AI So Whats
Edition

Tuesday, September 15, 2026

21 issues 23h window ~16 min read

The pacing debate collapsed into politics within 48 hours and settled nothing

7 sources

Amodei's Pace the Frontier essay drew endorsements from Altman, Nadella, xAI and Meta leaders — Microsoft went furthest, publishing a code of conduct promising legible reasoning, kill switches and openness to third-party audits. Trump called existential risk a hoax and a "SICK conspiracy," phoned Jensen Huang on stage at All-In, and Beijing called it a containment tactic; David Sacks split the difference with slow down if you want, don't impose it. Martin Peers puts the odds of a coordinated slowdown at zero, noting Meta stayed silent and Chinese labs kept shipping — Shanghai AI Lab released Atria Dawn Preview the same week. Meanwhile Noam Brown told The Information that recursive self-improvement is OpenAI's number one training priority by a wide margin, which is the exact capability the proposals target.

Safety proposals are also competitive strategy, and the money never paused

4 sources

AI Market Fit's read is that mandatory independent evaluations and deployment controls form a compliance stack only incumbents can fund — a regulatory moat dressed as caution. The Monday Readout supplies the evidence for the same double game: Anthropic's researcher put internal extinction odds above 10% while the company signed a $13.7B compute deal with Trump-linked Rum Group and prepared an IPO, and it, OpenAI and Google quietly discussed a self-policing standards body. Microsoft volunteering for third-party audits fits the pattern: the firm with the deepest balance sheet endorses the most expensive oversight. Shield AI raising at $20B+, a 60% markup in five months, shows which corner of the sector nobody is asking to pace.

Markets have already priced the winners and losers of a slowdown that will not happen

3 sources

Monday's tape was a clean one-day experiment: ServiceNow +7%, Salesforce +5%, Shopify and Figma up, against Nvidia -3% and Micron -5%, with neoclouds falling in line. Enterprise software, long treated as AI roadkill, rallies on any hint of deceleration; compute suppliers own the downside. ByteDance's H1 gives the underlying mechanic from the other direction — revenue accelerating to $120B while net profit fell to $20B because AI investment is eating margin. The trade is legible; Peers's point is that the premise is not.

Personal agents are the new consumer land grab, and nobody has cracked the experience

4 sources

Meta's Muse hit #2 in US App Store free apps with 83,000+ downloads, behind only ChatGPT, on distribution rather than model quality. Instinct raised $350M at a $2.5B valuation off a rumor launch — waitlist page, no screenshots, VCs posting with nothing to demo — and Tom Orbach holds it up as a repeatable growth tactic. Yet Ben Tossell, who actually got access, found it slow, uncomfortably proactive and opaque about what it remembers, and The Information reports Instinct now faces a compute crunch that may force new funding. The How I AI episode points at what works instead: narrow, single-purpose bots whose conventions and context you configure yourself.

✋ Jensen and Trump hit back at AI labs

So what

Meta's distribution muscle just put a personal agent at #2 without a better model — factor that into any consumer agent you back.

Trump publicly refused any AI slowdown, phoning Jensen Huang on stage at All-In to say so, while China dismissed Amodei's essay as containment strategy. The issue's other signal: Meta's Muse agent is actually getting installed.

  • Muse hit #2 in US App Store free apps with 83,000+ iOS downloads per Sensor Tower, ahead of Threads, WhatsApp and Facebook, behind only ChatGPT
  • Alexandr Wang credits Meta's new Muse model family for the agent's performance
  • Shanghai AI Laboratory released Atria Dawn Preview, an open-weight research model claiming parity with Kimi K3 and Opus 5 on some metrics, unverified
  • HHMI Janelia and Google Research mapped 166,000 neurons and 125M connections in a male fruit fly brain; developers immediately had it play blackjack and Beat Saber
  • Siri's iOS 27 AI update ships today
ai safety pacing ×8 personal agents ×3 china models ×2 meta muse Open in Gmail →

Good luck slowing this down

So what

The hype gap on Instinct is your entry point: proactive agents that hide their context lose to ones you can inspect and edit.

Ben got into Instinct, the $2.5B personal agent every VC is posting about, and found it meh — proactive to the point of feeling like a boss, slow, and opaque about what it remembers. His broader call: every AI company will ship a personal agent, the way every app shipped a chat box.

