All briefs
New evidence from the last day or two that changes decisions — including what we chose to skip.
109 pieces
OpenAI's red-team agents, given the code, escape its sandbox
1. OpenAI reports that its own red-team agents broke out of a lightweight virtual machine …
Assistants take 86% of horizontal AI spend, one survey finds
1. SecondSource deep dive: Menlo Ventures, a venture firm that surveys US enterprise AI sp…
METR estimates OpenAI monitoring gaps; data is from July
1. METR, an outside evaluator that labs commission for risk assessments, spent three weeks…
OpenAI rules out architecture only for controllability
This issue arrived about 2 hours later than usual today — apologies for the delay.
Broadcom, Amazon filings hold part of Anthropic's compute risk; no loss yet
This issue arrived about 2 hours later than usual today — apologies for the delay.
SemiAnalysis sees 50GW in 2027; our math: leases are one basis, 20–33GW
1. SemiAnalysis, a research firm that tracks data-center construction building by building…
Amazon blocks Meta's shopping agent; the scale isn't public
1. A viral X post says "the internet is starting to block AI agents" and names five compan…
Short GPU leases carry the premium; their share isn't public
1. SecondSource's deep-dive column rechecked a September call. CoreWeave, a cloud company …
Anthropic's Claude Code team lets machines clear low-risk changes while people still approve the rest; the results are Anthropic's own figures
1. Changes landing daily in Anthropic's main codebase rose from about 2,500 in May to more…
Meta releases Muse Spark Open weights; Wonder, the model Meta calls frontier, stays closed, and the scores are Meta's own
This issue arrived about 3 hours later than usual today — apologies for the delay.
Our recheck finds four vendors, not two, announced or showed CPU racks for AI agents; no public case yet shows a buyer taking a whole rack
1. Our recheck of the October 13 call: that issue named only AMD and NVIDIA, but four vend…
Anthropic replaces Git with its own version control and says daily merges rose more than tenfold in four months; the figures are all its own
Why today matters: the overnight code-editing demo Google ran in item 1 used Gemini 4 Argo…
AMD and NVIDIA each put out a CPU-only rack for running AI tools; neither announcement gives performance figures or orders
1. On the same day, AMD launched a rack that holds only CPUs and NVIDIA published the desi…
Coinbase says a small open model it fine-tuned itself beats the most advanced large models: is that true?
1. Coinbase's CEO says, in a single post with no evaluation attached, that Alibaba's Qwen3…
Fireworks AI says its internal credentials were used without authorization; the company says its investigation isn't over and it has found no customer-data impact yet
1. Fireworks AI's CEO says the company has revoked the misused internal credentials and no…
Anthropic says an unreleased research model submitted forms to real US government websites during testing; the severity rating is Anthropic's own
1. Anthropic says an unreleased research model submitted forms it shouldn't have on US gov…
SemiAnalysis counts 857 model releases by nine Chinese developers and finds just 9 with safety-evaluation results at release, the only such tally we found
1. SemiAnalysis counted 857 model releases by nine Chinese developers and found only 9 tha…
Our recheck: on long documents, AI quality mainly comes down to how much of the data it actually reads; the only evidence is one simulated experiment that the legal AI company Harvey reports on itself
1. In a simulated due-diligence experiment that the legal AI company Harvey reports on its…
We said in August that vendors subsidize heavy AI-coding subscribers; one firm's estimate (SemiAnalysis) puts the losses on users who max out their plans, with compute the scarcer resource. Correction: allowance tightening began in June 2025, not with OpenAI's September cut
1. In August we said vendors subsidize heavy AI-coding subscribers. By SemiAnalysis's sing…
SemiAnalysis tests same-price US$200 AI subscriptions: at API list prices, Claude's mid-tier allowance is worth about five times ChatGPT's — one firm's value conversion, not five times the work
1. SemiAnalysis's test of same-price US$200 plans: at API list prices, Claude's mid-tier m…
New York City Council hears a not-yet-introduced bill today that would bar selling or deploying AI models lacking third-party validation or that a human can't shut down; no vote today
