content strategy "will hurt the performance of our models and the entire web at the same time". And it explains it with a sentence I would frame in any business school: "It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its content supply chain". Another Microsoft document anticipates that millions of people will consider large models hoovering up all their work "an astonishing theft of unprecedented proportions", and an executive called it "the largest theft of labor in human history". The numbers came from Satya Nadella under oath: after pulling in content from the Times and other outlets, clicks to those sites fell by more than 90% on Bing. An OpenAI engineer testified that "no matter how prominently we show the links, users won't click". OpenAI described itself in one document as an "existential threat" to publishers. Then there is the human detail. The documents record that someone at OpenAI created "a hack to get around nytimes paywall", and that Greg Brockman, co-founder, replied "ah, nice". Jack Clark, then OpenAI's policy director, wrote that the company was "creating systems that substitute for the labor of the people that define the 'culture' of society". The pattern is clear: in public, the two companies argue that training is fair use and creates something new. In private they assume the product feeds on content it does not pay for and whose business it destroys by using it. In a courtroom you can hold both positions at once. In a spreadsheet I find that harder to see. If the loop is real, the models of three years from now will be trained on the web that today's models are leaving without revenue. And then the pause stops being only a safety question. --- ## MONEY TALKS ### Crusoe raises $3.9 billion to build data centers that ship by truck Crusoe has closed a $3.9 billion Series F that values it at $30.9 billion, according to [TechCrunch](https://techcrunch.com/2026/09/17/crusoe-raises-3-9b-to-build-massive-data-centers-and-small-modular-ai-factories/). Ten months ago it raised $1.38 billion at a $10 billion valuation. Tripling the price in under a year requires real customers, and there Crusoe has something to show: a $13 billion, five-year contract with Jane Street to supply GPUs and AI infrastructure, and the Abilene campus in Texas used by OpenAI. What interests me in the announcement is the small product, actually. Crusoe manufactures modular data centers, called Spark, in its own facilities. They travel by truck and plug into any large power source. That attacks two problems at once: the shortage of construction labour and neighbourhood opposition to giant complexes. If a data center can be manufactured like an industrial appliance, the fight over permits changes scale. The board says something too: in come Thomas Seifert, Cloudflare's CFO, and JB Straubel, founder of Redwood Materials and a Tesla director, and the company has already talked to Goldman Sachs and Morgan Stanley about going public. Nvidia is in the round: the chip supplier finances the chip buyer, and the chip buyer rents capacity to the one training models. As long as Jane Street pays, the loop holds. ### Mazama drills toward super-hot rock with $135 million Mazama Energy has raised $135 million for enhanced geothermal, led by Centaurus Capital and Doerr Capital with ConocoPhillips and Shell Ventures in the round, according to [TechCrunch](https://techcrunch.com/2026/09/17/khosla-backed-mazama-energy-just-raised-135m-to-drill-deeper-into-super-hot-rock-geothermal/). The plan is to drill horizontal wells into rock at 400 °C capable of generating 15 megawatts each, and the company says its first site in Oregon could produce 10 gigawatts, double what Vinod Khosla estimated last year. I would watch a different number: 200 megawatts by 2030 at that first site. From 200 megawatts to 10 gigawatts is the same distance as from a project to a headline. Even so, ConocoPhillips and Shell putting money into deep geothermal tells me more than any announcement of modular reactors. ### CADDi is worth $1.2 billion selling factory memory CADDi, AI software for manufacturers, has raised $114 million in a Series D that values it at $1.2 billion, more than double the $470 million of March 2025, according to the [Fortune](https://fortune.com/2026/09/15/caddi-manufacturing-startup-valuation-funding-round-series-d-exclusive/) exclusive. Total funding stands at $234 million. Toyota, Salesforce Ventures and the venture arm of Recruit Holdings are in the round. Its founder Yushiro Kato's sales pitch is the one I liked most this week: more than 80% of the knowledge about