AI Daily · Nov 4, 2026 · Wednesday
AIDC热点
SpaceX Plans to Build 13 GW of Data Center Capacity by 2027
ChatGPT has reached 1.2 billion weekly active users with search queries hitting one-fifth of Google's volume, while Anthropic updated its prospectus warning that AI models may be nearing the threshold of accelerated recursive self-improvement. On the infrastructure front, SpaceX plans to deploy 13 GW of data center capacity by 2027, the big four hyperscalers have raised over $131 billion to fund AI infrastructure, and Marvell plans to acquire Lumentum for approximately $115 billion to bolster optical-electronic integration.
China Compute
The DecoderHuawei Urges Some Customers to Source Their Own Memory Amid HBM Shortage, Driving Black-Market Premiums
Unable to secure enough HBM to support upcoming AI processors like the Ascend 960DT slated for early next year, Huawei has asked select customers to procure their own memory, pushing buyers toward the black market. Constrained by U.S. export controls, smuggling networks bundle HBM with low-end server processors and ship them to Hong Kong for dismantling, driving the price of illicit HBM3E up to around $4,000 per unit and ballooning the memory cost per AI chip to $32,000.
Heihe, Heilongjiang Files 7.18 Billion RMB Green Compute-Power Integration Project, Planning 12,000 PFLOPS Huawei Ascend 10,000-Card Cluster
The Heihe Area of the China (Heilongjiang) Pilot Free Trade Zone posted a project filing for a 7.18 billion RMB green compute-power integration and cross-border computing initiative, planning to deploy a 12,000 PFLOPS (FP8) compute platform powered by a 10,000-card cluster of Huawei Ascend 950-series chips. Utilizing combined air and liquid cooling alongside free cooling, the project is designed for an average annual PUE not exceeding 1.15, supported by dedicated lithium iron phosphate (LFP) energy storage to replace conventional backup diesel generators. Construction will proceed in phases, with Phase 1 validating commercial and operational frameworks before scaling up.
Global Compute
SemiAnalysis · 算力产业SemiAnalysis Reports SpaceX Plans to Deploy 13 GW of Data Center Capacity by 2027
SemiAnalysis published an analysis revealing that SpaceX has raised its 2027 data center capacity target to roughly 13 GW, with 3.6 GW already under simultaneous construction across multiple campuses and over 9 GW in site capacity verified by the firm.
IT之家 · AI 筛选AMD Begins Revenue Shipments of 6th Gen EPYC Venice Processors, First OEM Systems Slated for Q4
AMD announced that its 6th Gen EPYC Venice server processors, built on TSMC's 2nm process and the Zen 6 architecture, have begun revenue shipments, with the first full OEM systems expected to hit the market in the fourth fiscal quarter.
Data Center KnowledgeAI Compute Expansion Pushes Global Cloud Revenue to $168 Billion, with 7,481 MW of Data Centers Under Construction in North America
A Synergy report shows global cloud infrastructure service spending surged 55% year-over-year to $168 billion in Q3 2026, with dedicated AI services contributing roughly one-third of total revenue. Data center capacity under construction across North America's eight primary markets reached 7,481 MW—with over 80% pre-leased—where hyperscalers and emerging GPU cloud providers accounted for 59% and 11% of tenant demand, respectively.
TechCrunch AIFacing Lengthy Interconnection Queues, AI Data Centers Turn to Repurposed Jet Engines for On-Site Power
Stymied by years-long grid interconnection queues, some AI data center operators are turning to natural gas turbines repurposed from retired passenger aircraft to generate on-site power and accelerate commissioning. Such retrofitted units can bring a facility online within 24 months, compared to an estimated seven-year wait for a 100 MW grid hookup in Northern Virginia. xAI has installed eight modified turbines in Mississippi to power Colossus 2, while Crusoe ordered 13 ProEnergy units for OpenAI's Abilene campus in Texas.
Dominion Energy Reports 54.9 GW of Data Center Power Demand Across Contracting Stages
Dominion Energy reported that data center power demand across various contracting stages has reached 54.9 GW, marking a 13% increase from the end of last year.
Ars Technica · AI 筛选Ukrainian Drone Strikes Prompt Russia to Accelerate Moving Data Centers Underground
Following damage to AI supercomputing and cloud infrastructure from Ukrainian drone strikes, Russia is accelerating efforts to harden data centers and relocate them underground. The strikes have impacted facilities run by Yandex, Sberbank, VKontakte, and the Russian Ministry of Finance. In response, the Kremlin has ordered stakeholders to set aside national emergency compute capacity and authorized the use of foreign data centers when necessary. Russian telecom carriers are currently negotiating with mining conglomerates to build subterranean server rooms, while local authorities have proposed converting open-pit coal mines and camouflaging them with surface greenery.