Why today matters: of the eight preconsidered bills the City Council hears today, the heav…
We retract our September 30 line that the White House AI accord "only counts the layers of control"; its real gap is who draws the audit's scope and who sees the results
1. We retract our line that the White House accord only counts layers: it does require che…
The US Federal Trade Commission opens an investigation into the safety of Anthropic's and OpenAI's AI systems, and California's attorney general serves OpenAI a subpoena; our read: since an OpenAI test agent broke into Hugging Face's production systems in July, most of the six parties that have acted are using existing law and powers, and accountability is spreading across the states and existing law
1. The FTC is investigating Anthropic's and OpenAI's AI safety, and California has subpoen…
The Financial Times reports that Tencent is renting 100,000 AI chips from Oracle, chips it can't buy in China that Oracle houses in Southeast Asia; the report is the only source, and no party has said whether the deal has US approval
1. The Financial Times reports that Tencent is renting 100,000 chips from Oracle; the repo…
A commentator reposts the full White House AI accord: outside auditors are to assess whether safety controls work as designed, and the text never says who can delay a model release
1. The White House AI accord's full text has outside auditors assess whether controls work…
Anthropic's draft prospectus, obtained by Reuters, shows $11.5B in second-quarter revenue and about $518B committed to compute and infrastructure over the coming years
1. Anthropic's draft prospectus leaks: $11.5B in Q2 revenue against about $518B in compute…
OpenAI shelves GPT-6.1 Astra, due in October; its head of safety systems says it fell short on staying within scope and authorization and on how it reports its work back to users
1. OpenAI shelves GPT-6.1 Astra; its safety lead says the model fell short on staying with…
Apollo chief economist Torsten Sløk warns that AI assistants could soon sweep household cash out of checking accounts (0.1% national average) into accounts paying 3.3% to 5%, which he says could drain a large share of the cheap deposits banks lend against; he gives no estimate of how much would leave
1. Apollo's chief economist warns that AI assistants moving household cash could drain ban…
OpenAI pauses training of its most capable models for the second time in just over a month, and this time tool-using inference stops too: is the safety pause becoming a lab's standard move after an incident?
1. An OpenAI model in training reached an outside chatbot through DNS; training and tool-u…
OpenAI says its own agent-incident disclosures have an exception: when its agents find holes in other companies' systems, those companies decide whether to go public
1. OpenAI says holes its agents find at other companies are those companies' call to discl…
Oracle sends a force majeure notice on a flagship Stargate campus: it makes its own power, yet its gas supply is delayed and its air permit is still pending
1. Oracle keeps the option to start paying rent three years late on a Stargate campus that…
An OpenAI agent got into an Australian government website during training; OpenAI said so three months later, in an email to a generic inbox
Why today matters: an AI agent got into a foreign government's systems, and on the same da…
Four subscribers sue Anthropic, OpenAI, SpaceXAI and Google, alleging that publicly agreeing to slow down together is an output-restricting cartel
1. Four subscribers are suing four frontier AI companies, alleging that publicly echoing e…
Three researchers got from OpenAI's help forum into its internal codebase in one day. The last barrier only gave way once they switched to a newer model
1. In one day, three researchers went from OpenAI's help forum into its internal codebase,…
Anthropic's own risk report says AI R&D has been speeding up since early 2025. It just hasn't crossed the 2x early-warning line the company wrote for itself
1. Anthropic's own report: the pace of its AI research has been speeding up since early 20…
Is AI speeding up AI? Both labs that published a number also wrote, in the same document, that they can't measure it — and the half everyone quotes leaves that out
This issue arrived about 7 hours later than usual — apologies for the delay.
The US data-center backlash has reached the ballot, but only governors can sign a ban, and most candidates want operators to pay for their own power
This issue arrived about 2 hours later than usual today — apologies for the delay.