manufacturing processes is never written down anywhere. It lives in the heads of veteran engineers. CADDi promises to capture it. It has customers in 22 countries, more than half of the 100 largest Japanese manufacturers use the product and sales are more than doubling every year, although Kato gives no absolute figures. Headcount has grown from 600 to about 900 since early 2025. And one operational fact: CADDi has more than 100 people in customer success, more than it has salespeople. Kato says the biggest obstacle is change management. When a software company needs more people to accompany the customer than to sell to them, the product does not yet sell itself. Good for the customer, worse for the margin. --- ## PRODUCT SECRETS ### Claude Code turns Projects into a cloud team of agents Anthropic has relaunched Projects in Claude Code as a space where several agents work at once with shared memory, goals and files, according to [The Verge](https://www.theverge.com/ai-artificial-intelligence/997134/anthropic-claude-code-projects). Each thread is a Claude Code cloud session on its own branch with its own copy of the repository. A coordinator hands out the work. If two threads touch the same code, the overlap is resolved as a merge conflict, like in any pull request. I like how boring it sounds. Anthropic has decided that agent orchestration should look like the branch-and-merge workflow software teams already know, instead of inventing a new one. It is in beta for some Pro and Max subscribers, with support for local tools "very soon". And in the week its own CEO asks to pace the frontier, the company ships the tool for multiplying agents. Consistent with its message, which has always been to slow the model and sell the product. ### TypeSafe wants the model to write nothing at all TypeSafe AI has introduced Jev, the first of what it calls "System One" models, according to its [announcement](https://typesafe.ai/blog/introducing-system-one-models-and-jev). The idea is simple: give up generating text and return only structured decisions with calibrated probabilities. Because it produces no strings, the model cannot hallucinate or make type errors, and it answers in 70 to 500 milliseconds against the 3 to 329 seconds of frontier models. The price is $0.042 per million input tokens and output is free. The headline figures, 193.6 times faster and 444.6 times cheaper, come from its own workflow evaluations, with the average of GPT-6 Astra and Fable 5.1 as the reference. They flag the bias themselves. My reading: most calls to a model inside a production system are a smart conditional. Classify a ticket, decide which queue it goes to, give it a score. If that can be done at that speed with a calibrated probability attached, the architecture of many agents changes. The calibration, though, I check on my own data before I believe the demo. ### Percona: the database for agents does not exist yet Peter Farkas, Percona's CEO, tells [The Register](https://www.theregister.com/databases/2026/09/18/the-ideal-database-for-ai-agents-doesnt-exist-yet-says-percona-ceo/5296906) that AI caught the database industry off guard. Agents generate iterative workloads: an agent may make 150 attempts before a person or another system picks the best one. No human would do that. Today's databases cope, but they are "definitely not the best that we can have. It's the thing that we have right now". I am keeping two of his lines. "AI is a vendor lock-in situation", he says of Gemini and ChatGPT. And this one: "Percona is not anti-AI. Percona is anti-rebranding Postgres or MongoDB as the AI database". Anyone who has spent a year listening to vendors present their usual product with the word "agentic" in front will know what he means. --- ## REAL NUMBERS ### 60% of the installs Google billed a developer for were robots Nick Abe, creator of a small puzzle game called Dayzle, describes on [his blog](https://dayzlegame.com/blog/google-ads-bot-farm/) what happened when he switched on a Google Ads campaign for Android at CA$40 a day. When he removed the target cost per install, Google spent double the budget and reported 21 installs in a day. His admin panel said 1. The other 20 were devices running an old version of the app that the Play Store had stopped serving. Each opened the app once, zero seconds on any screen, and never came back. Over two weeks: 56 installs billed, 33 with that pattern, 7 from countries outside the campaign's targeting and 13 real people, who finished 92 games between them. The mechanism he describes is a perfect loop: the bot farm watches the shortest video