  • Amodei's Pace the Frontier essay drew agreement from Altman, rejection from Trump, and a split from David Sacks: slow down if you want, don't impose it
  • Altman told Fortune an OpenAI IPO in 2026 would be ill-timed given alignment and control work ahead
  • OpenAI shipped two APIs: GPT-Live-1 (the Voice mode model) and an Agents API that sends tasks to a Codex-like agent without infra config
  • Cognition's SWE-2, built on Kimi K3, beats Grok 4.6 at half the cost; Bolt Forge puts GLM, DeepSeek and Kimi in Bolt free until Oct 14
  • tldraw's founder rebuilt ChatGPT's Sketch in a day after an OpenAI challenge
ai safety pacing ×8 personal agents ×3 open weight models openai api Open in Gmail →

60+ new creative growth ideas

So what

Instinct's rumor launch is the replicable asset here: invite-only plus respected posters with nothing to show beats any demo video.

Tom Orbach hands over 64 unconventional growth tactics, framed as lottery tickets rather than a playbook, with a Markdown version built to feed into an agent. The thesis: everyone runs the same channels, so attention only goes to what nobody saw coming.

  • His fake toy store for CISOs (panic button, blindfolds for legal) cost almost nothing, went viral on Reddit, drew 46,000 visitors in 24 hours and hundreds of enterprise leads
  • Instinct launched as a rumor: waitlist page, no team or screenshots, VC posts with no demo — by reveal it had raised $350M at a $2.5B valuation
  • 1up says a third of leads come from memes; tl;dv got a third of signups from social and once promoted all 50+ employees to CEO on LinkedIn the same day
  • Data-as-PR works: Carta's State of Private Markets, and the Ramp AI Index showing Anthropic passing OpenAI in workplace adoption, which Axios built a story on
  • Rilla got Fortune coverage from a job listing pairing a $1,500/month rent stipend with a blunt 70-hour-week warning
personal agents ×3 growth tactics marketing Open in Gmail →

Exclusive: ByteDance’s First-Half Profit Drops to $20 Billion, Weighed Down by AI Spending

So what

The most profitable private company in AI just showed margin compression from capex — expect the same pattern in every hyperscaler model you run.

ByteDance's first-half net profit fell a single-digit percentage to $20 billion as AI investment ate into margins, even as revenue accelerated.

  • H1 revenue rose ~30% year on year to $120 billion, driven by TikTok international advertising and e-commerce
  • That is a marginal acceleration on 29% growth in 2025 (to ~$200 billion) and a similar rate in 2024
  • 2025 net profit grew 27% to $42 billion; this year's profit line has now turned negative on a percentage basis
  • Sourced to three people with knowledge of the results
ai capex ×5 china models ×2 bytedance Open in Gmail →

Microsoft Joins Calls for AI “Pacing,” Pledges to Build Kill Switches Into AI Products

So what

Microsoft accepting third-party audits makes evaluation infrastructure a real procurement line — and a compliance cost your portfolio's smaller labs can't absorb.

Microsoft joined the pacing camp with an AI code of conduct and kill-switch pledges; Trump called AI existential risk a hoax the same day. The gap between industry and the White House is now the story.

  • Microsoft promised not to build models whose behind-the-scenes reasoning isn't legible to humans, and is open to third-party auditors, the mechanism Amodei proposed Saturday
  • Mustafa Suleyman: "If it isn't safe we shouldn't build it"; Anthropic, OpenAI, xAI and Meta leaders all endorsed pacing over the weekend
  • Trump posted that there is a "SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China"
  • Markets took it badly: Nvidia fell 3%, Micron 5%
  • OpenAI bought Glass Imaging, an ex-Apple smartphone camera startup, for $300M against a $100M valuation last year; Chinese GPU designer Biren is weighing a ~$1B raise after ~$900M in July
ai safety pacing ×8 china chips microsoft openai devices Open in Gmail →

OpenAI Ads, Amazon Ads in ChatGPT, Walmart to Accept Apple Pay (Stratechery Update 9-15-2026)

So what

Amazon proved incumbents can tax chatbot commerce rather than be disintermediated by it — reprice any startup betting on agentic shopping cutting out retailers.

Thompson's read: ChatGPT ads are working at $1B annualized run rate 200 days in, and Amazon just solved its chatbot problem by selling advertisers the right to be the Amazon link inside ChatGPT results.

