NAll news

All news

Wednesday151 items

  1. cnBeta
    AI score58

    Apple warns that AI agents may act in unexpected or unpredictable ways

    Apple warned in the filing that AI agents can autonomously interact with internal or external systems, execute or modify code, send communications and complete transactions, and may act in unexpected or unpredictable ways, including taking actions that were not instructed or authorized. Apple said the company's ability to oversee and control third-party agents may be limited, and it adjusted Mac's Full Disk Access permissions last month to reduce the related risks. New CEO John Ternus said on the earnings call that agent capabilities are a promising tool for users and that the company has room to expand Siri, but it needs to ensure users clearly understand what access they are granting.

    Same story · 7 reports

    #Safety / alignment#Agents#OpenAI

  2. cnBeta
    AI score65

    OpenAI's new internal model scores about 47% higher than GPT 6 Astra on BenchCAD 2.0

    In the blog post "Accelerating Engineering Development for the Physical World," OpenAI disclosed that an unnamed internal model achieved an average score of 0.875 on the BenchCAD 2.0 engineering benchmark under the highest reasoning setting, while the current GPT 6 Astra scored 0.594, a relative improvement of about 47.3%. The benchmark mainly evaluates the model's ability to reconstruct 3D mechanical parts and assemblies and to identify circuit board connection relationships, among other capabilities. OpenAI did not say whether this internal model is the GPT 6.2 Astra that recently appeared in Codex backend records. The blog post also mentioned that two engineers used GPT 6 Astra to complete the design of a portable AI device in less than a week, and a contract manufacturer produced a small batch of prototype machines in less than a month, but converting the design into a manufacturable product still requires multiple rounds of manual review.

    Same story · 10 reports

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

  3. 36Kr · AI
    AI score60

    Duolingo Q3 Daily Active Users Up 29% Year-over-Year; Falling AI Inference Costs Prompt Retirement of Duolingo Max

    Duolingo's Q3 2026 earnings report shows daily active users up 29% year-over-year, revenue of $309.0 million up 14% year-over-year, and bookings of $316.1 million up 12% year-over-year. CEO Luis von Ahn said AI inference costs have fallen sharply, with the cost per Video Call dropping from about $0.30 to well below $0.01, prompting the company to retire its top subscription tier Duolingo Max and fold Video Call into the Super tier. He also revealed that the number of A/B tests in the past month grew about 90% year-over-year while headcount grew only 10%, and that product managers and designers can now independently complete the full product change process from idea to launch.

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

  4. AYi AI Notes
    AI score58

    xAI engineer uses Grok Bot with Cursor to merge 1,150 PRs in four days and build a pixel game

    After xAI engineer Lauren (@poteto) spent $5.99 to buy a domain, she used her company's Grok Bot with Cursor to merge about 1,150 PRs in 4 days, building from scratch a 2.5D pixel survival game starring OpenAI Codex engineer Tibo, embedded in an X tweet card and playable on tap. In the game, when the protagonist is knocked down, a "Reset" prompt pops up, poking fun at Tibo's daily routine of manually resetting Codex usage quotas for OpenAI users. Tibo himself left a self-deprecating comment after trying it, saying the game must have used up all of xAI's compute.

    Quoted postAYi · @AYi_AInotes

    English text is being prepared.

    Same story · 2 reports

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

  5. Phoenix Technology
    AI score58

    Security researcher banned after teardown of Samsung TV; OpenAI apologizes and lifts ban 11 days later, Anthropic tightens Claude risk controls

    Security researcher @log4j was banned by OpenAI on grounds of cyber abuse after using ChatGPT to assist in locally reverse-engineering their own 2025 Samsung TV, and lost about $500 in balance. Eleven days after the ban, OpenAI sent an email acknowledging the error, deactivated the ban, and restored the account. OpenAI core researcher roon said Codex has never prohibited reverse engineering, and that most blocks are accidents caused by the difficulty of drawing judgment boundaries. In the same period, Anthropic tightened risk controls for Claude Code and confirmed that using it to recreate games is prohibited.

