01News
Anthropic has updated its coordinated vulnerability disclosure (CVD) policy, reducing the disclosure window for high-risk and critical vulnerabilities from 90 to 30 days, and for other issues from 90 to 45 days. These changes apply to vulnerabilities reported to suppliers on or after January 1, 2027.
02News
Sakana AI has built the Japanese financial benchmark JFinBench using a dataset of 321 questions from six Japanese financial qualification exams (including Certified Public Accountants and Tax Advisors) for the 2026 version. The team made the dataset, code, and leaderboard public. Built with mostly coding agents and full human verification via a human-in-the-loop pipeline, it allows for low-cost updates in the future.
03News
Google has launched an AI travel planning feature in Gemini-powered Ask Maps on Google Maps. The tool generates daily or multi-day itineraries with routes and numbered markers, supporting sharing and export to desktop Google Docs. It is available in over 200 countries.'know before you go' Travel tips and booking search features launch first in the US, while Ask Maps expands to over 30 countries across Europe.
04News
Dreamteam, founded by a former Freshworks executive, has raised $53 million in seed and Series A rounds led by Together Fund and Accel, with participation from Premji Invest. The company aims to challenge Salesforce and HubSpot using an AI-native CRM.
05News
The King'The King's Trust and OpenAI have launched a £1 million partnership to help up to 25,000 young people in the UK develop AI skills over the next 18 months, with priority given to NEETs (young people not in education, employment, or training). OpenAI has donated £1 million for this initiative, which is part of a nationally recognized AI qualification that The King's Trust will roll out across the country in the coming months.'The King's Trust and OpenAI have launched a £1 million partnership to help up to 25,000 young people in the UK develop AI skills over the next 18 months, with priority given to NEETs (young people not in education, employment, or training). OpenAI has donated £1 million for this initiative, which is part of a nationally recognized AI qualification that The King's Trust will roll out across the country in the coming months.
06News
According to Artificial Analysis, since the first GTA VI trailer was released in December 2023, its Progress Index has risen from a high of 160 to 358, more than doubling.
07News
Why does the prompt 'You are the owner and long-term maintainer of the codebase' make AI more proactive about merging duplicate code? This article analyzes how this Claude programming prompt affects code refactoring, along with the increased token costs and functional change risks that follow.
08News
I wrote this post to share my experience using AI for assisted development. After several iterations, I believe this AI development workflow is quite complete. If you have different opinions after reading it, feel free to leave a comment.
09News
The author proposes a human-AI division of labor in development: let AI tools like Plaud and Granola record and analyze customer conversations to draft specifications, while humans write proof-backed guardrails, maintain machine-verifiable ontologies, and construct review harnesses including deterministic CLI interfaces, code pre-gates, secondary regeneration comparisons, and weekly drift audits.
10News
The speaker includes a DIY digital display with 16.8 million RGB colors, two 52mm full-range drivers, and integrates the Doubao large model for voice-based song selection and offline commands, allowing tuning for multiple scenarios.
11News
Tongyi Qianwen (Qwen) officially announced that Qwen4-Max is available from day one on Novita AI. The cited content describes it as a MoE model designed for complex coding, professional tools, and long-term agent workflows, featuring native visual capabilities, a 1M token context window, and enhanced tool usage and long-range task execution.
12News
MarkTechPost reports that Meta has advanced its LLM-generated test system, JiTTests, from a research prototype to production infrastructure. The system generates executable tests designed to catch failures in risky code changes. In the first half of 2026, it identified 74 privacy severity issues during the diff phase, reducing the incidence rate by 20%.
13News
The device features a laser light source and a short-focus lens, supports 4K resolution and a MEMC refresh rate of 240Hz, includes a dedicated golf mode, and is priced at 52,999 yuan.
14News
hidden
15News
The US has signed a new agreement with partners to compile lists of key AI technologies and supply chains requiring protection against Chinese competition. Initial signatories include Japan, Australia, Israel, and the UK. Reports indicate that South Korea, India, Qatar, and the UAE have also signed the San Francisco Declaration, while Germany, Sweden, the Netherlands, and the European Commission are expected to join next week. The deal was signed just before Trump's meeting with Xi Jinping in Shenzhen for the APEC summit.