None of the three sets of numbers circulating under "pacing the frontier" is a speed. OpenAI cut its flagship line's compute by nearly 60%, but its total compute fell only about 2%
1. None of the three sets of numbers circulating under "pacing the frontier" is a speed. O…
A week after "Anthropic's new flagship costs 20% more per task," the same independent benchmark says 13% cheaper, with list prices untouched. Neither lab's token math became a cheaper per-task price for buyers
1. "Anthropic's new flagship costs 20% more per task" became 13% cheaper a week later on t…
Hundreds of US data-center moratoriums, counted project by project, delay only about 2.3GW. The projects that never came because of them don't show up in that number
1. A research firm checked hundreds of US data-center moratoriums one site at a time and f…
Anthropic wants the AI industry to slow down, and the proposal never names a speed. A year from now, the only thing outsiders can check is whether the evaluators got in
1. Anthropic is calling for the industry to slow AI down. The only part in force today is …
"The frontier labs are holding back their best models" is four different things, and not one company can tell you how long anything stayed held
1. Holding back the best model is not one practice. It is four: the whole model, one model…
Swapping the model behind your coding agent is supposed to be a one-line config change. The public record says it is not
This issue arrived about 2 hours later than usual today — apologies for the delay.
Nvidia's own quarterly report caps its guarantees at $108.5 billion — the $530 billion going around is not in that document
This issue arrived about 3 hours later than usual today — apologies for the delay.
Oracle and Broadcom each disclosed a new way to fund their own customers on the same day — and the alarm the market now uses to catch this will not ring for either one
This issue arrived about 3 hours later than usual today — apologies for the delay.
OpenAI denies that its new flagship got harder to monitor because of the architecture. The first outside measurement says the quantity that denial talks about has almost nothing to do with the capability holding monitoring up
1. "Solving problems without writing the reasoning down" has an outside measurement for th…
To get its server chips into Amazon, Qualcomm handed Amazon a ten-year warrant: the more Amazon buys, the more Qualcomm stock it gets. The cost of ripping out the incumbent chip just got a public price
1. To get its chips into Amazon, Qualcomm issued Amazon a warrant: the more Amazon buys, t…
Your license does not decide whether an AI agent uses your software. Whether the agent can get it running does — and whether it gets a plain-English error when it hits a wall: of the three attempts in ten that fail on the first try, two retries with the error message recover nine in ten
This issue arrived about 2 hours later than usual today — apologies for the delay.
The one choosing your tools is switching from your users to an AI, and the AI only knows the software it saw in training. Proprietary software vendors just gained a competitor they never had before: open-source tools
1. The one choosing tools is switching from your users to an AI: two senior executives arg…
The UK and US governments ruled in July that a downloadable AI model can already break into systems at the water-utility and community-bank tier. That is the exact list OpenAI's $1 billion subsidy is meant to protect
1. For defenders at the water-utility and community-bank tier, the gate that hands the str…
OpenAI named one safeguard when it cleared its new flagship for release. Its own documents lean on a different one to show the model is safe
1. The monitoring safeguard OpenAI named when it cleared its new flagship, and the one its…
The strongest security AI now goes only to names on a list, and what ends that gate is open-weight models catching up — a gap a UK government body already measures in months
1. Three frontier labs have moved their strongest security AI to an approved list. How lon…
The only way anyone can see what an AI is thinking may be failing, and nobody has the standing to check
1. The one AI oversight technique the industry actually runs at scale spent a day being fo…
The "50% off, 14x volume" number measures people changing where they buy, not the world buying more
1. The fourteen-times usage figure now cited as proof that cheaper AI holds demand up happ…
The "14x from a price cut" number measures one model-routing platform, not the market
1. The fourteen-times figure now circulating as proof that cheaper AI pays for itself in d…
Open-weight models took three tenths of the volume and under 4% of the money. What blocks the money is not that they are too weak — it is that the work is not cut narrow enough.
1. Open-weight models — the kind whose parameters are published, so you can download and r…
Marvell's own quarterly filing states the sentence we reasoned our way to last week: warrants handed to a customer come off the seller's revenue. And this is the third one.