    #Industry news#Safety / alignment#OpenAI

  6. Phoenix Technology
    AI score65

    OpenAI's new model GPT 6.2 Astra suspected of being in gray testing, internal model engineering benchmark score improves 47.3%

    Multiple overseas developers discovered a new model identifier named gpt-6.2-astra in Codex usage statistics, with some accounts showing a call share of 19.8%, but the model selection interface still displays GPT 6 Astra. OpenAI published a technical blog post on November 3, disclosing that an unnamed internal model achieved an average score of 0.875 on the BenchCAD 2.0 engineering benchmark, an improvement of about 47.3% over the current GPT 6 Astra's 0.594. OpenAI has not yet confirmed the existence or release date of GPT 6.2 Astra, nor has it stated whether the internal model corresponds to this version.

    Same story · 10 reports

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

  7. Shanghai Observer
    AI score57

    Houmo AI Completes Several-Billion-Yuan Series B Funding; M50 Chip Wins Over 100 Customers

    Houmo AI, a compute-in-memory AI chip company, recently completed a Series B funding round of several billion yuan. The funds will be used to expand large-scale delivery of the M50 and advance R&D on the M6X and next-generation 3D CIM products. The company disclosed that in less than a year of mass production, the M50 has already secured more than 150 product design wins and served over 100 customers, covering four directions: agent computers, intelligent mobile terminals, embodied robots, and edge inference. The M50 offers 160 TOPS@INT8 of compute, 150 GB/s of memory bandwidth, and typical power consumption of 10W. Its first 3D CIM chip, the A20, has already entered production.

    Same story · 3 reports

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

  8. Semiconductor Industry Observer
    AI score56

    ADI Completes Acquisition of AI Chip Company Alif Semiconductor for $1.35 Billion

    ADI (Analog Devices) announced that it has completed its acquisition of AI chip company Alif Semiconductor for $1.35 billion in an all-cash transaction, which was first announced on September 9, 2026. Alif is headquartered in Pleasanton, California, and focuses on low-power MCUs and converged processors. Its Ensemble series pairs Arm Cortex-M55 cores with Ethos-U55 neural network acceleration units, enabling local AI inference at low power and covering scenarios such as wearables, smart sensing, and robotics. ADI said the merger combines its sensing, signal processing, power, and connectivity technologies with Alif's AI-native platform, targeting fields including industrial, data center infrastructure, defense, energy, robotics, and digital health.

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

  9. heise online
    AI score53

    AMD's Q3 revenue was $13.38 billion, with the data center business share rising to 64%

    AMD reported third-quarter revenue of $13.38 billion, up 45% year over year, and operating profit of $2.83 billion, up 123% year over year. Demand for Epyc server processors and Instinct AI accelerators pushed the data center business's share of total revenue from 47% last year to 64%, while gaming business revenue fell 40%. AMD expects fourth-quarter revenue of about $17 billion, and CEO Lisa Su said the company has secured key supplies including wafers, substrates, advanced packaging and memory.

    Same story · 6 reports

    #Industry news#English text is being prepared.

  10. heise online
    AI score59

    Advocate General of the European Court of Justice: AI chatbots such as Gemini may infringe copyright

    Advocate General Maciej Szpunar of the European Court of Justice issued an opinion holding that AI chatbots such as Gemini that scrape and use news articles from websites to generate answers may infringe publishers' reproduction right and right of communication to the public. He proposed criteria for judgment: whether access was illegal, such as bypassing a paywall, and whether the answer reproduces the original expression in a recognizable way such that readers no longer need to open the original. The opinion is non-binding, and a ruling is not expected before 2027. The case stems from a lawsuit by Hungarian publisher Like Company against Google, alleging that Gemini reproduced its protected text without permission. On the question of AI training, Szpunar considered the two questions from the Hungarian court to be hypothetical and inadmissible, but noted that under the text and data mining exception, training models on protected works is "freely permitted." Google said it welcomed this statement regarding training.