16News
Under the new agreement, LVMH will deploy OpenAI tools and ChatGPT Enterprise across all its magazines for tasks such as document retrieval, comparison, and preliminary translation.
17News
Alibaba's Qwen team launches Qwen4 series' first model, Qwen4-Max: a proprietary sparse MoE architecture with 2.4T backbone parameters, 71B activations per token, plus 328B N-gram embedding parameters. It supports text, image, and video inputs with a 1M-token context window and up to 393K output tokens.
18News
AI News: Jack Sonny Huang may return to NVIDIA; Huawei launches Embodied World Model; ChatGPT Wallet enables payments
Model & Technology
1. Huawei Cloud releases CloudRobo WM1.2, its first Ascend-native embodied world model
19News
IMF President Georgieva warns that the global economy is heavily reliant on AI investment. The scale of circular financing in the AI sector has already exceeded $1 trillion. A sudden drop in confidence could trigger risks to global financial stability.
20News
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'.
21News
Qwen has released its next-generation flagship model Qwen4-Max, featuring a new MoE architecture with 2.4 trillion backbone parameters and native visual understanding capabilities, supporting a context window of up to 1 million tokens. The model is now integrated into the Qwen platform, accessible via Token Plan or API for developers. Officially, it aims to enhance performance on complex tasks while reducing token consumption and overall operational costs.
22News
Hark has released the second-generation computer operation agent, Handoff², within Hark Pro. Average end-to-end speed is approximately double that of version 1, with flight bookings seeing a maximum acceleration of 3.8 times and long research tasks exceeding five times.
23News
🙌 Build with Qwen4-Max on Vercel's AI Gateway! @vercel_dev
24News
Built on a Windows mini PC base, this device integrates a gimbal camera and array microphones to run various composite AI workloads locally.
25News
Nubia officially announced today that the NaviX Ultra Sky Collection Edition is available everywhere. The device continues its classic sky design and features micro-nano relief craftsmanship. IT-Home reports the price is 6,999 yuan, with only a single 16GB+512GB configuration offered.
26News
Browser Use now allows agents to complete transactions via Evervault, ensuring no genuine card details are handled. Evervault generates tokens that pass Luhn validation and match the length and BIN of real cards, substituting them only for requests sent to whitelisted servers. Actual card numbers never reach models, agents, or browsers.
27News
Yang Yuanqing believes the competition in the AI industry will shift from comparing model parameters to demonstrating the deployment capabilities of hybrid AI. 'The next battle for AI won't depend solely on models; it will rely on data and hybrid AI,' he emphasized. He stressed that companies must recognize that data accumulated across various applications belongs to the users themselves, not to software applications or SaaS providers.
28News
Google's Gemini 4 Carbon takes the top spot on the general leaderboard at its debut, while domestic models occupy ranks 6, 7, and 8 on the video leaderboard. Kimi K3.1 Pro enters the front-end leaderboard at rank 12.
29News
Recently, there have been many popular articles on Juejin about using AI to create 3D dashboards and using Three.js for digital twin parks. Seeing the screenshots others took made them look incredibly smooth and sci-fi. I was fired up after reading them. I just got a project for an internal logistics and warehouse data dashboard, and I thought: now
30News
Elon Musk confirmed the changes described: text watermarks in the xAI API are now disabled by default and only added when legally required or explicitly enabled by users.
31News
Tencent Hunyuan launches Hy4 flash, a model positioned as the most affordable option for daily productivity. It claims performance comparable to Claude Haiku 5.5 (high) in human blind tests of daily tasks while reducing task costs by 18 times. API pricing is $0.003 per million cached tokens, $0.070 for input, and $0.281 for output. Link: https://hy.tencent.ai/research/hy4-flash
32News
In a recent interview, Hinton described Jensen Huang as an "upstanding person" and admitted he trusts him more than Snoop, Mark Zuckerberg, or Elon Musk. He also noted that Huang made one major mistake but believes NVIDIA can weather it.
33News
Amid an AI investment boom, concerns about circular financing are growing. TSMC has tightened its customer payment policies first, assessing credit limits according to financial status. Orders exceeding these limits require cash payment, and some clients must prepay for capacity retention. ASIC manufacturers are also rushing to raise funds to cope with the industry's changing landscape.