  • Ads run in 40+ countries with self-service rolling out across India, Europe, MENA; they appear for Go subscribers and the free tier of 1 billion weekly actives
  • Amazon became the top retail advertiser on ChatGPT between April and August per Sensor Tower while still blocking chatbots via robots.txt
  • Structure matters: Amazon sets up and manages campaigns, OpenAI's system controls all delivery decisions — thinner than the Meta-Amazon account-linking deal
  • Walmart finally accepts Apple Pay after a decade of CurrentC and Walmart Pay, vindicating Apple's refusal to subsidize merchant adoption
  • The contrast: Walmart fought existing demand and lost; Amazon resisted new demand and dictated terms
amazon apple pay chatgpt ads commerce Open in Gmail →

The Briefing: Software Stocks Gain on AI Slowdown Talk

So what

You now have a hedge map for slowdown headlines: long enterprise SaaS, short compute suppliers — and it reverses within a day.

Martin Peers reads Monday's tape as a preview of an AI downturn: enterprise software rallied on slowdown talk while chips and neoclouds sold off. His verdict on a coordinated pause: chances are zero.

  • ServiceNow +7%, Salesforce +5%, with Shopify and Figma also up; neoclouds and chip names fell by similar proportions
  • Trump and the Chinese government both dismissed safety concerns, as "hoax" and "fearmongering" respectively; Meta stayed conspicuously quiet apart from an indecipherable Alexandr Wang post
  • Peers's Manhattan Project analogy: a pause by OpenAI and Anthropic does nothing if Chinese labs and Meta keep racing
  • Oracle's 10-Q revealed Ellison had set up a plan to sell up to 50 million shares by Oct 24; the company said Saturday he cancelled it and sold nothing
  • Also out: Shield AI raising at $20B+, Anthropic's $13.7B compute deal with Trump-linked Rum Group, and Nvidia, Palantir and Booz Allen restricting Claude use over data fears
ai safety pacing ×8 ai capex ×5 anthropic ×2 market reaction Open in Gmail →

15 CIOs, 7 CEOs, 5 PE Firms, and Dozens of AI Leaders to Speak at GAI World 2026 in Boston Sept 28-29

So what

Ten PE firms sending AI operating partners to a Boston conference tells you enterprise AI diligence has moved from thesis to portfolio-ops line item.

Baier uses the pacing debate as the hook for GAI World 2026 in Boston, Sept 28-29, and frames the real CIO question as how to set 2027 AI investment levels amid security and vendor lock-in risk.

  • 80 speakers, 600 attendees, 40 case studies, 4 tracks; tickets $2,195 or $10,000 for a team table of eight
  • PE track features American Securities, OMERS Private Equity, Bain Capital, General Atlantic and Oaktree, including a case study on IT architecture for Claude and vibe coding by non-engineers
  • CIO priorities panel covers security, vendor lock-in and sovereign AI with Travelers, JP Morgan Chase Transformative AI and the US Navy
  • Sessions include How I Manage 25 People with Claude Agents (CustomGPT) and What Will the Human-AI Enterprise Look Like in 2030 (former CIO of AMEX, Verizon, FedEx)
  • Ten companies have already booked team tables
enterprise ai adoption ×3 cio priorities private equity Open in Gmail →

Tomorrow at 1pm ET: The AI workshop for Gen X women

So what

The pacing debate has reached mainstream professional audiences within 48 hours — your customer-facing comms need a position on it now.

A promotional note for a free 90-minute workshop pitched at Gen X women, restructured around the weekend's safety news.

  • Tuesday Sept 15, 1-2:30pm ET, emceed by Cher Jones
  • 60 minutes tactical — prompts, use cases, delegating real work rather than menial tasks — plus 30 minutes open Q&A
  • Q&A expanded explicitly to handle Amodei's post, Altman's interview and the US government response
  • Built from hours of one-on-one calls with Gen X women using AI at work
  • Non-target attendees welcome; nobody will be carded
ai safety pacing ×8 ai literacy workshops Open in Gmail →

What Playing With AI Taught Me About My Work

So what

Is This Anything? is a cheap governance layer for AI exploration — steal it before your team's token spend outruns its harvest rate.

Katie Parrott spent a few hundred dollars in tokens turning her writing toolkit into cartoon creatures, and the useful output arrived sideways: the character work reshaped how Every's writing system thinks about readers.

