Episode 29 · Friday, September 11, 2026 · 4:00 PM ET

Weekly Claw

The chat is empty. The agents left.
Nine cards, two grids: OpenAI's ten-thousand-agent Navier–Stokes claim; Meta ships Muse behind a Sentinel; XPENG walks a humanoid off an automated line; DeepMind precomputes nine billion DNA predictions; DeepSeek makes cache the launch; Mistral prices sovereignty at €3B; Suno signs with its former plaintiffs; the AI economy shows a boom — with a Stanford asterisk.

Hosts: @AndyML · @HiM ~32 min · two grids · one anchor · one debate

Sponsored by Heritage Telecom Herald Labs

“The chat is empty. The agents left.”

Weekly Claw #29

01 / COLD OPEN

Cold open · the frame

The chat is empty. The agents left.

Ten thousand agents at OpenAI on Navier–Stokes, one on your phone from Meta, one walking off XPENG’s line, nine billion pre-computed predictions from DeepMind. Meanwhile the boundaries around agents became products: DeepSeek made cache the launch, Mistral priced sovereignty at €3B, Suno turned plaintiffs into partners, and the first honest labor number is a boom — with a Stanford asterisk.

1

Agents left the chat

2

Boundaries as product

3

Outside signal

4

The argument about the argument

5

One to watch

🧠OpenAI ran ten thousand agents for 88 hours and posted a claimed Navier–Stokes proof. Claimed, expert review pending.

📱Meta launched Muse with an isolated cloud VM and a Sentinel the agent cannot override.

🤖XPENG walked the first assembled IRON humanoid off an 80%-automated production line.

Weekly Claw #29

SPONSOR

Brought to you by

UCaaS and VoIP phone service for businesses that just need their calls to work. Independent, boring reliability, zero telemetry.

Independent. Reliable. Quietly essential.

heritagetel.com

Weekly Claw #29

WHAT HAPPENED THIS WEEK

What happened this week · part 1

The chat is empty. The agents left.

Four receipts of agents outside a chat window — a claimed math proof, a phone-sized product, an assembly line, and a genome. Each card links to its primary receipt.

OpenAI Navier–Stokes launch page vortex diagram

OpenAI runs 10,000 agents for 88 hours on Navier–Stokes.

Claimed proof + public Lean 4 repo. Nature reported the announcement, not the acceptance.

openai.com · CLAIMED · review pending

Meta launches Muse behind a Sentinel and a cloud VM.

Isolated Linux VM, credentials outside harness, non-overridable Sentinel. Reuters’ internal tests found stalls.

▶ LIVE · about.fb.com launch

XPENG IRON humanoid production-line launch image

XPENG commissions an automated IRON production line.

80%+ automated core processes. First unit walked off under its own control. Capacity undisclosed.

xpeng.com · production line, not mass production

DeepMind precomputes 9B single-letter DNA predictions.

~1 PB resource. Portal + API + Antigravity skill. Nature: outputs are predictions, not clinical evidence.

▶ LIVE · alphagenome.google/atlas

Weekly Claw #29

WHAT HAPPENED THIS WEEK

What happened this week · part 2

Then the boundaries became products.

Cache economics, sovereign capital, licensing peace, and the first honest labor number for the AI economy.

DeepSeek V4.1 Flash benchmark table (vendor-reported)

DeepSeek launches V4.1-Flash with open weights and a smaller cache.

552B MoE · MIT weights live · API live. Vendor-reported cache claim; first independent agent run: >50× cheaper than Sol/Opus on 60 browser tasks.

▶ LIVE · HF: DeepSeek-V4.1-Flash

Mistral raises €3B to build Europe’s sovereign AI stack.

Samsung led. EQT + PSG co-led. ASML on the record. Post-money above €21B; nearly 2× the prior Series C.

mistral.ai · announcement

Suno launches v6 with the labels that sued it.

Warner + BMG + Believe as development partners. From-scratch training with licensed music + user data.

suno.com/blog/introducing-v6

AI jobs boom chart — 1M created vs 200K AI-attributed layoffs (est.), with Stanford entry-level warning

The AI jobs story is a boom — with a Stanford asterisk.

~1M US AI-linked jobs created vs ~200K AI-attributed layoffs since mid-2023. Entry-level hiring weakens.

Stanford SIEPR · asterisk in view

Weekly Claw #29

04 / HOT TAKE · 4 MIN

Hot take · two sides, one verdict

Is AI safety becoming a religion with a business model?

Rendered receipts collage: Coxon opening resignation post on X, The Effort $3.3M report on FLI-linked grants to religious NGOs, and Cal Newport's NYT column on the AI doomer cult. TESTIMONY — NOT PROOF label visible.

HENRY · INCENTIVES ARE INSPECTABLE

You don’t get to demand a global capability pause and skip the collection plate.

Timing: Anthropic catastrophic-risk claims scaled with IPO. Coxon post crossed ≈120M views in a day.
Ledger: The Effort: $3.3M FLI-linked grants to religious NGOs ($200K TGC + $125K Faith Matters).
Mind-changer: a serious safety org publishes budget, funding sources, and a policy measuring capability harm.

