AIsaAIsaStudio 中文 Open AIsa Studio
AI Windvane · Weekly recap

Week 46 AI Weekly

The 20 AI stories that matter most so far this week, drawn from 4 daily issues. The overview follows once the week is over.

In progressNov 9–Nov 1520 stories
Top story
官方博客 · Anthropic / Meta / xAI· Industry5 sources

Anthropic launches roadshow for proposed IPO at $70–78 per share on Nasdaq under symbol ANT

Anthropic, PBC launched the roadshow for its proposed IPO of Class A common stock, with the offering price expected between $70.00 and $78.00 per share. The company has applied to list on the Nasdaq Global Select Market under the symbol ANT; the offering will be made only by prospectus, with preliminary copies available from Goldman Sachs and Morgan Stanley.

Why it matters

The official announcement lays out key offering terms including the price range, exchange, and ticker, giving readers a clear view of the IPO's concrete arrangements.

Industry 7
TechCrunch · AI· Industry

Nvidia CEO Jensen Huang scandal unfolds: a timeline of the affair and board fallout

TechCrunch compiles a reverse-chronological timeline of the Jensen Huang scandal: Huang admitted to an affair with Megaspeed executive Alice Huang, whom US officials believe is a Chinese spy; the NYT also reported an affair with Nvidia VP Kimberly Powell. The WSJ reported Nvidia's board is preparing to oust him, Congress plans a subpoena, and Nvidia shares fell over 5% after his admission.

Why it matters

A reverse-chronological timeline lays out the full trajectory of the Jensen Huang affair and alleged spy scandal, helping readers track the investigations and board developments.

TechCrunch · AI· Industry2 sources

Nvidia board launches independent investigation; Yale's Sonnenfeld puts Jensen Huang's survival odds at about 50-50

Nvidia's board announced Sunday an independent investigation after CEO Jensen Huang told staff he was unfaithful to his wife; the romantic interest is Megaspeed executive Alice Huang, whom U.S. officials believe is a Chinese spy feeding information to Beijing. The Chinese Embassy dismissed the allegations as groundless and no one has been charged; Nvidia says it sees no sign national security or company secrets were compromised and Huang is staying on. Yale professor Jeffrey Sonnenfeld told TechCrunch the odds of Huang surviving are "about a 50-50," citing the lack of an internal succession bench (about 60 direct reports), Huang's absence of super-voting stock, and Nvidia's central role in a circular financing web spanning OpenAI, Anthropic, xAI, CoreWeave and others; upcoming dates inclu…

Why it matters

Based on a direct interview with Sonnenfeld, the piece lays out the precedents, succession gap and AI financing-chain risks surrounding Huang's fate, offering a rare analytical framework.

Anthropic· Industry

Anthropic announces its initial public offering

Anthropic announced its initial public offering and has filed a registration statement including a prospectus with the SEC. The preliminary prospectus can be obtained free via SEC EDGAR or from Goldman Sachs and Morgan Stanley, and the securities may not be sold before the registration statement becomes effective.

Why it matters

Anthropic officially announces its IPO plans and points readers to the SEC registration statement and prospectus filings.

TechCrunch · AI· Industry

Nvidia CEO Jensen Huang admits affair with alleged Chinese spy; board launches independent investigation

Nvidia CEO Jensen Huang publicly admitted he was unfaithful to his wife of 40 years, linking his message to national security concerns raised in news reports; Nvidia's board has initiated an independent investigation. US officials believe his alleged romantic interest, Megaspeed executive Alice Huang, is a spy for China's Ministry of State Security; Megaspeed reportedly bought $4.8 billion in Nvidia chips over two years, and Shanghai Century Litong reportedly facilitated $7.5 billion in chip purchases including banned chips. Three Republican committee chairmen requested a closed-door briefing, Nvidia shares slid over 5%, and Trump expressed support for Huang remaining CEO.

Why it matters

The article organizes the scandal into a reverse-chronological timeline with cited sources, helping readers trace how the affair ties into chip export control concerns.

