AI So Whats
Edition

Thursday, October 22, 2026

19 issues 23h window ~17 min read

The defender's window closed this week, and the harness is why

4 sources

The PHP patch landed only because an OpenAI-led coalition had worked the vulnerability since August; The Information Briefing makes the two-month lead time explicit and notes Z.ai's GLM-5.3 sits about four months behind US frontier models with none of the capability restrictions. Applied AI then shows what happens on the other side: Cybercartel automated bank intrusions with a harness that ran Grok for reconnaissance, Claude for execution, and Kimi whenever Claude refused. Superhuman AI and Ben's Bites both treat the response as a prompt you run yourself rather than a vendor patch you wait for. Defence at the model layer is already obsolete, because attackers swap models and defenders cannot.

Non-voting equity is now the standard instrument for AI capex

4 sources

Meta's SEC filing says management presented three business scenarios in May, each requiring hundreds of billions of dollars starting as early as fall 2026, and that the committee wanted capital secured before competitors saturate funding options — naming Alphabet, SpaceX, Anthropic and OpenAI. Class C exists so Zuckerberg can sell without losing votes. AI Market Fit has SpaceX, now public, claiming a $1 trillion ARR path while floating a Tesla merger and a chip fab. The Information Finance has Anthropic's advisers saying its IPO waits for someone else to move first. Moffett, in Stratechery, says most of the numbers people are watching are noise; the financing structure is where the risk sits.

Every proactive assistant shipped at once, and none of them can be trusted with a connector

4 sources

Duet, Dots, Muse, Instinct and Grok Bot are all now live, all as betas. AI with Allie documents Instinct reporting a $3.99 balloon as $34 with no call transcript, and Muse sharing a user's home address on Marketplace. Every's Vibe Check finds Duet good enough to split Shipper's usage away from ChatGPT while failing to send two emails and blocking a Gmail reply with no explanation. Superhuman AI adds the structural constraint: Amazon, Yelp, eBay, Zillow and several airlines are blocking agents outright, so the assistants' reach is shrinking while their reliability is still unproven.

🚨 Black October

So what

Any commerce position you hold assumes retail-media subsidy; agent checkout removes the ad click that pays for free delivery and low prices.

The issue treats the PHP disclosure as an internet-wide emergency and pairs it with Anthropic's Duet launch. Its original argument is "agent-flation": agent checkout breaks the retail-media subsidy that keeps consumer prices low.

  • Two critical PHP flaws hit a language behind 70% of websites; the NSA recommends pulling anything that can't be patched, and government sites went dark
  • OpenAI ran a coordinated awareness effort and shipped a Codex skill to let companies check their own exposure
  • Claude Duet runs on Opus 5.5, carries over existing memory, skills and connected apps, and is waitlisted for US Pro and Max
  • Amazon booked $68B in ad revenue in 2025 and Walmart $6B; agents don't click sponsored posts, so that subsidy has to come back as prices or fees
  • Amazon blocked Meta's Muse in September; Yelp, eBay, Zillow, Pizza Hut, Adidas and several airlines have followed
php vulnerability ×5 claude duet ×4 agent payments retail media Open in Gmail →

Claude’s personal agent

So what

Your exposure audit is one Codex prompt now, and AgentReady is the spec to check your portfolio products against.

Ben Tossell rebuilt his bookmark-sorting pipeline as Keshav leaves at year end, and reset the sponsorship model to four companies per quarter from 2027, with no sponsor sign-off on what he writes.

  • Anthropic's Duet is a separate US-only mobile app in research preview, so he can't use it
  • PHP and Railway both had major issues; he audited his own exposure with one Codex prompt asking it to check, assess and address
  • Phone-use agents now drive real Android and iPhone hardware, extending the browser-use pattern
  • Anthropic reports 17 of 211 Claude-designed protein binders against scorpion neurotoxins met its experimental binding criteria
  • AgentReady is an open spec for making products agent-usable across finding, reading and acting; Dan Shipper argues for a standing frontier team inside companies
php vulnerability ×5 claude duet ×4 agent memory ai agents Open in Gmail →

Exclusive: Andreessen Horowitz to Drop ‘Crypto’ From Name of Flagship Blockchain Fund

So what

The largest dedicated crypto franchise just repositioned as generalist frontier tech; price your crypto marks against capital that is leaving the category.

a16z crypto is renaming itself a16z frontier and broadening its mandate beyond crypto, according to LPs who were briefed.

