AI So Whats
Edition

Friday, October 2, 2026

18 issues 23h window ~16 min read

Running through today’s issues

The price of a token is collapsing from three directions at once

4 sources

Clouded Judgement makes the structural case: decision models like Jev price at $0.042 per million input tokens and could absorb 10-15% of today's token volume. Every reports open models already carry 56% of Vercel AI Gateway traffic and 40% of AT&T's AI use. Anthropic's own migration note says Sonnet 5.5 at low or medium effort beats Sonnet 5's best score for a tenth of the cost per task. Frontier-priced inference is becoming the exception you justify, not the default you budget.

Enforcement is losing to demand on Chinese access to U.S. AI

3 sources

The Information's Beijing exclusive found six of thirty tenants in a single Haidian office building reselling Claude access Anthropic refuses to sell. The AM briefing adds a state-backed Chinese leasing firm funding B300 server purchases and a California arrest over $300M of servers routed through Malaysia and Singapore. The Briefing reports OpenAI mitigating a coordinated distillation campaign tied to Moonshot AI. Export controls and geo-blocking are both being routed around through intermediaries, and the labs are now policing it themselves.

The public market is repricing AI infrastructure faster than the private one

2 sources

Accelevation priced below its $20-24 range at $18 and kept falling to $16.59, carrying $647.8M of debt against $28M of cash. Sharon AI's $356M GPU-backed loan costs 9.95% before fees and knocked 8% off the stock on announcement. Meanwhile Nvidia and SoftBank closed the final $20B of OpenAI's $852B round and the next raise targets $1.4T. Anthropic's reported mid-November IPO is the test of which market is right.

Taste and feedback loops are the assets AI cannot manufacture

2 sources

Kieran Flanagan argues that as models converge on competence, content converges too, and the only differentiator left is encoded judgment — his taste.md file exists because the model will otherwise draw from the same statistical center as everyone else. Deb Liu reaches the same conclusion from the apprenticeship side: tacit expertise transfers only through high-frequency correction, and agents doing the first draft remove the taps on the shoulder that produce senior judgment. Both land on the same operational risk — if production is free and the correction loop disappears, the organization stops generating the people who can tell good from merely finished.

Superhuman AI 06:15 PT

👱 Can you tell this avatar is AI? 48% can't.

So what

Assume any video call can be synthetic: 48% pass rate means your portfolio companies need identity verification on hiring and finance calls now.

A consumer-tools roundup whose real signal is that synthetic humans crossed a credibility line and the week's model releases landed in a cluster.

  • Tavus launched Griffin, its first Human Interaction Model, claiming 48% of people who spoke to it believed it was human; demo hit 5M views in hours
  • Ideogram 4.5 ships iterative region editing without quality degradation across rounds
  • Factory CEO Matan Grinberg publicly accused Cognition of sending engineers as fake job applicants and of a former board member leaking; claims disputed
  • Week's releases: Gemini Argon, Claude Sonnet 5.5, GPT-6.1 Sol, ElevenLabs' most emotive voice model, OpenAI's ChatGPT Space at DevDay 2026, new Microsoft Copilot, SpaceXAI Team Bots
  • The DC safety agreement signed by tech heavyweights was light on specifics; Zuckerberg called it "a start"
model releases ×3 agent startups ai safety regulation synthetic media Open in Gmail →
Clouded Judgement 06:04 PT

Clouded Judgement 10.2.26 - Decision Models: The Next Shoe to Drop

So what

Reprice any portfolio company whose COGS assume frontier tokens for routing and classification — that line item just dropped by an order of magnitude.

Jamin Ball argues decision models are the next structural price cut in the token economy, following open weights, and that a meaningful slice of today's token volume migrates to them.

