NAll news

All news

Monday595 items

  1. Ayi AI NotesOpinion
    AI score52

    Terence Tao criticizes the way AI proves mathematics: automated, predatory mining

    In a recent interview, Terence Tao criticized the way AI proves mathematics, describing the chaos surrounding the Navier-Stokes equations as a kind of automated, predatory exploitation and mining. He believes mathematicians want AI to serve as an on-call supercalculator, not as a competitor fighting over results; the value of humans solving difficult problems lies in refining conceptual systems and theoretical tools through long-term derivation, whereas AI, relying on computing power and black-box logic to directly take the answer, leaves behind theoretical ruins that are hard to understand and impossible to build on. He emphasized that a correct answer does not equal scientific progress.

    Same story · 2 reports

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

  3. Philipp Schmid
    AI score44

    Philipp Schmid on Agent Tools Converging into Code Mode

    Philipp Schmid raises the question: will agents soon be left with only one tool? He notes that the industry has gone from a small set of tools, to dynamic tool search, and now to converging everything into Code Mode. In his article, he explains how Code Mode works, why typed JS/Python sandboxes beat bare bash, and why MCP will make a strong comeback in 2027, along with a minimal example using Gemini 4 Argon.

    Same story · 2 reports

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

  4. Rohan PaulResearch
    AI score55

    Meta paper proposes TextSeal2: merging dual keys improves AI text watermark detection rate

    Meta's new paper proposes TextSeal2, which fuses two keys into one virtual key and adds new noise, so that every token carries the complete watermark signal while maintaining answer diversity. For speculative decoding, it selects replacement tokens from the remaining candidates of the draft's random competition, keeping the acceptance rate unchanged. On Qwen3.5-9B with speculative decoding enabled and diversity held equal, TextSeal2 achieves a detection rate of 89% on 200-token responses at a 0.1% false positive rate, compared with 71% for TextSeal and 67% for textGrain.

    #Papers/Research#English text is being prepared.#Meta

  7. Import AIOpinion
    AI score54

    Import AI 479: Realtime-Bench reveals LLM shortcomings on real-time tasks, DIA space security report, and…

    Import AI issue 479 introduces Realtime-Bench, a benchmark comprising 50 interactive games and web tasks, which shows that GPT-6 Astra (high) achieves the highest pass rate under real-clock conditions at only 36%, while with the clock paused it can approach 70% to 80%. This issue also excerpts content from the U.S. Defense Intelligence Agency's "Challenges to Security in Space" report on Chinese and Russian anti-satellite capabilities, space-based nuclear weapons, and on-orbit confrontation, and reviews features of Google DeepMind's Genie 4, including real-time modification, physics and character simulation, native audio generation, and the use of Gemini 3.8 Flash as a guiding agent.

    #Papers/Research#Evaluations / benchmarks#Agents

  8. Cailian Press
    AI score31

    Marvell Technology announces $115 billion all-stock acquisition of Lumentum, betting on AI optical interconnect

    Marvell Technology announced an all-stock acquisition of optical communications component manufacturer Lumentum, with an implied enterprise value of approximately $115 billion. Lumentum shareholders will receive 4.5 shares of Marvell Technology common stock per share, a premium of approximately 23.9% over its October 28 closing price. After the merger, Marvell Technology shareholders will hold approximately 67% of the combined company. The transaction is expected to close in the first half of 2028, and the two parties expect to achieve more than $1 billion in annual cost synergies. Through this deal, Marvell Technology moves from selling chips to vertical integration of "chips + lasers," moving closer to the Broadcom model, but it may also compete directly with optical module manufacturers that are its DSP chip customers.

    Same story · 7 reports

    #Industry news#English text is being prepared.

  9. Apple Machine Learning ResearchFirst-party
    AI score46

    Apple proposes an online dense reinforcement learning framework: a planning-grounding agent navigating real iOS applications

    Apple proposes an online reinforcement learning framework that directly trains a UI navigation agent on real iOS applications, splitting the agent into a Planning agent and a UI Grounding model, and uses an LLM-as-a-Judge to provide step-level dense rewards and task-completion terminal rewards, optimizing with PPO plus step-level advantage estimation. Its 8B agent achieves nearly a 7x improvement over the SFT baseline when generalizing out-of-distribution to the Android environment, and on real iOS applications surpasses Gemini-3.5-Flash and matches Gemini-3.1-Pro.

