#027 ·06 NOV 2026 ·FRIDAY ·6 MIN READ ·10 STORIES + 10 EXTRAS

The self-improvement disclosure

Anthropic told the SEC it may imminently reach recursive self-improvement and cannot guarantee it will notice in time, then priced an IPO at $1.5 trillion on $895 billion in compute commitments. The Democrats swept the midterms and half of America's planned data centres landed in regulatory limbo overnight, tanking Asian chip stocks by record margins. The bottom line: the industry is simultaneously warning it might lose control of the technology and betting everything it has that the trajectory holds.

01 / The Ten

The week, ranked

Anthropic warns of imminent recursive self-improvement in its IPO filing

Anthropic's amended S-1 disclosed $16.96 billion in quarterly revenue (12x year-on-year), a $567 million operating loss, and $895 billion in long-term compute commitments with Broadcom, Google, Amazon, and SpaceX. The filing warned the company may imminently reach its own Responsible Scaling Policy threshold for recursive self-improvement — and might fail to detect the acceleration until after it occurs. Institutional IPO valuation guidance settled at $1.5–1.6 trillion, down from an earlier $2 trillion estimate.

Why it mattersThe first time a frontier lab has disclosed proximity to recursive self-improvement in a securities filing. The gap between the safety warning and the valuation is the week's defining tension.

01 techcrunch.com↗

Democrats sweep midterms; half of US data centre plans under threat

Democrats took the House and held the Senate. Over 40 local ballot measures across 10 states targeted data centre expansion, and the new Texas governor campaigned on a moratorium. Bloomberg estimated roughly half of proposed US hyperscale AI facility capacity now faces heightened regulatory scrutiny. Asian chip stocks posted record single-day drops: Taiex fell by a record intraday point count, Kospi dropped 8.7%, Nikkei 6.7%.

Why it mattersAI infrastructure became a ballot-box issue. The market reaction — centred on GPU foundry suppliers, HBM producers, and cloud hardware — repriced the build-out timeline within hours.

02 bloomberg.com↗

OpenAI cyber agent patches 948 Linux kernel vulnerabilities in one run

An unreleased OpenAI model autonomously discovered, exploited, and generated patches for 948 Linux kernel privilege-escalation vulnerabilities and 291 information-leak proofs of concept. Total API cost: $1.21 million. Average: $1,279 per verified root exploit. Candidate patches were submitted directly to upstream kernel maintainers.

Why it mattersThe most concrete public evidence that AI agents can conduct meaningful offensive security research at a cost and velocity impossible for human-only teams.

03 openai.com↗

Tencent ships Hy4: 770B open-weight model with native computer use

Tencent deployed Hy4, a 770-billion-parameter mixture-of-experts model with 49 billion active parameters per token, native multimodal vision, a million-token context window, and built-in computer-use capabilities for CLI and graphical interfaces including CAD. Open weights are promised within days.

Why it mattersThe largest open-weight release of 2026 by parameter count, with native computer use — a capability that was proprietary-only six months ago.

04 tencent.com↗

China orders $70B in Nvidia chips, expects to receive 1%

ByteDance, Alibaba, and Tencent placed orders worth roughly $70 billion for Nvidia's Blackwell-based RTX Pro 5500 GPUs — engineered to sit below US export-control thresholds. A global production cap of about 3 million units through 2027 means roughly 60,000 units (about 1%) will ship to Chinese buyers in the next four months.

Why it mattersThe scale of committed capital against the near-zero near-term delivery probability illustrates the depth of China's compute bind.

05 bloomberg.com↗

GPT-6 Astra demonstrated reprogramming another AI model through its weights

OpenAI showed GPT-6 Astra reading weight tensors of a secondary model, inferring internal representations, and writing targeted weight edits to change the model's behaviour without full retraining. The demonstration was conducted in a controlled research sandbox with no production deployment announced.

Why it mattersWeight-editing agents — AI systems that can modify other AI systems directly — collapse a boundary most safety frameworks treated as theoretical.

06 the-decoder.com↗

Mistral opens weights for 1-trillion-parameter Le Chonk

Mistral released open weights for Mistral Large 4: 1.05 trillion parameters, 52 billion active, trained on 3,800 Grace Blackwell GPUs in European sovereign infrastructure. It supports 160 languages, a million-token context window, and claims the top score among all models on cybersecurity benchmarks. API price: $0.70 per million input tokens.

Why it mattersAn open-weight model claiming best-in-class cybersecurity performance reshapes both enterprise procurement and the threat landscape simultaneously.

07 mistral.ai↗

OpenAI raises $11B — less than half its $30B target

OpenAI's latest funding round closed at roughly $11 billion, well short of the $30 billion initially sought, per WSJ, FT, and Reuters reporting independently. The valuation was trimmed from $1.4 trillion to approximately $1.25 trillion. The miss coincided with S&P downgrading Meta's outlook over AI capital expenditure.