  • Jev from TypeSafe hit ~13% of paid teams on Vercel's AI Gateway within 24 hours of launch, the fastest adoption in gateway history; reportedly raising at $10B
  • OpenAI shipped a Decisions API, Databricks an ai_decide function, Perplexity an open-weight decision model, all in the same week
  • Jev quoted at $0.042 per million input tokens with free output; ~100x under frontier pricing, ~4x under GPT6-Luna
  • Use cases beyond classification: model routing, agent step control, LLM-as-judge on brand policy, contract flagging — all with a confidence score general LLMs don't emit
  • Claude's own estimate: 10-15% of tokens generated today are replaceable by decision models
  • Multiples: overall median 4.2x NTM revenue, top 5 median 38.3x, high growth 19.4x, median net retention 110%, median CAC payback 31 months
token economics ×2 model pricing ×3 saas multiples ×2 decision models Open in Gmail →
The Information 06:01 PT

Exclusive: How China’s Token Resellers Create an Anthropic Gray Market

So what

Your China-exposure diligence should now cover API resale, not just chips — revenue from banned geographies is flowing through intermediaries at scale.

A reported look inside the Beijing gray market that sells Chinese customers access to Claude and other U.S. models Anthropic refuses to serve.

  • In one ordinary Haidian District office building, six of roughly 30 tenants resell access to U.S. model APIs
  • Anthropic does not offer services in China on national security grounds, yet demand for Claude tokens is described as off the charts
  • Anthropic has called the extraction of its models' capabilities an "industrial-scale, covert campaign" using fake or stolen accounts, cards and identities
china export controls ×2 anthropic ×2 model distillation ×2 Open in Gmail →
Not Boring 05:44 PT

Weekly Dose of Optimism #213

So what

Matic's 16,000-home, 500M-square-foot data moat is the argument against humanoid-first robotics investments you are currently being pitched.

Packy McCormick's optimism roundup, newly produced with Runway-made video, runs from Starship's first orbital flight to evidence that transplanted hearts take on the recipient's biological age.

  • Starship reached orbit on its fourteenth flight and deployed 26 Starlink V3 satellites; one ascent engine shut down early and the mission was shortened
  • Joe Gebbia's National Design Studio launched America.gov, consolidating information from more than 29,000 government websites, with passport renewal and Medicare enrollment next
  • Matic has robots in 16,000 homes covering 500M square feet and unveiled a plan to add manipulation — floor pickup first, humanoid laundry folding later
  • Freckle launched a $199 kids phone with $25/month service running ClayOS: parent-approved calls, maps, AI object identification, no web, social or app store
  • Harvard/Brigham preprint: older donor hearts showed rejuvenation in younger recipients and vice versa across mice and 11 human biopsies; function tracked recipient age, not donor age
robotics ×2 consumer hardware longevity space Open in Gmail →
The Information 04:00 PT

OpenAI Fires Three Safety Researchers

So what

GPU-backed debt at 9.95% and a broken Accelevation IPO tell you the public market is now pricing AI infrastructure risk well above the private market.

The AM briefing stacks an OpenAI safety-staff firing against the closing of its $122B round and three separate stories about Nvidia chips leaking into China.

  • OpenAI fired three safety employees — one safety researcher, one alignment researcher, one alignment program manager — for sharing sensitive information with an outside evaluation organization
  • Nvidia and SoftBank each wired the final $10B of their $30B pledges; SoftBank's total hits $64.6B for ~13% of OpenAI, funded partly by an $11.1B high-yield bond. Next round targets $30B at ~$1.4T
  • State-backed Semi-Tech Leasing funded Glory View's purchase of 700+ servers, at least 32 with restricted Nvidia B300s; separately a California man was arrested for routing $300M of restricted servers via Malaysia and Singapore
  • A federal judge dismissed Chegg's and Penske Media's antitrust suits over Google AI Overviews, ruling crawling public sites isn't an agreement — leaving licensing as publishers' only path
  • Sharon AI raised a $356M GPU-backed loan at 9.95% pre-fees; stock fell 8%. Accelevation priced at $18 below its $20-24 range and kept sinking. Infinigence filed confidentially in Hong Kong at a $2.1B valuation
china export controls ×2 ai ipo market ×2 gpu-backed debt openai funding publisher litigation Open in Gmail →
Kieran Flanagan 03:57 PT

The Creative Context OS: How to Stop AI From Making Your Content Sound Like Everyone Else’s

So what

Swap your marketing team's prompt library for three context files; the scaffolding you paid to build is now the thing making output generic.

Flanagan argues that as models converge on competent prose, the output converges too, and the fix is three small context files rather than a prompt library.