34News
OpenBMB has announced that MiniCPM-OCR natively supports the vLLM and SGLang open-source inference frameworks, making it easier for developers to integrate MiniCPM-OCR into existing workflows. The author emphasizes its strong OCR capabilities, flexible deployment options, and a continuously expanding open-source ecosystem.
35News
This study evaluates how users interact with health AI. Participants can ask daily health questions and rate responses; data will not be used to train AI models. The research cycle exceeds 12 months.
36News
Guangzhou My Film Culture Media Co., Ltd. was fined over 68,200 yuan by the Guangzhou Panyu District Market Supervision Administration for violating regulations on medical advertising.
37News
Wallmappe's OpenBMB announces that MiniCPM-OCR 1.0 now has native support for vLLM and SGLang. The code is merged and ready for deployment, allowing integration into two leading large language model inference frameworks. The author states this enables simpler integration and more flexible deployment while keeping OCR capabilities unchanged.
38News
This new feature creates a personalized AI workspace for users, pre-configuring professional skills and tools. It also supports building custom workbenches, effectively lowering the barrier to AI use and boosting productivity.
39News
Many developers no longer open Android Studio, yet still find their computer running out of RAM? The Dart Analyzer is likely the culprit. Recent updates to the Dart Analyzer have made significant contributions to this issue.
40News
Alibaba unveiled its new base large model, Qwen4-Max, on November 16 with 2.4 trillion parameters and native visual capabilities supporting millions of tokens in context. The company states that based on a new architecture and training optimization, the model continues to evolve in key areas such as agents and programming, enabling it to move from information retrieval to professional delivery in fields like finance and law. During its launch period, Tongyi APP members can enjoy limited-time access; using Qwen4-Max does not consume membership quotas, with activity details subject to the product page.
41News
Wang Guanhua, the spokesperson for the National Bureau of Statistics, stated that AI is accelerating its transition from technological innovation to industrial implementation, continuously empowering industrial upgrades. In October, prices in sectors such as electrical machinery and computer communication rose 0.1% month-on-month, up 6.2% and 5.9% year-on-year respectively, collectively contributing about 1.18 percentage points to the year-on-year increase in the Producer Price Index (PPI).
42News
China and ASEAN have declared 2027 the year of artificial intelligence cooperation. The China-ASEAN Free Trade Area 3.0 protocol will take effect on December 1, 2026.
43News
Qwen4-Max is now live on @OpenRouter—use it to start developing.🙌
44News
The author explains why recently popular AI-generated pure HTML 3D games can be so small: in a 1.4MB game demo, about 700KB is the bundled Three.js library; the rest relies on procedural generation. Textures, sound effects, and all other assets are synthesized by code in real time without any image, model, or audio files. The browser renders these live via WebGL/WebGPU calls to the graphics card.
45News
Qwen has launched Qwen4 Max. According to Artificial Analysis data, its Intelligence Index score is 43, which is 9 points higher than Qwen3.8 Max (xhigh) and 3 points lower than Kimi K3.1 Pro. The cost per task is $3.27, roughly one-third of Qwen3.8 Max's price, with a 39% reduction in output token usage. It features a context window of 1M tokens and is priced at $1.50/$4.50 per 1M input/output tokens. The model is now available via Alibaba's first-party API.
46News
Qwen shares ValsAI rankings: Qwen 4 Max ranks 11th on the Vals Index with 63.20% accuracy, $9.33 per test, and a latency of 1 hour, 4 minutes, and 39 seconds, making it the highest-ranked Chinese model on the list. Its capabilities show a clear overall improvement compared to Qwen 3.8 Max (dated 0902), with particularly notable gains in coding and financial tasks.
47News
48News
Due to tight computing power, DeepSeek has imposed limits on its App and web versions for long conversations and daily usage quotas, capping single inputs at 40,000 characters. This move may be a response to the pressure of surging call volumes.
49News
Qwen4-Max's preliminary score of 1,674 places it ahead of seventh-ranked Claude Opus 5 (Max) at 1,664 and eighth-ranked GPT-6 Sol (Max) at 1,657.
50News
Huawei officially revealed more details on its first 'logic folding' technology τ chip — the Kirin 9050 series — today. New intelligence has been released regarding its smart audio and video processing capabilities. The chip features a new-generation audio-video processor, delivering a 30% boost in audio performance. It supports next-generation audio DSP, H.266 hardware video decoding, and low-latency AI heterogeneous collaborative computing.
