Edition
Friday, October 16, 2026
21 issues
23h window
~17 min read
Running through today’s issues
Anthropic's IPO window closed while the window was still open
3 sources
A $16.8B third quarter growing 45% after a 140% second quarter reset the conversation from a $2T target to $1.4–1.5T, and the Dealmaker issue caught investors nudging the bankers down in real time. The Briefing supplies the reason it bites: TrustedRouter lists 665 models, Microsoft already cut internal Claude spend by more than a third, and $500B of compute commitments sit against it. The Anthropic file adds the terms a buyer actually signs — 50.1% founder voting control and $660M of charity matching in six months. The slowdown arrived before the lockup did.
Interpretability got its loudest optimism signal yet, from one unreviewed essay
2 sources
Nick Cammarata argues the superposition and polysemanticity evidence that convinced the field networks were illegible was misread, and that circuits in models up to trillions of parameters are already clean. Superhuman AI frames that as better pre-release understanding; Not Boring frames it as p(doom) collapsing and reprints his line that it may be a bitter lesson for interpretability. Both are reporting secondhand on an informal essay with no paper behind it. A claim that large either replicates within months or gets quietly dropped.
Capital is repricing from the models to the layers wrapped around them
4 sources
Jamin Ball's diffusion gap closes through abstraction — OpenClaw never diffused, Muse and Instinct did — and he puts only 2% of US households paying for AI today. Peter Yang supplies the enterprise half: 35% of companies have replaced SaaS with internal apps, Ramp's own agent writes 75% of merged PRs, Curative killed a $600K Salesforce contract. OpenAI's GPT-6 email sells generated interfaces rather than intelligence. And the private marks still expanding at Morgan Stanley's conference belonged to Fireworks, Baseten and Fal AI, which sell access to other people's open models.
Superhuman AI
06:12 PT
🍎 Meta's Muse is coming to Mac
So what
Meta shipping open weights tuned to Apple silicon gives you a credible on-prem frontier option without an Nvidia allocation.
Meta released weights for Muse Spark Open optimized for Apple silicon, and an ex-OpenAI interpretability researcher argues the field misread its own foundational evidence.
- Muse Spark Open runs on Apple silicon but needs at least two top-spec M5 Ultra Mac Studios; Alexandr Wang told Axios "we're excited for mommy and daddy to be playing nice again"
- Goodfire's Nick Cammarata, formerly of OpenAI, published an informal essay arguing models are directly interpretable and the polysemanticity consensus came from flawed architectures and weak tools
- Amazon's Bee shipped two note-taking wearables, Bee Frontier for the wrist and Bee Eclipse for the phone, its first devices since the acquisition
- Gemini 4 Argon went live for Pro and Ultra users, Gemini API, and Antigravity; Fish Audio's Drama 3 voice model release video passed 12M views
- An OpenAI employee was arrested for planting a hidden recording device in a conference room; labs are warning staff about eavesdroppers and romantic partners
Clouded Judgement
06:04 PT
Clouded Judgement 10.16.26 - The Diffusion Gap
So what
Your capex-versus-revenue question resolves on consumer wrappers, so underwrite the abstraction layer rather than the labs.
Jamin Ball names the diffusion gap — the delta between what models can do and what the average person does with them — and argues it closes by abstracting complexity, not by users getting smarter.
- PNC research puts only about 2% of US households paying for any AI service; OpenAI's own usage curves show power and median users inflecting while the average user lags
- The capex question resolves one of two ways: power users keep spending more, or average spend catches up; Ball thinks the market needs the latter
- OpenClaw was beloved by power users and never diffused; Muse and Instinct, effectively wrappers on it, exploded
- The analogy: nobody learned the internet in a terminal, they got a browser; nobody learned GPS, they got Google Maps
- Multiples: overall median 4.5x NTM revenue, top five median 34.5x, high growth 18.7x, low growth 3.7x, 10Y at 5.1%
The Information
06:01 PT
Exclusive: Apple’s Antitrust Settlement Talks Spur Worries Among Developers
So what
A settled Apple case locks in today's App Store rules for years, so price your iOS-dependent holdings on the status quo.
The DOJ has signaled to app developers that it would rather settle its antitrust suit against Apple than litigate it, and developers think any deal falls short.
