Edition
Thursday, September 10, 2026
16 issues
23h window
~14 min read
Running through today’s issues
Anthropic's extinction warning has become an IPO problem, not a discourse problem
4 sources
Jacob Coxon's resignation thread passed 100 million views and Evan Hubinger, who leads alignment science at Anthropic, publicly put his own odds above 10% within the decade. The Information's Briefing goes straight to the consequence: Anthropic is about to market an IPO, its own August report raised model risk from very low to low, and executives will have to answer the extinction number on a road show. Paul Christiano endorsed the framing days after OpenAI named him to its Foundation board, which means the two most credentialed safety voices in the industry are now on the record together. AI Market Fit's counterargument — nuclear precedent, Lean proofs, build anyway — is the bull case, and it is notably an outsider's.
Apple's hub thesis depends on apps surviving agents, and Meta just bet the other way
4 sources
Ternus framed the iPhone as the intelligent personal hub with 300,000 Siri-compatible apps, but Thompson's read is that every one of those integrations requires developer work Apple must keep begging for, while browser-using agents get integration with everything for free. Meta's Muse — cloud VM, near-frontier models, runs continuously — fits the same hub description without the app dependency, and Ben's Bites and Superhuman both treat it as the week's other launch. The tell is in the failure too: Peers asked Muse to watch the Apple keynote and it couldn't handle video. Both hubs are unfinished, but only one is architecturally committed to apps.
Export controls stopped constraining Chinese AI revenue and started exporting Chinese hardware
2 sources
The China digest puts numbers on it: DeepSeek at $70M in July revenue, up tenfold; MiniMax at $800M annualized; Alibaba raising $10B, with DeepSeek ordering Huawei chips for a new data centre and Malaysia and Saudi Arabia buying into the stack. The AM briefing shows the political response — Washington accusing six Chinese firms of industrial-scale distillation, Beijing replying that US labs distill Chinese models too, ahead of a Trump-Xi meeting. The Ford story is the same dynamic in batteries, where Chinese suppliers dilute below 25% ownership to unlock US tax credits and Beijing selectively vetoes the deals it considers strategic. Controls are now shaping ownership structures rather than capability.
Better products are not moats, and the assistant category is about to find out
3 sources
Not Boring's argument is that Instinct and its peers are operationally better, not counter-positioned — nobody's profit pool is destroyed by shipping a helpful agent. The week's launches confirm the crowding: Meta's Muse, xAI's Grok Bot marketplace bundled into Cursor plans, Apple's Siri AI, all offering the same book-travel-handle-email surface. The Grok Bot marketplace hints at where differentiation actually lands — creator-named templates and distribution inside an existing developer subscription, not the assistant itself.
Ben's Bites
06:03 PT
Everyone needs a Muse
So what
Your open-model substitution plan stalls where intent inference matters — budget for frontier models on the fuzzy work, cheap weights on the specified work.
A grab-bag issue in which Ben admits he's hitting rate limits orchestrating agents with agents, and the headlines carry the weight: Meta shipped a personal agent, OpenAI claims a Millennium Problem, and an Anthropic safety resignation became the most-viewed AI safety post ever.
- Meta launched Muse, a cloud personal agent that books travel, handles email and connects FB/Instagram accounts; US adults only
- OpenAI says it solved Navier-Stokes, with authorship disputes, reportedly triggered by rumours Anthropic had cracked a Millennium problem first
- ChatGPT Images 2.5 does targeted element editing with everything else preserved, and runs up to 50% faster; new Sketch feature turns scribbles into images
- Jacob Coxon resigned from Anthropic on X over safety; Ben flags people asking whether it's a conspiracy, while another safety figure joins the OpenAI board
- Ben rates GLM 5.3 his favourite open model but says frontier models still read fuzzy intent better; he's shipping a design reference-sheet tool tomorrow
Not Boring
05:45 PT
An Ode to Counter-Positioning
So what
Ask every assistant startup in your pipeline who gets hurt when they win; if the answer is nobody, you're underwriting a feature, not a moat.
Packy Mccormick argues counter-positioning is the only one of Helmer's 7 Powers a young startup can actually wield, and that most AI assistant startups are confusing being better with having a barrier.