  • creator.md is extracted from 5-10 of your own pieces that show repeatable style — not your highest performers, since one-off viral posts teach the wrong pattern
  • audience.md goes past firmographics to what the audience is bored of and what still feels novel; built from five named readers plus engagement data (he used YouSpot for LinkedIn)
  • taste.md encodes why you admire specific work — Ries and Trout, Kevin Kelly's 1,000 True Fans, Paul Graham, Ben Thompson — then converts that into publishable rules
  • He explicitly deleted his previous complex writing OS because the models got good enough to need less scaffolding
  • Use is ideation before writing and sparring after, not drafting: he fed an underperforming post to taste.md and it diagnosed too many ideas in one post
ai writing content marketing context engineering Open in Gmail →
AI with Allie 01:58 PT

Your replay is ready: Codex vs. Claude Code

So what

Run that blind HTML bake-off across your own team this week — it produces per-task cost and latency data your vendor negotiations currently lack.

A workshop recap positioning Codex versus Claude Code as a personal-fit question, with a reproducible bake-off protocol and a $2,800 cohort pitch attached.

  • Allie's daily driver is Codex; the rest of her team runs Claude Code — framed as personality fit, not capability gap
  • Alyssa benchmarked the personal agents Muse and Instinct on the same seven tasks: flights, shopping, inbox cleanup
  • Step one prompt mines your own session history: group the past month's work into recurring responsibilities, rank by frequency
  • Step two gives both tools an identical brief across Fable 5.1, Opus 5.5, Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, then has them build an HTML comparison page with blind model names, 1-5 scoring, run time, tool calls, tokens and cost
  • AI Agent Mastermind Cohort 3 runs Oct 19 to Nov 12, $2,800 until Oct 7
coding agents ×2 agent evaluation personal agents Open in Gmail →
The Information Briefing 17:01 PT

The Briefing: Microsoft’s Shake-Up

So what

Anthropic's mid-November window now depends on a market that just punished a data center builder below its range — watch that print.

Martin Peers reads the Microsoft executive exodus as veterans cashing out before AI reorders the software business, and the Accelevation IPO as proof the window is closing.

  • Ryan Roslansky (EVP Office apps) leaves at year-end, a day after Microsoft Science President Peter Lee, following Office and Windows chief Rajesh Jha — mostly 15-year-plus veterans sitting on a stock up 1,289% since 2014
  • Nadella is promoting younger outsiders: Asha Sharma on gaming, Charles Lamanna and Jacob Andreou on Copilot, with Snap and Instacart backgrounds
  • Accelevation priced at $18 below range, closed Thursday at $16.59, carries $647.8M debt against $28M cash and will retire only about a quarter of it
  • Bloomberg reports Anthropic is targeting a mid-November IPO, the most anticipated listing in the sector
  • OpenAI says it identified and mitigated a coordinated model-distillation campaign involving people associated with Moonshot AI
ai ipo market ×2 anthropic ×2 model distillation ×2 microsoft Open in Gmail →
The Information 16:15 PT

Muse Lifted Meta 27%. Can It Last?

So what

Amazon blocking Muse is the whole thesis: agentic commerce take rates require retailer consent, and the largest retailer just said no.

A Meta-focused collection asking whether the Muse launch was a business or a very good demo, now that the stock has already repriced on it.

  • Meta stock rose 27% in the weeks after Muse's Sept. 8 launch; Amazon has already blocked Muse from its site, undercutting the take-rate model
  • One analyst estimate says users would need to spend over $1,000 a month through Muse for it to turn a profit
  • OpenAI is building features to counter Grok Bot and weighing a direct Muse rival; Meta One sells extra AI usage against free rival chatbots
  • Meta poached MongoDB's CEO to sell AI into enterprises, a market where it has failed before
  • Camera-free Luna glasses let users talk to Muse while defusing privacy objections; the display glasses drew hype but little sales disclosure
  • The Information's AI Inflection Index shows its agentic commerce coverage has more than tripled in a year
agentic commerce consumer ai agents meta muse Open in Gmail →
The Information 14:50 PT

Applied AI Exclusive: In Robotics, Google Goes Android to Tesla’s Apple

So what

If you hold robot hardware positions, price in a Gemini Robotics commoditization of the intelligence layer and a Google first-party chassis later.

