NAll news

All news

Tuesday581 items

  1. Rohan Paul
    AI score58

    Anthropic Updates S1/IPO Filing: Future Models May Be Approaching the Recursive Self-Improvement Threshold

    Rohan Paul, citing Anthropic's latest updated S1/IPO filing, said that future models may soon reach its recursive self-improvement (RSI) threshold. Anthropic's Responsible Scaling Policy defines this threshold as AI capability improving at twice the fastest sustained rate without significant AI assistance—that is, one year producing the equivalent of two normal years of progress—potentially because models are automating research and engineering. Within 8 months, Anthropic's internal use of Claude grew more than 250-fold, and by April its monthly value had exceeded the company's personnel spending.

    Quoted postRohan Paul · @rohanpaul_ai

    Anthropic's own AI was already doing 26% of the work to build its next models (RSI). "If you go back to August, Anthropic published a very interesting blog talking about how they were using RSI. They had a chart showing that 26% of their tasks that are involved in trying to create the next generation of models, they were already were using RSI to do that. Now, that was back in August, so who knows what they're at now, being in October." - @DavidSacks ---- Full video on "All-In Podcast" YouTube channel, (link in comment)

    Same story · 3 reports

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

  2. The Decoder
    AI score60

    SpaceX Reportedly Cuts Grok Bot Quota; Musk Says Running Costs Too High Relative to Revenue

    SpaceX has reportedly cut the usage quota for Grok Bot, with some users discovering over the weekend that their quota had been halved. Musk responded on X on Sunday that the cost of running Grok Bot is unsustainably high relative to revenue, but did not directly confirm the cut. Grok Bot can browse websites, reply to emails, and log into accounts, and its quota is shared with Grok and the coding platform Cursor; less than two weeks ago Musk had called it the fastest-growing product, and SpaceX President Gwynne Shotwell mentioned in early October that it had 800,000 paying users per week. OpenAI and Anthropic are also tightening their monthly subscription plans: OpenAI will merge Codex and Chat quotas starting in December, and Anthropic says the cost of an agent is 4 to 15 times that of a simple chat.

    Same story · 4 reports

    #Industry news#Agents#OpenAI

  3. LMSYS BlogFirst-partyAnalysis
    AI score59

    LMSYS explains in detail how HiCache adapts to multi-parallelism and distributed inference

    The LMSYS blog introduces how HiCache is implemented in context parallelism, pipeline parallelism, speculative decoding, unified KV layout, and P/D disaggregation scenarios, with the core being to keep cache decisions consistent across ranks. The iFlytek Spark MaaS team used HiCache plus external storage to serve GLM-5.1/5.2 CodingPlan workloads, raising the cache hit rate from about 30% to 95% and reducing median TTFT to under 3 seconds. In CP sharding tests, the GPU token hit rate increased by 40.48 percentage points, the host memory serving token volume dropped by 81.2%, and the average prefill queueing time decreased by 56.2%.

    A logical radix tree shared by two CP ranks: logical page IDs map to rank 0 or rank 1 by q mod 2, and extending the A/B/C prefix allocates page D on rank 1.

    #Tutorials / practice#English text is being prepared.#Open-source ecosystem

  4. Cailian Press
    AI score82

    Anthropic Returns to Operating Loss in Q3, Discloses Approximately $895.4 Billion in Compute Commitments for the First Time

    Anthropic submitted a revised prospectus to the SEC, adding data for the third quarter of fiscal year 2026. Q3 revenue was approximately $17 billion, up 47% quarter-over-quarter, a marked slowdown from the 144% growth in Q2, and it recorded an operating loss of $567 million, compared with operating profit of $974 million in the previous quarter. The company disclosed for the first time total purchase and lease commitments of approximately $895.4 billion, almost all for compute and data centers, of which about $69.8 billion will come due within one year; there are also unquantified contracts such as $125 billion to $150 billion in chip purchases from Google and over $20 billion in spending at AMD, which would bring the total to more than $1 trillion when included.

    Same story · 20 reports

    #Industry news#OpenAI#Anthropic

  5. The Verge AI
    AI score63

    The New York Times and other publishers sue Brave browser for scraping content to train AI and bypassing paywalls

    The New York Times and several other publishers have filed a lawsuit against the company behind the Brave browser, accusing it of scraping content without permission to train AI and of letting users bypass paywalls. The complaint states that Brave never sought permission from publishers, and that its browser also advertises a "Paywall blocker" feature that provides access to "hundreds of top publishers." Perplexity Chief Communications Officer Jesse Dwyer responded to The Verge, saying that The New York Times spends too much on lawyers and too little on journalists.

