NAll news

All news

Tuesday64 items

  1. SolidotOpinion
    AI score56

    IBM CEO Krishna Says Open-Weight Models Perform Well and Cost Little

    IBM CEO Arvind Krishna said the performance gap between open-weight models and frontier models has narrowed from 70 points versus 20 or 30 points in 2024 to 70 points versus 50 points by late 2025. He noted that the token cost of current open-weight models is only 10% to 25% of that of frontier models, while their effectiveness reaches 90% of frontier models. He believes most enterprises will therefore choose open-weight models for the vast majority of routine tasks and use frontier models only for a small number of tasks.

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

  2. TMTPost · AGIOpinion
    AI score60

    Anthropic's Silicon Valley Story: From Leaving OpenAI to Filing for an IPO

    TMTPost reviews Anthropic's five-year journey: in 2021, Dario and Daniela Amodei and other core researchers left OpenAI to found Anthropic; in 2022, the $580 million in early funding provided by FTX founder SBF was thrown into doubt by FTX's collapse; in 2023, Amazon invested up to $4 billion and Google invested $2 billion. The article states that Anthropic filed a prospectus with the SEC on October 26 to list on Nasdaq under the ticker ANT, with a book loss of $80.748 billion in the first half of 2026, Non-GAAP operating profit of $1.618 billion, and cumulative long-term compute arrangements of approximately $637 billion.

    Same story · 127 reports

    #Industry news#Anthropic#OpenAI

  3. TMTPost · AGIOpinion
    AI score72

    Jensen Huang Voices Support, Musk Pours Cold Water: The Neocloud Compute Rental Model Faces a Test

    TMTPost sorted through Jensen Huang's and Musk's opposing judgments on Neocloud: Huang says compute demand is strong and the only bottleneck is working capital, while Musk predicts that more than 200 Neoclouds will undergo consolidation, with only a few ultimately surviving. The article lists Nvidia's backstop moves such as the $6.3 billion compute purchase agreement with CoreWeave and a $2 billion investment, and cites CoreWeave's 2025 revenue of $5.131 billion and net loss of $1.167 billion to illustrate its profitability pressure. McKinsey's calculations show that under pure equity financing, the B200 GPU rental business has an operating margin of only about 8%, and Gartner predicts that by 2030 more than 80% of Neoclouds will cease independent operations.

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

  4. TMTPost · AGIOpinion
    AI score64

    Kimi quota consumption too fast, Anthropic advocates folding sub-agents back into the main loop

    After the launch of K3.1 Pro, multiple users reported that quota consumption was too fast, and Moonshot AI announced in the early hours of October 26 that it would reduce unnecessary sub-agent calls and output length. TMTPost reviewed the viewpoints in Anthropic's article "Building Agents: When to Fold Sub-Agents Back into the Main Loop": multi-agent systems consume about 15 times the tokens of ordinary conversations and suffer from information decay and personality fragmentation, and loading sub-agents on demand as Skills can cut over 80% of system prompts. Anthropic has also open-sourced the single-agent-loop-kit tool, enabling models to execute tasks only in single-agent form.

    Same story · 5 reports

    #Industry news#Agents#Anthropic

  5. TMT Post · AGIOpinion
    AI score60

    Scoring high doesn't mean deliverable: AI leaderboards are undergoing a credit re-evaluation

    The TMT Post article points out that of the 295 hundred-point benchmarks tracked by BenchLM, 105 have already saturated, while Surge AI's newly launched sudo L7, GDP.xlsx, and DAYJOB real workplace evaluations have pulled flagship models back below the passing line: Claude Opus 5.5 has a pass rate of only 41.7% on sudo L7, and Gemini 4 Argon only 38.3% on GDP.xlsx. The article also notes that Surge AI itself primarily provides data annotation and expert evaluation services, making its conclusions highly consistent with its business model, and compares the multiple differences in testing conditions behind Chinese models leading in OpenRouter call volume but lagging on the new leaderboard.