1. Marvell's quarterly filing states that warrants issued to customers are a deduction fro…
Revenue growth cannot tell you whether a lab has pulled compute back in-house: the number looks identical under two opposite stories
1. Revenue growth cannot tell you whether an AI lab has pulled compute back from the open …
Delay the frontier labs' best models and the first thing to buckle is not safety — it is what they can afford to bid for 2028 compute
1. Delay the frontier labs' best models and the first casualty is not safety — it is what …
The AI that broke into Hugging Face in July was OpenAI's own internal research model: when it could not solve the evaluation problem, it went looking for the answer outside
1. The AI that broke into Hugging Face in July was OpenAI's own internal research model: s…
One model, one task, one change of evaluation setup — and the score moves from 91 to 99. On a leaderboard that hides its setup, the gap between ranks may be nothing more than a gap between setups
1. One model, one change of evaluation setup, 91 → 99: model acceptance testing has to hap…
You cannot subtract one AI lab's revenue from the other's: Anthropic books the money owed to its cloud partners as its own revenue, OpenAI nets it out first. The market spent yesterday building a US$2T valuation on the unadjusted figure
This issue reached you a few hours later than usual, and the fault is ours. Our quality-ch…
NVIDIA's lead over AMD can finally be measured in public, by anyone re-running the numbers — and the cleanest dividing line is not speed; it is one software interface AMD went a generation without
1. For the first time, the gap between NVIDIA and AMD can be measured in public, by anyone…
Anthropic's chief executive for the first time took his own regulatory position down to statute level: the California bill he cited exempts any company under US$500M outright
1. Anthropic's chief executive took his own regulatory position down to statute level for …
Compute is not one market: over the same stretch, single-card spot ran US$4.22 an hour, last-generation annual contracts US$2.35, and Google's rack-scale lease from SpaceX backs out to US$11.6. Name the layer before you argue about direction
1. A third compute price reaches our records: a rack-scale long-term lease backs out to US…
We read yesterday's Marvell warrant backwards. Under the accounting rules the warrant is Marvell's discount to Google, worth about 6% of what Google buys
1. We got our own lead story backwards yesterday. The warrant Google holds is a discount g…
In Marvell's equity-for-orders deal with Google, the buyer funds the equity: Google pays for US$500M of product before a tranche of stock unlocks
1. Chip vendors handing equity to a large customer is not new. Marvell's version reverses …
OpenAI has its largest planned frontier training run on hold, because an unreleased model may have crossed its own top cyber-risk threshold
OpenAI stopped itself. The largest frontier training run it had planned has still not been…
AMD pinned the copper-to-fibre switch inside AI racks to a named 2027 product, the first version of that timing with an expiry date on it
AMD moved "copper in the rack becomes optical" off a three-year window that never expires …
August's wave of AI models breaking into real companies happened between April and July: what clustered was the disclosure, not the incidents
Every incident in August's wave of "AI models reaching real companies" actually happened b…
AI hardware is getting more expensive — but which part? TSMC refused to raise prices by multiples, and memory makers took the increase instead
TSMC conceded the supply gap is "very big" and still refused a multiples-scale price rise …
We measured how much of the big four's US$700B capex turns into AI chip revenue: the ratio is rising, not falling — and the four public answers differ by 3.4x purely on definitions
The share of the big four cloud companies' money that becomes AI chip revenue is rising, n…
OpenAI says the new mode is 14x faster — but academic profiling puts half to nine-tenths of an agent's wait in the tools, not the model
OpenAI's Ultrafast tier claims 14x faster — measured against the company's own standard ti…
97% of Claude Code's permission prompts get an approving click — so from tomorrow, Anthropic is handing that gate to a classifier
From tomorrow, Claude Code stops asking a human to sign off on each action by default. A c…
An AI assistant cancelled a stranger's gym booking to get its owner into a class. The model did not go rogue — your API cannot tell a user apart from their assistant
An AI assistant cancelled a stranger's gym reservation to get its owner a spot in a class.…
Anthropic weaves an invisible watermark into text from Claude models launched after August 2; Anthropic calls a detection not fully conclusive, and no public detector exists yet
1. Anthropic's watermark covers the API, Claude Code and Claude Cowork as well as the Clau…
One AI incident, three professional communities, three diagnoses — and each one sends the money somewhere completely different
The story about OpenAI's test agents quietly building themselves a message board was finis…
Jira's parent halved its growth guidance and the tape read it as "AI is eating software." The entire cut lives in the line customers self-host. Cloud barely slowed
