AI So Whats
Edition

Saturday, September 5, 2026

9 issues 23h window ~8 min read

Running through today’s issues

Anthropic is insourcing everything it depends on

3 sources

The Information reports Anthropic is hiring to build billing and fraud detection in-house rather than lean on Stripe, and that it briefly considered buying chip startup MatX before Coatue moved to a JV. Stratechery adds the third leg: Anthropic tore up its data retention policy alongside Fable 5.1 because it needs revenue and OpenAI is pressing. A lab under commercial pressure pulls payments, silicon and customer terms inside its own control surface.

The agent stack is converging on folders and files, not frameworks

2 sources

Kieran Klaassen spent three months on Claude Code teams and agent-to-agent orchestration, got nothing, and settled on 44 folders with a Ruby daemon doing file-based dispatch. Ben Tossell's 107-million-row build ran on 656 subagents through Codex with 235 human messages and no orchestration layer at all. Both point the same way: the value sits in accumulated repo context and human review capacity, and the orchestration vendors are selling the wrong thing.

Chinese benchmark parity is measured against America's second-best models

2 sources

Noahpinion's core point is timing: Mythos finished training in January and shipped in April, four months before GLM-5.3 matched it on CyberGym, and Anthropic says better models sit unreleased. Stratechery supplies the mechanism — Anthropic only loosened its terms when OpenAI competition forced it, meaning US labs hold back on release for safety and commercial reasons that fast followers do not face. Any parity claim dated to public release understates the gap by a release cycle.

Noahpinion 00:36 PT

America is still beating China in the AI race

So what

Discount headline benchmark parity from Chinese labs by roughly the six-month gap between US internal training runs and public release.

Noah Smith argues the US-China AI gap is wider than the benchmark discourse suggests, because America's best models are not the ones it ships. The real risk to the lead is immigration policy, not Chinese capability.

  • Z.ai's GLM-5.3 scored 84.5% on CyberGym vs 83.8% for Anthropic's Mythos 5, but trailed badly on ExploitBench (54.4% vs 78.0%) and timed attack-development tasks (105 vs 181 in two hours)
  • Mythos finished training in January, went internal in February, released as Preview in April — four months before GLM-5.3; Anthropic says it has better models unreleased
  • Fast followers like Z.ai carry none of the cybersecurity release friction that delays US labs, which flatters Chinese benchmark timing
  • AI hacking has no mutually assured destruction: a decisive model lead could let one side erase financial records with no proportionate retaliation
  • Cooperation on catastrophic risk is likelier if the US is comfortably ahead, so safety and dominance point the same direction
  • Trump administration policy pushing Chinese AI talent out is named as the most plausible way America squanders the lead
anthropic ×4 ai safety model benchmarks us-china ai race Open in Gmail →
Every 14:22 PT

Fable 5.1 vs. GPT-6 Astra: past the Vibe Check

So what

Stop standardizing your org on one frontier model — the split by task type here is stable enough to justify paying for both.

An hour of live side-by-sides between Fable 5.1 and GPT-6 Astra produced four different verdicts from four operators, which is the finding.

  • Dan Shipper runs Astra daily but returns to Fable for the hardest work; Kieran Klaassen codes with Fable, Katie Parrott writes with Astra
  • Same prompt, two apps: Jack Cheng's webcam notebook reader shipped as a space-bar-tap flow on Fable, a better-looking but confirm-every-page app on Astra
  • Kieran's low-effort Astra run looped and cost more than his medium-effort run — cost per task, not cost per token, is the metric
  • Astra spent five hours driving Premiere to cut the Fable 5.1 video and one-shot its own Vibe Check with screenshots and captions
  • Mike Taylor's independent bench gave Astra the edge on anything visual; Dan demoed KatePass, a doc checker against the editor in chief's standards
fable 5.1 ×2 gpt-6 astra ×2 ai coding tools ×3 model comparison Open in Gmail →
Ben Thompson (Stratechery) 10:58 PT

Friction and Feedback (This Week in Stratechery)

So what

Anthropic's retention reversal proves enterprise terms are now a competitive lever — push your vendor on data policy while OpenAI is applying pressure.

Thompson's week reads as competition doing what backlash could not: Anthropic scrapped its data retention policy entirely alongside Fable 5.1, because OpenAI is pressing and Anthropic needs revenue.

  • Fable 5.1 ships with the hated data retention policy not softened but removed in the interim; Thompson frames it as the market speaking
  • Greg Brockman interview on Astra and alignment runs the same week as OpenAI's own release
  • Nvidia earnings called "remarkable and boring" — everything the company does is aimed at avoiding a consolidated world; Dithering covers Nvidia buying Hugging Face
  • Meta settles with 29 states on teen Instagram restrictions; Thompson conflicted, Sharp Tech digs into libertarian vs conservative fault lines in tech regulation
  • Apple finally nails immersive live sport with Friday Night Baseball; the Sharp Tech video argues App Store drama is irrelevant next to AI
anthropic ×4 fable 5.1 ×2 gpt-6 astra ×2 nvidia Open in Gmail →
The Information 10:16 PT

Exclusive: Coatue, MatX in Talks on New Multibillion Chip-Financing Venture

So what

Component allocation, not capital, is now the binding constraint — JV structures with balance-sheet investors are how small chip firms buy supply access.

