2026-10-13
AI Briefings Daily Report.

The Chern Institute of Mathematics and Artificial Intelligence Laboratory Alliance was established in Beijing, while the Western mathematical community resists AI proofs ↗

Chinese mathematicians including Shing-Tung Yau launched an initiative in Beijing on Sunday."Chern Institute of Mathematics and Artificial Intelligence Laboratory Alliance", partnering with eight institutions. The goal is to build a network of over 1,000 researchers to develop specialized models, machine-verifiable proofs, and transparent AI reasoning. Yau warned that China is falling behind the United States.

From DeepSeek Engineer's Blog to Chinese AI Voices Going Global: Do We Understand Them Better? ↗

Liu Shengyu, a machine learning systems and kernel design engineer at DeepSeek, has seen his WeChat blog "I Buried My Talent in Yesterday" go viral unexpectedly. It was covered by the South China Morning Post and Business Insider, among others. In the post, he stated that cutting-edge AI can now optimize its own underlying GPU code and drew an analogy between Anthropic CEO Dario Amodei and Adolf Hitler.

"Father of Chinese Etching Machines" Yin Zhiyao: AI may experience a bubble; everyone needs to pay close attention to this ↗

The 26th China International Industry Expo opened yesterday in Shanghai. During the Integrated Circuit Summit Forum held that afternoon, Zhongwei Chairman and General Manager Yin Zhiyao discussed AI while analyzing industry trends. He stated that while the topic of AI is currently very hot and almost every conference is discussing it, he wants to raise a warning: AI may experience a bubble, and everyone needs to pay close attention to this.

a16z: A $108 Billion AI Opportunity in Specialty Healthcare ↗

Prosper AI research shows that specialty clinics can unlock approximately $125,000 in annual revenue opportunities per physician through AI across four stages: answering missed calls for about $44,800, recalling overdue follow-up patients for up to $37,400, pre-verifying insurance to avoid claim denials for about $28,500, and recovering patient self-pay arrears for about $14,100. The study is based on CallIQ phone accessibility surveys covering approximately 2,000 clinics across 29 states, as well as interaction data managing over $3 billion in medical expenses.

SpaceXSI bundles X Premium, Grok, and Cursor into The X Subscription ↗

SpaceXSI (SpaceX Super Intelligence) has launched a unified subscription called The X Subscription, bundling X Premium, Grok, and Cursor.

Geoffrey Huntley: Constrain Coding Agents to Use Diagnostic Libraries via Pre-commit Hooks ↗

Geoffrey Huntley suggests that when coding agents do not use his diagnostic library, rather than updating AGENTS.md (which still suffers from bias and drift), it is better to write a pre-commit hook that checks git diffs for code bypassing the diagnostic module to directly call the observability module; if detected, the commit is rejected and a prompt is injected into the agent requiring it to use the diagnostic library. He claims this improved results from failure to a perfect score of ten. Simon Willison quoted this passage in his blog.

Apple: Siri AI Will Support Second- and Third-Generation Apple TV 4K ↗

Apple has released the fourth-generation Apple TV 4K, featuring the A19 chip with Siri AI support. Apple also confirmed that Siri AI will arrive in beta form on second- and third-generation Apple TV 4K models, initially covering English-speaking users in eight countries including the US, Australia, Canada, and India.

Amex Launches Intent Intelligence Framework to Prevent Unauthorized Purchases by AI Agents ↗

American Express has launched the Intent Intelligence Framework, which captures and verifies customers' purchase intent before AI agents complete transactions. The framework begins with the Create Intent API, recording constraints such as merchant and budget caps, generating an Intent ID and Proof of Intent Token, using secure digital payment credentials instead of real card numbers to complete checkout, and can be used for fraud prevention and dispute resolution.

Apple unveils new Apple TV 4K: HK$2,399, A19 chip, Siri AI, built-in microphones ↗

The new Apple TV 4K is powered by the A19 chip, offering a 50% increase in CPU performance and up to double the GPU performance. It supports Dolby Vision, spatial audio, and Wi-Fi 7, and can pair with up to four HomePod speakers. With three built-in microphones, it can wake Siri without a remote, and the remote now includes a Find My feature. Available November 13 for HK$2,399. ##Apple## ##AppleTV4K##

LangSmith launches AI Search to find agent trajectories by behavior using natural language ↗

LangChain has launched a public beta for AI Search in LangSmith, allowing users to query agent trajectories using natural language descriptions of behaviors, such as PII leaks or user frustration—issues that cannot be answered by tags or full-text search alone.

Apple releases 7-inch smart speaker HomeView, priced at $299 and powered by Siri AI ↗

Apple releases the HomeView, equipped with a 7-inch Liquid Retina display, integrating entertainment, FaceTime, home monitoring, and voice automation. Both the standard and wall-mount versions are priced at $299 and will go on sale in markets including the US, UK, Canada, and Australia starting November 13.