- Ramp counter-positioned against points-based corporate cards by helping customers spend less; now valued at $44B versus Divvy's $2.5B sale to Bill.com and Brex's $5.15B to Capital One
- Base Power's Zach Dell: don't sell a cheaper battery, sell electricity, so incumbents must rewrite their business model to respond
- Historical set: Google/Android vs Nokia, Amazon vs Barnes & Noble, Microsoft vs IBM, Facebook vs MySpace, Somos vs debt-laden telcos
- Helmer's own caveat: counter-positioning is a take-off phase power and only a partial source; Dell used its window to build Scale Economies
- The sharp end: Instinct and the assistant category are better products, not counter-positioned ones — nobody is damaged by shipping a helpful agent
Superhuman AI
04:09 PT
📱 Apple goes all in on AI
So what
Suno v6 removes the licensing objection that kept AI music out of commercial work — your ad and content budgets can now use it cleanly.
Apple's first event under John Ternus put AI into every product line, headlined by a $1,999 foldable, and Suno's v6 arrived with actual label licences behind it.
- iPhone Duo at $1,999 with an AI-designed hinge; iPhone 18 Pro/Pro Max add a 48MP variable-aperture camera and Reference Image, which signs every pixel against later edits
- Siri AI ships in beta with iOS 27: conversational back-and-forth, onscreen awareness, personal context from apps, plus a dedicated cross-device Siri app
- Suno v6 is its first model trained on licensed content from Warner Music Group, BMG and Believe, with multimodal prompting and plain-language song editing
- Anthropic's economic simulator spans modest productivity gains to explosive GDP growth with unemployment near 14%
- Watch Series 12 gets Live Rewind (last 15 seconds transcribed) and Siri Recap; AirPods 5 add real-time Live Translation
The Information
04:07 PT
Apple Launches Foldable iPhone Duo
So what
Treat Chinese humanoid revenue lines as unverified until CSRC guidance clears; the fraud call is coming from a listed insider, not a short seller.
The AM briefing runs Apple's foldable alongside a hardening US-China AI fight and a rolling Anthropic safety story that now includes a fourth disclosed security incident.
- iPhone Duo starts at $1,999 and reaches $3,199 for 2TB; iPhone 18 Pro rises $100 to $1,199, which Apple pre-warned was driven by memory and storage chip costs
- Ternus pitched the "intelligent personal hub" and took a swipe at cloud data collection — pointed at OpenAI, which Apple sued in July over trade secrets and poaching
- OpenAI put Paul Christiano, White House AI advisor and ARC founder, on the OpenAI Foundation board and made him a non-voting observer on the for-profit board
- Mech-Mind's CEO says Chinese embodied-AI firms fake revenue via data collection centres and related-party deals, naming Galbot, after CSRC window guidance raised the humanoid IPO bar
- Beijing rejected the US accusation that DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.ai run industrial-scale distillation, calling out US firms for distilling Chinese models
Ben Thompson (Stratechery)
03:01 PT
The iPhone Duo, The Intelligent Personal Hub, Apple Watch Audio Intelligence (Stratechery Update 9-10-2026)
So what
If browser-use agents commoditize integration, Apple's 300,000-app API advantage becomes a liability, and your app-integration startups lose their reason to exist.
Thompson credits Apple's hardware-software integration on the Duo but argues the intelligent personal hub pitch is undermined by Apple's unexamined faith that apps remain the unit of computing.
- iOS was genuinely redesigned for the 7.6-inch fold — side controls, split view, second windows — and Federighi fronted the segment, but developers must do the work via new APIs
- Netflix, which refused to support Vision Pro, showed up prominently for Duo: developers work with Apple out of pure self-interest now
- Siri AI claims 300,000 compatible apps, but each new capability still requires developer adoption; browser-using agents get integration with everything for free
- Meta's Muse fits the hub description better on the merits: cloud VM, near-frontier models by default, runs continuously without you
- The deeper miss is that Apple frames Intelligence as augmenting app workflows — updating calendar and reminders — rather than absorbing them into the agent
The Information Briefing
17:02 PT
The Briefing: Anthropic’s Death Risk
So what
Anthropic's IPO now carries a disclosure problem its bankers cannot script around; price that governance risk before the roadshow does it for you.
Martin Peers treats the Anthropic extinction post as the story, and notes the awkward timing: the company is about to market an IPO with staff publicly putting double-digit odds on human extinction.