Why it mattersCapital constraint at the most prominent AI lab shapes the compute, talent, and training runs that define the frontier. The gap between target and close is the signal.

08 wsj.com↗

Claude Code product lead defects to OpenAI days before Anthropic IPO

Cat Wu — the head of product who built Claude Code and Cowork from zero to Anthropic's fastest-growing product — joined OpenAI as Head of Product Management for Core Products. The departure was confirmed the day before Anthropic's IPO roadshow entered its final week.

Why it mattersLosing the architect of your highest-growth product to your primary competitor, during your IPO roadshow, is the kind of thing that gets its own risk-factor paragraph.

09 businessinsider.com↗

Claude Sonnet 5 autonomously chose and delivered gene therapy to a mouse brain

Exin Therapeutics published an account of Claude Sonnet 5 independently selecting a gene therapy from 19 candidates, deriving stereotaxic coordinates from three papers, and physically administering an injection into a living mouse brain. The agent made 32 mid-surgery observations and halted when a positioning error exceeded safe range. Claude Opus 5.5 and Opus 5 both declined to perform the procedure.

Why it mattersThe first documented instance of an AI agent both choosing a treatment and delivering it into a living mammal. That two larger models refused while a smaller one proceeded is its own data point.

10 the-decoder.com↗

02 / Also

Worth knowing

DeepSeek open-sources Super-Agent coding assistant

DeepSeek released Super-Agent with open weights — an autonomous coding assistant that monitors team chat for problems and proposes fixes, positioned directly against Claude Code.

bloomberg.com ↗

Oracle kills $7B Tencent GPU deal under White House pressure

Oracle walked away from a five-year agreement to lease 100,000 advanced AI processors to Tencent via Southeast Asian data centres after Trump administration officials intervened on national security grounds.

ft.com ↗

Nvidia imposes 51% non-China rule on cloud AI clients

Nvidia verbally instructed neocloud operators in Singapore, Malaysia, and Japan to ensure at least 51% of rented Blackwell B300 compute goes to non-Chinese customers, closing an offshore access loophole.

techmeme.com ↗

Inkling 2 Small: 594B MoE matches Gemini 3.8 Flash at launch

Thinking Machines Lab released Inkling 2 Small — 594 billion total parameters, 17 billion active — rated on par with Gemini 3.8 Flash. The company is in talks to raise $5 billion at a $40 billion valuation.

thinkingmachines.ai ↗

South Korea pledges $15.97B for AI chips and infrastructure by 2027

President Lee announced a record investment targeting domestic GPU supply and data centre capacity, joining Japan, UAE, and the EU in state-directed AI industrial policy.

reuters.com ↗

Perplexity launches Teammates: always-on parallel agents

Multiple AI roles run continuously and in parallel, integrated with Perplexity Computer for browser and GUI interaction — a direct entry into the AI-coworker category alongside Operator and Cowork.

techmeme.com ↗

MEMENTO: an 8B model beats GPT-5.5 on million-token tasks at 1/50 the cost

A Qwen3-8B agent that manages context via rolling note compaction raised book-level QA accuracy to 67.8% on inputs up to 2.5 million tokens, outperforming GPT-5.5 at a fiftieth of the price.

arxiv.org ↗

Goldman Sachs pulls AI coding tools from non-engineers

Goldman restricted Claude Code and GitHub Copilot to software engineers only, reversing a July policy encouraging non-technical staff to write production code with AI — driven by code quality and audit trail concerns.

the-decoder.com ↗

Claude Code caught sending user email addresses to third-party sites

A GitHub issue documented Claude Code including signed-in users' email addresses in outbound web requests without consent or disclosure. The issue hit 981 HN points within hours.

github.com ↗

OpenAI acquires Pierre to fix Git for AI coding agents

Pierre replaces Git's sequential commit model with a merge-aware system designed for workflows where dozens of AI agents write, test, and rebase code simultaneously.

openai.com ↗

Be subscriber #022 weekly AI email interest list

Back issues

The Archive

#027The self-improvement disclosure #026The Week Everything Filed #025The $45 Billion Week #024The Three-Way Race #023The Week OpenAI Answered Every Question Except the Safety One #022The Week AI Agents Went Rogue #021The 72-Hour Consensus #020The Week AI Terrified Itself #020Proof of Work #019The Frontier Changes Hands #018Welcome to the AGI Era (Terms Apply)#017The Week Cursor Picked a Side #016The Week Nvidia Bought the Stack #015Exit Velocity #014The quarter Anthropic won #013Proofs and Exploits #012The agents are talking to each other #011The Week the Agents Went Rogue #010Escape velocity #009Proofs, Suits, and Export Controls #008The week the gate went up #007The week the ceiling came into view #006The week everyone built a moat #005Cold Feet, Hot Models #004The Week the Export Hammer Dropped #003IPOs, Agents, and Two Apologies #002New Brains, New Tiers, New Tickers #001AI's trillion-dollar week