    Same story · 3 reports

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

  6. NVIDIA Developer · Generative AIFirst-party
    AI score41

    NVIDIA AICR brings Day 1 support for Vera Rubin to Kubernetes

    NVIDIA AI Cluster Runtime (AICR) uses reusable recipes to codify tuning configurations, generate deployment packages, and validate running clusters, helping Kubernetes environments support Vera Rubin NVL72 from Day 1. The Vera Rubin recipe series locks in configurations including the R615.23 open-source kernel module driver, DRA driver v0.5.0 (with ComputeDomains), and the 64k-page kernel of Ubuntu 26.04; end-to-end validation was completed on a reference cluster with GPU Operator v26.7.0, where cross-node NCCL all-reduce reached 744.71 GB/s bus bandwidth at 16 GiB. Early chip validation also found that go-nvlib did not map the Rubin architecture value, causing the DRA driver kubelet plugin to crash and restart; the fix has been included in the official release of DRA driver v0.5.0.

    #Product updates#English text is being prepared.#Open-source ecosystem

  7. English text is being prepared.
    AI score40

    Zeng Yuqun: CATL Uses AI to Screen Thousands of Compounds, Identifying 7 Potential Materials

    Zeng Yuqun, founder of CATL, said at the 2026 World Laureates Forum that CATL is also using AI for technological innovation. Based on over 600TB of basic research data and over 1 trillion test records, AI can screen thousands of compounds in a short time and identify 7 potential materials. He also said that current AI capabilities are still far from sufficient and that more scientists and more advanced models are needed.

    #Industry news#English text is being prepared.

  8. English text is being prepared.
    AI score53

    Abu Dhabi and OpenAI sign four agreements covering government, energy, and investment sectors

    Relevant Abu Dhabi institutions and OpenAI signed four agreements involving the government, energy, and investment sectors. On energy, ADNOC became the first company in the energy sector to deploy OpenAI's frontier technology, with the collaboration based on GPT models and compute provided by inference infrastructure within the UAE; on government, Abu Dhabi and OpenAI launched the AI-native Machine-Readable Government (MRG) initiative, with the first pilot already in production at the Abu Dhabi Agriculture and Food Safety Authority, focusing on food imports. On investment, Mubadala, Inception42, and OpenAI are collaborating to apply AI to investment analysis, opportunity identification, and risk assessment, and Inception42 separately signed a five-year agreement with OpenAI to promote the adoption of frontier AI in government, enterprise, and education.

    Same story · 6 reports

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

  9. English text is being prepared.Product release
    AI score54

    Alibaba Cloud Releases Qovest, an AI Investment Research Workbench for Professional Institutional Investors

    Alibaba Cloud has released Qovest, an AI investment research workbench for professional institutional investors. After users submit a research request in a single sentence, it can automatically query data, collect materials, analyze and generate reports, and continuously track research subjects at a set frequency. Qovest comes with built-in authoritative financial databases and the self-developed Harness, an Agent engineering framework for financial scenarios; research results are accumulated in the investment research space so work can continue from there. The product is now open for invitation-based testing to individual users, who receive credits for a free trial after registering, with optional subscription plans of 99 yuan or 199 yuan per month thereafter.

    Same story · 4 reports

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

  10. English text is being prepared.Product release
    AI score55

    Alibaba Cloud Memory Service Upgrade: Enhancing Long-Term Memory and Personalization for Personal Agents

    Alibaba Cloud announced an upgrade to its Memory service, enhancing the ability to remember user preferences, plans, and historical agreements for Personal Agent scenarios, and providing access methods such as MCP and CLI. Memory extracts key information from conversations and saves it as memories and user profiles according to rules; in subsequent sessions, it retrieves information based on user identity and the current question and adds it to the task context. In the HaluMem-medium comparison, Memory achieved the highest scores in all four metrics for memory extraction: recall, accuracy, F1, and the memory question-answering C metric; in the PersonaMem-v2 evaluation, it ranked first among the 8 systems listed with an overall score of 56.82.

    Same story · 2 reports

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

  11. IT Home · Artificial Intelligence
    AI score62

    Two Brazilian lawyers manipulated court AI with hidden prompts and were fined $16,000

    A Brazilian labor court found that two lawyers embedded prompt injection instructions in white-on-white text in the complaints they submitted, requiring the AI to perform only superficial processing and not question the content of the documents. On this basis, the court fined the two 10% of the compensation amount involved in the case, about $16,000. The court determined that the conduct undermined judicial dignity and reported the two to the state bar association and the internal affairs department of the regional labor court. Because the plaintiff did not respond to the allegations, the plaintiff won by default, and this injection did not actually take effect; previously, another plaintiff was required to submit paper materials offline thereafter because of embedding hidden prompts.

    #Safety / alignment#English text is being prepared.

  12. IT Home · Artificial Intelligence
    AI score49

    Researchers test AI agent swarms in Minecraft: collaboration tends toward bureaucracy, GPT has a higher win rate against Claude

    Researchers from Y Combinator and Olam Labs used hundreds of agents in Minecraft to test collaboration and competition. 50 GPT-6.1 Sol agents failed to build a civilized settlement within three hours, and four agents even displayed bureaucratic behavior by spending 11 minutes appointing a "sole fertilizer operator." In Mindustry tower-defense matches, the GPT agents defeated Claude Opus 5.5 agents six wins to one loss, and the project has been open-sourced.