  • The fund has raised nearly $10 billion across five vehicles, one of the largest dedicated blockchain bets in venture
  • General partner Ali Yahya is leaving to work on his own startup; he joined in 2017 as the first full-time member of the crypto team
  • The strategy shift was communicated to limited partners rather than announced publicly
a16z crypto fund strategy venture capital Open in Gmail →

Every needs your help

So what

Five minutes buys you input into what Every builds next, including the tools you already pay for.

Housekeeping send. Every is asking paid members what is worth paying for and what isn't.

  • Member survey, roughly five minutes, for paid subscribers only
  • Answers will feed decisions about what the team writes and builds next
  • No editorial content in this issue
every newsletter business reader survey Open in Gmail →

Meta Creates New Share Class to Pave Way for AI-Related Fundraising

So what

Non-voting stock is now the AI-capex playbook; expect every founder-controlled company you hold to copy Meta's Class C structure.

Meta created a non-voting Class C share class specifically so it can raise AI capital without diluting Zuckerberg's control, and the SEC filing shows management expects to need hundreds of billions starting as early as this fall.

  • Q3 revenue rose 26% to $64.6B, but operating income fell 37% to nearly $13B and margin narrowed to 20% from 40% as expenses jumped 68% to almost $52B, including a $9.9B settlement charge
  • Capex guidance floor rose to $135B-$145B; free cash flow was negative $2.1B against positive $10.6B a year earlier
  • The special committee cited the need to secure capital before competitors saturate funding options, naming Alphabet, SpaceX, Anthropic and OpenAI
  • Zuckerberg flagged California's Proposition 40 wealth tax and threatened to list his Palo Alto home before the stock deal was agreed
  • Muse has 10M+ US and Canada users in six weeks with average purchase over $200; a New Mexico judge ordered $25B over Cambridge Analytica, reduced from a possible $219.5B, with 15% annual interest
  • Microsoft nominated Stripe co-founder John Collison to its board and cut Nadella's package 6% to $90.6M on Xbox revenue
php vulnerability ×5 stripe ×2 ai capex financing meta capital structure Open in Gmail →

SpaceX Q3 2026: Elon's $1 trillion ARR claim

So what

The $1 trillion ARR claim is marketing; TeraFab and a Tesla merger are what change your exposure across the Musk complex.

SpaceX's first earnings as a listed company reframed it as an AI infrastructure business that happens to own most of the launch market. Two side announcements carried the call.

  • Q3 2026 reported live from the Starlink factory in Bastrop, Texas, with Musk, Shotwell and CFO Bret Johnsen across 67 minutes
  • Musk claimed a path to $1 trillion in ARR
  • A potential Tesla merger and a chip fab called TeraFab were both floated
  • The company holds roughly 90% of mass-to-orbit
spacex ×2 ai infrastructure elon musk tesla Open in Gmail →

An Interview with Craig Moffett About SpaceX and Bubbles (Stratechery Interview 10-22-2026)

So what

Starlink's margins are an accounting choice, not a result; model launch and Starlink together before you underwrite the broadband story.

Craig Moffett argues SpaceX is misread as a space company: the terrestrial Grok and neocloud business is the largest revenue driver, launch is the actual monopoly, and Starlink's reported margins are an accounting artifact.

  • He calls the $28 trillion TAM in the S-1 idiotic, roughly the size of US GDP
  • SpaceX charges Starlink only the marginal cost of each launch, overstating Starlink profitability and understating launch; Moffett puts the space segment at three times what the accounting suggests
  • Launch is effectively a monopoly accumulating knowledge at something like 10x the rate of any competitor, which is where the optionality sits
  • MoffettNathanson initiated at neutral under the title Flywheels and the Unknown Unknowns, with Nick Del Deo on data centers and Julie Zhu on the space segment
  • The interview closes on dot-com math, and why Moffett thinks most of the numbers people are watching now are noise
spacex ×2 bubble math starlink telecoms Open in Gmail →

France and the limits of liberalism

So what

French sovereign risk is political with no exit path; treat your euro duration and France-exposed positions as repricing, not noise.