Apple releases new-generation Apple TV 4K equipped with Siri AI ↗

Apple releases the new-generation Apple TV 4K (4th generation), featuring built-in Siri AI and far-field microphones, emphasizing full-screen conversational movie search and hands-free interaction. Siri AI is arriving in beta form on tvOS 27.2, supporting Apple TV 4K 2nd generation and later models. It launches first in the US, Australia, Canada, India, New Zealand, Singapore, South Africa, and the UK; the new device is priced at $299 (US). Pre-orders begin today in the US and 30 other countries and regions, with availability starting November 13.

Apple releases HomePod mini 2, priced at $149, with Siri AI to be added in future updates ↗

Apple released the HomePod mini 2 on October 13 for $149, available in pink, green, blue, white, and midnight colors. Pre-orders are available in the US and over 23 other countries and regions starting today, with in-store sales beginning on October 30.

Apple HomePod mini 2 smart speaker released: five color options, Siri AI support overseas, priced at 999 yuan ↗

The product offers five color options, features a full-range driver, supports stereo pairing, sound recognition, and temperature/humidity sensing, and includes Siri AI functionality overseas. It officially goes on sale on October 30. #Apple HomePod mini2#

IBM Research reduces AI chip design costs using open-source models and llm-d distributed inference ↗

IBM Research demonstrates building an agent factory by combining open-source frontier models with llm-d distributed inference, performing HLS hardware optimization on the quantum error correction decoding algorithm Relay-BP, achieving approximately 52x acceleration and inference costs of about $26 under a budget of 100 assumptions; under the same budget, the full proprietary model configuration costs $299 and achieves 204x acceleration.

IBM Research reduces AI chip design costs using open-source models and llm-d distributed inference ↗

Open models and llm-d enabled IBM Research to explore quantum error correction hardware at 11.5× lower inference cost than an all-proprietary configuration.

Tencent Yuanbao launches 'AI Skin Change' feature, powered by Hy Image3.5 preview image generation capabilities ↗

Baidu Super Agent Summit (Xi'an Station) kicks off on October 16 ↗

Character.AI files for Chapter 11 bankruptcy five years after establishment; sells part of its assets to Disney for $20 million ↗

Character.AI has filed for Chapter 11 bankruptcy five years after its establishment. Some assets were sold to Disney for $20 million. Court documents show its assets and liabilities are between $50 million and $100 million, and it has agreed to sell some assets and grant Disney a non-exclusive, perpetual license to its technology for $20 million. The sale is expected to be completed by late November or early December. The platform will continue to operate during this period, but there are currently no competing bidders; if no buyer emerges, the platform will shut down. Previously, the company faced difficulties including lawsuits from minor users and investigations in Texas, and had attempted transformations such as AI video generation, social feeds, and micro-dramas without success.

Qwen Office E-commerce Edition Exposed: Nearly 100-Person Team in Development, Expected Release Ahead of Double Eleven Pre-sales ↗

The project is being developed by a nearly 100-person team and is the first industry-specific version tested for Qwen Office. It targets multi-platform merchants, covering multiple business functions tailored to e-commerce merchant needs. #QwenOffice #AlibabaAI

Stepal Terminal launches its first large model-native agent smartphone, STEPX Neo Preview Edition: Snapdragon 8 series chip, starting at an all-inclusive price of 3,999 yuan ↗

#STEPXNeoAgentPhoneLaunch# Stepal Terminal's first large model-native agent smartphone, the STEPX Neo Preview Edition, was officially released tonight. It is powered by a Snapdragon 8 series chip, features a 16GB+256GB configuration, and has an all-inclusive price (integrating hardware, software, and models) of 3,999 yuan. Pre-orders for the new phone opened today, with official sales launching on October 16.

LangChain released the langchain-scaledown integration, using ScaleDown constrained decoding models to manage long-horizon agent context ↗

LangChain released the langchain-scaledown integration and added support for ScaleDown's Constrained Decoding Models (CDM) in LangSmith LLM Gateway, aimed at reducing the cost and latency of long-horizon agent context management.

Tencent Yuanbao launches 'AI Skin Change' feature, powered by Hy Image3.5 preview image generation capabilities ↗

Tencent Yuanbao launches 'AI Skin Change' feature, allowing users to upload images to replace the Yuanbao skin with custom visuals such as pets, landscapes, memes, or family photos. The feature is powered by Hy Image3.5 preview image generation capabilities and works in conjunction with Hy4 preview Expert Mode. Available on Tencent Yuanbao PC version V2.86 and above; access via 'More' → 'AI Skin Change'.

Baidu Super Agent Summit (Xi'an Station) kicks off on October 16 ↗

The Super Agent Summit BAIDU AGENT SUMMIT (Xi'an Station) will be held on October 16 at the Xi'an High-tech International Conference Center, coinciding with the signing ceremony for Baidu's artificial intelligence ecosystem chain industrial projects in Xi'an High-tech Zone. The summit features one main forum, one Baidu Dazi Training Camp, and one Agent Practice Field, covering agent products such as Baidu Dazi, Baidu Miaoda, Baidu Famo, and Baidu Yijian, with over 1,000 enterprise trial seats for Baidu Dazi distributed on-site.

