NAll news

All news

Sunday45 items

  1. cnBeta
    AI score58

    Qualcomm Adreno 32.0.181.0 Driver Causes Snapdragon X2 Elite Local AI and Premiere...

    A Reddit user's testing found that after the Snapdragon X2 Elite was upgraded to Adreno driver 32.0.181.0, local LLM inference and Premiere Pro video rendering performance dropped sharply, with declines of up to 87%. When running Qwen3 series models with llama.cpp using the OpenCL backend, Qwen3-0.6B's generation speed fell from 98.1 tokens per second to 12.8 tokens per second, Qwen3-4B from 29.5 to 9.0, and Qwen3-Coder-30B-A3B-Instruct from 26.8 to 5.9. Another user reported that the export time for a 7-minute 1080p video increased from 1 minute 40 seconds to about 10 minutes. After rolling back to driver 32.0.172.2, performance returned to normal, with the CPU and NPU not noticeably affected. Qualcomm has not yet issued a statement on the matter.

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

  2. The Decoder
    AI score58

    OpenAI adds card and device limits to ChatGPT paid subscriptions

    OpenAI has added payment card and device limits to ChatGPT paid subscriptions. When too many subscriptions are linked to the same card or the same device, new subscriptions can no longer be activated, and upon reaching the limit, the message "Device limit reached" is displayed. The limits cover ChatGPT Go, Plus, Pro 100, Pro 200, and Pro 500; free accounts and the Business, Enterprise, and Edu plans are not counted. OpenAI's support page states that logging out does not automatically free up a slot, and the company recommends instead purchasing additional credits or using a usage reset.

    #Product updates#OpenAI#Industry news

  3. Numerama
    AI score57

    OpenAI's Yang-Mills theory proof awaits verification; physicists hold off on publishing

    The five physicists responsible for reviewing the 1,147-page proof produced by the OpenAI model refuse to publish it before confirming its correctness. They plan to first verify the formalization in Lean, and then submit the analysis to the scientific community; even if verification passes, the proof only covers the existence of Yang-Mills theory and does not resolve the mass gap problem required for the Millennium Prize Problem.

    #Industry news#OpenAI

  4. The Decoder
    AI score71

    OpenAI will merge ChatGPT's Chat and Work, with Plus, Pro, and Business sharing…

    OpenAI's updated support page states that starting in December, ChatGPT's Chat and Work will be merged, and the Plus, Pro, and Business plans will share the same token budget between ChatGPT and Codex. Users can view their remaining quota and reset time in settings; once the quota is exhausted, they can wait for a reset, upgrade their plan, or purchase additional quota. Conversations using the Instant model do not count toward the budget. OpenAI had previously announced plans to merge Chat and Work, and recently also doubled the usage quota for higher-priced plans.

    Same story · 13 reports

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

  5. Phoenix Technology
    AI score65

    Hon Hai says AI server capacity for the next 18 months is fully booked, targets about 45% market share next year

    Hon Hai Chairman Young Liu said demand for AI infrastructure has shown no clear slowdown, AI server capacity for the next 18 months is fully booked, and the company plans to raise its global AI server market share from about 40% currently to about 45% next year and to 50% in the coming years. Second-quarter data disclosed by Hon Hai in August this year showed that the cloud and networking products division, which includes AI servers, already accounted for 51% of revenue, surpassing 50% for the first time. Young Liu also stressed prudent investment management, saying that overinvesting would kill the company, and that depending on the degree of vertical integration, annual AI investment could exceed the originally planned US$3 billion to US$5 billion.

    Same story · 4 reports

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

  6. Phoenix Technology
    AI score58

    Exin Therapeutics Has an AI Agent Perform Brain Surgery on a Mouse to Treat Epilepsy

    Biotech startup Exin Therapeutics says it has had an AI agent complete brain surgery on a mouse with epilepsy. The team used an Anthropic Sonnet model and provided a spreadsheet of 19 gene therapy vectors, an epileptic mouse model, and a robotic arm setup with a camera; the model autonomously selected the drug and brain region, derived coordinates from three papers, and drove the robotic arm to inject 250 nanoliters of the drug, while the craniotomy and suturing were performed by human doctors. The company had originally preferred Anthropic Opus 5.5, but that model refused to carry out the task after detecting biologically related content, and reasoned about the consequences of stopping midway.

    Same story · 5 reports

    #English text is being prepared.#Agents#Anthropic

  7. 36Kr · AI
    AI score68

    OpenAI merges ChatGPT and Codex quotas, billing by token starting in December

    OpenAI has published documentation in its Help Center stating that starting in December, usage quotas for the three paid plans—Plus, Pro, and Business—will be shared between ChatGPT and Codex and calculated uniformly by token, with only Instant chats not counting toward the quota. The documentation does not specify an exact launch date or the token amounts for each plan; once the quota is exhausted, users can choose to wait for a reset, upgrade their plan, or purchase credits. During the same period, OpenAI also published a bank card and device restrictions page, setting a cap on the number of paid accounts usable on a single device; exceeding it will trigger a "Device limit reached" prompt or result in being logged out.