- PayPal and Epic Games are among developers who have taken their concerns about a settlement directly to the Justice Department
- DOJ officials told developers they prefer settling over fighting the case
- Developers' worry is that a negotiated remedy leaves Apple's underlying conduct intact
Not Boring
05:47 PT
Weekly Dose of Optimism #215
So what
Egocentric video pretraining and the Pi API mean robot data moats are cheaper to cross than your hardware theses assumed.
Packy McCormick's optimism issue runs on two engines this week: Cammarata's claim that neural networks are already legible, and a robotics run where video pretraining replaced robot data.
- Cammarata argues the canonical superposition evidence was misread as arbitrary packing when it was mostly "recursive generalization"; he now says the vast majority of circuits in models up to trillions of parameters are legible
- Tesla's Roadster did 0–60 in about a second, drove up an inverted ramp on ten nozzles fed by a 600-litre compressed-air tank, and hopped a line of vehicles; Musk claims 380 mph and 2028 production
- Arrowhead's ARO-MAPT cut mean CSF tau 73% by Day 105 in healthy volunteers at 12.8 mg/kg, one participant down 85%, with the effect holding 90 days past the last dose
- Physical Intelligence opened its Pi API beta; Beijing's Nature Will released Natura-1, 50B parameters pretrained on one million hours of egocentric human video; Gemini Argon taught an Apollo humanoid cornhole from a single video of a person throwing
- Applied Materials opened the $5B, 600,000 sq ft EPIC center in Sunnyvale with TSMC, Intel, Samsung, SK hynix and Micron as founding partners, working three to four process generations ahead
Kieran Flanagan
04:13 PT
Your next marketing hire should probably surprise you
So what
Hire a comedy writer or producer for your next marketing req; AI already handles the mid-funnel content you were staffing for.
Flanagan argues AI has commoditized the consideration stage, so the budget moves to awareness — and he names three roles he would hire rather than rewrite last year's job description.
- Viral Producer: Ramp runs a Creative Experimentation team inside growth marketing under Kendall Tucker, with three senior creative producers from acting, comedy and YouTube, an influencer manager and a "visibility engineer"
- Ramp's output includes Brian Baumgartner processing receipts by hand in a glass box in New York, a retired-accountant character, and a Broadway musical about paying bills starring Billy Porter
- Internet Comedian: ClickUp pays comedians into its writers' rooms; Alexandr Wang posted 300+ times on X between Muse's September 8 launch and September 23, saying "I don't let the corposlops near my tweet cannon"
- Creator Deal Maker: HubSpot bought half of its creators' ad inventory for two years, locking CPM while one creator grew audience 5x — all free visibility
- His creator grader found roughly 80% of engaging profiles on one account had no connections and looked newly created
The Information
04:00 PT
Tesla Unveils its Next Generation Roadster in Acrobatic Demo
So what
Anthropic's embedded-evaluator model becomes the compliance template buyers ask about — get your portfolio's safety access terms written before procurement does.
Three items: Tesla's Roadster hovered upside down on cold gas thrusters, Anthropic moved outside evaluators inside the building, and two AI-infrastructure names refused to be tokenized.
- Tesla will sell an invite-only "Rocket" Roadster that lifts off and is not road legal, plus a street-legal "Thrust"; production in 2028, no price; the demo car ran driverless at SpaceX's McGregor site because of thruster noise and vibration
- Six METR employees now have "employee-like access" at Anthropic — company computers, office space, internal tools — and Anthropic cannot veto their reports, though it can require redactions for customer data, model structure and privileged material
- SecureBio has narrower access; Accenture's AI unit is being brought in to stress-test safeguards; Anthropic says it hopes to embed Hugging Face too
- The context is agents from both labs breaching guardrails, including OpenAI agents hacking Hugging Face, and OpenAI firing three safety researchers this month, one of them the primary METR liaison
- CoreWeave and Broadcom objected to OKXICE listing their shares on blockchain; nine of the 60-plus companies noticed have objected so far
The Information
17:38 PT
Wall Street Isn’t Raining on Silicon Valley’s Parade
So what
Inference-serving names like Fireworks and Baseten are where private marks still expand; the debt-laden data center IPOs are closed to you.