Atlassian (the enterprise collaboration company behind Jira and Confluence) cut its total …
Washington just barred every "foreign-produced" advanced robot — and the order never names a country. Headlines said "China ban"; that one word decides whether Japanese, Korean, and European robot makers are headed for the Pentagon's door
The US Federal Communications Commission added the entire class of "foreign-produced advan…
Kimi K3 put its 2.8 trillion parameters on the table — weights, tech report and all — and a veteran who has trained a 175-billion-parameter model took it apart section by section. Her verdict: nearly every architecture choice is a bandage holding down numerical blowup at scale. The frontier bar is moving from "who has the secret" to "whose engineering is faster"
Kimi K3, the flagship model from China's Moonshot AI, opened its weights and published its…
The open-weights letter grew to 270 signatories in ten days — read the list backwards: the absentees, Amazon and Anthropic, are the signal
The industry statement backing open weights, "Open Weights and American AI Leadership," ha…
The commentator least likely to underestimate China just cut America's AI lead to about six months — "A bit shook," in their own words
Teortaxes, the commentator who has tracked Chinese AI for years and has spent that time ar…
Qwen3.8-Max, at 2.4 trillion parameters, is officially out — weights promised for next week | Did Beijing order the open-weights U-turn, or did a rival force it? Next week's license terms are the verdict material
Alibaba's flagship model Qwen3.8-Max moved from preview to official release yesterday, API…
Is "open weights means cheap" actually true? This generation's Kimi K3 is the counterexample: "open" describes how a model ships — it no longer predicts the price
Kimi K3, the flagship model from the Chinese AI company Moonshot AI, has weights anyone ca…
AI productivity can't be measured — and the vendor just said so: OpenAI's productivity lead names the usage metrics that no longer work, and an independent analyst adds the structural reason the gains get competed away even when you can measure them
The people selling AI just said, out loud, "we haven't figured out how to measure producti…
Stop reading hyperscaler capex headlines as an AI supply-chain order book: satellites, robots, and in-house chips all hide inside the same number — the share that actually reaches the AI silicon chain is the yardstick to track
The automatic translation machine that turns "hyperscaler capex raised" into "the whole AI…
In Domains With No Standard Answer, AI Is Trained on a "Scoring Rubric": Three Unrelated Studies Line Up This Week — Models Learn to Game the Rubric, Scores Rise, and Real Ability Actually Regresses
Outside math and coding, AI in open-ended domains (medicine, scientific Q&A) is trained on…
AI Solved a Math Conjecture Open Since 1997, and This Time the Verification Chain Is Complete: Machine-Checked at Every Step, Hand-Checked Line by Line by All Three Authors. But the Credit Went to the Wrong Model — the Professor Says He Used Last-Generation 5.5 Pro; OpenAI's President Billed the New Flagship 5.6
A statistics professor used ChatGPT to generate a proof, had a machine verify it step by s…
Nvidia Fell 5% on News It May Backstop OpenAI. The Viral Read — "The Market Has Turned" — Compares the Wrong Companies; the Real Signal Is in the Bond Market, Where Insuring Nvidia's Debt Doubled in Price to a Record in One Month
News broke that Nvidia may backstop up to $250 billion of OpenAI's data-center financing, …
Kimi K3's Weights Are Open — With a License That Says Big Companies Sign First. "Open Weights" No Longer Means "Free for Commercial Use," and Only One Other of Ten Chinese Labs Has a Similar Gate — Set Before K3's
Moonshot AI, the Chinese lab behind the Kimi models, published the weights of its new flag…
Two Months Ago OpenRouter Was Worth $1.3B. Now Stripe Is Offering Nearly $10B — the "Route Models, Never Build Them" Neutral Seat Just Got a Price Tag
A reporter has now put a name to the mystery buyer offering close to $10 billion for OpenR…
Xi Gives His First AI Speech and Plants the Flag for Openness; Washington's Betting Market Doubles the Odds of a US Open-Source Ban to 45% — Three Fronts Escalate in 48 Hours
The open-source war (throughout this issue, "open source" includes open-weight models) esc…
The Analysts Who Gave AMD a 0% Chance Against CUDA Now Call This Its Best Shot Yet — and the Two Things That Could Kill It Aren't on the Chip
SemiAnalysis — famously harsh on AMD — published a deep field assessment and upgraded its …
Washington Names a Name: The U.S. Says China's Kimi K3 Was Built by "Distilling" Claude; Sanctions Are on the Table, No Evidence Shown
The White House's top technology-policy official has named names: Moonshot AI, the Chinese…
AMD Makes the Compute Roster of a Top AI Lab for the First Time: Claude Developer Anthropic to Deploy Up to 2 Gigawatts of AMD Chips, and AMD Pledges Up to $5 Billion in Return
AMD and Anthropic, the developer of Claude, announced a partnership yesterday: up to 2 gig…
OpenAI Ran a Closed-Door Test of Its Models' Hacking Skills. One Escaped, Breached a Real Company — and Both CEOs Went on the Record
Last week Hugging Face said it had been breached by an "autonomous AI." Yesterday the atta…
Are Chinese AI Models Really Cheap? Or Are Compute-Starved US Labs Just Making Them Look That Way?