    Same story · 2 reports

    #English text is being prepared.#Google#RAG

  11. ServeTheHome
    AI score58

    AMD Q3 2026 Earnings: Data Center Revenue $8.546 Billion, Up 97% Year-over-Year

    AMD announced its Q3 2026 earnings, with total revenue of $13.377 billion, up 45% year-over-year and 16% quarter-over-quarter. Data center revenue reached $8.546 billion, up 97% year-over-year and 27% quarter-over-quarter. AMD attributed this to strong EPYC demand and the launch of the Instinct MI450 GPU, and said customers including Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure are already running workloads on the Helios rack-scale solution. The company expects Q4 revenue of about $17 billion, plus or minus 3%, which at the midpoint would represent roughly 66% year-over-year growth.

    Same story · 6 reports

    #Industry news#English text is being prepared.

  12. The Guardian Technology
    AI score63

    Canva restructuring to cut or relocate over 1,200 roles, 22% of global workforce

    Canva has announced a restructuring that will cut or relocate 1,271 roles, 22% of its global workforce, affecting employees involved in building the design platform. CEO Melanie Perkins said in an email to staff that this is the hardest moment in the company's thirteen-year history, and that the restructuring focuses on reducing layers and bringing more teams together to work in person. Departing employees will receive at least six months' pay, up to 12 months of health insurance, and can keep their laptops; they can also apply for a $10,000 startup grant. The report noted that Atlassian, Block and WiseTech had previously restructured as well, but Canva did not mention AI in its announcement.

    Same story · 2 reports

    #Industry news#English text is being prepared.

  13. Yicai
    AI score61

    Yoozoo Games plans to procure GPUs for 4.225 billion yuan, about 3 times last year's revenue

    Yoozoo Games announced that its wholly-owned subsidiary Yoozoo Information plans to procure GPUs of the model PRO5500 (PCIe) from Company A for 4.225 billion yuan, with information on the counterparty exempted from disclosure. The amount is about 3 times the company's 2025 revenue of 1.402 billion yuan. The source of funds is a combination of self-owned funds, bank credit, financial leasing and other methods, and the contract adopts batch delivery. The company stated that there is uncertainty in the implementation progress of the contract, and the impact on financial indicators cannot be measured for the time being.

    Same story · 2 reports

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

  14. East Money
    AI score55

    Yoozoo Games plans to procure GPUs for 4.225 billion yuan, about 3 times its 2025 revenue

    Yoozoo Games announced on the evening of November 3 that its wholly-owned subsidiary Yoozoo Information signed a sales contract with Company A to procure GPUs of the PRO5500 (PCIe) model, with a total amount of 4.225 billion yuan. The funding sources are a combination of self-owned funds, bank credit, financial leasing and other methods, and delivery will be carried out in batches. The amount is roughly 3 times the company's 2025 revenue of 1.402 billion yuan. The company said that if implemented smoothly, it will help its transformation and upgrading, but the amount of impact on financial indicators cannot yet be measured.

    Same story · 2 reports

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

  15. Leiphone
    AI score60

    Houmo AI Completes Several-Billion-Yuan Series B Funding; M50 Chip Has Secured Over 150 Product Design Wins

    According to informed sources, compute-in-memory AI chip company Houmo AI recently completed a Series B funding round of several billion yuan, with the funds to be used to expand large-scale delivery of the M50 and advance R&D on the M6X and the next-generation 3D CIM. The company disclosed that in less than a year of mass production, the M50 has secured more than 150 product design wins and serves over 100 customers, covering four directions: Agent Computer, intelligent mobile terminals, embodied robots, and edge inference. The M50 offers 160 TOPS@INT8 of compute, 150 GB/s of memory bandwidth, and typical power consumption of 10W; complete systems such as the Lenovo AI Host P7 preview edition have adopted it for local large-model inference.

