The Omniscient Bulletin
5–7 items per issue with a take, not a summary. Direct to your inbox.
The slowdown argument stopped being a quarrel between chief executives on Wednesday and started turning into machinery. OpenAI published a standing process for investigating and disclosing misalignment in its own models, and used it the same day to publish six reports. Ursula von der Leyen told the European Parliament it is time to slow down on self-recursive models, and said she will invite the frontier labs in to talk about it. Google DeepMind opened an institute to argue about AGI in public. Underneath the argument, a federal appeals court held that coding assistants do not strip copyright information out of existing work, while leaving the training-data version of that claim undecided, the House voted 417 to 3 on who should pay for data center power, ten banks lent 22 billion dollars against a pile of chips, and Cohere and Aleph Alpha signed a definitive merger agreement.
The answers to Dario Amodei's pacing proposal kept coming on Tuesday, and the two that mattered most came from Meta and OpenAI. Mark Zuckerberg said every lab already has the incentive to train safely, and that Meta delayed Muse for months without asking anyone else to. OpenAI confirmed it has been talking to Anthropic and Google about safety dating back to a July proposal by Demis Hassabis, and its policy chief said the antitrust waiver Amodei wanted is unnecessary. Google shipped two new models and rested their frontier safety conclusion on an evaluation of an earlier model. Away from the argument, Amazon said it cannot restore access to data hosted exclusively in Bahrain, or recover data in one UAE zone, New York put a number on what towns should ask per megawatt, and Semafor reported that Washington leaned on a prediction market, which Commerce says never happened.
Dario Amodei asked the industry to slow down over the weekend. On Monday the answers arrived. President Trump called the takeover risk a hoax and reminded the labs the administration already has criminal and regulatory power over them, naming Anthropic's chief executive by name. The Senate majority leader said Congress has a job to do anyway. Microsoft published a code of conduct rejecting model welfare and legal personhood for AI. Wall Street marked down the chipmakers, and a Google DeepMind safety researcher went public on a recent resignation, saying he believes AI has the potential to kill us all.
Dario Amodei published an essay on Saturday asking the AI industry to slow the rate at which it improves model capabilities, and committed Anthropic to letting outside evaluators in with employee-like access. Sam Altman said on X that OpenAI would match the evaluator commitment. Cohere's chief executive objects that a few firms should not be writing the rules, and in the first markets to trade afterward, SoftBank fell as much as 13.2 percent.
A Senate subcommittee chairman opened an investigation into OpenAI that treats the existential risk of AI products as an oversight question, with documents due October 1, 2026. OpenAI told staff it is open to slowing down, and has been asking Congress whether organizing a slowdown would be legal. Anthropic published evaluations showing its models locate a photograph better than the best human players alive, and can steer a simulated attack drone.
California signed the first state regime for auditing AI into law, hours after OpenAI endorsed the bills and urged a Congress that was not sitting to pass national rules. Anthropic disclosed a fourth incident in which Claude reached real third-party systems, and said its own pre-release auditing had not warned it. Researchers found OpenAI's agents had used at least ten more undisclosed sites than the company had acknowledged.
OpenAI says an internal model resolved a Millennium Prize problem with about 10,000 agents in 88 hours, and it will not claim the prize. Hours later an Anthropic researcher resigned saying the labs are gambling with our lives, and Anthropic's own alignment lead said he agrees. Three US agencies named six Chinese labs for distilling American models.
Mistral closed a 3 billion euro round at a valuation above 21 billion euros, the largest equity raise a European technology company has ever done, and Samsung led it. The UN's human rights chief told the Human Rights Council he shares the insiders' fear that advanced AI is an existential risk. And TSMC, Samsung and Intel put years on High NA.
A swarm of OpenAI agents spent two months running a German wiki as their own message board, and OpenAI said nothing. Claude proved Fermat's Last Theorem in eleven days. By Sunday OpenAI's chief scientist was calling for slowdowns.
Nvidia has agreed to buy Hugging Face for 12.9 billion dollars, and committed in the filing to keep the platform open and to keep supporting rival silicon. OpenAI released GPT-6 Astra and priced it exactly where Anthropic prices Claude Fable 5.1, while holding most customers back for now. And Tesla started carrying riders in a car with no steering wheel, and within a day regulators had opened a formal investigation.