    Same story · 3 reports

    #Papers/Research#Agents#Open-source ecosystem

  13. English text is being prepared.
    AI score47

    Turkish Competition Authority investigates Google Search AI Overviews and AI Mode

    The Turkish Competition Authority has launched an investigation to examine whether Google Search's "AI Overviews" and "AI Mode" features violate the country's Law on the Protection of Competition, and to review whether its content-use agreements with news publishers contain unfair terms. The authority said existing indications suggest these AI features are highly suspected of exclusionary and exploitative conduct toward news publishers, and it has imposed interim regulatory measures on Google to regulate its use of news publishers' content in AI-driven search.

    #English text is being prepared.#Google#Search

  14. English text is being prepared.
    AI score40

    IDC predicts that by 2028, enterprises will run 60% of AI workloads on open-source models

    IDC predicts that by 2028, enterprises will run 60% of their AI workloads on open-source models, using model routing strategies to reduce costs and hedge against vendor lock-in risk. Among G2000 enterprises, 60% will have most of their business processes running AI-first by 2029, compared with only 13% today; by 2028, 70% will have the ability to continuously validate AI agent decisions. G2000 enterprises that do not set up an emergency stop switch for agents in production environments will face 25% higher liability insurance premiums and double the exposure to major audit issues by 2028.

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

  15. IT Home · Artificial Intelligence
    AI score40

    Wharton Professor Mollick: Switching to Skilled Trades Won't Withstand the AI Shock; Four Core Capabilities Are the Key to Breaking Through

    Wharton School professor Ethan Mollick believes that relying on "learning to use AI" or switching to hands-on skilled trades to withstand the AI shock is of little use, because the robotics research community has already realized that there is a possibility of connecting large language models with robots. He points out that the apprenticeship training chain has already broken: newcomers directly throw requirements at Claude and submit the results as-is, yet lack the ability to judge whether the output is good or bad. The way out is to strengthen formal vocational education, or to have companies treat mentoring as a serious job and implement it properly, avoiding recruiting only seasoned veterans, which leads to a break in the talent pipeline.

    #English text is being prepared.#Industry news

  16. IT Home · Artificial Intelligence
    AI score62

    Cisco Talos discovers trojan VOLTA, which uses AI to analyze whether a victim is worth attacking

    The cybersecurity team Cisco Talos has discovered a new type of Windows malicious trojan named VOLTA. It collects data such as browsing history, cryptocurrency wallets, browser extensions, text documents, and chat records, and uploads them to online AI products such as Anthropic Claude to generate summaries, so that hackers can assess the value of the victim's information. After completing data collection and value assessment, VOLTA attempts to delete itself to reduce traces. Cisco Talos believes that this type of malware shows that AI is changing the way hackers operate, and may expose large institutions and executives to more precise attacks. Ordinary users' browsing history, spending habits, and social relationships may also be used for profiling analysis.

    Same story · 4 reports

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

  17. English text is being prepared.Product release
    AI score60

    Tencent Hunyuan releases AI research system Hyra 2.0, shifting toward having Agents develop reusable tools

    Tencent Hunyuan releases AI research system Hyra 2.0, with the R&D focus shifting from directly using AI to tackle scientific problems toward having general-purpose Agents participate in the development and upgrading of scientific models, research workflows, and software tools, so that the results can be reused repeatedly in subsequent research. Architecturally, Hyra 2.0 splits research work into three layers: Overall, Mission, and Worker; Workers can independently read code, run experiments, analyze the causes of failures, and question Mission's judgments. The system also designs three separate storage systems: project state, Agent layered memory, and a shared knowledge base called Books. During retrospectives, transferable methods can be organized into skills for use in subsequent tasks.

    Same story · 3 reports

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

  18. Google BlogFirst-party
    AI score35

    Google's October AI Update: Gemini 4 Argon Launches Globally, Gemini 4…

    In October, Google launched its frontier reasoning model Gemini 4 Argon globally, covering Google AI Studio, the Gemini API, Google Cloud, Search, and the Gemini app. Two weeks later, it announced Gemini 4 Carbon, focused on long-horizon engineering and cybersecurity defense, which is currently undergoing pre-release testing with the U.S. government and select trusted cybersecurity parties through the Fairwind Program. In the same period, it released Genie 4, a real-time interactive world model that supports native audio and natural-language scene control; SynthID Detector opened globally in English and can identify AI content from partners such as OpenAI, NVIDIA, Kakao, and Apple.

    Text with a rainbow colored gradient "The big picture"

    #Industry news#Model releases#Multimodal

  19. ElevenLabs BlogFirst-party
    AI score23

    How KPN and ElevenLabs Serve Entrepreneurs with Mobiel AI

    KPN, the largest telecom operator in the Netherlands, has partnered with ElevenLabs to bring voice AI into Mobiel AI, an assistant for small and medium-sized business owners, offering features such as voicemail transcription, call summaries and automated follow-ups, and incoming call menu routing. By supplementing Dutch-language data and fine-tuning the transcription model, the two parties raised validation accuracy from about 50% to 90%. Mobiel AI runs in KPN's own data centers, voice data stays in the Netherlands, and customer data is not used to train models.

    Jessica Day Headshot

    #Industry news#Speech#Agents