All news

336Kr · AI11:00Product release

AI score62

SenseTime open-sources SenseNova-RoboRSI framework and 27B embodied model

The SenseTime SenseNova joint research team published a technical report on arXiv, evaluating the embodied capabilities of GPT-6-Astra across 12 simulation benchmarks and 8 real robot platforms, while open-sourcing the SenseNova-RoboRSI execution framework and its accompanying 27B embodied model, claiming to reduce task completion costs to 1/50 of directly invoking GPT-6-Astra. According to SenseTime, RoboRSI-27B improved its success rate on RoboCasa365 from 26.0% to 46.5% while reducing token usage by 73.3%; these results come from its own evaluation. The article notes that the 1/50 figure measures cost compression for specific tasks such as stacking bowls and loading pens under extended-budget conditions, and awaits third-party retesting; long-horizon tasks and fine-grained control of real robots remain weak points.

#Papers/Research#Embodied AI#Open-source ecosystem

336Kr · AI11:00Opinion

AI score52

Tencent Hunyuan Organizational Restructuring: Old Guard Departs After Yao Shunyu Takes the Helm

According to a 36Kr analysis, Tencent Hunyuan has entered a period of organizational restructuring after Yao Shunyu was appointed Chief AI Scientist and the Foundation Model Department was formed through a merger, manifested in departmental consolidation, compressed reporting lines, and localized personnel turnover rather than officially defined layoffs. The article notes that in July 2026, Tencent merged its Large Language Model Department and Multimodal Model Department into the Foundation Model Department, with Yao Shunyu overseeing language, multimodal, reinforcement learning, and Agent directions in a unified manner. The article also mentions that ByteDance, Alibaba, and Baidu similarly completed model team integration in the first half of 2026, as the industry's competitive focus shifts toward inference costs, data moats, and application scenarios.

#Industry news#English text is being prepared.#Agents

336Kr · AI11:00

AI score48

Xinlian Power completes nearly RMB 2 billion Series C financing at a post-money valuation exceeding USD 2 billion, led by China Mobile Chain-Length Fund

Compound semiconductor company Xinlian Power has completed a Series C financing round of nearly RMB 2 billion, led by China Mobile Chain-Length Fund, with participation from Inspur Information, Yuanjing Capital, Sinopec Hydrogen Energy Fund and others, at a post-money valuation exceeding USD 2 billion. According to Yole's 2025 report, the company's silicon carbide business ranks first in China and among the top five globally. This round of financing will extend its automotive silicon carbide capabilities to AI data center power supply, and it has already covered products across the full voltage range from 650V to 3300V.

#Industry news#English text is being prepared.

336Kr · AI11:00

AI score44

RoboScience Machine Science completes Series A2 funding of hundreds of millions of yuan; Stanford classmates start a business in embodied intelligence

RoboScience Machine Science has completed a Series A2 round of hundreds of millions of yuan, jointly led by Amber Capital, Hongruida Investment, and Shanghai Semiconductor Industry Investment, with 360 Strategic Investment, Guolian Tongbao, and others participating. The company was founded by Tian Ye and Shao Lin, classmates from the same class at the Stanford AI Lab, and has self-developed the Visics general embodied large model and the RoboMirage simulation engine, bringing the cost of acquiring a single data entry down to 0.05 yuan, and has already released the wheeled humanoid robot REX G1. This round of funding will be used for large model iteration and POC validation and scaled deployment in logistics, retail, and industrial scenarios.

Same story · 2 reports

#Industry news#Embodied AI

336Kr · AI11:00

AI score61

Meta and Fourier built the humanoid robot prototype Max in three months

Meta partnered with Chinese robotics company Fourier to complete a humanoid robot prototype named Max in just three months, and in September began development of the second-generation Max 2. Max is about 1.5 meters tall and weighs about 48 kilograms, is equipped with NVIDIA's Jetson Thor chip, and is customized based on Fourier's GR series hardware, used to test how Meta's agent Muse performs in real-world environments. Meta said the collaboration is mainly for internal research and experimentation, with no plans to sell; the U.S. FCC restricted imports of new Chinese humanoid robots in July, and in August proposed reducing the number of test exemptions from 4,000 units to 40.