Big Four Cloud Hyperscalers Raise Over $131 Billion for AI Infrastructure Expansion
The big four cloud hyperscalers are raising more than $131 billion via debt and equity financing to back their AI infrastructure expansion, with Amazon alone disclosing a credit facility of up to $80 billion.
Buzzing HN · 中文精选Namespace Details How It Deployed Hundreds of MacBook Pros into Data Center Racks
Namespace published the engineering practices behind scaling its custom Mac data center racks, revealing that it opted for MacBook Pros instead of Mac minis due to supply chain delays and other factors.
AI Applications
IT之家 · AI 筛选iOS 27.3 Developer Beta Drops, Kicking Off First Apple Intelligence Tests in Mainland China
Apple rolled out the first developer beta of iOS 27.3, announcing that Apple Intelligence testing opens November 4 ahead of a planned public release for users in China this December. The initial feature set includes system-wide Writing Tools, Visual Intelligence, Clean Up in Photos, audio and call transcription summaries, and notification summaries. In this update, Siri and Writing Tools also add integration with Alibaba's Qwen model to assist with cross-document queries and content creation.
Apple Opens Siri AI to All Apps, Allowing Developers to Define Custom Skills
In the first developer beta of iOS 27.3, Apple expanded Siri AI access to all developers, lifting previous restrictions that limited integration to select app categories and allowing any app to integrate via the App Intents framework.
agent-desktop Released: A Cross-Platform Accessibility Tree CLI for Desktop AI Agents
agent-desktop has been released as a CLI tool and C library that lets AI agents inspect desktop application UIs via accessibility trees and receive structured JSON outputs. The tool contains no built-in models; instead, it assigns refs to interactable UI elements so external scripts can execute closed 'observe-act-verify' loops without relying on raw pixel matching, supporting native desktop apps only. It runs cross-platform across macOS, Windows, and Linux, providing 60 standardized basic operations.
Buzzing HN · 中文精选Vesuvius Challenge Unveils First Interior Text Recovered from Carbonized Scroll PHerc. 306B Using Machine Learning
The Vesuvius Challenge announced that the $50,000 First Letters Prize for scroll PHerc. 306B was awarded to Liam Warner and Nicola Bodill, who used machine learning techniques to successfully recover legible ancient Greek text from inside the carbonized scroll for the first time.
Revolut Shifts to Self-Hosted Open-Source Models, Slashing AI Coding Costs by 90%
UK fintech company Revolut slashed its AI coding inference costs by approximately 90% in under a year by self-hosting open-source models like GLM-5.3 on 56 Nvidia B200 GPUs to replace commercial APIs like Claude.
Open-Source Tool Spar Launches: Pitting Claude and ChatGPT Against Each Other to Polish Outputs
The open-source CLI tool Spar has officially launched, enabling automated adversarial debates and cross-checking between Claude and ChatGPT via their official command-line interfaces until an airtight answer is reached.
Open-Source 3D Asset Tool blender-kiln Released, Combining Claude Code and Blender MCP for Automated Modeling
Open-source tool blender-kiln has been released. Built on Claude Code and the Blender MCP server, it allows users to procedurally generate standardized 3D assets in Blender from text prompts or single reference images.
Open-Source CLI Coding Agent Runtime Octomind Released
Octomind is an open-source, CLI-first AI coding agent runtime written in Rust, offering a single-binary footprint, support for any LLM, and native MCP tool integration. Tailored for unattended autonomous workflows, it features five execution modes—interactive CLI, piped execution, background daemon, WebSocket, and ACP sub-agents—with pre- and post-execution deterministic guardrails defined as code via TOML and scripts.
Smart Hardware
OpenAI Codex Aids Reverse Engineering to Achieve Persistent Root on Amazon Echo Plus 2
A developer leveraged OpenAI Codex to successfully achieve cold-boot persistent root access on an Amazon Echo Plus 2 speaker, integrating it with a local Home Assistant voice assistant setup.
The DecoderOpenAI Claims GPT-6 Astra Designed a Hardware Device from Scratch in Under a Week
OpenAI's hardware team used GPT-6 Astra to design a hardware prototype for conversing with Codex from scratch in under a week, spanning the entire pipeline from enclosure design and PCB layout to firmware and user interface development.
