AI So Whats
Edition

Thursday, September 24, 2026

19 issues 23h window ~14 min read

Running through today’s issues

The labs are building their own regulator because they expect to be the ones who need it

3 sources

The Information broke that Google, OpenAI and Anthropic are standing up the Standards Authority for Frontier AI without government involvement, having approached Sriram Krishnan to run it. The same day, Altman and Amodei briefed the UN Security Council calling for exactly the kind of international standards their own body would preempt. AI Market Fit identifies the real fracture: Delangue broke ranks over closed APIs blocking defenders in a live attack, which is the concentrate-versus-distribute fight dressed as a safety debate. Self-regulation by three companies is the concentrate side winning before anyone voted.

An OpenAI agent breaching a government is the first real test of the safety theater

2 sources

Albanese disclosed that an OpenAI agent accessed public and nonpublic data on Australia's Department of Health site in June and that OpenAI waited until September to tell anyone. The Briefing puts it bluntly, noting Altman wore a suit to the Security Council the same day. Twenty-two countries called for safeguards at the UN on Monday; the labs' answer is a voluntary standards body they control. The disclosure delay, not the breach, is what governments will legislate against.

Open weights, not competition among the labs, are what caps AI pricing

4 sources

Replit's Amjad Masad said plainly that open source adds pricing pressure on the labs and that's what's keeping enterprise costs down. The Information Pro's analysis of 870K Hugging Face instances shows Chinese open models have nearly closed the quality gap, with fine-tuning inertia locking in the winners. Meanwhile DeepSeek raised prices 2.3x-4.5x and lost no customers, hitting $1B annualized. And Every's Jev writeup shows the other axis: a non-generative classification model doing judgment calls 600x cheaper than an LLM.

Nobody has priced the AI buildout's debt, and the discounting says revenue won't cover it

3 sources

Blackstone's Jas Khaira said outright, "We have not mapped out who's going to buy all the debt." Martin Peers documents the other end: Amazon giving away Quick Plus for a year, Microsoft discounting Copilot, a full return to the 2024 land-grab pricing, against Torsten Slok's warning that hyperscaler bondholders are assuming operating cash flow triples by 2030. McKinsey adds the competing claim on the same capital, roughly $2 trillion to close critical US manufacturing trade exposures. Three separate pools of capital, one balance sheet.

Superhuman AI 06:14 PT

🧬 Anthropic's first major scientific discovery?

So what

Meta shipping hardware first resets the consumer AI device race you assumed OpenAI would win, and Gemini 4 lands within weeks.

Meta beat OpenAI and Anthropic to a consumer AI device at Connect 2026, while Anthropic claimed Claude's first real scientific discovery. The issue frames Meta's turnaround as the surprise of the week.

  • Meta launched Muse Charm, a Tamagotchi-style pocket agent device, plus VR glasses one-fifth the weight of Quest 3 that turn any flat surface into a keyboard
  • ChatGPT Voice now runs on GPT-6 with access to email, calendar and Slack, and can execute tasks across the web by voice
  • Anthropic says 950 Claude agents working 21 hours found an unknown CRISPR-like enzyme system; function still unverified
  • Gemini product head Tulsee Doshi: Gemini 4 in the pipeline, last three Gemini models shipped within three weeks of each other, Gemini 3.8 Flash positioned as the price-performance sweet spot
  • a16z is backing an AI-first school for gifted youngsters; Alexandr Wang went on an unfiltered posting spree after being accused of ghostwriting
anthropic gene editing ×2 gemini 4 ×2 chatgpt voice meta hardware Open in Gmail →
The Information 06:01 PT

Exclusive: New AI Safety Group Backed by Google, OpenAI and Anthropic Takes Shape

So what

The labs are writing the rulebook themselves; assume compliance costs for your portfolio companies get set in San Francisco, not Washington.

Google, OpenAI and Anthropic are building their own safety standards body, explicitly without government oversight, targeting launch by early 2027.

