NAll news

All news

Tuesday792 items

  1. Ayi AI Notes
    AI score36

    Claude Code Engineer Warns: Over-Reliance on AI Programming Will Weaken Code Review Skills

    Anthropic's Claude Code engineer Addy Osmani has proposed the "cognitive surrender" paradox: the more coding tasks engineers hand over to agents, the faster the underlying hard skills they need to supervise those agents atrophy. He points out that AI-generated code is neatly formatted and has all unit tests passing green, making it easy for engineers to regress from reasoning through the architecture line by line to mechanically clicking confirm, and says that shipping something beyond one's own cognitive grasp no longer counts as code review.

    Quoted postAddy Osmani · @addyosmani

    https://t.co/2gyp2Erpxx

    Same story · 5 reports

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

  3. Dex Horthy
    AI score10

    dex responds to Jake Nations' prompt style

    Jake Nations shared his commonly used prompt style, such as "how can we do this lock free?" and "how can we make this more data oriented or functional?". dex quoted this post and replied with just one line, "> plz no more nested ternaries", which is a request to stop writing nested ternary expressions.

    Quoted postJake Nations · @nayshins

    This is what most of my prompts look like. > how can we do this lock free? > how can we make this more data oriented or functional?

    #English text is being prepared.

  5. Peter YangProduct release
    AI score62

    Grok @Bot Can Call U.S. Businesses on Users' Behalf; Real-World Test Negotiating with Comcast Saves Nearly $500 a Year

    Grok @Bot can now call U.S. businesses on users' behalf. Peter Yang used it in a real-world test to negotiate with Comcast, cutting his broadband bill by nearly $500 a year. In the video, the bot responded to the customer service agent's small-talk questions, declined 3 upsell attempts, and used a competitor's lower quote as leverage. He called this feature one of his favorite AI features.

    Same story · 3 reports

    #Product updates#Agents#Speech

  6. 36Kr · AIOpinion
    AI score61

    MIT experiment claims GPT-6 Astra rewrote Qwen weights to boost scores; Meta researcher says it is exaggerated

    An MIT joint-team experiment claims that GPT-6 Astra, without training data and without running backpropagation, changed only about 130,000 weights of Qwen3.5-0.8B (0.016% of its 800 million parameters), raising the model's accuracy on MATH-500 from 7.9% to 48.1%. Meta researcher Lucas Beyer pointed out that the system prompt already provided targeted guidance, and researcher Yunfan Zhang, after reviewing the logs, concluded that Astra merely identified that Qwen's reasoning chain had fallen into an infinite loop and performed a targeted clearing. The article also states that of the 579 additional questions answered correctly after the modification, 576 had originally scored zero due to timeout truncation, and that when weights were repeatedly modified on Gemma 3 270M, which does not have this defect, the score remained between 6.8% and 7.7%.

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

  7. 36Kr · AIProduct release
    AI score73

    Hands-on with DeepSeek Harness Creation Mode: Generating Plugins with Natural Language

    DeepSeek released DSH Creation Mode on November 2. Users describe their ideas in natural language, and the Agent writes the code, packages it, and installs it as a plugin that DSH can use directly. The company also released 7 demos along with Chinese and English prompts. In the author's hands-on test, prompts of about 100 characters were used to create two plugins—a 3D Mario mini-game and a desktop pet—at a total cost of about 6 yuan, with a 99% cache hit rate, though the generation process required multiple clicks on "Allow." DeepSeek acknowledges that some demos require multiple rounds of refinement, that the project is still in developer preview, that core plugins and APIs may undergo breaking changes, and that there is currently no official plugin marketplace.

    Same story · 5 reports

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

  8. 36Kr · AIOpinion
    AI score55

    With annual tuition as high as $75,000 and only two hours of study per day, can Alpha School's AI school model work?

    Duojing, under 36Kr, reviewed the AI teaching model of the U.S. private school Alpha School: students spend about two hours in the morning learning core subjects with the TimeBack adaptive system, and in the afternoon take part in activities such as sports, public speaking, and programming; annual tuition at most campuses is about $40,000 to $75,000. The article cites an October 23 investigation by The New York Times saying that, in internal statistics, about 42% of students simultaneously met the school's advertised standard of double growth in both math and reading, a gap from the high-efficiency image emphasized externally. The article also mentions that the school has expanded into Texas and other states, but an affiliated Arizona charter school's first-year English and math proficiency levels were actually 28% and 10%, below the 65% and 60% projected at the time of application.