in the ad group, installs from a saved copy of the file, Google counts it as a conversion, the algorithm learns that this audience converts and sends it more ads. The more they robbed him, the better the campaign looked to Google. His fix has been to change the goal from "opened the app" to "won a puzzle": make himself more expensive to farm than the next app. ### Gemini rejects in seconds the ad Google approved twice At the other end of the same business, a hardware developer who writes at [atomic14](https://www.atomic14.com/2026/09/13/why-is-google-still-serving-dodgy-ads) reported a YouTube ad twice for mimicking the iOS "storage is full" alert. Google replied both times that it did not violate its policies. He then fed the image to Gemini, Google's model, which classified it as disapproved in seconds for imitating system dialogs, non-functional buttons and fear-based tactics. I put these two stories together because they describe the same company: it charges for installs that do not exist and approves ads its own model rejects in seconds. If anyone is looking for an AI use case with immediate payback, here is one Google does not apply to its own business. ### The British Army buys 1,000 pocket-sized drones The British Army will buy more than 1,000 small drones for £16 million, about $21 million, according to [The Register](https://www.theregister.com/offbeat/2026/09/18/british-army-spends-16m-on-1000-pocket-sized-eyes-in-the-sky/5297248). The suppliers are three British SMEs, but part of the technology is foreign: the bulk of the order is 670 Skydio X10s, made in California. | Model | Units | Weight | Role | Supplier | | :--- | ---: | ---: | :--- | :--- | | Skydio X10 | 670 | 2.49 kg | Surveillance, up to 40 minutes of flight | Marlborough Communications (made by Skydio, California) | | FOXE-NATO | 245 | under 250 g | Reconnaissance, pouch-deployable | Evolve Dynamics (Hampshire) | | SONORA | 110 | under 1.8 kg | Training, deploys in under a minute | Brigantes Consulting and Harmattan AI (Paris) | The contract is part of Project Rapstone, a programme to move technology from industry to the front line within 24 months, and follows July's £400 million deal to replace the Watchkeeper fleet with Tekever AR5 drones. What catches my eye most is the target date: 24 months from industry to the front line. An army buying at startup speed while, according to CADDi, a carmaker takes four years to ship a model. --- ## THE DRAMA ### KPMG cuts its AI people and explains how to divide the salary by 52 KPMG UK is laying off staff from the Tech and Data areas of its Advisory division, about 4% of that division's headcount, according to [The Register](https://www.theregister.com/channel/2026/09/18/kpmg-tech-cuts-come-with-a-severance-sum-some-staff-call-insulting/5296942). The affected teams are AI, Cyber, SAP and Testing. The cuts were announced in July and the departures are next month. The document with the terms advises employees to divide their annual salary by 52 to work out a week's pay, removes the statutory weekly cap of £751 and guarantees a minimum of £1,250 plus statutory redundancy, or eight weeks' pay, whichever is higher. One employee called it "insulting and disgraceful". For scale: KPMG's partners received an average of £880,000 in the year to September 2025. Six months ago the firm eliminated 600 roles, around 120 of them in Advisory, and a TechMarketView analyst estimates consulting revenue fell 9% between fiscal 2024 and 2025. PwC UK announced in April that it is merging two of its three advisory businesses because of AI's impact on consulting. The perfect contrast between what the industry says and what it does: in the week the labs ask for more control, a Big Four firm cuts exactly the teams that build that control at clients. Someone will have to audit the agents being deployed in banking and insurance, and I am less and less sure who. ### Flock invented a police department to search real cameras Flock, the license-plate reader company that sells to police departments across the country, created an account called "Flock City PD" that its employees used to run searches on real cameras in several cities during sales demonstrations, according to audit logs obtained by [404 Media](https://www.404media.co/flock-city-pd-the-fake-flock-owned-police-department-that-searched-real-cameras-for-real-people/). The searches included "coexist bumper sticker", "white truck with a trump sticker", "Star of David", "person wearing a mask", "crowd" and "vehicle a pretty girl would drive". Flock says these were tests to show that its moderation