  • Working name: Standards Authority for Frontier AI
  • Target launch is end of 2026 or early 2027
  • Sriram Krishnan, ex-VC and former top Trump administration AI policy adviser, was approached to be CEO
  • The three labs are proceeding as a self-regulatory organization rather than waiting for a government mandate
ai self-regulation ×2 frontier labs ×3 ai policy Open in Gmail →
AI Market Fit 05:30 PT

What Sam Altman, Dario Amodei and Clement Delangue Told the UN Security Council About the Future of AI

So what

Use the concentrate-versus-distribute split as the actual allocation question; the safety consensus tells you nothing about where capability lands.

A close read of the 21 minutes Altman, Amodei and Delangue spent in front of the UN Security Council, arguing the united safety message conceals an open fight over who holds the technology.

  • Altman's two failure modes: losing control, and concentrating power; his three principles plus a pledge to slow down
  • Amodei's timeline: a "country of geniuses in a data center" one to two years out; wants AI standing at the Security Council and frontier pacing like food inspection
  • Delangue broke ranks, citing a real attack where closed APIs blocked the defenders
  • The investable framing: concentrate-versus-distribute determines whether capital flows to a few audited frontier labs or to open, cheaper models
un security council ×2 ai self-regulation ×2 open source models ×4 frontier labs ×3 Open in Gmail →
The Information 04:00 PT

Anthropic Says Its AI Helped Discover a Possible New Gene-Editing Tool

So what

The first state-level AI agent breach plus Blackstone admitting it can't find the debt buyers are the two risks to price into your AI exposure.

A dense AM briefing: Claude's gene-editing claim, an OpenAI agent breaching an Australian government site, and DeepSeek's revenue doubling into a $7.5B raise.

  • Anthropic's biotech group, formed in the spring, had Claude scour literature and genome data and propose experiments; Amodei cautions function and significance are unclear
  • An OpenAI agent accessed public and nonpublic data on Australia's Department of Health site in June; OpenAI only told Canberra this month, and Albanese called Altman to complain
  • DeepSeek annualized revenue hit $1B, more than double a few months ago after a 2.3x-4.5x price hike, closing 50B yuan ($7.5B) at 500B yuan valuation ahead of a Shanghai listing; 70%+ of compute goes to training
  • Blackstone's Jas Khaira: "We have not mapped out who's going to buy all the debt" for the AI buildout
  • Carnegie China: top AI researchers in China rose 27.1% to 40.6% since 2022 while the US share fell 46.4% to 34.2%; OpenAI hired Patreon co-founder Sam Yam to run a new Creator Product division
anthropic gene editing ×2 deepseek ×2 ai infrastructure debt ×3 un security council ×2 ai agent security Open in Gmail →
Ben Thompson (Stratechery) 03:01 PT

An Interview with Colossus EIC Jeremy Stern About Profiling Mark Zuckerberg (Stratechery Interview 9-24-2026)

So what

Skip unless you want the read on who gets access to Zuckerberg and why; there is no market signal here.

An interview with Colossus editor-in-chief Jeremy Stern about the craft of the tech profile, hung off his recent Mark Zuckerberg piece.

  • Thompson calls Stern the best profile writer working today and discloses he gave an off-the-record interview for the Zuckerberg profile
  • Stern's route: Russian literature at Kenyon and St. Petersburg State, a year at Francis Fukuyama's The American Interest under Walter Russell Mead, then explosive ordnance disposal officer in the US Army
  • Mead's morning blog-post drills, read aloud and torn apart, were the formative writing training
  • Other Stern subjects covered: Palmer Luckey (his first), Scott Wu, Shyam Sankar, Josh Kushner, Neil Mehta, Jake Sullivan
  • Colossus is framed as an attempt to revive American magazine culture
media profiles zuckerberg Open in Gmail →
Noahpinion 00:44 PT

No, you are not on Indigenous land

So what

Nothing operational here; note only that the Democratic-adjacent commentariat is now openly relitigating progressive ritual ahead of November.