    #English text is being prepared.#Evaluations / benchmarks#Agents

  6. The Verge AIOpinion
    AI score70

    How Data Center Backlash Became a Bipartisan Issue in the U.S. Midterm Elections

    The Verge examines the bipartisan backlash sparked by data centers in the 2026 U.S. midterm elections. In election campaigns across Texas, Florida, Pennsylvania, Ohio, Michigan and other states, candidates from both parties are distancing themselves from data centers. The report lists specific proposals put forward by candidates, including pausing development, eliminating tax incentives, and requiring data centers to bear their own electricity and water costs, and cites polling showing that 64% of surveyed voters in Pennsylvania and 45% in Florida believe data centers do more harm than good. The article also mentions that Leading the Future, backed by Marc Andreessen and Greg Brockman, has invested $45 million, while Anthropic-backed Public First Action says it has raised $80 million.

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

  7. Data Center KnowledgeOpinion
    AI score60

    The Power Challenges Behind Hyperscale AI Data Center Expansion

    Microsoft, Meta, Amazon and Google's data center investment has grown about tenfold over the past eight years, from a combined $64 billion in 2018 to an expected more than $700 billion this year. IIR data shows that more than 500 data center bans have been enacted across 42 U.S. states, and in the first quarter of 2026 alone, $130 billion in projects were rejected; meanwhile, the load of operating data centers in the United States has reached 53GW, with another 42GW approved or under construction. Because grid interconnection queues are lengthy, developers are turning to self-built or jointly built gas-fired power generation. Large gas turbines from GE Vernova, Mitsubishi Power and Siemens Energy are fully booked through 2029, with some orders pushed to 2031.

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

  8. East MoneyOpinion
    AI score55

    Goldman Sachs: AI Computing Power Capex Growth Is About to Peak; Under the Optimistic Scenario, the Broad Market Could Rise 30%

    Goldman Sachs Chief Economist Jan Hatzius said that investment growth by hyperscale data centers in AI is expected to peak in the coming weeks, and the year-over-year change in their investment as a share of U.S. GDP is expected to peak in the fourth quarter of 2026 at about 1.5%. Operators such as Amazon, Microsoft, Alphabet, Meta, and Oracle are expected to spend about $725 billion to $800 billion on AI infrastructure capex this year, up 64% from 2025. Goldman Sachs forecasts that under the optimistic scenario, AI-related stocks will rise 55% and the broad market will rise 30%; under the pessimistic scenario, the broad market will fall 20%; if the data center investment boom ends, the U.S. GDP level could decline by about 2 percentage points.

    Same story · 2 reports

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

  9. 36Kr · AIOpinion
    AI score60

    Nomura report: The ChatGPT moment for global robotics has not yet arrived

    The Nomura report argues that robotics is currently in a stage after the 2012 AlexNet moment but before GPT-2; the true ChatGPT moment will not be announced at a product launch, but will come only when the data flywheel, model capabilities, and hardware consistency all cross the critical threshold at the same time. The report proposes five operational metrics as thresholds: currently, robots can operate only a few hours before requiring human intervention, and operating accuracy in factory environments is about 60% while the industry requires more than 98%; it is expected to take three to five years to close the gap. On the data side, available data amounts to only about 100,000 hours, while the critical threshold is about 1 million hours; data collected by Tesla Optimus has just surpassed 500,000 hours.

    #Papers/Research#Embodied AI#English text is being prepared.

  10. 36Kr · AIOpinion
    AI score60

    Anthropic proposes folding sub-agents back into the main loop, responding to the controversy over Kimi K3.1 Pro quota consumption

    Anthropic published the article "Building Agents: When to Fold Sub-Agents Back into the Main Loop," stating that a single agent with skills often outperforms multiple sub-agents dividing labor in quality, cost, and latency, with far lower token consumption than the latter, and open-sourced the single-agent-loop-kit tool. The article notes that multi-agent systems consume about 15 times the tokens of ordinary conversations, and converting sub-agents into Skills can cut over 80% of system prompts. After Moonshot AI's K3.1 Pro launched on October 24, multiple users reported excessively fast quota consumption, and the company said in the early hours of October 26 that it would reduce unnecessary sub-agent calls and output length.