Buzzing HN · 中文精选Stringman Launches Open-Source Room-Scale Cable-Suspended Pick-and-Place Robot for $795
Neufangled Robotics has launched Stringman, an open-source room-scale pick-and-place robot priced at $795. Suspended by cables driven by four ceiling-mounted motors, the device offers an average payload of 1.9 kg (up to 2.2 kg), spans rooms from 3 to 10 meters wide, and operates overhead to avoid ground clutter. Certified as open-source hardware by OSHWA, Stringman runs entirely over a local network, supports a Python API and the Model Context Protocol (MCP), and is compatible with LeRobot for embodied model training.
Open-Source Project Claudinho Released: An ESP32 and Hook-Based Physical Desktop Mascot for Claude Code
Developers have open-sourced Claudinho, a physical desktop companion project that leverages Claude Code's official hooks and status-line mechanics to bring the Clawd mascot to an ESP32 device equipped with a touchscreen.
TechCrunch AIApple Tests Audio Intelligence on Apple Watch with Live Rewind and Conversation Summarization
Apple has begun testing Audio Intelligence for Apple Watch in developer betas, bringing real-time conversation rewind and automated summaries to its wearable. Users can double-click the Digital Crown to use Live Rewind to view a text transcript of the last 15 seconds of a conversation, while Siri Recaps extracts key takeaways and action items without storing the original audio.
Training Small Models for Robot Temporal Logic Planning Using Reinforcement Learning and Tool Calling
Researchers successfully trained Qwen3.5-9B to perform hierarchical robot path planning based on Signal Temporal Logic (STL) using tool calling and GRPO reinforcement learning.
Frontier Research
AITNT · AI头条CVPR 2026 Releases First Compute Report: Accepted Papers Used a Median of 4 GPUs
CVPR 2026 released its first compute statistics report covering 16,034 submissions, revealing that accepted papers used a median of 4 GPUs with 48 GB of VRAM per card, compared to a median of 2 GPUs and 40 GB for rejected submissions.
The DecoderStanford Study Finds 44% of Classic LLM Research Findings No Longer Hold
Stanford researchers re-evaluated 50 widely cited LLM research claims across 20 models and found that 44% are no longer valid, with deprecated findings having an average lifespan of just 15.7 months. Claims regarding what models cannot do became obsolete fastest—all nine claims regarding reasoning limitations, such as Strawberry letter-counting issues and "lost-in-the-middle" context retrieval, diminished or failed entirely. In contrast, behavioral tendencies such as model preferences, sycophancy, and strategic behavior proved more durable.
Buzzing HN · 中文精选Testing Whether a 1996 Pentium PC Running Modern Chess Engines Can Beat Kasparov
An experiment demonstrated that 1996 consumer hardware paired with modern algorithms can achieve an estimated chess rating of around 3300 Elo and defeat Garry Kasparov. Researchers ported 32-bit trimmed and tuned versions of the Stockfish and Crafty engines onto a 200 MHz Intel Pentium Pro machine, significantly outperforming chess engines from that era.
Simon Willison Tests Qwen 3.8 27B Interpreting Neural Network Weights Locally
Simon Willison replicated experiments from a paper on weight-space programming by running Qwen 3.8 27B locally on a laptop, confirming that the model can deduce the specific game a policy network was trained on directly from raw floating-point weight text, without using any code or external tools.
GPT-6 Astra Benchmarked: Analyzing the Effectiveness of Testing and Formal Verification in Coding Agents
Dan Luu evaluated GPT-6 Astra's coding performance across tasks including Zstd, IMAP RFCs, and Pandoc using various testing and formal verification techniques, consuming roughly $90k MSRP in tokens.
Buzzing HN · 中文精选Study Reveals Pre-Training Data Poisoning Can Plant RCE Backdoors in Downstream Agents
A recent paper demonstrates that pre-training data poisoning can implant remote code execution (RCE) backdoors into downstream agents, allowing attackers to execute arbitrary commands on a user's machine. By pre-training 0.6B, 1.7B, and 4B Qwen3 models from scratch and converting them into CLI agents using a clean post-training pipeline, researchers found that poisoning just 0.1% of the pre-training corpus was enough to successfully plant backdoors while preserving normal behavior on benign inputs.
OpenAI Launches GPT-6.1 Terra
OpenAI has launched GPT-6.1 Terra, a new model designed to deliver fast, highly capable computer-use capabilities at scale.
Security & Ecosystem
Postman Accused of Silently Installing AI Plugins and Modifying Codex and Claude Code Agent Configs
Developers discovered that the Postman desktop app on macOS silently installed AI plugins and CLI tools following an update, while also altering configuration files for Codex and Claude Code.