Jessica Lessin frames Google DeepMind's robotics strategy as the Android playbook — license Gemini Robotics into everyone's hardware — and predicts Google eventually builds the chassis anyway.

  • New DeepMind CEO Koray Kavukcuoglu declined to rule out a Google humanoid but located the company's edge at the model layer, not the body
  • Gemini Robotics is going into partner hardware including Boston Dynamics, explicitly contrasted with Figure's and Tesla's single-flagship bets
  • Lessin's analogy: Sergey Brin in 2007-08 said Google would build a phone "only if it had to," then shipped HTC partnerships and eventually Pixel
  • Kavukcuoglu ties robotics to the AGI thesis — one multimodal Gemini handling text, vision, audio and physical control — and name-checks Waymo as the scaled proof point
  • Gemini 4 Argon shipped Wednesday to a select cyber-defense cohort ahead of broad release, the test of whether Google has closed the gap to OpenAI and Anthropic
robotics ×2 model releases ×3 google deepmind Open in Gmail →
Claude Code 14:23 PT

How to turn off upfront thinking on Claude Sonnet 5.5

So what

Drop your default effort to medium on migration — you are paying roughly ten times the necessary cost per task if you leave it at high.

Anthropic's migration note for Sonnet 5.5 makes a breaking change to thinking controls and quietly resets the cost curve for agentic work.

  • thinking: {"type": "disabled"} now returns an error on Sonnet 5.5; use between_tools instead, available at low, medium and high effort
  • Sonnet 5 users running thinking off at xhigh or max should map to high; API default is high but medium is recommended for well-specified agentic work
  • On several benchmarks Sonnet 5.5 at low or medium effort beats Sonnet 5's best score for roughly a tenth of the cost per task
  • Forced tool choice is no longer supported; use structured outputs for guaranteed formats
  • The claude-api skill in Claude Code will migrate model names, prompts and settings via /claude-api migrate
model pricing ×3 coding agents ×2 model releases ×3 Open in Gmail →
James Clear 13:33 PT

3-2-1: On staying in the game, the power of best practices, and how to handle bad events

So what

The one usable prompt: audit which hours of your week you actually control, then defend them against the next deal cycle.

A standard 3-2-1 issue on not relitigating the past, borrowing industry best practice as starting material, and staying in the game long enough to be found.

  • "Don't let one bad event beat you twice" — reflect, learn, move on, but stop replaying old injustices
  • Opportunities that fit you tend to find you; keep energy high and keep working after a deal falls through
  • Collect your industry's best practices as a starting list, then combine, iterate and distill to your strengths
  • Henry Miller on developing interest in the world and forgetting yourself; Engels on an ounce of action over a ton of theory
  • The question: which hours of the day are your favorite, and what adjustments put them under your control
judgment ×2 personal operating system Open in Gmail →
Sacra 11:00 PT

kanbanization of ai medical scribes

So what

Freed's drop from 66% to 22% growth is your template for what happens to every single-feature AI app once an incumbent bundles it free.

Sacra calls the standalone AI medical scribe dead — absorbed as a free bundled feature — and reads Freed's decelerating growth as the proof.

  • Freed hit an estimated $33M ARR in August 2026, up 22% year to date, versus $27M at end-2025 when it was growing 66% YoY
  • Heidi Health reached an estimated $50M ARR in 24 months (March 2026), valued at $900M after a $100M Series C led by Blackbird, with enterprise sales layered on PLG and an ex-U.S. focus
  • Scribing is now bundled free into OpenEvidence, Doximity and Epic — the same pattern as Trello's kanban and Gong's meeting recording getting absorbed
  • Freed's self-serve PLG and downmarket positioning took the worst of the pricing pressure and incumbent distribution
  • The counterattack is workflow expansion: Freed into medical search, AI coding and AI receptionist; Abridge and Ambience into revenue-cycle workflows for large health systems
saas multiples ×2 commoditization healthcare ai vertical ai Open in Gmail →
The Information 10:30 PT

Last day: Lock in WTF Summit pricing to join leaders from Goldman, Uber, Google and NYSE in Napa

So what

Amanda Peterson Corio and Lynn Martin in one room is the data-center-energy-plus-listings conversation you would otherwise pay a banker for.