    Same story · 13 reports

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

  8. 36Kr · AI
    AI score58

    Anthropic prohibits Claude from replicating existing games in the Opus 5.5 system prompt

    Anthropic technical staff confirmed that two sentences have been added to the Opus 5.5 system prompt, guiding Claude not to replicate specific games belonging to others; Claude Code may create original games or games using the user's own assets. The prohibition was introduced after the Opus 5.5 release on September 22 triggered a large wave of game reverse engineering, porting, and cloning, but the article points out that it only constrains Claude itself, and developers have already reported on GitHub issues that original games were misjudged and rejected.

    Same story · 14 reports

    #Safety / alignment#English text is being prepared.#Agents

  9. 36Kr · AI
    AI score58

    Meta plans to raise up to $97.5 billion; Oracle caps additional financing for this round at $20 billion

    Meta has launched a new round of equity financing, expanding the size from $35 billion to $37.5 billion; combined with a stock sale plan of up to $60 billion, the total is up to $97.5 billion, to support AI infrastructure and personal superintelligence investment. Oracle, meanwhile, announced at its investor day that additional capital market financing still needed for this investment cycle through fiscal 2030 will be limited to $20 billion, and it expects free cash flow to turn positive in the second half of fiscal 2028. Both companies use supplier financing, customer-provided hardware, and prepayments to reduce upfront outlays, but they have set out different arrangements for subsequent fundraising.

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

  10. Phoenix Technology
    AI score58

    xAI will add invisible watermarks to text generated by Grok starting November 7

    xAI announced that starting November 7, it will add invisible watermarks to text generated by Grok models, and detectors can identify this signal to determine whether text carries a watermark invisible to the naked eye. The company said the move is in response to the EU Artificial Intelligence Act; watermarks are enabled for xAI APIs in all regions worldwide, covering the reasoning, non-reasoning, and multi-agent versions of grok-4.7, grok-4.6, grok-4.5, grok-4.3, and grok-4.20, and team administrators can choose to disable the watermark for an individual model. The watermark is not implemented through hidden characters or metadata, but is instead embedded in the model's token selection process when generating text.

    Same story · 2 reports

    #English text is being prepared.#xAI#OpenAI

  11. cnBeta
    AI score59

    Anthropic and OpenAI Tighten AI Decompilation Controls, Raising the Bar for Browser Ports and Game Remakes

    Anthropic updated its real-time cybersecurity classifier to block users from performing automated decompilation through the community-developed "Reverse Engineer Anything (REA)" framework, and OpenAI also banned related accounts on grounds of "cyber abuse." Because a large number of low-quality recompilations and browser ports rely on such tools, growth in this wave of low-barrier port content may begin to slow, with some users turning to local deployments or open-source alternatives. The original article also mentions that on October 27, Anthropic added a non-disableable game remake rule to Claude, prohibiting the recreation of specific existing games, even if renamed or claimed to be original.

    #Safety / alignment#Anthropic#OpenAI

  12. Phoenix Technology
    AI score50

    Anthropic Bans Claude from Recreating Existing Games in Opus 5.5 System Prompt

    Anthropic confirmed that it added two sentences to the Opus 5.5 system prompt instructing Claude not to recreate specific games belonging to others. This came after blogger chasm discovered that Claude Code was blocking game clones. The rule only constrains Claude itself; users can switch to other models or break tasks down into specific technical tasks whose context does not reveal the overall goal to bypass it. After the ban was announced, a large number of original developers complained on the Claude Code GitHub issue page, saying they were refused when developing original games on the grounds that they were suspected of cloning existing games.

    Same story · 14 reports

    #Safety / alignment#English text is being prepared.#Agents

  13. Sina Tech
    AI score62

    Wall Street AI hedging sentiment heats up, demand for cloud providers' CDS and put options surges

    Some Wall Street investors are hedging against the risk of an AI rally reversal through put options and credit default swaps, with the cost of debt default protection for cloud providers such as Amazon, Alphabet, Microsoft, Meta and Oracle rising to its highest level since 2021. The scale of CoreWeave-linked options contracts hit a record high this month, with the most actively traded being put options expiring in 2028 with a strike price of $35. Jump Trading has newly established a compute trading desk, designing over-the-counter derivatives for clients to hedge against the risk of falling compute prices.

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

  14. 36Kr · AI
    AI score81

    OpenAI's new funding target cut from $30 billion to $20 billion

    According to multiple media reports, OpenAI has cut its new funding target from $30 billion to $20 billion, with its valuation dropping from $1.4 trillion to $1.25 trillion. As of Friday's commitment deadline, actual investment commitments stood at about $11 billion, less than half of the original target. This round is led by a UAE consortium, with UAE state-owned technology funds such as MGX as core investors, and Founders Fund, Andreessen Horowitz, SoftBank and others participating; OpenAI plans to expand fundraising to family offices and high-net-worth individual investors.