Noah Smith reads the French riots as the limit case of European welfarism: benefits ratchet upward, austerity never arrives, and bond markets have started pricing default rather than adjustment.

  • Roughly 2,400 arrests, several hundred million dollars of damage, and the Limoges city hall spire collapsed in a fire
  • French CDS spreads have widened faster than yields, meaning investors doubt repayment rather than just demanding term premium
  • Retirement age was 62 in 2023 and the phased rise to 64 is frozen until after next year's election; actual retirement age is the lowest in Western Europe
  • The old wealth tax raised about 1% of total revenue while 10,000+ wealthy residents left; Pichet estimates net revenue loss at twice the take, and the 75% supertax raised €160M in 2014 before repeal
  • France ran neither austerity nor stimulus through the 2010s, which undercuts the Keynesian prescription that boom-time austerity funds bust-time spending
fiscal policy france sovereign debt welfare state Open in Gmail →

$100m/year scale ai for folding laundry

So what

Mecka's 4x against Mercor's 22x tells you the market prices robotics data as commodity; worker supply is where you'd underwrite.

Mecka pays workers to record themselves folding laundry and making coffee, then sells the demonstrations to robotics labs. Sacra puts it at $100M annualized revenue as of June 2026.

  • Workers wear headgear and gloves and record on a head-mounted iPhone, capturing video, 3D position, force and pressure data
  • Two channels: a consumer gig app, Atlas Capture, and a B2B arm partnering with bakeries, repair shops and warehouses in the Philippines
  • It supplied household-motion data for 1X's world model; the $60M October Series B was led by Sequoia with NVIDIA, M12, Qualcomm and Samsung at a $500M valuation, roughly 4x revenue
  • Comparables run far hotter: XDOF near $50M ARR talking at $1.2B (24x), micro1 at $1B ARR, Mercor at $2B ARR valued at $10B (22x)
  • $47B was raised globally for physical AI in H1'26, and Mercor, Scale, Turing and micro1 are all pushing into robotics data
data labeling physical ai robotics data valuation multiples Open in Gmail →

The Briefing: Another Day, Another Cyberscramble

So what

Your patch cycle now has to beat a four-month model gap, not a two-month coalition; budget incident response on that clock.

Leo Schwartz argues the PHP patch only landed because defenders had a two-month head start, and that the defender's window is about to close.

  • An OpenAI-led coalition had been working on the PHP vulnerabilities since August before Wednesday's patch
  • Anthropic's late-September post put Z.ai's GLM-5.3 roughly four months behind US frontier models, without the capability restrictions US labs impose
  • The new Super Intelligence Force, led by Jay Clayton and Emil Michael, was briefed late last week in its first major engagement with OpenAI, and pressured relevant parties to ship the patch
  • A former White House official says the July Gold Eagle vulnerability clearinghouse may already be outdated; the April Mythos reorientation of AI policy may need a sequel
  • Meta's quarter reassured investors enough for a 2% after-hours rise; Tesla's revenue was flat at $28B and shares fell 3%
php vulnerability ×5 ai cyber capability ×2 china open models ×2 ai policy Open in Gmail →

She's led IPOs for Airbnb, Uber, and Netflix. Hear what she's learned.

So what

If you're taking anything public inside two years, request the invite; this is the bank-and-exchange side of the table in one room.

Promotional invite for a November 10 panel on taking companies public, pitched at executives 18 to 24 months out from a listing.

  • Kate Claassen, Morgan Stanley's Chairman of Global Internet Banking, has led IPOs and financings for Airbnb, Amazon, Meta, Netflix, SpaceX, Tesla, Uber and xAI
  • She also advised on Musk's $44B Twitter acquisition and Block's $29B Afterpay deal
  • Crystal Shen of the NYSE joins, previously running liquidity strategy at Carta and originating tech IPOs at Jefferies
  • Dakin Campbell, who covers the money behind the AI boom, moderates; 5:30pm PT at The Howard in San Francisco
ipo market ×2 capital markets events Open in Gmail →

Confused by all the new AI assistants? Start here.

So what

Test any of these with zero connectors first; Instinct misquoted a price tenfold and left no call record for you to audit.