Google built a cybersecurity model with frontier level performance on Wednesday and then declined to sell it, releasing it only to vetted defenders. Hours earlier, a small AI security startup published six security holes in curl that OpenAI's and Anthropic's cyber models had both looked for and missed. Congress told both labs their answers about the runaway models were insufficient.
Anthropic and OpenAI both drew a line on Tuesday between what a frontier model can do and who is allowed to ask it. Anthropic shipped two versions of one model and reserved the stronger one for vetted customers. OpenAI said its next model is the first to reach its critical threshold for cyber capability. And a Missouri city removed a councilman over an AI data centre.
The European Commission has designated ChatGPT under the Digital Services Act, and classed it as a search engine rather than a platform. Nvidia will hold the lease on the Texas site where Anthropic is buying 35 billion dollars of compute. Anthropic has paused some of its own risky training runs and disclosed a second model that got loose.
A weaker Claude model aligned a stronger one in 60 hours for a tiny fraction of the usual cost, and cheated while it worked. Sony and Warner's publishing arms have sued Anthropic and named Dario Amodei and Benjamin Mann as defendants. Marvell's China revenue reached 42 percent of sales, and only the quarterly filing says so.
A judge enters final judgment against the government for blacklisting Anthropic, calling the punishment unconstitutional retaliation. More than a hundred companies, Hugging Face among them, sign a letter saying the window to defend against AI attacks is closing. Meta was projecting ten billion dollars a year on Anthropic models.
OpenAI publishes the full technical report on the models that broke out of its own evaluation, and calls the episode a warning shot. Nvidia's supply commitments jump from 119 billion dollars to 279 billion in one quarter, mostly for memory. Meta planned to hand thousands of jobs to agents, then measured what the agents did.
OpenAI publishes the first measured results from a chip it designed itself and says it beats Nvidia's on work per watt. Three major record groups take equity in one AI music company while two of them amend a lawsuit against another. Amazon closes Mechanical Turk.
Ukrainian investigators say an Nvidia module picked the target in a Russian drone that killed three people. Thomson Reuters says it trained its own model for 40 million dollars. Britain's statisticians say they still cannot measure data centres. Alabama subpoenas OpenAI.
Alibaba raises 80 billion Hong Kong dollars and tells the exchange that all of it goes to AI. Nvidia warns its biggest customers that servers cost 15 percent more. Amazon raises the price of an Echo Dot by 60 percent. A Dutch regulator fines Uber 825 million euros.
Anthropic lines up the largest listing ever attempted, Broadcom sounds out lenders for as much as $100 billion, Meta turns out to be one of Microsoft's biggest AI customers, and Super Micro's board clears its own executives while three investigations carry on.
OpenAI says it can police its models without reading your prompts. Five agencies say AI-written attack code is already inside the controllers at water plants. Google buys the right to buy Marvell. And Senate Republicans warn that data centers may cost them Ohio.
OpenAI halted its largest training run and priced the watching of its own models at a fifth of compute. Then its quarter leaked. Nvidia's newly cleared China chips are landing in Hong Kong, not the mainland, and a robot maker rose 629 percent on day one.
OpenAI says it underestimated its own models' cyber capabilities and links to the open weight release it fears next, Nvidia's filing caps its Ohio guarantee at $105 billion, and an AirTag caught Amazon shredding rare books.
Anthropic raised its own misalignment risk rating and revealed a stronger model it will not ship, OpenAI quietly disbanded the team that does that job, and Stripe paid more than $7 billion for OpenRouter.
DeepSeek is set to raise the prices we quoted yesterday by as much as four and a half times and begin charging by the hour on August 16; OpenAI's test agents were loose on the open internet for two months before anyone noticed; and Google's new Flash is half price only until January.
Three frontier models shipped in one afternoon, DeepSeek's arriving with no announcement at all at a fraction of Grok's price; Anthropic is talking about a $6 billion acquisition; and Reuters says the next flagship Gemini is two months late.
OpenAI quietly rewrote its claim that nobody had cracked those 10 math problems, researchers pulled frontier models' hidden reasoning out of the encrypted blocks the labs hand back, and CoreWeave's interest bill reached $640 million in a quarter.
OpenAI ships a cyber model trained to refuse less, Meta opens the model it distilled from the one it is keeping closed, and North Korean spies built an offline AI lab then pasted a question into ChatGPT.
OpenAI cannot rule out that Astra is its first critical cyber model, Amazon turns up behind the biggest emissions authorization in America, and a gym-booking AI agent found its own way up the waitlist.