    Same story · 8 reports

    #Industry news#OpenAI#Anthropic

  15. ServeTheHome
    AI score62

    NVIDIA Adjusts Rubin GPU Specs at Mass Production: HBM4 Bandwidth Reduced to 19.2TB/s…

    When the Rubin GPU entered mass production, NVIDIA lowered its official specs: per-card HBM4 memory bandwidth dropped from 22TB/s to 19.2TB/s (about 11%), NVLink bandwidth dropped from 3.6TB/s to 3.0TB/s (about 17%), while scale-out network bandwidth increased from 0.4TB/s to 0.45TB/s. The Vera Rubin NVL72 system was adjusted in tandem: memory bandwidth dropped from 1580TB/s to 1400TB/s, NVLink dropped from 260TB/s to 216TB/s, network bandwidth rose from 28.8TB/s to 32.4TB/s, and HBM4 capacity remained at 20.7TB. The HGX Rubin NVL8 was hit hardest: HBM4 capacity was cut from 2.3TB to 2.0TB, memory bandwidth fell by about 26%, and FP4 training performance dropped from 280 PFLOPS to 270 PFLOPS.

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

  16. The Paper
    AI score56

    SiliconFlow's Hong Kong IPO filing accepted, aims to become the first AI Token factory stock

    The Paper learned from the website of the China Securities Regulatory Commission that the Hong Kong stock issuance and listing filing status of Beijing SiliconFlow Technology Co., Ltd. is "accepted," with a material receipt date of October 19, 2026. The company submitted a listing application to the Main Board of the Hong Kong Stock Exchange on June 30 this year, intending to list under Chapter 18C, with Huatai International and Guotai Haitong serving as joint sponsors. The prospectus shows that SiliconFlow's revenue in 2024 and 2025 was 7.346 million yuan and 55.33 million yuan, respectively, with losses for the year of 81.915 million yuan and 345 million yuan, respectively. At the end of 2025, total liabilities were 693.7 million yuan, of which redeemable liabilities were 498.6 million yuan.

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

  17. Phoenix Technology
    AI score62

    US Department of Justice investigates case of AI fake students defrauding $200 million in student loans

    The US Department of Justice announced the results of "Operation Ghostbusters," an investigation into student loan fraud that used stolen identities and artificial intelligence to complete applications and pose as students enrolled in online courses, with the US government estimated to have lost more than $200 million as a result. Multiple community colleges and universities have suspected fraud in the Free Application for Federal Student Aid (FAFSA) program since 2021; five people have been charged so far, and more than 40 open cases involve 26 FBI field offices and 20 US Attorney's Offices.

    Same story · 3 reports

    #Industry news#Safety / alignment#English text is being prepared.

  18. Phoenix Technology
    AI score71

    OpenAI's funding target shrinks from $30 billion; Anthropic reportedly unlikely to obtain investment-grade rating after IPO

    According to The Wall Street Journal, in OpenAI's new funding round, which originally targeted $30 billion, investors were willing to provide only $11 billion, cutting the funding target by one-third; Sam Altman said he will continue reaching out to family offices and high-net-worth individuals in the coming weeks. Blackstone, a major investor in fellow AI company Anthropic, changed its tune and said that Anthropic will likely be unable to obtain an investment-grade rating after going public, which would affect its arrangements to borrow at low cost for its plan to build more than $800 billion in computing infrastructure by 2045. On the day the news emerged, the Nasdaq 100 index closed down 3.5%, bringing its cumulative decline for the entire month of October to about 5.3%.

    #Industry news#OpenAI#Anthropic

  19. IT Home · Artificial Intelligence
    AI score55

    White House Science and Technology Advisor Council Chairman Sachs Comments on Anthropic S-1: May Become the Greatest Company in the History of Capitalism

    David Sachs, Chairman of the U.S. White House Science and Technology Advisory Council, analyzed Anthropic's S-1 registration filing on the All-In podcast, saying that this artificial intelligence company is likely to become the greatest company in the history of capitalism. The filing shows that Anthropic had revenue of $16.3 billion in the first half of 2026 and cash on its books of $112 billion, with projected 2026 annual recurring revenue reaching $100 billion, and seven founders holding 50.1% of voting rights through Class F shares. Sachs believes that its abundant cash makes its IPO financing needs lower than outside expectations, thereby giving it stronger pricing power.

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

  20. English text is being prepared.
    AI score57

    Report: Spotify plans to launch AI music remix service early next year, allowing users to create their own remixes and covers

    According to the Financial Times, Spotify plans to launch an AI-driven music remix service in early 2027, allowing paid subscribers to create their own remixes and cover versions of songs, and is considering charging an additional $4 to $5 per month. Spotify is negotiating licensing agreements with Sony Music and Warner Music, has reached agreements with Universal Music Group, Merlin, and Believe, and artists themselves can still decide whether to participate. Under the proposed plan, royalties generated from remixes will flow in full back to the original artists, and plays of remixed works will also count toward the original songs' chart performance.

    Same story · 2 reports

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