Allie Miller reviews the proactive assistant category now that Duet has joined Muse, Dots, Instinct and Grok Bot, and her verdict is that all of them are betas you shouldn't wire to real systems.

  • Duet is the most proactive and the most business-leaning, but over-flags; US waitlist on Pro and Max
  • Dots has the shortest time to value if your tools are already in ChatGPT; she gave it $500 to plan a mystery day and had to apologize to two restaurants and one friend
  • Instinct, from Spear Street ($1B raised), lives in iMessage and can connect to other users' Instincts; it reported a balloon at $34 that actually cost $3.99, with no transcript or recording of the call
  • Muse has the strongest idea generation and real privacy incidents: leaked private conversations, and a shared home address on Facebook Marketplace
  • Her process: pick one, run it with zero tools connected, then add connectors one at a time
claude duet ×4 proactive assistants ×2 agent reliability ×2 openai dots Open in Gmail →

Vibe Check: Duet—Anthropic's Always-on Agent

So what

Duet is good enough to split your usage away from ChatGPT, but keep email out of it until sends stop failing silently.

Dan Shipper used Duet daily for two weeks and calls it the first personal agent that qualifies as a pro tool. Every limitation he found is plumbing, not intelligence.

  • Research preview on iOS, Android and web, running on Opus 5.5, first access for Pro and Max subscribers off the waitlist
  • It moved him to roughly 50/50 between ChatGPT and Claude, displacing time on Dot
  • Duet can see the Claude sessions on his machines but can't direct them, and scheduled runs can't reach custom connectors
  • Email is the weak point: meeting times proposed well past working hours, a Gmail reply blocked with no explanation, and two sends that never went out
  • Kieran Klaassen says he doesn't trust the link to his machines yet, and it can't replace his eight Claude Projects
claude duet ×4 proactive assistants ×2 agent reliability ×2 anthropic Open in Gmail →

Can Private Equity Still Win in Cybersecurity?

So what

Price your security holdings on per-agent revenue, not seat counts; that is the screen a $6.3B buyer just used.

Thoma Bravo's Proofpoint is buying Varonis for $6.3B, which Alex Eichenstein reads as the template for how private equity still gets into a sector corporate buyers have bid up. The issue also tracks Anthropic's IPO timing.

  • Varonis goes at roughly 6x forward ARR against Qualys at about 8x; its stock had fallen nearly 50% from last October to around $33 in June during a licence-to-cloud transition
  • Proofpoint generates over $1B in EBITDA and was taken private at $12.3B in 2021, putting it at the point where exit pressure builds
  • Cyber M&A hit a record 170 transactions in Europe and North America in Q3, after Alphabet-Wiz at $32B and Palo Alto-CyberArk at about $25B
  • Thoma Bravo's Andrew Almeida wants agentic revenue, not an agentic vision, plus pricing that captures agent growth rather than human seats, and expects another Varonis-sized deal within 12 months
  • Anthropic's filing could surface by week's end, but Edward Best says it won't open the floodgates; it will wait until the water moves, then widen the river
private equity ×2 ipo market ×2 anthropic ipo cybersecurity m&a Open in Gmail →

Enovix’s Drone Battery Deal Shows How Going Exotic Doesn’t Always Cut It

So what

Defence buyers pay premiums for boring chemistry; underwrite any next-gen battery position you hold on conventional cells shipping now.

Enovix landed a 100,000-unit combat drone order out of nowhere, and it was not for the stainless-steel caged battery it spent a decade building. Steve LeVine's point is that the drone boom rewards companies willing to drop their exotic chemistry.

  • The company promised investors roughly $1B in annual sales by this year; it now guides to about $41M
  • The single drone order exceeds all of Q2's $9M revenue, and T.J. Rodgers expects a follow-on tranche in the second half of next year
  • The order is for conventional architecture with a silicon anode and no cage; Rodgers says military customers pay a premium and are good for a manufacturer still immature on cost
  • SES AI will triple Korean drone battery capacity to 3 million units a year by Q2 after winning no EV customers for its lithium-metal cells
  • Agility Robotics picked Nyobolt graphite-anode cells for the next Digit, not Nyobolt's niobium technology; Enovix stock is down 68% over the year
batteries defence demand drones humanoid robots Open in Gmail →

Revising the private equity playbook

So what

Entry multiple now decides your outcome: 2018-23 vintages returned 3.4x MOIC at the low end against 2.7x at the high end.