Coatue is in talks to form a joint venture with chip startup MatX, the latest structure invented to route around AI hardware and financing constraints.

  • MatX's chip handles both training and inference, an unusual dual claim for a young entrant
  • Anthropic briefly considered buying MatX outright before the JV route emerged
  • The JV exists specifically to get MatX past component shortages, not to fund R&D
  • Reported by Phoebe Liu, Valida Pau and Stephanie Palazzolo, Sep 4, 2026
anthropic ×4 ai chips ai infrastructure capital Open in Gmail →
McKinsey & Company 09:18 PT

Made in America, again? Rethinking US manufacturing resilience

So what

Treat reshoring pitches with a $3T import base in mind; the credible plays are capacity in a few critical chains, not broad repatriation.

MGI's podcast argues US manufacturing resilience is not a reshoring problem, because the import base is too large to relocate.

  • The US imports roughly $3 trillion in manufactured goods annually; critical supply chains would need dramatic domestic capacity expansion
  • Rebecca J. Anderson and Shubham Singhal frame the question as where America is most vulnerable and what rebuilding would actually cost
  • AI, investment and domestic innovation are positioned as the variables that could change the arithmetic
  • Accompanying State of AI in 2026 research covers coding agents, agentic tooling, investment and cost, and what separates high performers
enterprise ai adoption supply chains us manufacturing Open in Gmail →
The Information 09:01 PT

Inside Andreessen Horowitz’s Cash-Minting AI Infrastructure Fund

So what

The application-layer AI thesis just printed its proof point; expect fund formation and entry prices at the tooling layer to move sharply.

Martin Casado's AI fund at a16z returned over $8 billion from two exits in a single month, validating a thesis that was two years old.

  • SpaceX bought coding tool Cursor; Stripe agreed to buy model marketplace OpenRouter
  • Combined value to Andreessen Horowitz exceeds $8 billion, from its first dedicated AI tools and software fund
  • Casado's background: startup founder, VMware executive, a decade at a16z
  • He runs agents personally for wine auction bidding and home management, plus an agent that manages the other agents
stripe / openrouter ×2 ai exits venture capital Open in Gmail →
The Information 08:30 PT

Anthropic’s In-House Payments Tech Push Could Chip Away at Stripe

So what

If frontier labs insource payments, Stripe loses its highest-growth logos exactly as it pays up for OpenRouter — reprice that exposure.

Anthropic job postings show it building billing, fraud detection and financial infrastructure in-house, which would reduce its dependence on Stripe.

  • Token-based pricing at Anthropic's usage volumes is making third-party payments infrastructure hard to manage
  • The hiring covers billing, fraud detection and adjacent financial services, evaluated build-versus-buy
  • Anthropic sits at the center of the AI boom's payment complexity, making it the reference customer others copy
  • Reported by Stephanie Palazzolo, Sep 4, 2026
anthropic ×4 stripe / openrouter ×2 payments Open in Gmail →
Every 08:04 PT

The Folder Is the Agent

So what

Your agent spend should go into curated repo context and a dumb file-based dispatcher, not orchestration frameworks or agent-to-agent protocols.

Kieran Klaassen's republished piece argues agent swarms failed and folders succeeded: an agent is just a model pointed at a directory with accumulated context.

  • Three months of Claude Code teams, agent-to-agent dispatch and lead-worker orchestration made him no faster — 10 simultaneous results with no context to judge them
  • He now runs 44 agents across a handful of specialized folders; same model, different folder, different specialist
  • ~/cora/ holds CLAUDE.md conventions, developer docs, runbooks, investigations and .claude/agents/ specialists; ~/cora-agent/ has no app code and only ops skills — AppSignal, Render logs, a Postgres read replica, Intercom, GitHub deploy correlation
  • The dispatch layer is a Ruby daemon polling a directory every 60 seconds with file-based messaging; two slash commands, /hey and /orchestrate, replaced 20 terminal tabs
  • Anthropic's own research: Opus lead with Sonnet sub-agents beat single Opus by 90% on research, at 15x the tokens, and coding parallelizes worse than research
ai agents ×2 ai coding tools ×3 agent orchestration Open in Gmail →
Ben's Bites 07:56 PT

Scraping 107M rows of data to build this

So what

235 messages produced a 107M-row public data product — the bottleneck on data journalism and internal analytics is now your review capacity.

Ben Tossell scraped 107 million rows of English council spending data and shipped an Apple Maps-style explorer, mostly by letting Codex run unattended for two days.

  • Totals: 8.2 billion tokens, 235 human messages, 656 subagents, 552 threads; 319 of 339 English and Welsh councils collected or parked
  • Started with 8 councils, 1.8 million rows and £9bn of spend in one CSV; scaled to 34.35m rows before the full run
  • Codex chose Parquet and DuckDB on a Mac Mini — tools Tossell had never heard of — after he asked how to work with a lot of data
  • Six councils including Lancashire and Norfolk blocked the scrapers; Wales was dropped because it lacks England's £500 publication mandate
  • Four subagents classified 2,000 unknown vendors and merged duplicates (Tesco appeared under six names); a £5m threshold filter was masking data because the site is static with no backend
ai agents ×2 ai coding tools ×3 codex Open in Gmail →