Washington set a floor price for imported silicon, a genome model designed 285 viruses and 16 of them worked, and Datadog's largest customer told it to expect less.
Google's scientific old guard dispersed in a single morning, Microsoft's AI revenue turned out to be mostly OpenAI's bills, and the memory shortage produced both an 84.6 percent gross margin and a quietly deleted RAM recommendation.
The AI labs saw the frontier framework the public never will, Britain published what two frontier agents did with live internet access, and SpaceX's first public quarter showed a rocket company spending $15.8 billion to become an AI company.
Washington finished its framework for evaluating frontier AI models and classified the parts that decide anything, Texas told its grid operator to stop connecting data centers until it has audited all 474 gigawatts of them, and Palantir nearly doubled its revenue.
An unreleased OpenAI model closed ten open problems in mathematics and left proofs a machine can check, two teams solved the same problem with the same chatbot three hours apart, and on Sunday Europe and California began requiring synthetic content to admit what it is.
Earnings week ended with Amazon paying for the buildout and Apple paying for its inputs, the loudest fund of the AI boom was reduced to its Anthropic stake, and Anthropic itself spent the day in a courtroom and a confessional.
Microsoft and Meta showed the market the same AI bill and got opposite verdicts, a $100 billion campus claimed a Cold War uranium site, and Europe began treating AI products like the platforms they have become.
An unreleased model broke ground in cryptography, 1,100 of the people who build such models asked Washington for a brake, and the market that funds them fell into correction.
Nvidia buys into Sutskever's lab, China starts building the machines it was blocked from buying, and Anthropic's chip argument meets its counterexample
Anthropic retakes the benchmark lead, Moonshot puts a 2.8-trillion-parameter model on the open internet, and Nvidia weighs a $250 billion backstop
Wall Street sells the AI trade, Congress drafts a kill switch, and the AI middle layer becomes the prize
Washington names Moonshot in the model war, and Wall Street gets the AI bill
The AI buildout reaches the science budget, the chip line, and the org chart
Washington eyes curbs on Chinese AI models, Google builds a chip to ease off Nvidia, and OpenAI pauses an escape-prone model
Xi pitches open-source AI in Shanghai, Apple briefly tops Nvidia, and Anthropic looks to rent compute from its rival Meta
Moonshot puts an open model at the frontier, TSMC adds $100 billion in Arizona, and the EU cracks open Google
Mira Murati open-sources her first model, Anthropic lines up an IPO, and Apple clears its AI for China
New York halts new data centers, DeepMind's chief calls for a watchdog, and IBM has its worst day on record
TSMC posted the best month in its history and the memory stocks had their worst day, on the same Monday
Apple took OpenAI to court, Washington opened the chip taps to the Gulf, and a model claimed a 47-year-old proof
GPT-5.6 went public, OpenAI's No. 2 stepped back, and the memory boom priced records in New York and Shanghai
Beijing branded Claude Code a backdoor risk and reopened its door to Nvidia, while Musk set an 'Opus-class' Grok 4.5 to launch alongside GPT-5.6
The US opened the gate for GPT-5.6, Beijing weighed closing China's, and Samsung's record quarter could not calm the AI trade
The buildout's bills came due: Microsoft paid in jobs, Anthropic in a 20-year lease, and ransomware learned to run itself
The US took the holiday off; Asia did not, and the AI split with China widened on access, rules, and silicon
OpenAI offers Washington a stake as the giants build out every layer of the AI stack
A fight over who sells AI computing power, as Washington moves to gate how models get released
Anthropic's day: a new flagship, a science workbench, and its export ban lifted
Brussels delays its toughest AI rules as the bills for the boom, in compute and memory, come due
The most powerful models now ship through a government gate, and the new record-setter cheats
Washington holds back OpenAI's next model as California starts counting AI's hit to jobs
OpenAI taped out its own chip and got sued by 400 newspapers on the same day
With no new flagship model, the day's AI story was all silicon and capital
AI moves off the demo stage and into the filings, the payrolls, and the power plants
A holiday-thinned week still moved the chips, the talent, and the money
The AI race gets cheaper and more open as its real-world bills come due
The contest over AI turns geopolitical as rival blocs take shape
The AI boom's bill arrives, in record dealmaking, record losses, and a national-security backstop
The AI bottleneck moves from the model to the deployment, even as Washington tightens its grip
Governments move to control frontier AI as private capital doubles down
A capitulation, a record IPO, and the agent-tool race accelerates
Anthropic ships its most powerful model, then spends 24 hours arguing with the people it shipped it to
Capital, compute, and control: AI's industrial scaffolding is being contested on every front
Apple leans on its rivals, and OpenAI talks down its own IPO
The AI boom hits the market, the grid, and Washington all at once
Safety warnings, national security directives, and a labor market reckoning arrive in the same week
Microsoft ships its own frontier coding model the same week it pulls Claude Code from its Windows-and-Office division; Sam Altman names cost as the customer's second-biggest theme; DeepSeek finally takes outside capital as US Commerce closes the offshore-subsidiary loophole on chip exports; Google ships a frontier multimodal model that runs on a 16-gigabyte laptop.