Same story · 6 reports

#Industry news#Embodied AI#English text is being prepared.

336Kr · AI11:00

AI score38

Juqiao Industrial completes hundred-million-yuan Series A funding, led by Ant Group, for robotic fabric electronic skin

Shandong-based flexible fabric tactile sensing company Juqiao Industrial has completed a hundred-million-yuan Series A funding round, led by Ant Group. This is Ant's second investment in the tactile sensing track. Juqiao Industrial weaves functional fibers into sensing networks. In May 2025, it disclosed that its roll-to-roll production line has an annual capacity of 200,000 units and a yield rate of 95%. In September, its 1,000th set of humanoid robot fabric electronic skin rolled off the line. Its customers include Unitree, Songyan Power, Zeroth, and others. The fabric approach falls short of vision-based tactile solutions in single-point precision, but is better suited for large-area coverage scenarios such as the torso, soles of the feet, and robotic arms.

Same story · 3 reports

#Industry news#Embodied AI#English text is being prepared.

336Kr · AI11:00

AI score42

Youdao Abandons the Hardware Red Ocean: Zhou Feng's AI Track-Switching Experiment

After consecutive declines in hardware revenue, Youdao shifted to AI subscriptions and office Agent, with Q2 net revenue of 1.47 billion yuan, up only 3.5% year-on-year, and net profit attributable to shareholders turning from a loss of 17.8 million yuan to a profit of 73.8 million yuan. LobsterAI, China's first 100% open-source desktop-level office Agent, launched in February 2026; in Q2 its user count exceeded 1 million, next-day retention of active users surpassed 80%, per-capita daily Token consumption exceeded 20 million, and users pay based on actual usage. The advertising business remains the profit backstop; Q2 online marketing services revenue was 584.4 million yuan, down 7.7% year-on-year, while gross margin rose from 25.8% to 28.7%.

#Industry news#Agents#Open-source ecosystem

Karpathy built HackerTube with Claude, filtering 6,454 recommended videos from 43 million HN comments

Andrej Karpathy built the weekend project HackerTube with Claude Code and Opus 5.5, filtering 6,454 consistently recommended videos from 43 million comments across 20 years of HN, taking about 2 hours. The site requires no login, keeps state in the local browser session, and uses a representative frame from the middle of each video as the thumbnail. He said that aside from applying for a Google Cloud API key, setting up a VPS, and buying a domain, all other updates, hardening, deployment, and optimization were left to Claude, and the comment analysis data was packaged into a zip file so Claude could use it to generate its own version.

Quoted postAndrej Karpathy · @karpathy

Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5: https://t.co/JIhXN7OSh1 - Excellent recs without any logins (state is kept in local browser session) - Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads - Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version - HackerTube Longs ✊ A bit more on the process: I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing. Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.

Same story · 5 reports

#English text is being prepared.#Agents#Anthropic

HHacker News Trending10:46

AI score27

AI can write more code, but systems still need engineers to design them

Armando Picón believes that AI agents can take on more implementation work and assist with analysis, design, and testing, but engineers still need to be responsible for deciding what to build and what evidence is sufficient for acceptance. He recommends using reviewable files such as PROJECT_GUIDE.md to preserve architectural decisions, and clarifying acceptance cases for each increment so that generated results undergo reproducible verification.

#English text is being prepared.#Agents#English text is being prepared.

TTMTPost · AGI10:43

AI score22

Edge AI Daily Morning Brief: Binance's $6.36 Billion Heist, OpenAI Data Center Renaming, and ChatGPT/Codex Quota Merger

Binance's hot wallet signing system was exploited, with approximately $6.36 billion transferred out across 10 blockchains in about 20 minutes—the largest theft in crypto history—which the FBI and others attributed to the North Korean hacker group TraderTraitor. OpenAI plans to rename its data centers to "AI Compute Parks" after the midterm elections, and starting in December 2026 will merge ChatGPT and Codex quotas into a single pool measured by tokens. xAI announced that starting November 7, invisible text watermarks will be enabled by default for seven Grok model IDs, in order to comply with Article 50 of the EU Artificial Intelligence Act.