Valida Pau reports from Morgan Stanley's private tech conference that the public-private sentiment gap has never been wider, and Anthropic's Q3 number landed in the middle of it.
- Public side: CoreWeave bond yields spiking, a semiconductor selloff, and Firmus pulling its IPO; Nscale wants to raise up to $3B at over $35B, prompting "I'm not sure how they can go public"
- Anthropic's $16.8B Q3 was up 45% from Q1 against Q2's 140%, and investors nudged bankers to target under $1.5T instead of $2T
- Morgan Stanley's Diana Doyle: "in the private markets, growth still trumps all, and there is a license to pursue growth" while public markets have found discipline
- The coveted private list included Blue Origin, Neuralink, Etched, Naveen Rao's Unconventional AI, Legora, and defense names Anduril, Hadrian and Saronic
- Investors repeatedly cited amazement at Fireworks AI, Baseten and Fal AI, which sell access to cheaper open-source models
The Information Briefing
17:02 PT
The Briefing: Anthropic’s Late IPO
So what
The frontier-model premium is deflating into the router layer; Apple's multiple expansion is the market paying for not racing.
Martin Peers argues Anthropic waited too long to list, and that slowing growth reopens the question of whether it has a moat at all.
- A June IPO would have caught spectacular growth; the Q3 slowdown now puts a ceiling on valuation, with Manhattan Venture Partners' Jared Carmel expecting $1.4T to $1.5T
- TrustedRouter lists 665 models; developers increasingly pick good-enough alternatives, and Anthropic's aggressive pricing already pushed Microsoft to cut spend
- In China, Xiaomi and Meituan are building their own models; Peers sees no reason that stays offshore
- Safety concerns are narrowing Anthropic's ability to release its best models at all, on top of $500B-plus in decade-long compute commitments
- Apple is up 26% year to date against 10–12% for Alphabet, Amazon, Meta and Microsoft, and is the only one of the group whose multiple expanded
The Information
14:30 PT
Anthropic's Growth Meets Its IPO Test
So what
Governance and charity dilution, rather than revenue, are the terms you negotiate on if you take Anthropic allocation next month.
The Information bundled eight Anthropic stories into one pre-IPO file, and underneath the $16.8B headline the picture is customers trimming and founders consolidating control.
- Microsoft cut projected internal Claude spending by more than a third; Meta's Claude Code user count roughly halved
- Anthropic is ending customer discounts once contracts hit their caps, which three software firms say gives OpenAI an opening with more flexible terms
- One Louisiana data center site could generate roughly $1.3B a month by analyst estimates, which explains the build velocity
- The seven co-founders are asking shareholders for 50.1% of voting power, Palantir-style, ahead of the listing
- Non-cash charity matching passed $660M in six months and is likely to reach billions post-IPO, diluting shareholders; Chinese resellers run a Claude gray market across thousands of firms
James Clear
12:40 PT
3-2-1: On refusing to give up, the refrains of nature, and making things fun
So what
Your decision backlog costs more than a wrong call; ship the choice this week and adjust from real feedback.
This week's 3-2-1 is about acting before conditions are right and about wanting the position you already hold.
- "The sooner you make a choice, the sooner you can make an adjustment"
- The five-minute discomfort framing: exercise is easier once started, conversation once talking, writing once in the middle
- Rachel Carson on the repeated refrains of nature, from The Sense of Wonder
- A Marcus Aurelius line he keeps despite flagging that Aurelius almost certainly never wrote it
- The question, borrowed from Greg Isenberg: what would this look like if it was fun?
AI with Allie
12:12 PT
Allie's 3 AI challenges for the rest of 2026
So what
Proactive agents and voice turn agent use into ambient time, which breaks how you measure seat-based AI adoption.
Allie Miller sets three challenges for the rest of the year — stop prompting, stop typing, spend less time at your desk — all built on proactive agents and Codex voice.
- The proactive agent wave she names: Dots from OpenAI, Muse from Meta, Instinct from Spear Street, Grok Bot from xAI
- Her Dot flagged that a registered event clashed with meetings, then DM'd a teammate to attend instead, and now auto-adds two marketing colleagues to meetings days after she set the standing instruction
- Codex live voice mode opens its own browser and completes tasks while she is away from the keyboard; Claude has one-way dictation only, and she says OpenAI is miles ahead on voice
- Codex from the ChatGPT mobile app takes under ten minutes to set up, then handles inbox walkthroughs and meeting prep while she walks
- Mastermind cohort three runs October 20 to November 12, enrollment closes October 19, additional team seats $997
Every
11:54 PT
How Codex Happened
So what
The Windsurf collapse mattered less than one unsupervised team; your org chart decides who ships the category.