    Bottom banner

    #Papers/Research#Agents#Reasoning

  10. Sina Tech
    AI score77

    Anthropic's third-quarter revenue was $16.96 billion, with Claude business revenue at $17 billion

    The S-1/A filing released by Anthropic shows that revenue for the three months ended September 30, 2026 was $16.96 billion, a year-over-year increase of 1127%, compared with $1.38 billion in the same period last year, of which Claude business revenue reached $17 billion. The net loss in the third quarter was $23.3 billion, compared with $12.1 billion in the same period last year; adjusted operating profit was $2.5 billion with a margin of 15%, and the operating loss on an accounting basis was $567 million. As of September 30, total contractual obligations were $320 billion, and there were more than 11,900 customers with annualized revenue of at least $100,000. The company also committed to not placing ads in Claude.

    Same story · 18 reports

    #Industry news#Anthropic

  11. Sina Tech
    AI score78

    Anthropic's third-quarter revenue neared $17 billion, 12 times the same period last year

    As of September 30, Anthropic's third-quarter revenue was close to $17 billion, about 12 times the same period last year. Its net loss for the quarter was $23.3 billion, and operating expenses rose to $17.5 billion, up 292% year over year. The company filed for an IPO on October 26 and is expected to go public before the U.S. Thanksgiving holiday, seeking to match or exceed the $86.2 billion raised in SpaceX's listing. The filing also shows $895 billion in committed spending, of which about $448 billion is expected to occur before the end of 2030.

    Same story · 4 reports

    #Industry news#Anthropic

  12. Leiphone
    AI score48

    Li Jiaqi sells Qwen Office Tokens in livestream: 100,000 monthly cards sold in half an hour, about 2 trillion Tokens

    On the evening of November 2, Qwen Office Tokens went on sale in Li Jiaqi's livestream, selling 100,000 personal membership monthly cards in half an hour, totaling about 2 trillion Tokens. Qwen Office CEO Chen Yusen appeared in the livestream, demonstrating AI office scenarios such as organizing a child's homework and test papers, generating a business analysis report from livestream replays and bullet comments with one click, and generating a global fishing map through conversation. Li Jiaqi proposed on the spot making an AI clone for companion chat, and Chen Yusen responded that AI cannot replace Li Jiaqi, and that he would rather help fans use AI to solve trivial matters.

    Same story · 5 reports

    #Industry news#Agents

  13. Shanghai Observer
    AI score55

    Six AgiBot Genie G2 Humanoid Robots Work 24-Hour Shifts at a Shanghai Convenience Store

    Starting at 10:00 on November 2, six AgiBot Genie G2 humanoid robots clocked in at a 24-hour convenience store in Xuhui, Shanghai, taking on roles including greeter, barista, stocker, cashier, and cleaner. They worked continuously for a full day, with the whole process livestreamed for viewers across the internet. On-site experience officers set up challenges such as handing over the wrong-sugar cola, disarranging shelves, and human roadblocks, and the robots responded through voice recognition, visual inventory checks, and path replanning. The six robots share AgiBot's embodied Agent framework Genie Master as their scheduling and learning hub; each day they summarize successful and failed trajectories and store them in long-term memory, converting effective adjustments into training data.

    Same story · 5 reports

    #Industry news#Embodied AI#Agents

  14. Shanghai Observer
    AI score25

    Global Teacher Professional Development Forum discusses AI's impact on teacher professionalism: beware of education to the "AI power"

    At the Sixth Global Teacher Professional Development Forum held at Shanghai Normal University on November 2, multiple experts discussed AI's impact on teacher professionalism. Cheng Jieming criticized some schools for using AI cameras to count the number of times students raise their hands as "very absurd." Wu Di's research pointed out that AI not only reduces teachers' transactional burdens, but also dissolves knowledge authority, erodes professional judgment, and weakens teacher-student emotional interaction, calling for "tailored training" for teachers. Zhou Zuoyu cited a BBC survey saying that teachers were listed as the profession most affected by AI, but the unemployment rate was only 0.4%.

    #English text is being prepared.#Industry news

  15. TechCrunch AI
    AI score37

    Omen AI acquires Aquatic Labs for $16 million to complete its water quality sensor technology

    Omen AI has acquired competitor Aquatic Labs for $16 million in cash and stock, combining the latter's alkalinity detection technology with its own elemental detection sensors into a real-time water quality sensor expected to launch early next year. Omen just closed a $31 million Series A round in June this year; its sensors will monitor chips with 250 to 300 MW of power by the end of the year, with a goal of covering the equivalent of 1 GW of chips by the end of 2027. U.S. data centers use 17 to 19 billion gallons of water each year to cool chips, and real-time monitoring can extend the service life of coolant and reduce downtime and water replacement costs.