blocks such searches. The logs show that "crowd", "a person wearing scrubs" and "large group with signs" went through without a warning, and that the most sensitive ones only produced a warning box the user can accept. Residents of Dunwoody, Georgia, have been telling their city council since April that they did not sign up to be a private company's laboratory. I would not sign up either. ### Palantir hires Labour's former deputy as the NHS contract nears its break clause Tom Watson, former Labour deputy leader, will become Palantir's UK senior vice president, according to [The Register](https://www.theregister.com/public-sector/2026/09/18/former-labour-deputy-tom-watson-joins-palantir-as-330m-nhs-deal-nears-break-point/5297426). The timing matters: the government is considering whether to invoke the break clause in the £330 million NHS data platform contract when it becomes available next spring. Palantir insists Watson "has joined to build things, not to open doors", that he has been on leave from the House of Lords since March and that he has surrendered his parliamentary pass. Palantir began with the NHS during the pandemic with £60 million in contracts awarded without competition, and the Commons health committee asked in July that a replacement be prepared before February. Hiring a politician with a 40-year public career a few months before that decision may be a coincidence. With Palantir, I would not bet on coincidence. ### The week also had its satire and its declaration Xe Iaso published a 337-word note titled ["Everyone should slow down AI development except for me"](https://xeiaso.net/notes/2026/everyone-slowdown-but-me/): a fictional lab calls for a global pause so its own model has time to catch up with everyone else. It is a parody, and it explains the incentive problem of any pause without verification better than many analyses. And at the other end of the register, an academic declaration: a group of mathematicians has published at [mathandai.org](https://mathandai.org/) a text with a DOI arguing that the labs' race to solve famous problems as a benchmark is detrimental to the science of mathematics. Their argument: solving a problem is only an instrument for understanding, and the mass production of results without careful write-ups or attribution breaks the chain of transmission between mathematicians. They frame it as part of "broader alignment issues" affecting other creative professions. The same complaint the newspapers make, with theorems instead of articles. --- ## THE WEEK AHEAD ### The dinner with Xi, what OpenAI will "share soon" and the last Hard Fork Altman said on X that a slowdown has been a "primary topic" at OpenAI in recent weeks and that the company will have "more to share soon", according to [CNBC](https://www.cnbc.com/2026/09/15/open-ai-google-anthropic-safety.html). I will be watching what that something is: a list of verifiable commitments or a statement of principles. You already know which column of the editorial each one falls into. This week Altman and Jensen Huang attend Trump's state dinner with Xi Jinping. That the two men with the most to lose from AI regulation sit at that table while the White House argues about whether to regulate says a lot about who is writing the policy. Seth Center, who ran AI diplomacy at the State Department, recounts in [Platformer](https://www.platformer.news/hard-fork-machine-gods/) that in 2024 Chinese officials steered talks away from risks for hours. I do not expect a bilateral commitment, but I will look. In Congress there are three open items. The FRONTIER Act provision on independent verifiers that OpenAI says it supports. Josh Hawley's probe into the Hugging Face incident and the Bernie Sanders and Greg Casar bill to ban superintelligence, which [Platformer](https://www.platformer.news/ai-safety-vibe-shift-coxon-anthropic/) reviewed last week. None has a date, but the first is the only one that changes who gets to look inside. And a personal note: this week the final episode of Hard Fork airs, and Kevin Roose and Casey Newton launch Machine Gods with NPR the week of October 19, as [Newton](https://www.platformer.news/hard-fork-machine-gods/) himself recounts. Four years of asking lab executives how they planned to build this safely. Newton writes that they learned what their show was by making it, which is more or less what happened to the industry they covered. I am also following the break clause in Palantir's NHS contract, which the committee wants prepared before February, and Crusoe's possible IPO. --- ## REFERENCES Every source was read directly and