51News
Arena has announced that Qwen4-Max ranks sixth in Code Arena: WebDev with an initial score of 1750. This is 74 points higher than Qwen3.8-Max (0902)'s score of 1676. Qwen4-Max also achieved the highest scores among Chinese models in HTML (1748), Brand & Marketing (1705), and Reference-Based Design (1770). & Cited content indicates Qwen4-Max debuted at rank 14 overall in the Agent Arena, based on conversations with 14.7k real AI agents. This represents a significant improvement over Qwen3.8-Max's rank of 39. It is cited as China's top-ranked model on the list, making Alibaba the fifth-ranked lab globally.
52News
The Arena leaderboard shows Qwen4-Max at rank 14 overall in Agent Arena with a net improvement of +7.5%, making it the top Chinese model. This positions Alibaba as the leading Chinese lab overall and fifth globally, trailing only Anthropic, Google, OpenAI, and Meta.
53News
Arena announced that Qwen4-Max has entered the Arena and released a test video by Peter Gostev. The video compares outputs from Qwen4-Max against Opus, GPT, Kimi, and Fable in areas including 3D scenes, games, creative prompts, website design, and interactive explanations. Prompt details are available via links in the post.
54News
In the Agent Arena, Qwen4-Max surpassed Qwen3.8-Max with a net improvement of +7.5%, moving from rank #39 to #14 overall. As China's top-ranked model on the list, it helped Alibaba become the leading Chinese lab and the fifth-best lab globally, trailing only Anthropic, Google, OpenAI, and Meta. This assessment is based on 14.7k real AI agent conversations across four signal categories showing comprehensive improvements, including a +13% increase in Praise vs. Complaint and a +9% rise in Confirmed Success, with the largest gains in user satisfaction and controllability. By category, it ranked #12 in Work, #14 in Chat, and #16 in Code, all among the top Chinese models.
55News
Artificial Analysis reviewed Alibaba's newly released Qwen4 Max. It scored 43 points in the Intelligence Index using the xhigh reasoning tier.
56News
The SSD's AI Cache can store AI model weights, context, and local inference intermediate data, offering more flexible capacity for running local models.
57News
Qwen's Qwen4 Max is now live on OpenRouter as the first model in the Qwen4 series, designed for coding, office work, and long-horizon agent tasks.
58News
Qwen releases Qwen4-Max based on a new MoE architecture with native visual capabilities and a 1M token context window. Official internal benchmarks show approximately 50% fewer tokens compared to Qwen3.8-Max. Pricing is set at $1.5/M for input and $4.5/M for output tokens. The API is now live on QwenCloud. Unsloth has reposted the news and indicated that smaller open-source Qwen4 models will launch soon, with Qwen4-27B expected to run on devices with 16GB of memory.
59News
OpenRouter has announced pricing for Qwen4 Max at $1.50 per million input tokens and $4.50 per million output tokens, a reduction from the $2/$6 rates of Qwen3.8 Max. Cache hits are priced at $0.075 per million, with a context window of 1M. The model supports text, image, and video inputs, along with tool calling and structured outputs. See https://openrouter.ai/qwen/qwen4-max for details.
60News
Tencent Culture published live notes from the General Office LinkTime podcast, calling for core businesses to adopt an attitude of "starting from zero" and fully embrace AI. The focus is on building human-centric AI. It hints that the AI pyramid is already taking shape and urges breaking down organizational silos to accelerate transformation.
61News
Alibaba's Tongyi Qianwen unveils its latest flagship model, Qwen4-Max. Built on a new MoE architecture with native visual understanding and a 1 million-token context window, it reduces task costs by over 60% compared to the previous generation. The API is now open for access, with plans to open-source smaller models later.
62News
Apple has partnered with clinical AI platform OpenEvidence to launch the Apple Health AI Usage Study. The project aims to understand how people use conversational health AI tools and evaluate the accuracy, detail, and relevance of their responses.
63News
Qwen has launched its flagship model Qwen4-Max with a new MoE architecture, native vision support, and a 1M token context window. Internal evaluations show it reduces token usage by approximately 50% compared to Qwen3.8-Max, lowering average cost per task by over 60%. Pricing is set at $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching—a 70% reduction.