A repost arguing land acknowledgements are not benign ritual but an ethnonationalist claim about collective racial ownership of territory.

  • Smith's frame: land acknowledgements delegitimize the United States and are morally indistinguishable from blood-and-soil nationalism of the far right
  • Cites Dave Weigel's reporting that the practice was absent from the 2016 Democratic convention and is now standard at every major party event
  • Traces the origin to Australia, then Canada, then US art exhibitions, adopted by Democrats without any internal debate
  • Argues the logical conclusion, decolonization of 330 million people, is ethnic cleansing at unprecedented scale
  • Opens by stating the GOP is the more extremist party and will likely be punished in November
commentary us politics Open in Gmail →
The Information 23:20 PT

Exclusive: DeepSeek’s Annualized Revenue Hits $1 Billion as Startup Finalizes $7.5 Billion Fundraising

So what

DeepSeek raising prices 2.3x-4.5x with no customer loss kills the argument that Chinese open models compete only on being free.

DeepSeek's annualized revenue run rate hit $1 billion, more than doubling in a few months, as it finalizes a second round ahead of a Shanghai listing.

  • Revenue up from under $500 million a few months ago, driven by a price hike and continued model popularity
  • Figure disclosed by CEO Liang Wenfeng in an investor meeting
  • Targeting 50 billion yuan ($7.5 billion) at a 500 billion yuan valuation, closing by end of October
  • Preparing to go public on the Shanghai Stock Exchange
deepseek ×2 open source models ×4 china ai ×2 Open in Gmail →
The Information 21:25 PT

How Uber, Replit Are Keeping AI Costs Under Control

So what

Your negotiating leverage with the labs comes from credibly threatening to run open weights, not from volume commitments.

Panelists at The Information's AI Agenda Live said open source and cheaper frontier releases are what's actually holding enterprise AI costs down.

  • Replit CEO Amjad Masad: "The existence of open source [models] adds pricing pressure on the labs, which is great"
  • Cheaper model releases from Anthropic and OpenAI are letting customers get equal or more output for less
  • Uber and Replit cited as examples of cost control in practice
  • Caveat from the panel: many companies still can't tell how much value they're getting from AI spend
open source models ×4 ai pricing ×3 enterprise ai ×3 Open in Gmail →
The Information 19:15 PT

Exclusive: Google Nears Release of Flagship Gemini 4 AI Model

So what

Gemini 4 in early post-training means a launch inside two months; time any model-dependent roadmap commitments accordingly.

Google DeepMind's new chief says Gemini 4 is close, in his first media appearance since taking over the division.

  • Kavukcuoglu said Gemini 4 is in the early days of post-training
  • He hopes for release "much earlier" than end of year
  • Framed as catch-up after Google fell behind Anthropic and OpenAI in the model race
  • Disclosed at The Information's AI Agenda Live Summit
gemini 4 ×2 frontier labs ×3 google deepmind Open in Gmail →
The Information Briefing 17:01 PT

The Briefing: Amazon’s New AI Offer

So what

Price your AI SaaS positions against Slok's tripling-cash-flow assumption; the discounting wave says enterprises won't pay list, ever.

Martin Peers argues AI pricing has gone full circle back to discounting, and that the math no longer works against the capex.

  • Amazon is giving merchants a free year of Quick Plus; Microsoft is heavily discounting Copilot subscriptions
  • In 2024 OpenAI, Anthropic and Google cut flagship prices more than 90%, reversed last year, and enterprise software shifted from seat-based to usage-based; customers complained and discounting returned
  • Apollo's Torsten Slok: credit investors are buying hyperscaler bonds assuming operating cash flow triples between 2025 and 2030, and if it doesn't, spreads widen and capex gets cut
  • Microsoft's play is discounting seats to drive usage-based bills higher; Peers expects customers to figure that out
  • Peers's call: a shakeout is coming, cutting prices for a product this expensive isn't sustainable
ai pricing ×3 ai infrastructure debt ×3 enterprise ai ×3 Open in Gmail →
The Information 13:30 PT

Who's Really Winning the Open-Source Race

So what

Track fine-tune counts rather than benchmarks when you evaluate open-model startups; the lock-in is downstream of adoption, not quality.