- Anthropic's own August report raised its model risk assessment from "very low" to "low" — the direction, not the level, is the tell
- Peers asks how executives handle the extinction number on an investor road show; jokes about analysts asking Amodei about it circulated widely
- He tried to make Meta's Muse watch the Apple event for him; Muse replied it can't watch video streams
- Apple's foldable landed as expected at $1,999–$3,199, with no Face ID
- Also: OpenAI working with Samsung on next-gen chips, six Chinese AI firms accused of distillation, Block applying for a national trust bank for crypto custody
The Information
16:19 PT
Exclusive: Jeffrey Katzenberg Teams Up with Former OpenAI Sora Head on New AI Video Startup
So what
A Sora-native founder plus Katzenberg's studio relationships is the credible entrant in pro AI video; expect your Runway and Luma marks to be tested.
Jeffrey Katzenberg is backing a new AI video startup with the former head of OpenAI's Sora app, aimed at filmmakers rather than consumers.
- The company plans to train its own video models rather than build on top of someone else's
- Andreessen Horowitz is among investors being approached for significant funding
- The pitch is winning over filmmakers, a different customer than Sora's social-feed audience
- The Sora app lead departing to compete directly is a notable talent leak from OpenAI's consumer video effort
The Information
15:53 PT
Why An Anthropic Researcher's Terminator-Style Warning Caught Fire
So what
Two of the most credible safety voices now endorse the extinction framing publicly, which hands regulators a citation they did not previously have.
The Anthropic extinction warning has escalated from an X thread into a Washington and Wall Street problem, with OpenAI's newest nonprofit board director publicly endorsing it.
- Evan Hubinger's agreement made this an internal company view rather than a disgruntled exit, since he leads alignment science
- Paul Christiano, just named to the OpenAI Foundation board, said that without more robust alignment "most people could die"
- The piece flags direct complication for Anthropic's upcoming IPO and its efforts to repair relations with US political leaders
- Some Anthropic employees defended their harm-minimisation work while conceding how much remains undone
Masters of Scale
14:27 PT
Three places to aim your AI
So what
Score your portfolio companies on coordination cost reduction, not seat-level AI adoption; the second is easy and the first is where the margin is.
Zapier's Wade Foster argues the gap is no longer AI adoption but organisational design: plenty of 10X individuals, almost no 10X companies.
- Foster called a code red when GPT-4 landed, paused the company for a week and ran an all-hands hackathon; daily AI use went from 10% of employees to over half in five days
- His model: customers come from R&D and go-to-market; everything else a company accumulates is coordination overhead
- Three targets follow — 10X go-to-market efficiency, 10X R&D speed or quality, cut coordination costs to a tenth and redeploy that time
- "We have the technology, but we don't know how to build the organizations around it quite yet"
- His verdict on most companies: haphazard tool distribution to individuals, hoping something productive emerges
The Information
14:15 PT
Alibaba, DeepSeek and Tencent Didn't Wait for Nvidia
So what
Chinese labs now clear real revenue without Nvidia, so your China exposure thesis should shift from sanctions risk to price competition on inference.
The China digest makes the case that export controls failed at their stated purpose: Chinese model businesses are compounding revenue on domestic silicon, and the workarounds are now an export product.
- DeepSeek hit $70M in July revenue, a tenfold jump from 2025; MiniMax is at $800M annualized; Alibaba is raising $10B and expects $10B AI annualized revenue by September
- Z.ai's API sales surged on a new low-cost model, and Tencent's new flagship showed major progress
- DeepSeek is planning a major Huawei chip order for a new data centre; CXMT broke through on advanced memory days after suing the US over blacklisting
- Malaysia is eyeing Huawei chips for its sovereign AI program despite US warnings; the administration is drafting a rule to curb China's remote chip access
- Saudi firm Humain built an Arabic model with MiniMax; Moonshot AI filed confidentially for a Hong Kong IPO
The Information
13:55 PT
Exclusive: OpenAI Cuts Off Adobe, Others From Advertising Some Competing AI Products in ChatGPT
So what
OpenAI is shrinking its own ad TAM to protect first-party features — discount the ad revenue line in any OpenAI secondary you're pricing.
OpenAI has quietly stopped accepting ChatGPT ads for image and audio generation products, blindsiding advertisers including Adobe.
- The policy is unpublished and not reflected in OpenAI's stated ads policy
- The excluded categories are precisely where OpenAI ships its own features
- Extending the logic to other competing products would materially narrow the addressable ad pool
- OpenAI has given investors aggressive ad growth projections, positioning ads as the monetisation path for non-paying users
AI Market Fit
13:26 PT
Will AI Kill Humanity?
So what
The insider critique is now specific enough to use as a diligence lens: ask labs whether race logic or stakes assessment governs their release decisions.