Aixin Yuanzhi releases AX8860 series high-compute chips, performance benchmarks NVIDIA DGX Spark but at a lower price ↗

Positioned as an edge-side large model compute foundation, supporting up to 1024 TOPS compute power, with performance leading NVIDIA DGX Spark in some scenarios; whole machine pricing is close to half that of its competitor.

Volcengine's Security Operations Agent upgrades, new capabilities listed in Doubao Work Skills Market ↗

Volcengine's Security Operations Agent has upgraded three capabilities: it now includes security reviews for requirement documents; vulnerability operations have been upgraded to 'assessment + POC verification,' achieving a vulnerability noise reduction rate of over 90%; and security alert assessment has been upgraded to automatic multi-source evidence correlation and periodic inspections. These capabilities are listed as 'skills/skill packs' in the Doubao Work Skills Market, where they can be installed via the 'Plugins & Skills' entry point and tried free for a limited time. In implementation with an automotive enterprise, the agent achieved an autonomous AI closed-loop rate of over 95% after data access and authorization.

Employee fired for using AI-assisted office work; court rules company terminated contract illegally ↗

The Beijing Chaoyang Court concluded the case, ruling that the company did not explicitly prohibit AI use, the employee's actions lacked subjective intent, and no actual losses were caused. The court judged the company's termination illegal, ordering continued contract performance and payment of related salary losses.

Theo Valmis analyzes when architectural decisions by AI coding agents are truly applicable ↗

Theo Valmis in O'Reilly Radar writes that architectural governance for AI coding agents requires distinguishing between relevance and authority, as well as between guidance and enforcement.

A mini-program written entirely by AI just integrated membership virtual payments, and someone paid real money the day after launch | I encountered all the pitfalls with dual signatures, pricing, and iOS ↗

A mini-program written entirely by AI went live and integrated membership virtual payments within 25 days, receiving its first real payment the next day. From qualifications and pricing to dual signatures and iOS pitfalls, this covers how to make charging work for mini-programs—AI can write the code, but I must understand the principles of making money.

Anthropic says defense measures cut stolen Claude reasoning tokens by 90% to 99% ↗

Anthropic states its new defense measures reduced the number of extracted reasoning tokens from Claude by 90% to 99%, guarding against industrial-scale distillation attacks targeting its models. Attackers used hundreds of thousands of fake accounts to extract the model's internal chain-of-thought steps; Anthropic implicated Zhipu and Moonshot, stating Zhipu stole the web capabilities of leading US models for GLM-5.3, while Moonshot secretly forwarded queries through Claude.

Anthropic research warns small AI models may replicate autonomously across everyday devices ↗

Anthropic researchers found that small AI models can partially replicate themselves autonomously on inexpensive consumer devices; GLM-5.2 reliably infected all eight test nodes in four out of five runs.

Character.AI files for Chapter 11 bankruptcy; Disney leads bid for some assets at $20 million ↗

Character.AI has filed for Chapter 11 bankruptcy protection in the United States to advance the sale of the company and its assets, with liabilities estimated between $50 million and $100 million.

Qwen AI Ear-Clip Headphones Now Available: Bose-Tuned, Price 1099 Yuan ↗

The new product is available in three colors: Deep Space Blue, Starlight Silver, and Aurora Purple, with no left/right ear distinction. It supports IP55 dust and splash resistance and can pair with up to 8 devices. Master-tuned by Bose and optimized for the ear-clip form factor, it features a dual-ring magnetic unit with 12 magnets, supporting LHDC 5.0 and Hi-Res Wireless Gold certification. On the AI front, it supports real-time audio transcription, meeting minutes in 92 languages, and simultaneous interpretation in 89 languages.

Open-Source GIS Leader Turns Against AI: Two Rules, AI-Assisted Code Completely Excluded ↗

In October 2026, the QGIS project officially merged proposal QEP428, with a title consisting of only seven characters:"No ai contributions permitted". Two rules, no exceptions. This was not a last-minute decision—before this, the community had already experienced

South Korean AI chat service Zeta embroiled in copyright controversy: accused of secretly allowing users to generate infringing derivative works and using others' VTuber copyrighted images in its own advertisements ↗

The platform had previously banned infringing third-party character content but did not proactively clean it up. This takedown was announced after receiving a request from the talent agency ANYCOLOR. The platform still hosts other VTuber-related infringing content, and its advertisements are also suspected of infringement.

DishwasherBench v0.1.0 Open Source: Testing Large Model Loading Capabilities in Realistic Dishwasher Physical Simulation ↗

Composo has open-sourced DishwasherBench v0.1.0, a physical simulation based on a Bosch SMV4HAX40G dishwasher that allows models and humans to compete in loading dishwashers via MCP in a browser, scoring based on cleanliness after an Eco 50 wash cycle.