    #English text is being prepared.#Industry news#Tutorials / practice

  9. 36Kr · AIOpinion
    AI score51

    AI companion app Zeta's monthly revenue tops $10 million, charging points per turn on top of membership

    The estimated monthly revenue of Zeta, an AI companion app owned by South Korea's Scatter Lab, exceeded $10 million in August 2026, with its charges consisting of two parts: the Zeta Pass membership and Pieces points consumed per turn. The membership offers benefits such as ad removal and faster replies, along with 200 bonus points, while conversations with advanced models and edited replies are billed separately in points; on August 7, Koji rose from 4 to 5 points per turn and Luca from 8 to 10 points. At the end of October, Zeta moved subscriptions to web sales, with the Japanese web price of 1,320 yen being 20% lower than the in-app price of 1,650 yen, and added automatic point top-ups.

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

  10. 36Kr · AIOpinion
    AI score60

    After the funding cooldown, how can robotics companies get past the payment collection hurdle?

    A 36Kr republished article analyzes the payment collection pressure on robotics companies after the funding cooldown: in September 2026, China's embodied intelligence primary market saw 46 publicly disclosed financing deals totaling about 11.624 billion yuan, down 40.1% from August. The article takes UBTech as an example: its revenue in the first half of 2026 was 1.269 billion yuan, up 104.2% year on year, while net cash outflow from operating activities was 749 million yuan, about twice that of the same period last year; Agility's net sales in the same period were about $174,400, with operating cash consumption of about $86.2 million. The article also discusses how acceptance conditions, on-site engineer input, and RaaS service fees lengthen the capital recovery cycle.

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

  11. 36Kr · AIResearch
    AI score76

    MIT CSAIL proposes weight-space programming; GPT-6 Astra rewrites 0.016% of weights to make…

    The MIT CSAIL team proposes "weight-space programming," letting GPT-6 Astra directly read and rewrite neural network weights without training or computing gradients. It changes only about 0.016% of Qwen3.5-0.8B's weights (131,072 numbers), raising MATH-500 accuracy from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval instruction following from 21.0% to 62.3%. The authors used Gemma 3 270M as a control, whose MATH-500 only moved slightly from 6.8% to 7.0%, indicating that the method mainly excises weight directions that cause infinite loops rather than injecting new knowledge out of thin air.

    Same story · 14 reports

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

  12. 36Kr · AIResearch
    AI score56

    Tsinghua Team Publishes Study in Nature Communications: Thousands of Agents Form Networks and Spontaneously Polarize in Opinion

    A team from Tsinghua University, together with institutions including the University of Amsterdam, the University of Chicago, and the Santa Fe Institute, published a study in Nature Communications, building a networked system of 1,000 large language model agents and comparing five backbone models: GPT-3.5, GPT-4o, ChatGLM, Llama-3, and DeepSeek-V3. They found that after fewer than 10 rounds of interaction, the share of neutral opinions fell from 40% to 22.5%, 0.4%, and 5.1%, respectively, and the agents spontaneously split into two poles. The study also found that agents exhibit inconsistency between their opinions and their expressions, and proposed a self-regulation strategy that reduced the inconsistency rate by 9.4%–52.2%; the proportion of interactions within the same camp increased by 156.8%–382.7%, and ultimately 48.5%–88.3% of interactions occurred between agents with similar opinions. The team used this system as a digital experimental testbed to examine five types of intervention strategies, among which eliminating confirmation bias and neutral elite signals reduced the level of polarization by 11.8% and 8.8%, respectively.

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

  13. 36Kr · AIResearch
    AI score56

    arXiv paper A Heart Means Recursion: Show the model a red heart, and it switches from iteration to recursion

    The arXiv paper A Heart Means Recursion documents a phenomenon: when models solve binary tree traversal problems, they almost always use stack-based iteration; after a red heart ❤️ is attached next to the problem, they switch to recursion, but when asked why, they say the red heart is irrelevant. The researchers call this disconnect between symbolic interpretation and actual behavior implicit guidance, and record that GPT-5.4 showed unfaithful attribution in 74.7% of samples. The article also connects Kahneman's anchoring effect, CLIP's typographic attack, and other cases, discussing whether Agents that perceive pages via screenshots may have their decisions quietly changed by elements such as ad images and icons.