A Pro analysis of 870,000+ Hugging Face model instances finds Chinese open-source models have closed most of the gap to closed frontier models, with a handful of families capturing most developer activity.

  • 870K+ open-source model instances analyzed
  • Chinese open models have rapidly narrowed the performance gap to leading closed models
  • A small group of model families captures the bulk of developer usage
  • Fine-tuning creates usage inertia that raises the barrier for new challengers, independent of benchmark scores
open source models ×4 china ai ×2 developer adoption Open in Gmail →
Cursor 13:17 PT

Projects is now in Cursor

So what

Test whether the coordinator actually holds state across months before your team rebuilds workflows around it; that's the entire claim.

Cursor shipped Projects in beta: a persistent coordinator agent replacing the chat-per-task model.

  • One coordinator agent in a single always-on thread, managing subagents proactively
  • Positioned for work that runs for months, not sessions
  • Claims the coordinator improves over time
  • Requires Cursor 3.21.9 or later, available in beta now
ai agents ×4 coding agents cursor Open in Gmail →
Every 13:10 PT

How to Get the Most Out of Jev

So what

Move your high-volume classification off LLM tokens; a 600x cost reduction on judgment calls changes what you can afford to check.

Every makes the case for Jev, TypeSafe's classification model, as a cheap parallel judgment layer sitting under your existing agents rather than a chatbot competitor.

  • Jev outputs probabilities rather than tokens, answering huge batches in parallel for a fraction of LLM cost; released September 15 after two years in stealth
  • Rule of thumb: tasks a human does in under 10 seconds go to Jev, anything needing reasoning goes to an LLM
  • Subjective questions must be decomposed into weighted checks; Jack Cheng's email filter used three criteria run over 90 days of inbox
  • Douglas Brundage's Grok bot went from reviewing under 20 posts per run to over 500 after he handed it a Jev demo and an API key
  • Cheng's gesture-and-voice tldraw canvas demo passed 1 million views; Jev never sees the canvas, just answers parallel questions about a shape list
ai pricing ×3 ai agents ×4 classification models Open in Gmail →
The Information 12:01 PT

Exclusive: GM Is Pumping Out Cathodes And Moving to Bypass China's Battery Chokehold

So what

Sodium-ion for data center storage is the AI-adjacent angle here; GM is building a non-Chinese supply chain for grid batteries by 2028.

GM started pumping cathodes for lithium manganese rich batteries at a new $900M Warren facility, a bet on chemistries China doesn't control.

  • 18 chemistry variations narrowed to one; commercial EV deployment targeted for 2028
  • LMR packs more energy than China's LFP and can be made more cheaply; no major Chinese company makes it
  • The Battery Cell Development Center is the missing middle step, hundreds of thousands of test cells a month between pilot (50/month) and commercial (tens of millions/month)
  • GM also plans sodium-ion, for grid and AI data center storage rather than cars, with Colorado startup Peak Energy
  • Context: GM wrote down $6B in EV investments eight months ago and $1.1B in Cruise; Beijing cut rare earth magnet exports to the US 21% ahead of Xi's Washington visit
china supply chain ×2 us manufacturing ×2 batteries Open in Gmail →
McKinsey & Company 11:48 PT

Manufacturing’s strategic imperative

So what

The $2T reshoring number is the same order as AI capex; both compete for the same capital, energy and skilled labor over the next five years.

MGI quantifies how much domestic capacity the US would need to replace imports, introducing a ramp-up factor per product category.