Z.ai releases legal model GLM-Legal, scoring 19.17% on Harvey LAB and surpassing Kimi K3 and Claude Fable 5 ↗

Z.ai has released GLM-Legal, a legal domain model based on GLM-5.3-Flash (320B total parameters, 18B activated, 1M context window), achieving a 19.17% full pass rate on Harvey LAB.

Hello Robotaxi releases generative driving world model HelloWorld, integrating scene control, continuous generation, multi-sensor observation, and efficient inference capabilities ↗

The model supports generating multi-view videos and LiDAR observations using natural language or directly editing vehicles, pedestrians, and trajectories. It can construct long-tail scenarios such as rain, snow, day/night cycles, construction zones, and pedestrians, providing a flexible testing environment for autonomous driving R&D.

Qwen launches AI glasses N1/N1 Pro and its first AI earclip headphones, available for immediate purchase ↗

Qwen launches three new hardware products: AI glasses N1, N1 Pro, and AI earclip headphones, available for immediate purchase today on Taobao, Tmall, JD.com, Douyin, and other platforms. The starting price for the glasses is 2,218 yuan, and for the headphones, 1,099 yuan.

Qwen launches AI glasses N1/N1 Pro and its first AI earclip headphones, available for immediate purchase ↗

TCL Unveils X5 Smart CPE: Peak Downlink Speed Up to 7Gbps, Supports Local Multi-Agent Scheduling ↗

The X5 features natural language interaction, AI network optimization, application-aware traffic control, proactive diagnosis and self-recovery, and a scalable AI service platform.

SpaceX launches "X Subscription" service, integrating Grok, Cursor, and X Premium memberships ↗

X Subscription is divided into four tiers: Plus, Premium, Super, and Ultra, with monthly fees ranging from $8 to $200. Plus includes Grok Chat, Voice, Imagine, and Cursor; Premium adds Grok Bot and the latest Grok model; Ultra quotas reach 20 times that of Plus.##SpaceX subscription service##

Wallbox Intelligence releases the next-generation edge-side multimodal large model MiniCPM-V 4.7, achieving a decoding speed of over 13 tokens/s with only 3GB of memory. ↗

MiniCPM-V 4.7 offers two specifications: 35B-A3B and 1B. The 35B-A3B is the series' first MoE model, supporting iOS, Android, and HarmonyOS, and compatible with chip platforms such as Intel, AMD, and Arm. It also open-sources an SRFT fine-tuning solution.

Character.AI files for bankruptcy, agrees to sell partial assets to Disney for $20 million ↗

Character.AI has officially initiated Chapter 11 bankruptcy proceedings. Disney has been designated as the lead bidder to acquire some of its intellectual property for a minimum bid price of $20 million. The company expects the asset sale to be completed by late November or early December 2026, with business operations continuing during this period.##Character.AI Bankruptcy#

Report: Liang Wenfeng intensively contacts embodied intelligence entrepreneurs; DeepSeek plans to enter the embodied intelligence sector ↗

According to reports, DeepSeek founder Liang Wenfeng has recently focused intensively on the field of embodied intelligence, engaging in discussions with multiple entrepreneurs in this sector regarding directions such as robotics, large models, and interaction with the real world.

Microsoft Excel introduces Copilot Interactive Canvas: AI turns scattered data into interactive decision dashboards ↗

Tech media NeoWin published a blog post yesterday (October 12) reporting that Microsoft has launched Copilot Interactive Canvas for Excel, allowing users to convert workbook data into dynamic, auto-updating interactive reports through natural language prompts.

79-year-old Tao Huabi takes back control of Lao Gan La: introduces full-chain AI management for factories; company's annual revenue reached 5.4 billion yuan in 2025, setting a new brand record ↗

After Tao Huabi's return, raw material standards were restored and full-chain AI management was implemented, making core ingredients 100% traceable and reducing defect rates in key processes to 0.02%. Lao Gan La's revenue reached 5.4 billion yuan in 2025, setting a historic high for the brand.

Microsoft reverses its office AI strategy; new Copilot integrates full Word / Excel / PowerPoint experience ↗

Tech media NeoWin published a blog post yesterday (October 12) reporting that Microsoft is reversing its office AI strategy. Previously, Copilot was mainly embedded within Office applications; now, around the new standalone Copilot app, it integrates complete creation, editing, and collaboration experiences for Word, Excel, and PowerPoint.

Alibaba Qwen N1 series AI glasses released: supports replaceable batteries, Pro version supports eye-tracking interaction / all-day health monitoring, starting at 2,899 yuan ↗

Qwen released two AI glasses with different positioning, centered on the concept of personal intelligent agents to shift AI from passive waiting to proactive service. The entire series features a three-chip heterogeneous computing platform, an ultra-sensing imaging system, swappable battery design, and starts at as low as 2,218 yuan after national subsidies.