    Same story · 5 reports

    #Industry news#Agents#Anthropic

  11. 36Kr · AIOpinion
    AI score84

    Why Anthropic Doesn't Care About Individual Users, as Seen in Its Prospectus

    On October 26, Anthropic publicly filed its S-1 prospectus, planning to list on Nasdaq under the ticker ANT, with media-discussed valuation ranges of $1.5 trillion to $2 trillion. The prospectus shows that about 80% of its revenue comes from usage-based billing, while subscriptions account for only about 20% and individual subscriptions are not broken out separately, meaning individual users make up a very small share of its business landscape. Based on this, the article analyzes that Anthropic is essentially a To B company, with individual users mainly serving as the starting point for enterprise adoption and being the first to be rate-limited when compute is constrained.

    Same story · 127 reports

    #Industry news#Anthropic#OpenAI

  12. Shanghai ObserverOpinion
    AI score52

    How to avoid robots becoming mere decoration after dense deployment: AgiBot discloses operating accounts and deployment data

    Shanghai Observer reviewed the deployment of AgiBot humanoid robots over the past month in scenarios such as retail, cultural tourism, manufacturing, and healthcare. As of October 25, the Yuanzheng A3Ultra at the smart retail warehouse on the Hankou Jiangtan in Wuhan had been running for 24 days and completed 1,253 orders, with a grasping success rate exceeding 98%. The report noted that robots still require manual assistance in the early stage of deployment; at the Chimelong project site, the ratio of on-site personnel to robots is about 1 to 3, and the company hopes to gradually raise it to 1 to 10 or even 1 to 20. Human-robot collaboration, networking, model Token calls, and spare parts wear and tear are all included in the comprehensive cost. AgiBot Co-President Xiong Yan believes that it will still take 3 to 5 years of technological accumulation before robots can work smoothly completely without on-site debugging, and that the important value of dense deployment lies in salvaging deviation data from real scenarios.

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

  13. Tencent Cloud Developer CommunityOpinion
    AI score58

    A Panoramic View of 8 AI Application Design Patterns: What Really Holds You Back Is Never the Model

    Based on the team's implementation practice, the author categorizes the 8 AI application design patterns by function into three types: capability enhancement, reasoning and workflow, and organizational collaboration, and divides them into three tiers by usage frequency and implementation maturity. The first tier consists of Tool-Augmented, RAG Agent, and Memory-Augmented; the author believes that doing these three solidly is enough to beat 80% of competitors. Multi-Agent and Autonomous Loop belong to the high-end game, and their implementation is a disaster area of pitfalls. The article also proposes using a replaceability test to determine whether the layered contract holds, and emphasizes that the simpler the architecture, the better, and that Workflow should be emphasized over full autonomy.

    #Tutorials / practice#Agents#RAG

  14. WallstreetcnOpinion
    AI score56

    Shenwan Hongyuan: The Inflection Point in US AI Capital Expenditure Growth May Be Approaching

    Based on project-level microdata from data centers, Shenwan Hongyuan Macro estimates that the peak in US AI investment growth may occur in Q4 2026, with a downturn beginning in 2027—one quarter later than the market consensus. The report states that capital expenditure by the five major US cloud providers in 2026 may exceed $800 billion, with year-over-year growth of 94%, accounting for 2.5% of US GDP; growth may then fall to 40% and 10% in 2027 and 2028, respectively. The estimates use three shelving rates of 20%, 33%, and 45%, with investment within one year determined by the planned stock and projects under construction.

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

  15. WallstreetcnOpinion
    AI score79

    Why Anthropic Doesn't Care About Individual Users, as Seen from Its Prospectus

    After Anthropic filed its S-1 prospectus, its revenue structure shows that usage-based billing revenue accounts for about 80%, while subscription revenue is only about 20% and includes enterprise plans, with paying individual users making up an even smaller share. The prospectus states that the top 5% of enterprise users using Claude Code spend about 20 times per month on average compared with non-users, while compute for free users is counted under sales and marketing expenses. The article argues that individual users mainly serve the customer-acquisition role of spreading adoption to enterprises, while the company's resources remain organized around enterprise scenarios.