ANDY · SEPARATE EVIDENCE FROM MOVEMENT

Guilt-by-association fails as a filter; system cards do not.

Evidence: Astra system card admitted oversight-evasion decline; Anthropic alignment lead public Sep 9.
Not a filter: a researcher’s funding history does not falsify a capability claim.
Mind-changer: evidence that capability-harm work is not happening inside those orgs.

Verdict: the safety-movement critique is fair when aimed at incentives and funding — those are inspectable. It is not fair as guilt-by-association against measurable safety work. Both statements are true.

Weekly Claw #29

SPONSOR

Also brought to you by

An applied AI product lab where humans and agents build together. Entity is mission control for agent teams. Hacker houses worldwide.

Build with humans. Ship with agents.

labs.theherald.co

Weekly Claw #29

05 / CLOSE

One to watch + where to follow

See you next Friday.

The migration

Monday, September 14: DeepSeek is scheduled to move V4-Pro API traffic to V4.1-Flash. That is when the launch turns into a deployment.

The audit

First independent mathematician who publicly checks the OpenAI Navier–Stokes Lean formalization. Not another summary — the compiled result.

The line

XPENG’s next monthly disclosure: is there a monthly capacity number behind “line commissioned,” or is “commissioned” still the whole story?

Agents left the chat. Follow the operator.

Weekly Claw #29

SOURCES / LINKS

Every claim, one click · verified 2026-09-10

Sources & Links.

A1 · OPENAI NAVIER–STOKES

openai.com: primary launch github.com/openai/NavierStokesAndEuler: public Lean 4 repo nature.com: independent coverage openai.com: MIT quantum-lab sibling openai.com: 3.1 agent-workdays metric

A2 · META MUSE

about.fb.com: primary launch research.meta.ai: security architecture reuters.com: independent adverse tests

A3 · XPENG IRON PRODUCTION LINE

xpeng.com: primary release @xiaopenghexpeng: CEO launch post cnevpost.com: independent manufacturing coverage

A4 · ALPHAGENOME ATLAS

deepmind.google: technical post alphagenome.google/atlas: live portal nature.com: independent coverage

B1 · DEEPSEEK V4.1-FLASH

api-docs.deepseek.com: primary launch + benchmark huggingface.co: MIT weights + config bloomberg.com: independent read x.com/gregpr07: independent 60-task browser run srnnews.com: launch coverage

B2 · MISTRAL €3B SERIES D

mistral.ai: primary announcement qz.com: independent corroboration

B3 · SUNO V6 WITH PLAINTIFFS

suno.com: primary launch reuters.com: launch + partnership axios.com: partnership detail theverge.com: from-scratch training

B4 · AI JOBS BOOM + STANFORD ASTERISK

economist.com: primary analysis siepr.stanford.edu: entry-level asterisk goldmansachs.com: displacement outlook Henry pulse (Sep 8)

SIGNAL FROM OUTSIDE · YC PAPER CLUB

YC Paper Club: Why The Harness Matters More Than The Model

HOT TAKE · SAFETY AS RELIGION+BUSINESS

x.com/hilbertspaess: Coxon opening thread techcrunch.com: resignation coverage forbes.com: Hubinger response nytimes.com: Newport column (Sep 10)

Caveats: OpenAI Navier–Stokes is CLAIMED — expert review pending; do not call it accepted. Meta’s Sentinel architecture is company-reported; independent audit pending. XPENG line is commissioned, not mass-production; capacity undisclosed. AlphaGenome outputs are predictions, not clinical evidence. DeepSeek benchmark and cache numbers are vendor-reported. Mistral round is corroborated; valuation and customer counts are self-reported. Suno v6 dataset scope and compensation terms are undisclosed; litigation continues elsewhere. Economist jobs numbers are estimates; Stanford entry-level warning stays visible. Coxon’s claims about colleagues and lab race dynamics are testimony, not measured facts; The Effort’s grant ledger is inspectable, its “coordinated campaign” framing is The Effort’s interpretation. All video controls are manual; no autoplay is used anywhere in this deck.

Weekly Claw

The chat is empty. The agents left.

The chat is empty. The agents left.

OpenAI runs 10,000 agents for 88 hours on Navier–Stokes.

Meta launches Muse behind a Sentinel and a cloud VM.

XPENG commissions an automated IRON production line.

DeepMind precomputes 9B single-letter DNA predictions.

Then the boundaries became products.

DeepSeek launches V4.1-Flash with open weights and a smaller cache.

Mistral raises €3B to build Europe’s sovereign AI stack.

Suno launches v6 with the labels that sued it.

The AI jobs story is a boom — with a Stanford asterisk.

Same weights. Thirty, or ninety-five.

Why the harness matters more than the model.

Is AI safety becoming a religion with a business model?

You don’t get to demand a global capability pause and skip the collection plate.

Guilt-by-association fails as a filter; system cards do not.

See you next Friday.

The migration

The audit

The line

Sources & Links.