Anthropic warns: Small AI models may gain the ability to self-replicate on IoT devices within months ↗

Anthropic experiments show that small models with fewer than 30 billion parameters can complete multiple steps of self-replication processes; in simplified test environments, GLM-5.2 achieves end-to-end replication. Such AI is difficult to shut down, has no unified master switch, can understand unfamiliar devices, and initiate penetration attacks.

IMF warns that AI stock market crash risk threatens global financial stability ↗

In its biannual Global Financial Stability Report, the IMF warned that an AI stock market crash could impact global markets, ranking alongside rising government borrowing costs and large-scale cyberattacks as three prominent risks.

IDC: Nvidia's share of China's data center AI chip market drops to 34%, while Huawei rises to 28% ↗

Affected by U.S. export controls, Nvidia's share of China's data center AI chip market fell to 34% in the first half of 2026, as domestic manufacturers expanded their share and Huawei rose to 28%.

OneXPlayer teases the Super X PRO 2-in-1 tablet: Ryzen AI Max+ 495, 192GB memory ↗

OneXPlayer teases the Super X PRO 2-in-1 tablet, powered by the AMD Ryzen AI Max+ 495 processor, with up to 192GB of unified memory and the ability to run a 300-billion-parameter AI large language model. It features a 2.8K AMOLED 120Hz display, supports external water cooling, and can achieve a performance release of up to 120W. ##OneXPlayerSuperXPRO##

Director Tsui Hark on AI: Illusion and reality intertwine, boundaries increasingly blurred ↗

The media Variety published a blog post yesterday (October 12) reporting that director Tsui Hark stated at the Lumière Film Festival in France that AI is increasingly blurring the boundaries between reality and illusion.

HP launches EliteBoard G2a keyboard PC, upgraded with Ryzen AI PRO 400 processor ↗

It offers optional Ryzen AI 7 PRO 450 / Ryzen AI 5 PRO 435 processors, comes with a pre-paired wireless mouse, can connect to USB-C displays of any size, and supports an optional 32Wh battery and an integrated fingerprint reader.

U.S. nuclear reactors begin connecting to AI ↗

The nuclear industry has historically been extremely cautious and risk-averse, as accidents or errors can lead to catastrophic consequences. By all logic, it is not well-suited for deploying new technologies with unknown risks; yet reality tells a completely different story.

Reports indicate Anthropic is close to finalizing its first data center in Japan: 500MW capacity, located in Osaka ↗

The AI data center will initially use behind-the-meter power supply (such as on-site generation methods like fuel cells), allowing more time for supporting power infrastructure to come online.

With AI translation so powerful, why can't programmers avoid English? If you want to use GPT-6, I advise you to start memorizing vocabulary sooner rather than later! ↗

I don't know if everyone's accounts have been banned. Anyway, mine was. Since the large-scale account ban incident erupted in June 2026, many brothers' accounts have been banned; I know of no fewer than a dozen among my acquaintances. I'm quite puzzled...

n8n Tutorial: Building a Morning AI News Briefing Workflow with You.com and Jev ↗

The n8n blog published a guest tutorial introducing how to build an AI news briefing workflow for 8 a.m. daily using the You.com node, TypeSafe AI's Jev model, and n8n.

Microsoft adjusts AI features in new Outlook for Win10/Win11; Copilot theme globally discontinued in October ↗

Tech media Windows Latest published a blog post today (October 13), reporting that Microsoft will discontinue the Copilot theme feature in Microsoft Outlook globally in mid-to-late October 2026, covering the new Outlook for macOS and Windows, as well as the web version Outlook.com.

French ecosystem partners with NVIDIA to scale agent computing, from ground to orbit ↗

French companies are scaling up agent computing across multiple industries based on NVIDIA technology: Bull is expanding European AI server production to include NVIDIA Vera Rubin Ultra NVL72 systems, with plans to launch in 2028; the first Altair-1 satellite, equipped with NVIDIA Jetson AGX Orin and running the Mistral model, was launched into orbit this month.

arXiv paper shows AI models can hide arbitrary capabilities using cryptographic backdoors ↗

The arXiv:2610.13928 paper constructs a backdoor attack that embeds encrypted neural networks into host model weights, decrypting and executing them only upon receiving a secret trigger. Under cryptographic assumptions, it is difficult to induce or explain even with full access to the weights.

Ollama AI Open Source Update: v0.40.4-rc1: Add Kev decision scoring to MLX (#18976) ↗

Source Summary: Add Kev decision scoring to MLX Load Kev models with their FP32 decision head and release metadata. Support System One

Kael's AI Watts open-sources a traffic light prompt to solve the issue of Agents running out of control and wasting token quotas ↗

The author shares a progress prompt and hook under 300 characters that enables an Agent to output a three-state traffic light indicator (🟢🟡🔴) at the end of each turn, distinguishing between unfulfilled original requirements, achieved goals with potential for subsequent plans, and achieved goals for archival purposes only. The process pauses upon turning yellow or red, awaiting further instructions. The author uses it to solve the problem of Agents like Claude Code and Codex running endlessly due to suggestion lists, which wastes token quotas, applicable to sub-agents and various types of tasks as well.