    Same story · 2 reports

    #Industry news#English text is being prepared.

  16. 36Kr · AI
    AI score29

    The Seven "Not A, But B" Points in Tencent Cloud ADP Course Materials: A Checklist of Trade-offs for Enterprises Deploying AI Agents

    Course materials for Tencent Cloud ADP's Forward Deployed Engineer (FDE) program summarize the key trade-offs for enterprises deploying agents using seven "not A, but B" formulations: the first step is not to open the platform and build an application, but rather to first determine whether the business problem can be handled reliably by AI, which ADP mode should carry it, and what results will prove it effective. Enterprise-grade Agents need controllability, testability, auditability, operability, reusability, and integrability; actions such as deleting customer information must wait for human approval. The materials also emphasize that agents gradually become stable through debugging, optimization, evaluation, and regression, and that going live is not the endpoint but the start of continuous operations.

    #Tutorials / practice#Agents#RAG

  17. 36Kr · AIOpinion
    AI score67

    Bernstein Statistics: Only About 35% of the 488GW of U.S. Data Center Plans Is Likely to Materialize

    Bernstein statistics show that the cumulative planned capacity of announced data centers in the United States has reached 488GW, but only about 170GW can be confidently realized, accounting for about 35%. BloombergNEF statistics in July said that among the 100 largest data center plans in the United States, 52 come from developers without construction experience, totaling 87GW; Goldman Sachs research in May estimated that the on-time delivery ratio of planned capacity in the next one year and two years may drop to 60% and 50%, respectively. On the financing side, conditions tightened simultaneously. From October 6 to 20, the combined market value of eight AI infrastructure companies, including CoreWeave, Nebius, and IREN, evaporated by about $51 billion, with a median stock price decline of 28%.

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

  18. 36Kr · AIResearch
    AI score50

    Meta FAIR proposes routed continual learning architecture Routed CL, API costs can be reduced by 65% to 72%

    The Meta FAIR team proposes a routed continual learning architecture, Routed CL, which groups historical trajectories into different clusters and maintains memory for each, with an LLM-based router deciding whether to reuse an existing cluster or create a new one. On 301 and 97 instances from the mixed CL-Bench and τ³-Bench, this architecture improves normalized reward relative to Global ICL by an amount equivalent to 15% to 30% of the maximum achievable value, with estimated API cost reductions of 65% to 72%. The authors also note that concurrent updates, reading old-version states, and different models reusing the same cluster have not yet been evaluated.

    #Papers/Research#Agents#English text is being prepared.

  19. 36Kr · AIOpinion
    AI score54

    NVIDIA, Google DeepMind, and Intel Make Successive Moves as Physical AI Competition Shifts to Infrastructure

    In late October 2026, NVIDIA expanded its Isaac robot development platform, Google DeepMind released the Genie 4 world model, and Intel and SAP demonstrated an autonomous warehouse proof of concept. Together, these three developments point to competition in physical AI shifting from model intelligence to infrastructure. NVIDIA's Isaac covers data, simulation training, runtime, and edge computing; Google attempts to use Gemini to break down tasks and have Genie 4 generate interactive simulated environments in real time; and Intel and SAP dispatch warehouse business tasks to robots and receive execution status back. The article argues that the key to deploying physical AI lies in embedding intelligent capabilities into enterprise business processes, forming a closed loop from business decisions to result feedback.

    #Industry news#Embodied AI#Google

  20. 36Kr · AI
    AI score37

    Meta and Fourier's Max Humanoid Robot: Just an ODM Collaboration for a Small Number of Prototypes

    Meta and the Chinese robotics company Fourier jointly created a humanoid robot prototype named Max within three months, but multiple industry insiders say this is not a large commercial order or long-term joint development—it is merely Fourier supplying a small number of prototypes to Meta and AWS under an ODM model. Max stands about 1.5 meters tall and weighs 48 kilograms, is equipped with NVIDIA's Jetson Thor chip, and features an expressive display screen and five-fingered dexterous hands. It is currently used only for internal experiments, with no mass-production timeline.

    #Industry news#Embodied AI