Guillermo walks through the Coxon resignation and Hubinger's endorsement, concedes the warning is credible and unverifiable, then argues the nuclear precedent justifies building rather than freezing.
- Coxon, 27, spent three years in pretraining at OpenAI (GPT-4o) then Anthropic; his thread passed 100 million views and was picked up by Fortune, WaPo, CNBC and others
- His distinction between labs: at OpenAI the civilizational stakes aren't internalised; at Anthropic they are, but the race logic overrides them
- Hubinger, Anthropic's alignment science lead, put his own odds above 10% within the decade
- Anthropic's August report rates current-model catastrophic risk as low while warning future models "may cause unbounded harm"
- Counterweight cited: Claude completed the first formalized Lean proof of a major mathematical result, work that used to take years
The Information
12:22 PT
Ford’s Exchange of Words With a Trump Lieutenant Reflects Tension Over Chinese Content
So what
Ownership-dilution structures below 25% are the working template for keeping Chinese tech in US supply chains — and Beijing is selectively vetoing them.
Transportation Secretary Sean Duffy attacked Ford's patriotism over its CATL battery licence and Geely tie-up, and Ford concluded he was freelancing — but the underlying decoupling is real and bidirectional.
- Duffy's letter demanded Ford prioritize "technological sovereignty"; Ford comms chief Mark Truby: "He's well out of his own lane"
- A House committee bill would outright ban imports of Chinese internet-connected vehicles and hardware, superseding the existing 100% tariff
- The White House report The Great Transshipment Scam claims over 40 countries act as hubs, costing up to $26B a year in lost tariffs
- Chinese suppliers cut stakes below 25% to qualify US customers for tax credits — Ronbay to 24.9%, Gem 55% to 21%, Huayou to 24%, CNGR below 25%
- Beijing bars licensing fourth-generation LFP equipment abroad; CRU's Sam Adham says many US buyers still prefer Chinese cells on quality
Cursor
09:49 PT
The Grok Bot marketplace is live
So what
Templates plus creator names is how agent distribution actually works; if your agent startup ships a blank canvas, you're losing on onboarding alone.
xAI opened a marketplace of pre-built Grok Bots, distributed through Cursor plans with usage on top of the existing Cursor pool.
- Featured bots come from recognisable operators: Lenny Rachitsky's Overheard mention digest, Claire Vo's Tradbot for family admin, Eric Zakariasson's Notion-driven Projects Manager
- Lauren Tan's dr eggbot designs other Grok Bots from a few preference questions — agents configuring agents
- Positioning is explicitly anti-blank-canvas: start from a job someone already made work
- Bundling into Cursor plans with separate weekly usage is a distribution play, not a pricing one
Every
09:34 PT
What Writers Who Use AI Want You to Know
So what
The reusable pattern here is agents as research and interrogation infrastructure, not drafting — point your content team's spend there.
Laura Entis interviewed five professional writers about their AI workflows and found no shared recipe — the same tool that sharpens one writer's thinking degrades another's prose.
- Kevin Roose delivered a 448-page AGI book off 150+ interviews in about a year, estimating three to five without AI; an agent transcribed, extracted details, suggested five follow-up sources and flagged artifacts to request
- Roose's agent found contact details for hard-to-trace sources, in one case exploiting a security flaw to extract an email from a GitHub account
- Alexandra Samuel wrote 52,000 words in 10 days by converting research into podcasts and having ChatGPT interview her on walks; almost none of the words survived, but the momentum did
- Maggie Appleton uses Claude Code and Codex to surface literature she's overlooked — it re-introduced Lucy Suchman's work into a speech argument
- Emilia David fine-tuned a custom GPT on a reader profile and editor feedback to pre-screen story pitches
AI with Allie
08:45 PT
Steal this prompt to see what AI can take off your plate
So what
The interview-first prompt pattern beats prompt libraries for internal enablement — steal it for your own team's automation discovery sessions.
Allie Miller's pitch is that most AI teaching hands people two prompts and no principles, and she offers a single interview-style prompt for finding the one task worth automating.
- The prompt has Claude Code or Codex interview you one question at a time about your week — what you dread, what you defer, what only you remember
- Explicit instruction to avoid 2019 assumptions about work and to reason from 2027 tech assumptions backward
- Her framing: the hard part is describing what you want, so use AI to do the describing too
- From hours of interviews with Gen X women, almost none asked about productivity; they asked how to get their expertise into the model and delegate what they never wanted to do