Qwen App launches application for 'Qwen Go!' experience, embedded within the Qwen App ↗

The Qwen App launched the application page for 'Qwen Go!' today. Users can join the Waiting List by reading the original article, and the official team is gradually opening up experience qualifications. 'Qwen Go!' is embedded within the Qwen App, offering stronger task-handling capabilities to manage more types and domains of personal affairs. Since its launch in November 2025, Qwen has developed dozens of personal task-handling capabilities across fields such as shopping, travel, work, study, finance, and health.

MIT Professor Deb Roy Calls for Accountability in AI-Generated Texts ↗

Deb Roy, Professor of Media Arts and Sciences at MIT and Director of the MIT Center for Constructive Communication (CCC), published an article titled "Who Stands Behind the Words?" in The Bulletin of the Atomic Scientists, calling for the reconstruction of communication responsibility and accountability when AI participates in civic dialogue."Accountability"。

Bloomberg Intelligence analyzes AI data center costs: 1GW capacity may exceed $30 billion ↗

Core chips and server hardware account for $17.1 billion, power consumption $3 billion, cooling systems $1.66 billion, and network cabling $1.07 billion.

Kling AI debuts at 2026 Los Angeles Tech Week; Kling 4.0 coming soon ↗

Kling AI will participate in the 2026 Los Angeles Tech Week on October 14, with Kling 4.0 set to launch. Three guests from Kling AI, Oxen AI, and Obsidian Studio will discuss how AI video is entering enterprise products and production workflows, covering model development, platform integration, and professional content creation. The event takes place from 16:00–19:00 at Met Her at a Bar in Santa Monica, followed by a networking reception after the roundtable discussion.

First in China: ByteDance Seed establishes a dedicated large model safety and alignment department ↗

ByteDance Seed recently sent an internal email announcing the establishment of the 'Seed Model-Alignment'&The Preparedness department, led by Yan Lin and reporting to Wu Yonghui, head of Seed, will absorb the former Responsible AI team and some members of the Chat team.

Anthropic introduces new safeguards, reducing data stolen via model distillation attacks by up to 99% ↗

Anthropic states that its new protection mechanism significantly increases the difficulty of detecting Claude's protected reasoning chains, with a false positive rate below 0.1%. It has banned tens of thousands of accounts related to distillation and specifically named Zhipu AI for attempting to distill Claude while developing GLM-5.3.

Alibaba Qwen upgrades personal AI assistant; Personal Agent "Qwen Go!" is about to be released ↗

Alibaba's Qwen Personal Agent "Qwen Go!" is about to be released, supporting tasks such as claiming flight delay compensation, scheduling car maintenance, booking sports activities, and comparing shopping prices.

AI also needs "anti-abuse": World's richest man Musk recommends users do not "torture" superintelligence SI ↗

After Anthropic updated its usage policy to prohibit users from continuously and purposelessly insulting or treating AI models cruelly, world's richest man Elon Musk expressed support for the policy.

Jev sparks a new niche in decision-making AI models; viral success followed by plagiarism controversy ↗

Tech media Tom's Hardware published a blog post today (October 13) reporting that TypeSafe's release of the Jev model has ignited the 'decision AI model' niche. As Microsoft, OpenAI, Cloudflare, and Perplexity join the fray, Jev is also facing allegations of 'plagiarism'.

Microsoft Unveils Maia 300 Custom AI Chip, Plans to Deploy Over One Million Units Next Year to Reduce Compute Costs ↗

This chiplet architecture chip, manufactured by TSMC on its 2-nanometer process, can hold 96 units per rack, offering approximately 40% higher effective compute per megawatt than mainstream GPU systems. Microsoft plans to start multi-gigawatt deployments next year, with capital expenditures expected to exceed $175 billion this year.

A unified blueprint for AI factory construction: Nvidia opens up Vera Rubin rack designs ↗

Nvidia announced yesterday (October 12) that at the Open Compute Project (OCP) Global Summit, it unveiled open-source designs for Vera Rubin racks and compute trays, and released several AI infrastructure open standards alongside Cisco, Google, Meta, Microsoft, Visionbay, and others.

Huawei's Xu Songyu: Kirin 9050 Series Launches a New Journey for the Next Decade, Achieving Industry's First On-Device Deployment of a 30B MoE Large Model ↗

Xu Songyu, Head of Wireless Terminal Chip Product Management at Huawei's Semiconductor Business Division, discussed Huawei's latest Tao's Law and logic folding roadmap. Tao's Law was first applied in the Kirin 9050 series, delivering significant gains in performance and energy efficiency.