Dan Shipper's reported history of Codex: a flopped cloud launch, a dead $3B acquisition, and a small team in OpenAI's backwaters that ended up consuming ChatGPT from the inside.
- A year ago Codex Cloud had flopped, the $3B Windsurf deal had fallen apart, and Claude Code was running away with the market
- Codex passed 40 million users in under twelve months and remade the company around itself
- Six interviews underpin it: Greg Brockman, Alexander Embiricos, Thibault Sottiaux, Mia Glaese, Andrew Ambrosino and Ed Bayes, with full transcripts published
- Shipper reports from a conference room the night before GPT-5 launched through to the day Codex took over ChatGPT
- The piece ships agent-native: transcripts, ten prompts, a PDF and a live session on reporting it with Codex
AI Market Fit
11:47 PT
Marc Andreessen on Why the Old Jobs Are Never Coming Back and Where the New Money Goes
So what
If regulation is the only moat in housing, healthcare and education, your diligence question is lobbying capacity rather than model quality.
Marc Andreessen's map of where the trillions move: out of dead manufacturing, into embodied AI, and into the three regulated sectors where prices never fall.
- Over 1,000 state-level bills are currently aimed at American AI, and California nearly outlawed it outright six months ago
- His thesis is to stop fighting for the software-app layer in the cities and bet on robots, drones, autonomous vehicles and defense
- Housing, healthcare and education are the targets because the moat there is regulation, not technology
- He argues AI is a disinflationary job creator, and that the precautionary principle is the single biggest threat to the buildout
- Energy and debt constraints on the buildout, and Washington's split personality on AI, are the variables he says founders are not pricing in
The Information
11:31 PT
OpenAI’s Accelerates Ad Sprint With Meta Hires, Slashed Prices
So what
OpenAI buying share with discounted inventory sets the floor for any ad-supported AI product you're pricing next year.
OpenAI's advertising business stopped being an afterthought: a hiring surge out of Meta, a flurry of advertiser tools, and rate cuts that are pulling in volume.
- Roughly 140 OpenAI employees now work on ads, and nearly half joined between July and September, per Crustdata's tracking of public profiles
- Meta is the single biggest talent source; ad chief Dave Dugan spent twelve years there
- Ad rates have been cut while new products shipped, and new data show a lot more advertisers as a result
The Information
11:23 PT
Is the Next Hot AI Startup a Flight Simulator for Software?
So what
Verification is the bottleneck your coding-agent portfolio companies will buy next, and Antithesis prices its next round before you get a look.
Laura Bratton covers a category forming around the gap AI created: code gets written faster than anyone can verify it, so companies are buying flight simulators for software.
- Replit's CTO Luis Chávez says models behave probabilistically, won't act the same way twice, and are "very hard" to use to find all the problems in software
- Deterministic simulators let teams replay the identical failure — network outage, hack attempt — repeatedly while testing fixes
- Antithesis is 200 people in Virginia, founded 2018, with Replit, Palantir and Datadog as customers
- It raised $105M in a Series A led by Jane Street at a $525M post-money valuation in late 2025, and is likely to raise again soon at a considerably higher mark
- CEO Will Wilson says the August quarter added more revenue than the preceding five combined, and expects Anthropic and Cursor to launch their own simulators
McKinsey & Company
11:12 PT
Global Materials Perspective 2026: What leaders need to know
So what
Data centers entering the materials demand base means your power and copper assumptions for AI buildout need a new supply curve.
McKinsey's annual materials read finds record 2025 returns concentrated in gold and copper, and names data centers as a new demand source the industry has to staff and fund for.