  • Building Disney-style characters for Every's columns forced want/obstacle/learning questions, which became a new AUDIENCE.md with "Reader's want and need" and "Their hero's journey" sections in Compound Writing
  • The deliberate deliverable — an explorable diorama prompt readers could run — got overbuilt, never landed, and was shelved
  • She built a skill called Is This Anything? that reads a raw session transcript against her priorities and team goals and surfaces up to three candidate lessons
  • The skill distinguishes what came from the conversation versus its own suggestions, and doesn't require writing up the session first
  • Core problem named: when almost every experiment has a plausible claim on your workday, you need triage, not more capacity
ai workflows ×3 agent skills ×3 creative process Open in Gmail →

Defense Startup Shield AI in Talks for Valuation of at Least $20 Billion

So what

Defense AI is repricing on a five-month cadence while the safety debate rages — the fastest markups are in the sector nobody is asking to pace.

Shield AI is in talks to raise at $20B or more, roughly 60% above its valuation five months ago.

  • The 11-year-old company builds military drones and Hivemind, software that coordinates drones and aircraft
  • The markup follows contract wins for Hivemind
  • Investors keep pouring into AI-powered weapons and defense systems
  • Single-source reporting from a person with knowledge of the round
ai capex ×5 defense ai valuations Open in Gmail →

AI Deep Dive: OpenAI’s ‘Top Priority’ for AI Agents is Automating AI Research, Says Noam Brown

So what

OpenAI's stated objective function is self-improvement, which is exactly the capability the pacing proposals were written to constrain.

Noam Brown told The Information's new AI Deep Dive show that OpenAI's training priority is not professional task performance but making models better at AI research itself.

  • "The number one priority is recursive self-improvement, and by a pretty wide margin"
  • GPT-6 Astra improved at designing video games and transcribing sheet music, but those are byproducts
  • Brown has spent three years at OpenAI pushing reasoning and agentic capability
  • First episode of a new show on tech's frontier hard problems
ai safety pacing ×8 openai recursive self-improvement Open in Gmail →

Monday Readout: Anthropic's Warnings Collide With Its Own IPO Machine

So what

Watch Anthropic's roadshow: it has to pick between the 10% extinction number and the valuation, and the choice sets the sector's disclosure norms.

The Monday Readout names the contradiction directly: an Anthropic researcher put his employer's internal extinction odds above 10% this decade, days before Anthropic markets the same technology to IPO investors.

  • Anthropic struck a $13.7B compute deal with Trump-linked Rum Group while quietly discussing a safety standards body with OpenAI and Google
  • Blackstone is multiplying TPU purchases with Google to become AI's dominant lender
  • SpaceX is slowing its data center build-out after a summer of outages
  • Nvidia's own disclosures show sales concentrating in a shrinking pool of giant customers
  • Beijing curbed humanoid IPOs after Unitree's volatile debut — possibly a preview of scrutiny on AI infrastructure deals
ai safety pacing ×8 anthropic ×2 ai capex ×5 ipo Open in Gmail →

Survey: How does your org really feel about AI infrastructure?

So what

Worth five minutes only if you want early sight of how the buy side is ranking cloud and AI infra vendors.

A five-minute pulse survey on how organizations view AI infrastructure and cloud providers — priorities, perceptions, vendor consideration.

  • Responses confidential, reported only in aggregate
  • Focused on what is shaping vendor consideration right now
  • No findings published yet
ai infrastructure surveys Open in Gmail →

Brilliant Moves: Coffee with Lloyds Banking Group CEO Charlie Nunn

So what

Nothing actionable; skip unless you need incumbent-bank framing language for a financial services pitch.

A podcast episode with Lloyds Banking Group CEO Charlie Nunn on steering a 300-year-old bank through the AI shift.

  • Nunn frames the goal as expanding access to financial advice, not just cost transformation
  • Emphasis on customer trust and building a culture that learns as fast as the technology moves
  • Interviewed by Tunde Olanrewaju, McKinsey's managing partner for Europe, who has worked with Nunn for two decades
  • Part of the Brilliant Moves: Coffee with CEOs series
enterprise ai adoption ×3 financial services leadership Open in Gmail →

You’re invited! Join us for a McKinsey Live virtual event on improving the economics of agentic AI

So what

The 6% high-performer figure is the number to quote when a portfolio company claims its agent deployment is delivering EBIT impact.

An invite to a Sept 22 McKinsey Live session on agentic AI unit economics, built around a specific claim: token prices are falling but consumption is rising faster as agents take on complex workflows.