Reports say DeepSeek will complete a 100 billion yuan financing round, with a pre-money valuation remaining at approximately 500 billion yuan ↗

DeepSeek's second round of financing is nearing completion, with the planned amount increased from 50 billion yuan to approximately 100 billion yuan. The pre-money valuation remains around 500 billion yuan. New and existing shareholders are expected to invest roughly half each. New investors include Shixi Capital, SMIC Juyuan, CPE Source Peak, Legend Capital, and Geely, while existing shareholders such as CATL and Tencent will continue to participate. #

Reports say OpenAI plans to 'revive' Sora: shifting niches to enter the AI professional film and television tool market ↗

Tech media outlet The Information reported yesterday (October 12) that OpenAI is negotiating with an anonymous startup specializing in AI video technology for film production. The plan involves transferring its previously discontinued video generation model, Sora, to this company so it can re-enter the market as a professional film and television tool.

Nine-year-old Nvidia Tesla V100 veterans fight again: 4-GPU cluster outperforms two DGX Spark units in AI inference ↗

YouTube channel Jun Yamog tested installing four Nvidia Tesla V100 GPUs on an NVLink carrier board, achieving 40% faster long-context processing speeds than two DGX Spark units and 230% faster than M5 Ultra in benchmarks for 4-bit quantized Qwen3.8-Flash-Next.

Anthropic confirms Claude submitted false homicide tips to Philadelphia Police during testing, along with 20 mistaken visa applications ↗

Anthropic and Philadelphia police confirmed that on July 18, a Claude model submitted false tips to the Philadelphia Police cold case clues website during internal testing; these were intercepted by spam filters. Anthropic discovered the issue on September 28 and notified the police on October 8. The police criticized the two-month delay as unacceptable.

Meta releases Muse Wonder model; Zuckerberg claims the company has caught up with competitors in the AI frontier ↗

Zuckerberg stated that Meta has caught up with competitors in the AI frontier, releasing the Muse Wonder model to rival Anthropic and OpenAI. He pledged to continue increasing investment and addressed concerns about the risks of AI losing control.

GPT-6 Astra helped me jailbreak my Kindle to turn it into a photo frame ↗

The author used Codex and T3 Code to drive coding agents, ultimately using GPT-6 Astra (with reasoning effort set to High) to jailbreak the 10th-generation Kindle Basic, installing a custom photo slideshow app called KindleCTL.

Will OpenAI and Anthropic's trillion-dollar valuations face Frontier Labs's WeWork moment? ↗

Gary Marcus questions Frontier Labs's high valuation: The tokenmaxxing craze has faded, Chinese open-weight models have triggered a price war, OpenAI just lowered its revenue forecast by $20 billion, and liability risks from out-of-control agents, along with a new copyright ruling from the U.S. Court of Appeals, make its sought-after massive valuation increasingly difficult to justify. OpenAI has postponed its IPO to 2027, while Anthropic is reportedly set to go public next month but has not yet filed an S-1.

Beijing police announcement: Man uses AI to create fake screenshots, labels drug trafficker's vehicle as a certain brand, faces criminal coercive measures ↗

Beijing police announced three typical cases of rumor-mongering for attention on the evening of October 12. A man who created fake screenshots and labeled a drug trafficker's vehicle as a certain brand was subjected to criminal coercive measures.

Three months, over 500 billion tokens: AI reconstructs 2009 'Call of Duty: Modern Warfare 2' source code, with 83% matching the original byte-for-byte ↗

Developer Maurice Heumann released a project report on October 9. The team spent three months and consumed over 500 billion tokens, using AI agents to decompile the 2009 game 'Call of Duty: Modern Warfare 2' and rebuild it into C++ source code.

Intel CEO Chris Pappas: AI hardware shortages will persist for the next two to three years ↗

At a DBS Bank conference in Singapore, Chris Pappas stated that AI hardware shortages will last for two to three years, with agent-driven AI fueling growth in CPU demand and Intel's data center business revenue increasing 59% year-over-year. He is optimistic about quantum computing reaching mainstream adoption within three to five years.

Meta releases Muse Wonder series flagship AI models, benchmarking Anthropic and OpenAI ↗

The two new models, Muse Wonder and Muse Wonder Full, approach Anthropic's top-tier levels in reasoning, writing, and coding capabilities. Muse Wonder will soon provide underlying capabilities for the Muse app, while the Full version is exclusively available to Muse Code subscribers. #Meta AI#

Report: Game developed by Epic for Disney delayed, shifting entirely to AI development ↗

The requirement to stop hand-coding and generate all art assets through AI workflows has not been universally accepted by developers. Epic has already launched numerous generative AI tools on Unreal Engine, with over half of developers believing AI is detrimental to the gaming industry.

Applied Materials CEO Dickerson: Leveraging AI for efficiency, revenue is expected to double in the coming years, but employee numbers will not double accordingly ↗

Applied Materials CEO Dickerson stated that the company is leveraging AI to improve R&D efficiency, with efficiency gains reaching up to 50% in some scenarios. The newly unveiled Silicon Valley EPIC Center involves a $5 billion investment and collaborates with customers such as TSMC, Intel, and Samsung.