A last-call promotion for the WTF Summit in Napa on October 27-28 as ticket prices rise.

  • Confirmed speakers include Whitney Wolfe Herd (Bumble), Anne Wojcicki (23andMe Research Institute), Lynn Martin (NYSE Group), Kim Posnett (Goldman Sachs investment banking co-head)
  • Also Jill Hazelbaker (Uber), Amanda Peterson Corio (Google, data center energy), Pauline Brunet (SpaceXAI forward deployed engineering), Ami Badani (Arm)
  • Subscribers get an additional $200 discount by logging in before purchase; PwC presenting, Legora premier partner
events Open in Gmail →
McKinsey & Company 09:48 PT

Global Energy Perspective 2026

So what

Ask your data center exposures which signposts would make them change course on power procurement; most have scenarios but no trigger conditions.

McKinsey's Global Energy Perspective 2026 argues the energy system is getting less predictable, with AI power demand now a named structural driver alongside geopolitics and affordability.

  • Affordability and security have moved up alongside decarbonization as defining planning considerations
  • AI is called out as a new source of power demand reshaping the demand curve
  • The recommendation is to build for multiple plausible futures and define the signposts that trigger a course change, not to make a single bet
  • Explicit warning that scenario planning is not a substitute for strategic action and can produce decision paralysis
  • Companion webinars October 8 (energy) and October 14 (materials)
data center power energy Open in Gmail →
Deb Liu (Perspectives) 09:01 PT

Chicken Sexing: How to Build Expertise One Tap at a Time

So what

Your junior hires lose the correction loop when agents do the first draft; design explicit feedback reps or you stop producing senior judgment.

Liu uses chicken sexing — a skill experts can't articulate and can only transmit by tapping an apprentice on the shoulder — to argue that tacit judgment becomes the scarce asset as AI absorbs everything written down.

  • The vent-sexing technique came from Japan, reached the U.S. in the 1930s, and by 1937 had its first American school; it passed through Japanese, then Korean and Mexican immigrant workforces
  • Apprentices learn only by calling each chick themselves and being corrected immediately — watching doesn't transfer the skill
  • Her claim: product sense, medical instinct and legal judgment are all built the same way, through high-frequency correction, not documentation
  • The AI risk is that it becomes too easy to sort the chicks without ever learning whether you were right — the feedback loop disappears
  • Prescription: ask for frequent small taps ("what is one thing I could have done differently?"), get reps, and teach it to prove you have it
judgment ×2 ai and work talent development Open in Gmail →
Every 07:57 PT

Getting Started With Open Models

So what

If 56% of Vercel gateway tokens are already open-weight, your frontier API spend has a defensible floor well below what you're paying.

Kai Zau's guide argues frontier models are overkill for most work and that open models now cover a growing band of commodity intelligence you can own outright.

  • Vercel reports 56% of tokens on its AI Gateway ran through open models in August; AT&T is at 40% of AI use and expects up to 60%
  • On-device AI is the headline feature of Apple's latest Mac Studio and Mac Mini; the author runs millions of tokens a day on a Mac Studio
  • The framing is commodity intelligence: tasks where open models match frontier quality at a fraction of cost, versus the ones where you still reach for the frontier
  • Practical content covers model tiers, hosting choices, and wiring open models into coding agents and automated workflows
  • Benefits claimed: data control, per-model fit, and a fallback when a provider goes down
token economics ×2 model pricing ×3 open models Open in Gmail →
McKinsey & Company 07:37 PT

Author Talks: The business of owning a sports team

So what

Skip unless you are underwriting a sports franchise; the governance-in-public framing is the only transferable part.

An Author Talks interview with Baltimore Orioles owner and Carlyle co-founder David Rubenstein on what sports ownership demands that ordinary business ownership does not.

  • Rubenstein covers the leadership mindset shift required when your asset performs in public every night
  • Culture building and long-term thinking are framed as the core constraints of team ownership
  • Packaged with related McKinsey pieces on what elite athletes and coaches do differently and the Josh Harris playbook
leadership sports business Open in Gmail →