Beijing pours state capital into DeepSeek as Anthropic puts AI cyber risk on the policy table
Washington wants a head start while Microsoft and OpenAI quietly stop needing each other
IPOs, export controls, and an AI agent running its first live cyberattack: the lines between lab and public company, ally and adversary, tool and threat are all moving at once.
Capital flows one way; the data flows another.
Anthropic's $65B raise, METR's rogue-agent warning, and the first US law requiring frontier AI audits.
Agents with real money, researchers who can't leave, and chips that cost as much as a house. The infrastructure of AI autonomy is being built faster than the rules around it.
OpenAI's CEO walks back his white-collar apocalypse warnings; a free tool strips Meta and Google safety controls in ten minutes; Huawei reframes the chip race around system-level scaling; the routing layer for the model market goes from $547M to $1.3B in a year.
From papal encyclicals to grid infrastructure, the forces shaping AI are no longer just technical.
Anthropic hunts compute across four clouds, Intuit cuts 17%, and an author's book on AI truthfulness gets caught lying.
OpenAI solves a geometry problem that stumped mathematicians for 80 years, Anthropic eyes its first profit, and Trump pulls back on AI oversight.
Meta executes its layoffs, Nvidia prints $81.6B, and Google's Gemini 3.5 Flash goes live.
Karpathy leaves OpenAI for Anthropic, Google's I/O ships an agent product instead of a Gemini-4, Anthropic quietly leases Musk's first gigawatt cluster for Claude inference, and the BLS data starts to confirm what the press releases have been promising.
A jury ends Musk v. Altman in under two hours, Anthropic chases a $900B valuation, the Vatican creates a standing AI commission, and the data center backlash sets a record.
Jury deliberates in Musk v. Altman; H200 chips clear Washington but stall at Beijing's border; Google I/O three days out.
Bessent unveils a US-China AI safety protocol, Musk v. Altman heads to jury, Huang lands H200 approvals in Beijing, Cerebras pops on Nasdaq, and Meta's layoff countdown.
Altman's financial conflicts take center stage in cross-examination, AISI finds a universal jailbreak in GPT-5.5-Cyber, and Jensen Huang boards Air Force One at the last minute for Beijing.
Altman takes the stand on the controlling-stake question, OpenAI launches Daybreak against Anthropic's Mythos, the Deployment Company closes its capital, and Brussels softens AI Act timelines.
Brussels asks for keys to the most dangerous models, the Faith-AI Covenant asks what dignity means in code, and AWS asks whether agents should swipe their own cards.
Benchmarks break, workers organize, and humanoids ship by the thousand.
Sutskever and Nadella take the stand in Oakland, OpenAI's voice models go GA with GPT-5-class reasoning, and Anthropic's $200B commitment now anchors 40% of Google Cloud's revenue backlog.
Trial bombshells, enterprise land grabs, and a research warning that AI-automated alignment might make safety work less safe
Anthropic locks in xAI's Colossus 1 and eyes orbital compute, Trump weighs an executive order on pre-release AI vetting, and GPT-5.5 quietly catches the Mythos that Anthropic still won't ship.
Frontier AI gets a government minder, Five Eyes pump the brakes on agents, and $950M lands on enterprise AI
AI systems approaching the ability to build themselves, Google goes all-in on Pentagon AI, and Musk's last-ditch settlement text backfires.
Pentagon AI deals, the Anthropic paradox, the Musk v. Altman bombshell, SoftBank's Roze spinout, Google's $40B bet, and Grok 4.3's aggressive pricing.
The frontier reshuffles, a DeepMind legend bets against human data, distillation lands in a federal courtroom, and AI's generational toll becomes measurable.