64News
Qwen officially states that Qwen4-Max outperforms Qwen3.8-Max in token efficiency and cost. In Work360 benchmarks, token usage dropped by approximately 24% (Xhigh) and 35% (High), while Terminal-Bench 4.0 scores saw reductions of about 48% and 49%. Average single-task costs in Work360 fell from $4.43 to $1.67.
65News
Recently, during a conversation with JPMorgan Chase CEO Jamie Dimon, Anthropic CEO Sam Altman stated that computing power is the most profound commercial challenge facing the company. He noted that the IPO was designed to build cash reserves to address uncertainty in computing demand and mentioned Anthropic's $900 billion commitment related to computing power.
66News
The Qwen team has released Qwen4-Max based on the new Qwen4 architecture. It boasts a 2.4T backbone, 71B activated parameters per token, and 328B N-gram embedding parameters. The model features native visual capabilities and a context window of 1 million tokens.
67News
The author translated the five components of Anthropic's resident agent guide (inbox, timer, memory, permissions, and outbox) into a single prompt for their own Grok bot to prevent token waste from idle scheduled tasks.
68News
In this update, Huawei integrates XiaoYi and Work paid features into a three-tier membership system. Plans range from $49 with 1,000 points to $199 with 6,000 points, both including cloud storage. Existing XiaoYi Work subscribers can upgrade at no cost.
69News
Apple has partnered with medical AI company OpenEvidence to launch a research app studying how users employ clinically grounded AI for health Q&A. The app targets iPhone 15 or later, iOS 26 or later, and Apple Intelligence-enabled users aged 21 and above in the US, aiming to recruit up to 200,000 participants over more than 12 months.
70News
Anthropic's Claude Fable 5.2 does not use traditional diffusion models to generate images. Instead, it writes Python code to simulate brushes, paint, and canvas textures, completing drawings stroke by stroke while automatically correcting errors. The model excels in programming, research, and robot control, scoring far higher than OpenAI's GPT-6 Astra on robot control tests.
71News
Tencent WorkBuddy International Edition introduces Google's Gemini 4 Carbon large model, focusing on complex reasoning and long-task handling with multimodal input and a million-token context limit. Access is restricted to Pro plan subscribers at this time.
72News
Physicist Nathaniel Craig shares insights into collaborating with GPT-6 Astra to develop a candidate lattice definition for the Standard Model. This article marks the first post on OpenAI's newly launched science blog; the full piece is available at https://openai.com/index/step-toward-defining-the-standard-model.
73News
The author dissects a recently circulating 1.4 million-character pure HTML 3D game shared in groups, noting that approximately 700KB of Three.js is bundled within the file, while everything else is procedurally generated by code without images, models, or audio.
74News
A paper from Google DeepMind and Columbia University introduces the Back-to-the-Future (BTTF) framework, which converts over 10GB of chip simulation waveforms into compact SQLite databases. Multi-agent systems then translate natural language queries into SQL to check design rules and audit results.
75News
Databricks has released the Genie One Skills guide. Skills are workflow recipes written in an open Agent Skills format, enabling Genie One to produce results with fixed team steps, checks, and formatting.
76News
Hugging Face CEO Clément Delangue reiterated that the concentration of power, capability, and wealth is the greatest risk in AI, noting this tendency is extremely strong. He warned that without a more robust open-source ecosystem than what exists today, the world could face unprecedented asymmetry and inequality. Citing an IPO valuation estimate of $150 billion, Delangue noted that among Anthropic's approximately 5,300 employees, about 4,250 would become millionaires, with roughly 50 individuals worth more than $100 million—a figure exceeding Anthropic's current target of around $1 trillion.
77News
Following a controversy involving Nvidia executives, the U.S. Congress is seeking to further tighten export controls on AI chips. Both Democrats and Republicans have expressed support for legislation aimed at restricting these exports and have called for Nvidia executives to appear before Congress. Senate Banking Committee Chairman Tim Scott has sent letters to both Nvidia executives and Commerce Secretary Howard Lutnick, requesting their public testimony. House Speaker Mike Johnson stated that there is now a consensus within Congress regarding limiting chip exports to China.
78News
Custom themes are essentially mods, so your background can even react to the actual conversation! I configured Claude to snow as my context fills up (this mod hooks into turn.step), but you can get very creative with it ☃️