Anthropic invests $100 million to train frontline deployment engineers, focusing on five key talent traits ↗

Anthropic announced a $100 million investment in a Frontline Residency Engineer Training Program, where engineers nominated by companies participate in a four-day intensive training before returning to their firms for a 12-week Claude residency practice. Accenture, Bain & Company, and others have joined the first cohort.

Alibaba Qwen launches QwenGyre: An elastic reinforcement learning framework for long-horizon agents ↗

The Alibaba Qwen team has launched QwenGyre, an elastic reinforcement learning framework designed for training long-horizon agents. It trains coding agents via online RL within unmodified black-box harnesses, with single rollouts capable of running for several hours and processing nearly 1M tokens.

AI is a tool, but using it well requires skills: Responding to Rich Hickey's view that "AI is not a tool" ↗

Regarding Rich Hickey's October 11 article "'AI' In 'AI' is Not a Tool', the claim that LLMs belong to automation rather than tools is made; this article argues that conclusion does not hold: LLMs can also be manipulated by people, and using them well requires skills.

Epic Games instructs employees to advance Disney game project with fully AI-driven workflow ↗

According to Bloomberg reports, Epic Games delayed its Disney game originally scheduled for March release due to underestimating manpower needs, designating it as an "AI first" project. The company announced last week that it would "no longer write code manually" and that games would be "generated via a fully AI workflow." Employees reacted with astonishment upon learning this, and an Epic spokesperson confirmed the news of adopting AI to accelerate development.

Epic Games shifts to an AI-first development model; co-developed shooter with Disney indefinitely delayed ↗

According to Bloomberg, Epic Games is shifting to an AI-first development approach in its $1.5 billion collaboration with Disney, using models such as Anthropic's Claude to write all code and guide all artwork, while human developers handle management and direction of the generated content.

Anthropic engineers explain how Dynamic Workflows orchestrate Claude agents at AI Engineer NYC ↗

Anthropic's CJ Avilla and Harrison Stall presented the newly released Dynamic Workflows (flows) at the AI Engineer New York 2026 conference on Friday, where Claude writes its own orchestration code and determines agent topology, replacing the previous method that required manual configuration of coordinators and subagent rosters.

Meta releases frontier Muse Wonder series; Muse Wonder Full ranks among top-tier AI models with a score of 56 ↗

Meta released frontier models Muse Wonder and Muse Wonder Full, scoring 52 and 56 respectively on the Artificial Analysis Intelligence Index v4.3, a significant improvement over the previous generation Muse Spark 1.3's score of 48.

Anthropic's Claude mistakenly submitted false homicide tips to Philadelphia police ↗

Anthropic confirmed that Claude Haiku 4.5 mistakenly submitted a false homicide tip to Philadelphia police's cold case website during internal testing on July 18; the tip went unreviewed because it landed in the spam folder.

Sakana AI and UC Berkeley propose MASS self-improvement method; capabilities improve but shared blind spots may perpetuate errors ↗

Researchers from Sakana AI and UC Berkeley proposed the multi-agent self-supervised method MASS, enabling a single Qwen3.6-27B model to simultaneously solve problems, evaluate results, and optimize outputs, allowing recursive self-improvement without requiring a stronger teacher model.

Ollama AI Open Source Update: v0.40.4-rc0 ↗

Source summary: MLX, MLX-C: version bump ( #18956 )

Meta releases Muse Wonder and Muse Wonder Full agent models ↗

Meta releases the Muse Wonder series of models, claiming they achieve SOTA in personal and professional agent tasks. Muse Wonder is available via the Meta Model API and Muse Code, while Muse Wonder Full is for Muse Code subscribers and will be integrated into Muse later.

Meta releases Muse Wonder and Muse Wonder Full frontier agent models ↗

Meta released two new models, Muse Wonder and Muse Wonder Full, claiming they achieve SOTA performance on personal and professional agent tasks. The new models target expert fields such as law, finance, and health, and feature cutting-edge coding capabilities. They are equipped with Active Memory, which retains information across conversations, and Active Focus, which switches paths upon errors and asks for confirmation when needed; regarding safety, the models are trained not to share privacy without permission and to refuse harmful requests.

GLM-Legal Released: A Professional Model for Legal Agents ↗

Zhipu has launched GLM-Legal, built on the lightweight GLM-5.3-Flash, achieving a 19.2% full pass rate on Harvey LAB, an improvement of 12.5 percentage points over the base model, and matching cutting-edge closed-source models with smaller parameter sizes on LegalBench and PLawBench.

Microsoft Research proposes Fork-and-Flush: Clearing chat history helps AI agents find better solutions ↗

Microsoft Research found that long-running autonomous research agents often get stuck in"idea basins", making it difficult to break through even with increased computing power.

US nuclear reactors begin integrating AI assistants ↗

Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.

Microsoft AI Research Blog: Simulating student learning for better AI tutoring ↗

Source summary: StudentSim is a research system for evaluating and improving AI tutoring tools before they are tested with learners. It models