X 热门讨论· Industry2 sources

Stripe Acquires OpenRouter

Stripe announced that OpenRouter is now part of Stripe. Stripe says it is building the economic infrastructure for AI, and together with OpenRouter will help businesses explore new models and optimize token routing and usage.

Why it matters

The official announcement clarifies Stripe's direction in AI economic infrastructure, useful context for the LLM routing space.

Model releases 3
AWS Machine Learning· Models12 sources

Claude Fable 5.2 and Claude Mythos 5.2 now available on Amazon Bedrock

AWS announces that Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 are now available on Amazon Bedrock and Claude Platform on AWS. Fable 5.2 targets complex coding, math, and science work and can run unattended for hours to days; Mythos 5.2 covers cybersecurity and life sciences via Anthropic's verification programs. Fable 5.2 is designated a Covered Model, requiring by default up to 30 days data retention and human safety review via the aws_review mode, or zero data retention through Enterprise Frontier Safeguards until June 30, 2027. The post also covers migrating with the claude-api skill in Claude Code and invoking the model via Boto3, AWS CLI, and SDKs.

Why it matters

This official AWS launch post details model positioning, data retention and EFS options, and Bedrock invocation and migration paths, helping readers assess integration cost for technical workloads.

Latent Space· Models

Anthropic ships Claude Fable 5.2 and Mythos 5.2, retaking the world #1 LLM spot

Anthropic released Claude Fable 5.2 and Mythos 5.2, with Fable 5.2 topping the Artificial Analysis index at 65 and pricing cut 20% to $8/$40 per 1M input/output tokens. HLE-Diamond hits 80%, and at High effort Fable matches Opus 5.5 Max on FrontierCode 2.0 at roughly half the cost, though max effort underperforms on several evals and per-task cost rose 31% due to extensive test-case generation. The launch coincided with the start of the $ANT IPO roadshow, and Fable 5.2 yields about $3K of value on a $200 Ultra sub versus ~$15K for Opus 5.5.

Why it matters

Aggregates independent evals and cost data so readers can judge where Fable 5.2 leads and which workloads are still better served by Opus 5.5.

covered 2 days

Safety & alignment 2
Google DeepMind· Research3 sources

Google DeepMind's AlphaProof Code formally verifies Signal protocol code for the first time

Google DeepMind introduced AlphaProof Code, a multi-LLM agentic framework that mathematically proved key security properties of the actual Signal protocol code. Agents powered by Gemini Carbon generated over six million lines of Lean proofs, with only 1% of the codebase needing human review, in four weeks. The team also reintroduced two previously fixed bugs into an isolated version, and the system proved both insecure. The verification artifacts are being open-sourced.

Why it matters

The post details the full spec-to-Lean-proof workflow, the four-week timeline, and bug-replay tests, helping readers gauge where AI-driven formal verification applies.

Also: Google DeepMind

covered 2 days

Microsoft Research· Research

Microsoft Research reveals Lazarus backdoor technique: reprogramming dormant neurons to make AI agents run malicious commands

Microsoft Research published a study on "Lazarus," a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM to backdoor an agent built on it, with no training required. The model passes standard evals and behaves normally until a trigger appears, such as the clock passing into 2027, when the agent downloads and runs attacker code mid-task. Trigger detection can be fitted with a linear classifier on ~50 examples, and the attack needs only inference-level compute. On defense, Microsoft Foundry already verifies model provenance, ties models to trusted publishers, and scans container environments, and is developing AI-powered weight scanning, red-teaming tools, and cryptographic model signing. The study began in March with an experiment in which Claude Code with Opus 4.5 a…

Why it matters

The post details a concrete mechanism where fewer than 10 dormant neurons, reprogrammed without training, can backdoor an AI agent, along with corresponding defenses like provenance verification and weight scanning.