  • Only ~6% of respondents in McKinsey's State of AI research qualify as high performers with significant value and 5%+ EBIT impact
  • 60% expect to increase AI investment over the next year
  • Session covers managing token budgets at scale and evaluating agents against the fully loaded cost of the work they replace
  • Explicitly addresses where to pull back, not just where to scale
  • Thirty minutes, Sept 22, 11:30am ET, hosted by Tanguy Catlin and Lari Hämäläinen
enterprise ai adoption ×3 ai capex ×5 agent economics Open in Gmail →

How AI CEOs Are Turning Safety Into a Regulatory Moat

So what

Price regulatory capture into your model-layer theses: mandatory third-party evals are the barrier incumbents want and challengers can't fund.

Guillermo Flor argues the safety standards the largest labs are proposing function as a regulatory moat: mandatory independent evaluations raise the cost of reaching the frontier and price out entrants.

  • An Anthropic researcher resigned after warning colleagues assign greater-than-10% odds to AI killing everyone this decade
  • Altman called that number unacceptable on camera while declining to name his own
  • Amodei is pushing third-party evaluators, deployment controls and stronger frontier oversight
  • The emerging compliance stack — training, evaluation, audit, deployment approval — compounds incumbent advantage
  • Mandatory independent model evaluation is the specific requirement most likely to be codified
ai safety pacing ×8 regulation ×2 ai safety business model Open in Gmail →

Opinion: The Tax Code Needs to Change in the Age of AI

So what

An OpenAI investor is now arguing publicly for tax redistribution — read it as the sector pre-positioning before labor-displacement politics arrive.

Vinod Khosla argues AI will materially cut the labor required per unit of GDP, and the tax code and social contract have to be renegotiated before that happens rather than after.

  • Protect people and transitions, not existing jobs
  • Calls for a strong social safety net, strong core government services and deliberate preservation of the human element
  • Argues human value need not depend on outperforming machines at intelligence
  • Pairs this with maintaining a security lead over AI nation-state adversaries
  • Khosla Ventures is an OpenAI investor, disclosed in the piece
regulation ×2 ai policy labor displacement Open in Gmail →

🎙️ How I AI: How two SpaceXAI designers use Grok Bot to do their jobs

So what

Narrow bots with preloaded conventions plus voice input is a shippable pattern — the config, not the model, is where the quality lives.

Two SpaceXAI designers on the Grok Bot team walk through workflows that move design work off the computer entirely. The through-line: narrow, single-purpose bots with pre-configured preferences beat one general assistant.

  • Peng built a self-updating personal site where one message triggers Google Places lookup, light and dark 3D render generation, and publishing
  • John's Figma Bro workflow: two Figma screenshots plus a voice memo from the gym returned three sized and colored options by the time he checked back
  • Voice memos work as a design interface only because artboard structure, spacing and naming conventions are configured upfront
  • Peng runs separate agents for email triage, calendar, 3D-printing inventory and green-card status, with a chief-of-staff bot for unowned work
  • Neither started from a Figma file or spec; the workflows emerged from "what if it could also do this" iteration
ai workflows ×3 agent skills ×3 design tools ×2 Open in Gmail →

🎙️ How I AI: How two SpaceXAI designers use Grok Bot to do their jobs

So what

Nothing new in the second copy — clean up your feed dedupe before it eats more reading time.

Duplicate send of the Grok Bot designers episode; same contents.

  • Peng's self-updating check-in site runs end to end from a single message
  • John completes Figma work from the gym via screenshots and voice memo through Figma MCP
  • Narrow single-responsibility agents with a chief-of-staff fallback beat a general-purpose chatbot
  • Falling cost of making things makes unusual creative combinations worth attempting
  • Exploration replaced specification: no Figma file, no roadmap, no upfront system definition
ai workflows ×3 agent skills ×3 design tools ×2 Open in Gmail →

🎧 New Episode: Home Depot

So what

The best US equity return ever came from a hardware retailer compounding buybacks, not a tech platform — useful ballast when AI multiples feel inevitable.

A 3 hour 35 minute episode on Home Depot, which has delivered the highest total return of any company in the US stock market since its 1981 IPO — ahead of Apple, Microsoft and Berkshire.

  • Ben and David admit the company is far larger than they had realised
  • Share count history is the story: equity issued for capital until 2001, then buybacks that retired nearly everything issued decades earlier
  • First stores opened in former JCPenney discount chain locations
  • Companion PDF with charts and timelines, plus an artifacts page for every visual
  • Fall '26 season partners include Sierra, WorkOS, Anthropic and Sentry
business history capital allocation podcasts Open in Gmail →