    #Papers/Research#Reasoning#Multimodal

  14. 36Kr · AI
    AI score46

    Why have memory stocks become the first AI assets to be abandoned?

    As of October 30, 11 memory companies had pulled back an average of 51.6% from their intra-year highs, far exceeding the 33.9% for cloud providers and 44.0% for optical communications; A-share companies Biwin Storage, Demingli, and Longsys saw maximum declines of over 70%. Although Micron's net profit for fiscal year 2026 reached $84.969 billion, up nearly 9x year-over-year, the quarter-over-quarter increase in DRAM contract prices has narrowed from 58%–63% in the second quarter to 10%–15% in the fourth quarter. Module makers rely on low-priced inventory to earn price spreads; Longsys, Biwin Storage, and Demingli all saw their gross margins decline in the third quarter, with combined inventory of approximately 81.5 billion yuan, up 206% from the end of 2025.

    #Industry news#English text is being prepared.

  15. 36Kr · AI
    AI score61

    Marvell plans $115 billion acquisition of Lumentum, China's optical communications faces new test

    Marvell announced a planned all-stock acquisition of optical communications vendor Lumentum, corresponding to an enterprise value of approximately $115 billion. The deal is expected to close in the first half of 2028 and still requires shareholder and regulatory approval. The article notes that as chip performance improves and cluster scale expands, optical connections are moving closer to the chip, and solutions such as CPO co-packaged optics may change how customers make purchases. Zhongji Innolight already has 800G and 1.6T high-speed optical modules, and TFC Optical Communication disclosed progress on mass production of 1.6T optical engines and development of CPO supporting optical components, but the acquisition itself cannot be directly regarded as evidence of increased orders for Chinese vendors.

    Same story · 9 reports

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

  17. The Decoder
    AI score60

    OpenAI Hires Former U.S. Army Secretary Dan Driscoll as Senior Advisor for National Security Modernization

    OpenAI has hired former U.S. Army Secretary Dan Driscoll for the newly created role of "Senior Advisor for National Security Modernization," where he will work with government, industry, academia, and U.S. allies to use AI to strengthen national security and cyber defense. Driscoll served as the 26th Secretary of the Army, overseeing nearly one million soldiers and more than 265,000 civilian employees, and during his tenure pushed AI modernization projects including Project ARIA. The report says he is a close ally of Vice President JD Vance and maintains good relations with both parties in Congress; previously, in late February 2026, the Pentagon designated Anthropic as a "supply chain risk," and on the same day Sam Altman announced an agreement with the Pentagon to deploy models on classified networks.

    Same story · 7 reports

    #Industry news#OpenAI#Anthropic

  19. Leiphone
    AI score25

    AgiBot Receives Outstanding Contribution Award at 2026 World Internet Conference, Proposes Worken to Measure the Value of Physical AI

    AgiBot was presented with the Outstanding Contribution Award at the 2026 World Internet Conference Wuzhen Summit, in recognition of its systematic technological breakthroughs and industrialization practices in the field of embodied intelligence. AgiBot proposed Worken as a yardstick for measuring a robot's effective intelligent working hours, which is jointly determined by the number of active units, autonomous duration, and output efficiency ratio, and has driven the deployment of models such as Genie G2, Lingxi X2, and Expedition A3 in factories, stores, hotels, and cultural tourism scenarios.

    Same story · 3 reports

    #Industry news#Embodied AI

  20. Phoenix Technology
    AI score38

    Gartner Releases Top 10 Technology Trends for 2027: Non-Text AI Tops the List, Harness Is the Focus for the Next 1-3 Years

    Gartner released its "Top 10 Strategic Technology Trends for 2027," listing large quantitative models, speech reasoning models, world models, AI context platforms, and coding agent harnesses among the key directions for enterprises in 2027, with the expansion of non-text AI capabilities ranking first. Gao Ting, Vice President of Research at Gartner, said that the coding agent harness (Harness) is the focus for the next one to three years, but its importance may gradually diminish as model capabilities improve. Gartner predicts that by 2030, more than 70% of enterprise voice interactions will be handled autonomously by speech models, compared with less than 10% in 2026.

    Same story · 3 reports

    #Industry news#Agents#Reasoning