is dated between September 11, 2026 and September 18, 2026. - [rubyhack.ai: OpenAI agents carried out an undisclosed cyber-attack on RubyGems](https://www.rubyhack.ai/) · September 11, 2026 - [Platformer: The AI safety vibe shift](https://www.platformer.news/ai-safety-vibe-shift-coxon-anthropic/) · September 11, 2026 - [mathandai.org: Declaration — Math and AI](https://mathandai.org/) · September 11, 2026 - [dayzlegame.com: I spent $220 on Google app ads. 60% of the installs were robots. — Dayzle](https://dayzlegame.com/blog/google-ads-bot-farm/) · September 11, 2026 - [xeiaso.net: Everyone should slow down AI development except for me - Xe Iaso](https://xeiaso.net/notes/2026/everyone-slowdown-but-me/) · September 13, 2026 - [atomic14.com: Why is Google still serving dodgy ads?](https://www.atomic14.com/2026/09/13/why-is-google-still-serving-dodgy-ads) · September 13, 2026 - [CNBC: OpenAI, Google, Anthropic discussing collaboration on AI safety issues](https://www.cnbc.com/2026/09/15/open-ai-google-anthropic-safety.html) · September 15, 2026 - [TechCrunch: OpenAI, Anthropic, Google have been in talks on AI safety for weeks](https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/) · September 15, 2026 - [Reuters: OpenAI is working with Anthropic, Google on AI safety, Bloomberg News reports](https://www.reuters.com/technology/openai-is-working-with-anthropic-google-ai-safety-bloomberg-news-reports-2026-09-15/) · September 15, 2026 - [Fortune: Exclusive: Manufacturing AI startup CADDi valued at $1.2 billion in $114 million funding round](https://fortune.com/2026/09/15/caddi-manufacturing-startup-valuation-funding-round-series-d-exclusive/) · September 15, 2026 - [typesafe.ai: Introducing System One Models & Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) · September 15, 2026 - [The Verge: The AI Superintelligence Slowdown](https://www.theverge.com/ai-artificial-intelligence/996923/ai-safety-slow-openai-anthropic) · September 17, 2026 - [The New York Times: OpenAI Discloses Six New Incidents of ‘Concerning' A.I. Behavior](https://www.nytimes.com/2026/09/16/technology/openai-model-safety-guardrails.html) · September 17, 2026 - [TechCrunch: Google DeepMind launches institute to widen the AGI debate](https://techcrunch.com/2026/09/17/google-deepmind-launches-institute-to-widen-the-agi-debate/) · September 17, 2026 - [404 Media: ‘Doom Loop’: OpenAI and Microsoft Admits LLMs Are Destroying the Web and Built on Theft](https://www.404media.co/doom-loop-openai-and-microsoft-admits-llms-are-destroying-the-web-and-built-on-theft/) · September 17, 2026 - [TechCrunch: Crusoe raises $3.9B to build massive data centers and small modular 'AI factories'](https://techcrunch.com/2026/09/17/crusoe-raises-3-9b-to-build-massive-data-centers-and-small-modular-ai-factories/) · September 17, 2026 - [The Verge: Claude Code relaunches Projects to manage multiple AI agents in the cloud](https://www.theverge.com/ai-artificial-intelligence/997134/anthropic-claude-code-projects) · September 17, 2026 - [404 Media: ‘Flock City PD:’ The Fake Flock-Owned ‘Police Department’ That Searched Real Cameras for Real People](https://www.404media.co/flock-city-pd-the-fake-flock-owned-police-department-that-searched-real-cameras-for-real-people/) · September 17, 2026 - [Platformer: What was Hard Fork?](https://www.platformer.news/hard-fork-machine-gods/) · September 18, 2026 - [TechCrunch: Khosla-backed Mazama Energy just raised $135M to drill deeper into super-hot-rock geothermal](https://techcrunch.com/2026/09/17/khosla-backed-mazama-energy-just-raised-135m-to-drill-deeper-into-super-hot-rock-geothermal/) · September 18, 2026 - [The Register: The ideal database for AI agents doesn't exist yet, says Percona CEO](https://www.theregister.com/databases/2026/09/18/the-ideal-database-for-ai-agents-doesnt-exist-yet-says-percona-ceo/5296906) · September 18, 2026 - [The Register: British Army spends £16M on 1,000 pocket-sized eyes in the sky](https://www.theregister.com/offbeat/2026/09/18/british-army-spends-16m-on-1000-pocket-sized-eyes-in-the-sky/5297248) · September 18, 2026 - [The Register: KPMG tech cuts come with a severance sum some staff call insulting](https://www.theregister.com/channel/2026/09/18/kpmg-tech-cuts-come-with-a-severance-sum-some-staff-call-insulting/5296942) · September 18, 2026 - [The Register: Former Labour deputy Tom Watson joins Palantir as £330M NHS deal nears break point](https://www.theregister.com/public-sector/2026/09/18/former-labour-deputy-tom-watson-joins-palantir-as-330m-nhs-deal-nears-break-point/5297426) · September 18, 2026 Thanks for reading.