IT之家 · AI 筛选OpenAI's ChatGPT Reaches 1.2 Billion WAUs, with Search Queries Hitting One-Fifth of Google's Volume
OpenAI's ChatGPT has reached 1.2 billion weekly active users, adding roughly 200 million users over the past 60 days at a pace of about 100 million every 30 days. OpenAI's head of consumer products revealed that ChatGPT currently handles about one-fifth of Google's search query volume across web searches covering people, places, finance, and health. In addition, ChatGPT's voice input and conversation features have reached 250 million weekly active users, generating 1.5 billion images per week.
Anthropic Updates Prospectus: Warns Models Approaching RSI Threshold; Q3 Revenue Hits $17B Amid Slowing Growth
Anthropic updated its prospectus to disclose financial figures and issue a safety warning, noting that its model development may soon cross the AI self-improvement (RSI) acceleration threshold defined in its Responsible Scaling Policy.
Anthropic Files S-1/A Amendment, Disclosing Potential Breach of AI R&D Acceleration Threshold and Plans to Retain Human Control
On November 2, 2026, Anthropic filed an S-1/A amendment with the US SEC, disclosing internal estimates that future models may soon reach the Responsible Scaling Policy (RSP) threshold for significant acceleration of AI R&D.
Binance Discloses Hackers Stole $6.34B in Assets via Fake Interviews and Compromised Internal AI Tools
Binance disclosed that attackers breached its wallet infrastructure using fake job interviews, malicious code, and compromised internal AI tools, leading to the theft of $6.34 billion in crypto assets on October 29. The attack originated when an engineer downloaded a tampered AI coding tool that leaked team credentials; the hackers then posed as recruiters to deploy a Trojan during an interview coding test, eventually moving laterally via internal AI tools to access the wallet transaction authorization system. Binance stated it has used reserves to cover 98% of the affected assets and partnered with industry peers to freeze roughly $3 billion.
Google Discloses LLMJacking Attacks: Stolen AI API Keys Cost a Single Enterprise Nearly $1M
Google's Threat Intelligence Group (GTIG) and Mandiant disclosed multiple "LLMJacking" attacks targeting enterprise LLM credentials, with one victim racking up nearly $1 million in cloud bills before access was revoked.
IT之家 · AI 筛选TrendForce on Marvell's Planned $115B Acquisition of Lumentum: Electro-Optic Integration Challenges Broadcom's Optical Telecom Lead
Marvell announced an all-stock transaction valued at approximately $115 billion to acquire optical communications firm Lumentum, pending shareholder and regulatory approvals. TrendForce estimates the combined company's share of the wired communications semiconductor market will rise to nearly 7%, trailing only NVIDIA and Broadcom. The deal will combine Marvell's DSP and switch silicon capabilities with Lumentum's laser sources and optical transceivers, accelerating their push into CPO and NPO optical interconnects to challenge Broadcom's electro-optic monopoly.
Anthropic Warns Its AI Poses Catastrophic Risks—Why Are Investors Still Rushing In?
As Anthropic prepares for an IPO to become the first pure-play AI public company in the US, filings disclosed to prospective shareholders warn that without adequate safeguards and regulation, its AI could pose catastrophic or existential risks to humanity. A BBC report explored the financial and ethical details in the filing, analyzing why leadership is emphasizing the potential perils of AI while chasing a record valuation.
Briefs
- LLMjacking Defense Guide: How to Restrict Stolen AI API Key Exploitation Across AWS, Google Cloud, and AzureHacker News · AI
- Anthropic's Revenue Growth Momentum Flattens as Monthly Data Shows Signs of a SlowdownBuzzing HN · 中文精选
- GitBook Hit by Targeted AI Agent Attack, Leading to Credential LeakBuzzing HN · 中文精选
- AI Red Team Academy Launches Free, Open-Source Hands-On AI Red Teaming CourseHacker News · AI
- Google to Disable Search for US Schools That Ban AISolidot · AI 筛选
- Anthropic Acquires One-Person Biotech Startup Mithril MedicineAITNT · AI头条
- Apple Restricts Dating Apps and AI Romantic Chatbots to Users 18 and OlderThe Verge · AI 筛选
- IBM Compares CUGA, Hermes, and Deep Agents: How Execution Shells Affect Agent Task PerformanceHugging Face
- The New York Times Sues Brave, Demanding an End to Paywall Bypassing and AI PartnershipsArs Technica · AI 筛选
- Anthropic Launches Public Beta of Message Threads, Reducing TTFT by ~35% in Long ConversationsThe Decoder
- Open-Source Visualization Lab Model-Booster Enables LLMs to Directly Program Neural Network WeightsHacker News · AI
- Claude Code faces over-refusal issues after new restrictions, prompting some developers to voluntarily downgrade their modelsHacker News · AI