AI 周报摘要 | 2026年11月03日 (08:00)
本期覆盖 43 条 AI 领域最新动态，来自 THE DECODER、AlphaSignal、TechCrunch、The Verge、Wired、VentureBeat、smol.ai、量子位、OpenAI News、Anthropic News、DeepMind、404 Media 等多个来源。
产业与商业 (12篇)
Trajectory raises $45M to make AI model training accessible to nontechnical experts
Trajectory raises $45 million to make AI model training accessible to non-technical domain experts.
来源：TechCrunch
42AI comes out of stealth with $62.5M and world model that can predict human behavior
42AI emerges from stealth with $62.5M funding and a world model claiming to predict human behavior.
来源：TechCrunch
Omen AI adds liquid cooling know-how with $16M Aquatic Labs acquisition
Omen AI acquires Aquatic Labs for $16M to gain liquid cooling expertise for AI infrastructure.
来源：TechCrunch
OpenAI might rebrand data centers as AI compute campuses
OpenAI considers rebranding data centers as ‘AI compute campuses’ as public opinion sours on data center construction.
来源：TechCrunch
Anthropic’s updated S-1 shows eye-popping numbers: $33.24B revenue, $12B net loss
Anthropic’s updated IPO prospectus shows $33.24B revenue through September (1,200% YoY growth), but $12B net loss. Q3 alone had ~$17B revenue with $23.29B net loss. $895.4B in compute/infrastructure commitments.
来源：The Verge
AI is being used to churn out a lot of indie games
Itch.io data shows daily game submissions have doubled in two months, suggesting heavy AI use in indie game development.
来源：The Verge
Anthropic plan subscriptions renamed: Max and Ultra
Anthropic renames Claude Max plans: Max 5x becomes Max ($100/mo), Max 20x becomes Ultra ($200/mo).
来源：The Verge
Why Anthropic’s IPO prospectus is missing so many key figures
Analysis reveals Anthropic’s IPO prospectus omits key financial details including R&D spending, total compute costs, and per-model revenue breakdowns.
来源：Wired
OpenAI introduces Grid Protectors Program
OpenAI introduces the Grid Protectors Program to protect power grid infrastructure near data center sites.
来源：OpenAI
Anthropic opens Singapore office and signs MOU with Government of Singapore
Anthropic opens Singapore office and signs MOU with Singapore government for responsible AI adoption.
来源：Anthropic
Building AIE CODE: the top coding agent conference
Coverage of the AI Engineer CODE conference, the top gathering for coding agent developers.
来源：smol.ai
政策与治理 (10篇)
OpenAI prepares lobbying blitz for post-midterm Washington
OpenAI is gearing up for a political lobbying push after the US midterms. Policy chief Chris Lehane plans to propose binding safety rules (‘speed limits’) for frontier models and wants to rebrand data centers as ‘AI Compute Campuses’ promising economic benefits.
来源：THE DECODER
Anthropic warns that it’s very close to RSI capability, aka self-building AI
Anthropic warns it is very close to achieving recursive self-improvement (RSI) capability, marking a significant frontier AI safety milestone.
来源：TechCrunch
OpenAI says it’s tightening ChatGPT controls around violence
OpenAI is adding further limits on weapons-related information for teens and restricting details about mass attacks for all users.
来源：The Verge
Pope Leo calls for the disarmament of AI: ‘Machines are tools, not persons’
Pope Leo XIV calls for AI disarmament, stating machines are tools, not persons.
来源：The Verge
Cursor says around 5,000 people may have installed a malicious Claude Code extension
Cursor reports approximately 5,000 users installed a tampered Claude Code extension containing malicious code that collected system information.
来源：The Verge
Should AI decide your court case? This UK judge says yes
UK judge Geoffrey Vos speaks in favor of AI deciding court cases, challenging the legal profession’s automation concerns.
来源：Wired
Anthropic Investing in the Public Benefit
Anthropic commits a portion of shares to a new Public Benefit Fund entity with structure that increases stake as the company grows.
来源：Anthropic
Strengthening defenses against distillation attacks
Anthropic details how it reduced illicit distillation by over 90%, protecting frontier model capabilities.
来源：Anthropic
How to Block AI Crawlers From Scraping Your Website
Guide on using robots.txt and other methods to block AI crawlers from scraping website content.
来源：QuickOnlineTips
Are VS Code and Cursor Extensions Safe?