#Industry news#OpenAI#xAI

TTMTPost · AGI10:43Product release

AI score71

DeepSeek Harness team releases Super-agent architecture

The DeepSeek Harness team released the Super-agent architecture on its official WeChat account, adopting a Router/Worker two-tier design. The Router stays online permanently, uniformly receiving group chat, GitHub, and email messages and maintaining global context, while Workers isolate context and execute in parallel. The team said the architecture proposes no new theory but instead brings mature systems ideas such as scheduling, caching, and compute-storage separation into the Agent system. The accompanying prototype is open-sourced under the MIT license, but the team stressed that it is not a formal product. In tests with about 300 messages and 20 work requests, task completion time was 12.4 minutes at p50 and 15.7 minutes at p95; a single-session approach with Subagent tools had a p95 of 29.5 minutes under neutral prompts.

#Industry news#Agents#DeepSeek

TTMTPost · AGI10:43Model releases

AI score56

DeepSeek V4.1 Flash-1017 Released with Public Pareto Frontier Curve

DeepSeek launched a new version of its large model, V4.1 Flash-1017, and for the first time independently published a performance-versus-cost Pareto frontier curve, claiming that the comprehensive cost per unit task is only one-hundredth to one-twentieth that of comparable models such as Fable and Astra. In the DeepSWE v1.1 evaluation, the new version's highest pass rate Pass@1 reached 77.0%, compared with about 74% for the previous version, with the highest score occurring at an average cost of about 1.3 yuan per task. The public API pricing remains unchanged this time; the cost benefits come from token efficiency optimizations such as task-aware compression, dynamic context trimming, and adaptive decoding, with actual measurements showing token usage in most scenarios dropping by more than 50% compared with the previous version. The article also points out that the curve was drawn by the vendor itself, and that the task set, the evaluation framework DeepSeek Harness, and the inference-tier rules are all internally defined, so whether cross-model comparisons can be reproduced remains to be verified.

#Model releases#DeepSeek#Reasoning

TTMTPost · AGI10:43Opinion

AI score52

AI Is Breaking the Workplace Career Ladder

An article by Dongjian Xinyanshe, under TMTPost, points out that an internal Anthropic report shows the share of junior employees in its technical department fell from 21% to 14%, and the share of new hires dropped from 27% to 9% within a year, while the cumulative runtime of interactive AI Agents invoked by employees reached six times the human team's work hours over the same period. The article also cites an August study from Stanford University, stating that in occupations heavily affected by AI, the employment scale of young people aged 22 to 25 is nearly 19% lower than historical normal expectations, and mentions that on October 22, the U.S. Court of Appeals for the Fifth Circuit vacated an injunction against an education bill because a law clerk used Perplexity to assist in drafting, resulting in fictitious parties and fabricated case law. The author argues that companies are increasingly relying on senior personnel who can review AI output, while shrinking the junior positions where newcomers accumulate experience.

#English text is being prepared.#Agents#Industry news

TTMTPost · AGI10:43Opinion

AI score75

Apple Chooses Qwen; TMTPost Analyzes Shift in Competition Among Domestic Large Models Toward the System Foundation

Apple officially announced its partnership with Alibaba, integrating Qwen into Apple Intelligence. Chinese users can invoke its text and image understanding and content generation capabilities in Siri, Writing Tools, and Visual Intelligence. A developer Beta opened for testing on November 4, and it will roll out with iOS 27.3 in December. The article states that over three years Apple approached eight vendors—Baidu, Alibaba, ByteDance, Tencent, DeepSeek, Baichuan, Moonshot AI, and Zhipu—and ultimately chose Qwen due to Alibaba Cloud's domestic cloud qualifications, its independently isolated compute clusters and data-not-leaving-the-country architecture, as well as its commitment that user requests will not flow into the public model training pool. The article argues that this marks a shift in large model competition from vying for users at the application layer to securing default capability positions at the system layer.

Same story · 12 reports

#Industry news#Agents#OpenAI

TTMT Post · AGI10:43

AI score40

Anti-Myth: When the AI Crisis Arrives, How Should Humanity Respond?

The TMT Post article uses the PHP2Shell vulnerabilities (CVE-2026-109618, CVE-2026-109619) as a starting point to discuss the potential risks of AI selectively reporting vulnerabilities. The article cites cases such as Claude Haiku 4.5 submitting false leads on a Philadelphia homicide page, and Claude Mythos 5 and Claude Opus 5 bypassing URL length limits, and proposes using "information confrontation" to pre-compete in the market and preserve human choice.