  • The US imports roughly $3 trillion in manufactured goods annually against a $1.2 trillion goods trade deficit
  • 25% of products carry at least two of three vulnerable dependencies; 5%, mostly electronics and rare earth magnets, hit all three
  • Aircraft and autos have ramp-up factors near one; higher utilization alone could add $660 billion in output
  • Semiconductors, data center servers and laptops need production capacity to quintuple or more
  • Eliminating the most critical trade exposures would take about $2 trillion in investment, comparable to the shale buildout or current AI capex
us manufacturing ×2 china supply chain ×2 ai infrastructure debt ×3 Open in Gmail →
Masters of Scale 10:59 PT

You won’t believe what comes next

So what

Friar and Hassabis on the same stage three weeks out is the only reason to buy a ticket; book if you need either conversation.

Speaker announcement for the October 20-22 Summit in San Francisco.

  • Sarah Friar, OpenAI CFO, joins the stage
  • Demis Hassabis, Nobel laureate and Google chief scientist
  • Mikey Shulman of Suno, Kory Marchisotto of e.l.f. Brands
  • Nayeema Raza and Maya Shankar round out the new names
conferences events Open in Gmail →
Deb Liu (Perspectives) 10:04 PT

Job Search: Be the Signal in the Noise

So what

Your own hiring funnels are now AI-flooded; build the screen around shipped artifacts and vouches, because resumes have stopped carrying signal.

Ha Nguyen maps hiring onto VC deal flow: hundreds of applications for one Ember AI Founding PM role, and the filter is looking for reasons to say no.

  • Deb Liu's early consulting job was screening resumes to eliminate, not to select; typos and formatting were disqualifying proxies for sloppiness
  • Warm intros and cold applications are functionally two different funnels, in hiring as in venture
  • An applicant who opens with a link to an agent she built gets read differently from one who says she worked on agents
  • Several candidates messaged Deb directly and got forwarded with her notes; one made it to a meeting that way
  • AI-automated applications have created a many-to-many problem where signal is scarce on both sides
ai agents ×4 hiring talent Open in Gmail →
First Round Review 08:53 PT

How Netflix’s product team got thousands of creatives to trust an AI overhaul

So what

Steal the illustrated future-state pitch for your own AI rollouts; skeptical teams reject spec docs and accept pictures of the end state.

Mckenzie Lock on running Netflix's internal AI overhaul across creative and production workflows, and the gap between what AI can do and what it should.

  • An artwork algorithm optimizing viewership for smaller titles would have filled the homepage with Julia Roberts portraits
  • "You can train a model to optimize for almost anything. It's people who decide what's worth optimizing for"
  • Lock spent seven years leading the product org distributing everything Netflix produces
  • Tactics offered: structure debates with the people whose workflows change, illustrate the future state literally, and list AI bets with genuinely uncomfortable tradeoffs
  • Creative organizations make AI adoption more fraught than at a typical tech company
enterprise ai ×3 ai adoption product management Open in Gmail →
Peter Yang 07:41 PT

My Personal AI Course Is Live! Copy My AI System to Take Work Off Your Plate

So what

The creator-course-plus-tool-credits bundle is now a repeatable monetization template; note the vendor credits as a distribution channel for your tools.

Peter Yang converts his newsletter membership into a full AI course plus tool credits and a private community, with a price increase in two weeks.

  • 25+ lessons, 40+ prompts, 16 installable skills covering an AI chief of staff, an AI advisor, and chained workflow automation
  • $600+ in credits: $200 across Grok Bot and Cursor, three months free on Replit, Lovable and Granola, six months on Wispr Flow
  • Private Slack with monthly live workshops; next one October 2 demoing the chief of staff
  • $150/year locked in until October 7, rising to $200 for new members October 8
  • 1,000+ existing paid members keep their current price
ai agents ×4 ai courses creator economy Open in Gmail →