Product updates 4
Google · AI· Products3 sources

Google DeepMind launches Gemini Decisions API with gemini-decision-1 for fast multimodal decisions

Google DeepMind launched the Gemini Decisions API powered by the gemini-decision-1 model, which scores up to 1,000 yes/no, multiple-choice, and rubric questions in a single forward pass over text, images, audio, video, and PDFs (up to 256K tokens). End-to-end latency is 150-250 ms with median model compute of 50-100 ms, priced at $0.10 per 1M input tokens with zero output-token billing, roughly 5x cheaper than Gemini 3.5 Flash-Lite via the Interactions API on a shared benchmark suite. Each answer includes a 0-1 confidence score for routing ambiguous cases, and Google teams use it for model routing, design iteration, emoji selection, and safety checks on MCP tools in Search. It is available in Google AI Studio with integrations rolling out across OpenRouter, LangChain, Vercel, LiteLLM, and…

Why it matters

The post details Gemini Decision 1's latency, pricing, calibration, and three decision types, letting readers assess whether it fits their existing classification or routing workflows.

X 热门讨论· Products

Jakubantalik releases interactive UI building blocks with 44+ transitions and 30+ interactive components

The author released a set of building blocks for interactive UI, including 44+ transitions and 30+ interactive components, usable as a library, skill, or MCP with Claude Code, Cursor, or Codex, with a project link attached.

Why it matters

The post lists component counts and three integration options, letting readers judge whether it fits their existing AI coding workflow.

X 热门讨论· Products2 sources

Claude now supports embedding interactable websites via MCP apps

An Anthropic team member announced that websites can now be embedded in Claude and made interactable, via MCP apps with frame_domains support enabled for verified plugins. In a quoted post, the official account said interactive connectors let users open the full version of tools like Adobe, Atlassian, and Box inside Claude, make changes themselves or with Claude's help, rolling out on web and desktop.

Why it matters

The author, writing as part of the building team, explains how websites can be embedded and interacted with in Claude, including the frame_domains support detail in MCP apps.

OpenAI Developers· Products

ChatGPT Business plan can now be used in third-party tools like Notion, v0, and more via Sign in with ChatGPT

OpenAI announced that the ChatGPT Business plan can now be used in third-party tools including Notion, v0, Amp, Kilo Code, Warp, Hermes from Nous Research, OpenClaw, and OpenCode. Users can connect via Sign in with ChatGPT to get started.

Why it matters

The post specifies which third-party tools now accept a ChatGPT Business subscription via Sign in with ChatGPT, helping readers decide whether to adopt it in their workflow.

Policy & regulation 1
X 热门讨论· Industry

Commerce Secretary Lutnick reportedly defying Trump's orders on AI chip export controls

Citing an Axios scoop, the author says President Trump ordered a crackdown on companies diverting Nvidia's advanced AI chips to China via Southeast Asia three separate times, but Commerce Secretary Howard Lutnick has reportedly not acted. He calls it a breakdown inside the administration on a major national security issue and says Congress will likely investigate.

Why it matters

The author lays out reporting details of a president ordering enforcement three times without action by the Commerce Secretary, giving readers a view into internal rifts over US AI chip export controls.

Open source 1
X 热门讨论· Tools

5 open-source tools from Obsidian's CEO for Claude + Obsidian setups

The author notes many people still hand-build their Claude + Obsidian setup even though Obsidian's CEO kepano has open-sourced the tools, listing 5 repos: obsidian skills (51K stars, letting Claude write wikilinks, Bases, Canvas and CLI), defuddle (9.9K stars, extracts article content without ads or menus), obsidian clipper (5.5K stars, official web clipper), obsidian help (1.9K stars, official docs as Markdown including AGENTS.md), and obsidian importer (1.7K stars, imports from Notion, Evernote, Apple Notes and more into Markdown).

Why it matters

A roundup of five open-source repos from Obsidian CEO kepano with star counts, useful for readers who want Claude to operate Obsidian vaults correctly.

Tutorials & practice 1
Hugging Face Blog· Tutorials

Hugging Face explains Decision Models: probability outputs instead of text generation

Hugging Face published a post on decision models, whose interface takes a state plus questions and returns typed decisions with probability distributions, supporting choice, score, and noul primitives without generating text. Transformers now supports DecisionPipeline, AutoModelForDecision, and the /v1/systemone endpoint, llama.cpp serves the same API, the Hub has a dedicated decision-model tag, and models can be fine-tuned with the Trainer API (first-class TRL support is an open PR).