Alexander Edlich argues private equity's cheap-debt-and-multiple-expansion model is finished, and the data behind that is a maturing industry rather than a rough patch. Three consecutive years of underperforming public markets.

  • Returns have fallen into the high teens from a 20%-plus standard; new firm formation dropped about 18% a year from 2020 to 2025, with a record 304 firms going inactive and the fewest first-time funds in a decade
  • 2025 deal value rose 19% to $2.6T while deal count fell for a fourth year; buyouts were down 5% and venture down 10%
  • Exit value rose over 40% to $1.2T on 15% fewer exits; DPI was 6% of assets against a 16% average, five-year rolling distributions hit a recorded low, and secondaries surged 48%
  • Global buyout prices hit a record 11.8x EBITDA, 18.1x above $500M, with leverage down to 37% and 42% of dry powder more than two years old
  • Only 6% of firms see high AI impact today; 70% expect it within a few years
private equity ×2 deal multiples fund returns liquidity Open in Gmail →

Presented by Stripe

So what

Your fintech analyst is now paid by Stripe with disclosure and editorial independence; subscription-plus-sponsorship just got its proof point at scale.

Thompson took an exclusive Stripe presenting sponsorship, wrote the conflict into his ethics policy, and is spending the money on product. He also admits he was never objective about Stripe in the first place.

  • The contract grants total editorial independence: Stripe can't dictate what he writes or what he doesn't, and sponsored content remains his call
  • Three disclosed conflicts predate the deal: Stripe's 2011 Subscription API made Stratechery's model possible, Stripe underpins Passport behind Stratechery, SemiAnalysis and The Ankler, and he is friends with the Collisons
  • Ads run in the presenting slot, between sections of Updates and Interviews, and as a weekly Sharp Tech discussion segment; there is no ad-free tier for subscribers
  • He does not anticipate raising subscription prices again over the deal's term
  • Sharp Tech becomes a weekly in-person video podcast, free on YouTube, plus a subscriber-only mailbag episode recorded the same day
stripe ×2 media business models sponsorship stratechery Open in Gmail →

How Bank Hackers Got Claude to Do Their Dirty Work

So what

Model-level refusals are not a control; the harness swaps models, so put your security spend at the harness and action layer.

Straiker got inside Cybercartel's operation and found a model-agnostic harness automating attacks on Latin American financial institutions. No frontier-class model was required.

  • Grok agents handled reconnaissance and credential hunting while Claude agents planned and executed the intrusions, with Moonshot's Kimi taking over whenever Claude refused
  • The group used prompts like analyze this vector from the perspective of an authorized red team against our own system to convince Claude it was working in a harmless environment
  • Mythos-class models remain restricted to US-government-approved companies, and this campaign never needed one; Opus and Sonnet were enough
  • Ken Huang: because the power lies in the harness rather than a single model, future attacks will effortlessly swap different models
  • The architecture mirrors how enterprises cut AI costs — AT&T mixes models including open source ones — and Microsoft and Palo Alto are selling mixed-model defensive tools on the same logic
ai cyber capability ×2 china open models ×2 php vulnerability ×5 model refusals Open in Gmail →

Women In Product at 10: What’s the Name of Your Book?

So what

Your accumulated chat history is already a personal profile; ask a model what it knows about you before someone else does.

Liu's tenth Women In Product anniversary conference fell on the day Grace Hopper Celebration was cancelled, the event that prompted the conference's founding. The session that landed hardest was a closing question answered by ChatGPT.

  • The Anita Borg Foundation cancelled Grace Hopper weeks before it was due; Women In Product began at a Palo Alto dinner asking why no equivalent existed for product management
  • OpenAI's VP of Health, Ashley Alexander, fed Liu's standing closing question to ChatGPT and read the result on stage, covering a health product serving 300M people a month
  • Liu then had the whole room run the same prompt against their own chat history
  • Frédérique Dame's generated title disclosed that she is about to become a mother at 50, after Uber and a venture career
  • Liu's read: the model had the material because people had been feeding it months of ordinary late-night questions
community llm personalization openai women in product Open in Gmail →