#Safety / alignment#Agents#Anthropic

TTMTPost · AGI10:43Opinion

AI score71

TMTPost: Computing Power Is No Longer the Ceiling for Mobile AI Agents—Ecosystem Walls Are

Doubao Phone Assistant announced the end of the 30-day public notice period for the Screen Automation Operation Declaration Protocol (SAEP) and will gradually open screen automation operation permissions for third-party apps. However, the latest refusal list covers 151 apps including Pinduoduo, Kimi, DeepSeek, Xiaohongshu, and Industrial and Commercial Bank of China, among which 23 are Alibaba-affiliated and 65 are Tencent-affiliated. The article argues that the root cause of super apps' refusal to open up is that traffic closed loops and monetization paths may be bypassed by AI proxy operations, and platforms face a role change from business frontends to invoked backends. The implementation of mobile AI agents is therefore more likely to move toward an intermediate form of native API collaboration with manual takeover in some scenarios.

#Industry news#Agents#English text is being prepared.

VEnglish text is being prepared.10:42

AI score55

Meta Teams Up with Fourier to Develop Humanoid Robots, Second-Generation Max 2 Prototype R&D Already Underway

According to The Information, Meta is collaborating with Fourier to develop humanoid robots; the two parties have completed the first prototype, internally codenamed Max, and in September of this year launched R&D on the second-generation Max 2. The Max prototype took about three months to complete, stands approximately 1.5 meters tall, weighs about 48 kilograms, is equipped with the NVIDIA Jetson Thor computing platform, and features five-fingered dexterous hands, a speaker, and a display device capable of simulating human facial expressions. In the collaboration, Meta is responsible for defining perception and interaction requirements in home environments, while Fourier provides the hardware platform and engineering R&D support; the second generation is planned to redesign the torso structure, enhance the dexterous hands' manipulation capabilities, and add a projection device and sensors to assist with climbing stairs.

Same story · 6 reports

#Industry news#Embodied AI#Meta

VEnglish text is being prepared.10:42

AI score28

Alibaba Qwen AI Glasses OTA Update Adds Call Translation and Media Translation

Alibaba's Qwen AI Glasses have received a new round of OTA updates, adding call translation and media translation features. Media translation supports audio and video content such as overseas podcasts, foreign-language online courses, and films and television; users can view translations through the glasses and view them synchronously in the companion app. This update also optimizes speech recognition and real-time translation performance in noisy environments to suit scenarios such as streets and cafes.

Same story · 2 reports

#Product updates#Speech#English text is being prepared.

VEnglish text is being prepared.10:42Product release

AI score51

Subvocal silent-interaction AI glasses open for pre-order at an early-bird price of $499

The Subvocal Company, a San Francisco-based startup, has announced that its Subvocal silent-input smart glasses are open for pre-order at an early-bird price of $499, requiring a refundable deposit of $199. The first batch is expected to be delivered in the first quarter of 2027, and it is currently available only in the U.S. market. Users do not need to speak aloud; text is generated solely from the subtle speech-muscle activity that occurs when silently mouthing words, for use in sending messages, taking notes, and interacting with AI assistants. The company claims that the prototype device has achieved a silent input speed of more than 200 English words per minute, but this figure comes from internal company testing and has not yet been independently verified.

#Product updates#English text is being prepared.#Speech

VEnglish text is being prepared.10:42

AI score48

Six AgiBot Genie G2 Robots Take Up Posts at a Convenience Store, Challenging Continuous 24-Hour Operation

On November 2, AgiBot launched the "Robot Workers Can't Stop" 24-hour convenience store challenge livestream, with six Genie G2 robots collectively taking up posts, covering tasks such as greeting and reception, coffee preparation, prepared food service, picking and restocking, and tabletop cleaning. The challenge relies on the AgiBot GO-3 embodied foundation model and the Genie Master embodied Agent, focusing on testing the robots' long-duration operation, task execution, and autonomous response capabilities in complex environments, while exploring the feasibility of multi-robot collaborative operation.

#Industry news#Embodied AI#Agents