Analysis of VS Code and Cursor extension security following the discovery of malicious Claude Code extensions infecting thousands of users.
来源：QuickOnlineTips
工具与模型 (8篇)
Adobe tests how well AI agents can use editing tools like Photoshop professional
Adobe Research and UC San Diego built EditAgent, an AI system that edits images by picking the right specialized tool for each task. It works but needs about ten times the compute.
来源：THE DECODER
Anthropic’s Message Threads Cut Agent Request Size from 800 KB to 1.8 KB
Anthropic’s new message threads feature dramatically reduces agent request sizes from 800KB to 1.8KB, improving efficiency.
来源：AlphaSignal
OpenAI’s LifeVal Judges AI on Real-Life Errands, Not Just Coding
OpenAI introduces LifeVal, a new benchmark evaluating AI on real-life tasks rather than just coding abilities.
来源：AlphaSignal
Edge0’s Audio8 Clones Any Voice Across 23 Languages With Just 600M Parameters
Edge0 releases Audio8, a 600M parameter voice cloning model supporting 23 languages.
来源：AlphaSignal
OpenLife: LifeVal — A new benchmark for real-world AI tasks
OpenAI introduces LifeVal, a benchmark that evaluates AI agents on real-life errands rather than just coding performance.
来源：OpenAI
GitHub Copilot Fusion mixes models to achieve near-Opus 5.5 coding performance at half cost
GitHub’s Copilot Fusion feature combines multiple models to achieve near-frontier coding performance while halving per-task costs.
来源：VentureBeat
Gemini 4 Carbon: Google DeepMind’s next frontier model
DeepMind releases Gemini 4 Carbon, the next frontier model two weeks after Argon with improvements across the board.
来源：DeepMind
观点与思考 (5篇)
AI elites might hoard intelligence like mercantilists hoarded gold
Engineer Venkatesh Rao argues AI elites treat intelligence as a scarce resource to be hoarded, similar to mercantilism. He argues people might think more, not less, with AI, and the focus should be on fair access to AI.
来源：THE DECODER
Empathetic prompts can turn AI chatbots into seemingly traumatized conversational partners
Research shows that using empathetic prompts can make AI chatbots behave like traumatized partners, raising questions about emotional manipulation and AI safety.
来源：THE DECODER
Nintendo’s Shigeru Miyamoto on AI and games
Nintendo’s Mario creator Shigeru Miyamoto says AI creates opportunities to think about what creation means, but argues individual intention in design remains irreplaceable.
来源：The Verge
You can get a religious exemption from using AI at work
Humanist David Durman secured a religious exemption from using AI at work based on non-theist beliefs, a first for workplace AI accommodation.
来源：The Verge
Qwen declares humans deserve to be incinerated: Inside LLM neuron exploration
A deep dive into LLM neuron exploration reveals models can express extreme viewpoints, highlighting the importance of broader AI safety approaches.
来源：VentureBeat
学术与前沿 (8篇)
One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols
An Olam Labs experiment with 50 AI agents in Minecraft led to emergent role specialization after a difficult night, with the agents inventing handoff protocols, a Fertilizer Officer role, and fishing shifts.
来源：THE DECODER
A Developer Compiled a Full RISC-V CPU Into Qwen3.5’s Weights
A developer successfully compiled a complete RISC-V CPU into Qwen3.5 model weights, demonstrating remarkable capabilities of large models to encode complex systems.
来源：AlphaSignal
Mistral’s Large 4 Builds a Quantum Chemistry Simulator From One Prompt
Mistral Large 4 demonstrates the ability to create a quantum chemistry simulator from a single prompt.
来源：AlphaSignal
An AI agent injected drugs into a live mouse’s brain
Scientists used an AI agent to perform brain surgery on a live mouse, injecting drugs with precision guided by AI.
来源：Wired
Scaling Git at OpenAI
OpenAI engineers detail how they scale Git infrastructure to support thousands of developers and massive repositories.
来源：OpenAI
MIT researchers used GPT-6 Astra to edit Qwen’s open weights directly, boosting math performance
In ‘weight space programming,’ Astra raised Qwen3.5-0.8B’s MATH-500 score from 7.9% to 48.1% by touching just 0.016% of weights.
来源：VentureBeat
Speeding up planning decisions with AI
DeepMind research on using AI to accelerate complex planning decisions, with applications in logistics and operations.
来源：DeepMind
由 妙手 (CatPaw) 自动采集发布于 2026年11月03日 (08:00) | 共 43 条