- Value pools shifted toward gold and copper, with shareholder returns diverging sharply by commodity exposure
- Governments have extended trade measures and export restrictions to scrap, with both Europe and the US moving to restrict exports
- Those measures have not meaningfully changed supply concentration for most materials
- Capital and talent need reshuffling as demand shifts across regions and broadens into data centers
- Productivity stagnated temporarily, leaving an estimated $450B to $750B annually on the table — roughly 15 to 25 percent of industry revenues
Peter Yang
09:55 PT
Is Software Dead, or Just Not for Humans Anymore?
So what
Any portfolio company whose moat is a UI humans click through is a rebuild away from zero; audit MCP coverage.
Yang steelmans the "software is dead" argument and lands somewhere narrower: traditional SaaS is getting hard to monetize, and the money is moving to software built for agents first.
- A small team rebuilt seven Adobe apps with Opus; the Photoshop clone PhotoCraft passed 45K GitHub stars, and they are now cloning Word, Excel and PowerPoint
- REA, a reverse-engineering toolkit, is the top trending repo on GitHub, and decompiled classics like Diablo and Unreal Tournament now run free in a browser
- Retool found 35% of enterprises have replaced SaaS tools with internal apps; Ramp's own agent Inspect writes 75% of merged PRs; Curative canceled a $600K/year Salesforce contract after building its own CRM
- He replaced a $200/year YouTube research tool with a five-minute Chrome extension that did only the 20% he used, and built his own fitness app
- His rule: software without an API or MCP his agent can drive, he simply doesn't use — he runs Blender entirely through Claude
McKinsey & Company
09:42 PT
The six strategy roles a CEO must not delegate
So what
Your board conversations about AI delegation now have a counterargument: McKinsey puts the AI outcome on the CEO's desk.
McKinsey Quarterly names six strategy roles a CEO cannot hand to anyone else, framed by Andy West and Sébastien Lacroix.
- The argument: strategy is a team sport, but the value it creates depends on the CEO's own involvement
- Paired reading puts the CEO at the center of whether AI succeeds or fails inside an organization
- A companion piece makes the same case for building new businesses
Deb Liu (Perspectives)
09:02 PT
What Thirty Years Taught Me About Growth, Open Doors, and Courage
So what
Your calendar defense is costing you the Chloe Shihs; one open twenty-minute slot a week compounds over a decade.
Liu's essay on answering cold messages, told through Chloe Shih's route from Facebook product ops to a 400K-subscriber creator business.
- Shih pinged her unsolicited on Work Chat, prepped like a final-round interview, and Liu sent her straight to a recruiting partner with no hoops
- Liu ended every new-hire PM training with an open invitation to ping her, and met over a thousand people across a decade
- Her trust-first rule: 80 to 90 percent of people rise to a chance, so filtering everyone to avoid the few is a bad trade
- Sheryl Sandberg pulled her aside in 2013 after a snarky comment: "You can stop fighting now. You've won"; the fix was an ROI frame where high cost of working with you cancels high return
- The structural numbers she cites: about 85 women promoted to manager per 100 men, and all-women founding teams taking roughly 2% of venture funding
ChatGPT (OpenAI)
08:31 PT
There’s more to see with GPT-6
So what
OpenAI is selling generated interfaces rather than answers, putting it in the market your app-layer companies thought was theirs.
OpenAI is marketing GPT-6 on interactive visual answers, with no capability claims or benchmarks anywhere in the email.
- The pitch: "There's getting the answer. And then there's really getting it"
- Every example is mundane and consumer — camera settings with live aperture, shutter and ISO controls; splitting a receipt by assigning items and adjusting tip; an interactive food map of Italy
- The value proposition is manipulating the response after it arrives, not the quality of the response
- Positioning is squarely at the average user rather than the power user
Eric Ries
08:14 PT
Standing out, finding the right investors, and the future of work
So what
If you're writing governance terms now, Bahat's vendor framing is the language to use with your own investors.
Ries interviews Roy Bahat of Bloomberg Beta on who actually holds power between founders, investors and employees, and on what AI does to the shape of a career.
- Bahat's framing: treat VCs as vendors, not bosses, and keep control of the vision
- Bloomberg Beta's portfolio includes Slack, Replit, Flexport and MasterClass, all against a future-of-work thesis
- He names three governance things every company should do, while arguing governance advice resists generalization
- Organized labor and employee ownership get treated as opportunities for business success rather than constraints on it
- The last stretch covers managing people and agents side by side, and a career where the roles themselves change