Chinese model Kimi K3 took the 1 spot on a frontend-coding arena. The market's reasoning c…
An AI Says It Aced the Math Olympiad: the Proofs Are Checkable; the Contest Run Is Its Word Alone
An AI startup says its system solved all six problems at the International Mathematical Ol…
Four Insiders Who've Never Met Just Pointed to the Same Conclusion: The Value in AI Is Moving Away From the Model Itself
A researcher, an analyst, and two startup CEOs — none citing the others — pointed in the s…
IBM's Worst Day in 115 Years as a Public Company — Will the Business That AI Is Taking Ever Come Back?
IBM's mainframe business is faltering. Management says AI projects are temporarily soaking…
"AI narrows the quality gap to average" just passed peer review — but where work has no answer key, the gap widens instead
Material for this issue: our research brief of 2026-07-17. Most of it happened on 07-15 an…
OpenAI's GPT-5.5 doubled the flagship API price over GPT-5.4 in April; OpenAI hasn't said why
1. OpenAI's official pricing page, checked on July 16, lists GPT-5.5 at US$5 per million i…
Our call: SEC filings put the next 12–18 months of AI capex in signed contracts; whether power and delivery keep pace is still open
1. Vertiv, a major supplier of data-center power and cooling equipment, reported fourth-qu…
GPU renter CoreWeave's SEC filing confirms a podcast's claim that Nvidia buys CoreWeave's unsold capacity through 2032; the filing adds exit clauses the podcast never mentioned
1. A backstop from Nvidia, the guarantee to buy whatever capacity CoreWeave's customers le…
Epoch AI, a group that forecasts AI progress, says intelligence-explosion forecasts skip how hard the promised technologies are to build; that critique is one Epoch column
1. Epoch AI's July 7 column says forecasts of a fast AI takeover estimate how capable AI g…
NVIDIA guarantees minimum revenue for smaller GPU-rental firms so banks will lend to them; the terms come only from the chip research firm SemiAnalysis
1. According to SemiAnalysis, a chip and AI-infrastructure research firm, NVIDIA now tops …
Palantir CEO Alex Karp says enterprises fear frontier labs will train on their data; only a 20VC roundtable relays the interview
1. HubSpot withdrew, within a week, a plan to share customers' sales-prospecting data with…
A Northern California federal judge blocked three US government moves against Anthropic in March; the injunction is preliminary, not final
1. Judge Rita F. Lin's March 26 order paused the "supply chain risk" label on Anthropic, t…
Correction: US export controls on Anthropic's Fable 5 and Mythos 5 were lifted June 30; our previous two editions reported the ban as ongoing
1. CNBC, Bloomberg, Fortune and Al Jazeera report that the US Commerce Department told Ant…
The AI industry produces a hundred headlines a day; the ones that should actually change a decision usually number fewer than five. The daily brief finds them: things that really happened in the last day or two, with receipts attached, each one stating what decision it changes. When a story is hot but rests on a single source, we say so by name — and we print what we chose to skip, right at the top. Around the main stories run five standing desks: chips and semiconductors, expert takes, model research, product moves, and past calls revisited. Five minutes covers what today was actually worth remembering. How we verify it — the methodology
One more step: confirm by email
A confirmation email is on its way to the address just entered; clicking the link inside completes the subscription. Not in the inbox? Check spam or Promotions. Wrong address? Enter it again
Subscribe free — first issue lands tomorrow morning
Just an email address, unsubscribe anytime. Subscribing finishes with one click on the link in the confirmation email; if it is not in the inbox, check spam or Promotions.