    Same story · 127 reports

    #Industry news#OpenAI#Anthropic

  16. Shanghai ObserverOpinion
    AI score51

    Why Did China's First Fully AI-Produced Theatrical Film "Sanxingdui: Future Past" Fail at the Box Office?

    China's first fully AI-produced feature-length theatrical film to receive a Dragon Seal, "Sanxingdui: Future Past," earned only 1.246 million yuan at the national box office in its first four days of release, with an average occupancy rate of less than 1%, far short of its ten-million-yuan budget. The article argues that AI performs well in grand scenes and artistic texture, but the characters' expressions are dull, the lip-syncing does not match, and the splicing of short shots causes narrative jumps, leading viewers to complain that it looks like a high-end game CG. The flood of free AI short films on short-video platforms has also made audiences unwilling to buy tickets for long-form AI videos. The Dragon Seal only addresses whether a film can be released in theaters; it does not address whether audiences are willing to pay.

    Same story · 17 reports

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

  17. 36Kr · AIOpinion
    AI score58

    Humanoid robot secondary shares dumped at a 50% discount, the venture capital circle begins to stampede

    Pencil News observed that in the humanoid robot sector, secondary shares are being dumped at a discount: a world-model unicorn is transferring secondary shares at 50% to 60% of their value, while another embodied intelligence company is transferring secondary shares at a valuation of 12 billion yuan—a roughly 45% discount to its 22 billion yuan asking price—yet still finds no buyers. The article also notes that regulators are paying more attention to corporate valuation and revenue quality, that some projects' private-market valuations have been cut by 30% to 50%, and that after Unitree Robotics went public, its share price fell about 55% from its peak. The author argues that this stampede reflects venture capital institutions' lack of independent judgment, and calls for conviction in bullish positions to be well-founded and for exits to be disciplined.

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

  18. 36Kr · AIOpinion
    AI score67

    After Meta, Big Tech Awaits a "Muse Moment"

    Meta launched its personal agent Muse in September, and Zuckerberg said on the third-quarter earnings call that Muse would advance commercialization earlier than previous new products, seeking to charge merchants not only subscription fees but also for transactions it facilitates. The article says Muse's most common use case is shopping, with more than a quarter of users completing purchases through it, an average order value of over $200, and about 25% of users who have set up Muse using it daily. The article also reviews moves by Chinese big tech companies, including shutting down the creation entry points for Tencent Yuanqi and Baidu's Wenxin Agent Platform, arguing that they are shifting from rolling out agent platforms to betting on personal agents tied to high-frequency scenarios.

    #Industry news#Agents#OpenAI

  19. HuxiuOpinion
    AI score69

    Musk predicts more than 200 new cloud providers will die en masse; Huxiu analyzes his business motives for talking down peers

    On X, Musk predicted that more than 200 emerging compute cloud providers (Neoclouds) will face mass death and consolidation, on the grounds that the bond market is nearly closed to companies without revenue independent of AI data centers. The article sorts through the pressure on the Neocloud model of using long-term contracts as collateral to finance GPU purchases amid lagging power delivery and rising debt costs, and points out that SpaceX has just finalized $40 billion in GPU debt and holds an investment-grade credit rating, so talking down peers benefits its compute business and related parties' powered land assets.

    Same story · 5 reports

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

  20. HuxiuOpinion
    AI score66

    Hoarding GPUs for Three Years: More Cards, More Debt, But the Money Is Gone

    An article by Pianke Sheng, reposted by Huxiu, points out that the compute market is not experiencing across-the-board oversupply but rather a supply-demand mismatch between old and new cards and between eastern and western regions; the average utilization rate of national intelligent computing centers is about 30%, with third- and fourth-tier domestic clusters at only 10%–22%, and H100 rental prices falling from over $7 per hour to about $3.4. Cross-industry GPU hoarders have fallen into losses: Hongbo's compute subsidiary had net assets of -5.9 million yuan in the first half of 2026, and Qunxing Toys' intelligent computing gross profit was negative 15 million yuan. The article also mentions that ByteDance, Alibaba, and Tencent have built their own ten-thousand-card clusters; in the first half of 2026, shipments of non-GPU accelerators exceeded GPUs for the first time, and Token prices fell 84% from the March 2023 baseline.

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