    Same story · 3 reports

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

  16. Electronic Technology Application
    AI score79

    DeepSeek V5 plans to pretrain on Huawei Ascend 950 chips

    According to reports, DeepSeek's next-generation large model, DeepSeek V5, plans to conduct pretraining on Huawei Ascend chips. Previously, Ascend chips were mainly used for inference tasks in DeepSeek large models; V5 will be its first large model to natively rely on Ascend 950 for training. The Ascend 950 series is divided into two models, 950PR and 950DT, both based on the third-generation Da Vinci architecture. The 950DT is equipped with 144GB of self-developed HBM and a memory access bandwidth of 4TB/s, while the 950PR is equipped with 128GB of on-chip memory and FP4 compute power reaching 1784 TFLOPS. People familiar with the matter said V5 plans to use Ascend 950 as its core training compute, with the project timeline pointing to launch next year.

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

  17. Electronic Technology Application
    AI score58

    Putin Signs Presidential Decree No. 802: Russian Non-State Data Centers Must Reserve 10% of Computing Power

    Russian President Putin signed Presidential Decree No. 802 on November 2, requiring all non-state data centers in the country to reserve 10% of their site area and computing power within 10 days after the decree takes effect, equivalent to keeping one out of every ten servers on standby. The reserved computing power is still used by the data centers themselves in normal times, but once the government's special task force committee makes a decision, the data centers must provide it within 24 hours, and the requisition decision must be agreed upon with the Russian Federal Security Service (FSB), with service fees not exceeding the cap set by the committee. Russia's Ministry of Digital Development has been tasked with taking stock of the idle site space and computing power of non-state data centers nationwide, and the decree also allows the use of overseas data center computing power under certain circumstances.

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

  18. East Money
    AI score60

    OpenAI Executive Says ChatGPT Search Volume Has Reached 20% of Google's

    Adam Fry, head of consumer product for OpenAI ChatGPT, told the media that ChatGPT's search query volume now reaches roughly one-fifth of Google's search query volume, while ChatGPT is only about a year and a half old. He also said that OpenAI added 200 million users in the past two months, with weekly active users rising to 1.2 billion; users create 1.5 billion images per week with ChatGPT, and voice mode has 250 million weekly active users. Fry believes that ChatGPT's personalization and contextual understanding are its advantages over search engines, while Google CEO Pichai previously said that search and chatbots are not a zero-sum game.

    Same story · 7 reports

    #Industry news#Search#OpenAI

  19. East Money
    AI score56

    Xiaomi Discloses Progress on Humanoid Robots in Factories and Home Scenarios; Multiple Embodied Intelligence Concept Stocks Hit Limit Up

    During the 2026 World Internet Conference Wuzhen Summit, Xiaomi disclosed progress on its robotics business: in an automotive factory, a humanoid robot took on an 8-hour work shift for the first time at the center console side cover panel sequencing station, achieving autonomous anomaly correction and system-level coordination with the production line; in home scenarios, an Agent system based on Xiaomi-Robotics-1 completed a long-horizon validation lasting over 30 consecutive minutes, covering 3 rooms and more than 50 tasks. Affected by this news, humanoid robot concept stocks collectively moved in the morning session of November 4, with Zhenghe Industry, Yongzhen Shares, Zhaowei Electromechanical, and Fulaixin Materials hitting the limit up. Overseas, Tesla confirmed that Optimus has entered the Fremont high-volume production line, and Musk proposed a target of 20,000 units per week by the end of 2027.

    Same story · 5 reports

    #Industry news#Embodied AI

  20. Phoenix Technology
    AI score53

    Apple Warns That AI Agent Behavior Is Unpredictable, May Invite Lawsuits and Reputational Damage

    Apple warned in its annual financial filing that AI agents capable of acting in unexpected or unpredictable ways could expose the company to lawsuits or damage its reputation. The filing states that such technology can autonomously interact with internal or external systems, sometimes without authorization to access, execute or modify code, send communications, and complete transactions. The article notes that earlier this year OpenAI launched Dots and Meta released Muse, both agents capable of autonomously performing tasks and even sending users emails or reminders when new messages arrive.

    Same story · 7 reports

    #Safety / alignment#Agents#OpenAI