Top Stories
5 stories + 9 briefsAgent Skills moves to a neutral foundation, and the skill standard enters shared governance
On October 22, Anthropic announced that it is donating the Agent Skills standard to the Agentic AI Foundation, part of the Linux Foundation. The transfer includes the name and logo, the GitHub organization, and the agentskills.io domain. According to the announcement, engineers from several companies will maintain the standard as individuals. Using or contributing to the standard does not require joining the foundation first.
A skill here is a folder that can hold a SKILL.md file, scripts, or reference material, so an agent gets the steps and domain knowledge for a kind of task. It is not a new capability of one model, and it does not certify every third-party skill as safe. The company's later roadmap includes versioned releases, a formal proposal process, and security guidance for writing and installing skills.
The news is that governance of a portable way of working is starting to move from a single vendor to a neutral organization. Teams will need to watch the specification version, the maintainers, and execution permissions, not only collect prompts. An open standard and whether a particular skill is safe to run are two different judgments.
Anthropic News | October 22, 2026
Anthropic discusses folding subagents back so shared context stays in the main loop
In an October 22 engineering article, Anthropic argues that some subagents split by business function can gradually be folded into skills and tools called by the main agent. With progressive loading, the system first offers a short description and loads the concrete procedure and tools only when they are needed, instead of stuffing a large context in at the start. The article is about agents built with Claude and Claude Code.
The authors say that once knowledge for a continuous task is split across several contexts, a summary handoff can drop detail, and a local evaluation may not represent completion of the final task. They suggest building an end-to-end evaluation first, then merging one boundary at a time, and comparing quality, latency, and specific failures. Tasks with a lot of shared context are the first candidates for merging. Independent research, components with a clear interface, and black-box checks may still be better kept apart.
This is architecture advice, not an experimental verdict that multi-agent systems have failed. The article also reports that a merge once broke a flow that had been working, so an average score is not enough, and a check should not simply be deleted. Permissions, budgets, and approval thresholds still have to be enforced by the tools and the runtime.
Claude Blog | October 22, 2026
Google expands a virtual-power-plant partnership to ease grid pressure through demand response
On October 22, Google and Voltus announced an expanded partnership to build a 1-gigawatt virtual power plant in the PJM grid region of the United States, ten times the earlier 100-megawatt agreement. This is aggregated adjustable capacity, not a 1-gigawatt generator already in operation.
Voltus aggregates flexible resources such as smart thermostats, batteries, and industrial equipment at businesses and homes, temporarily reducing demand at peaks. Participants receive bill savings or cash. Google is providing funding for the project. The text announces the expanded agreement. It does not give a timetable for when the full capacity will actually be in place.
This is another way to look at the expansion of AI compute: besides adding supply, shifting when electricity is used can ease pressure. Whether the callable load is really available, how long it lasts, and whether participants can respond still decide the practical value of this kind of resource. Agreement capacity should not be written as a solved power gap for data centers.
Google AI Blog | October 22, 2026
NVIDIA introduces NVIDIA Minerva, putting the machine room and the jobs on one view
On October 22, NVIDIA introduced NVIDIA Minerva, combining telemetry from GPUs, networks, storage, schedulers, and AI jobs so operations teams can locate anomalies and possible faults. It uses a dynamic knowledge graph to relate infrastructure and the jobs that are running, and it groups many related alerts into a more concentrated incident.
The publisher describes asking about cluster state in natural language and receiving repair suggestions. A team can keep a person in the loop, or connect automation through an infrastructure interface. IREN is listed as a user. Some other partners are planned deployments, so the whole partner list should not be counted as already running.
The focus is the useful work a machine room actually produces, not only how many GPUs were purchased. A training job that slows down may come from the network, storage, or a local hardware state. A cross-layer link explains more than an isolated count. The announcement does not prove the same fault prediction and repair benefit in every environment.
A global forest-type map at 10 meters separates natural forest from planted uses
On October 22, Google DeepMind, Google Research, and partner research groups released the Forest Typology 2020 global map. The Chinese edition also uses the short form ForTy. At a 10-meter scale it separates primary forest, naturally regenerating forest, planted forest, intensive plantation, tree crops and agroforestry, and other land. The team also shared an early year-by-year change sequence from 2017 through 2025.
The question is not simply whether trees are present. It is how the land is used, and what replaced a forest afterward. The model combines satellite time series with AlphaEarth Foundations representations, and uses planting patterns, growth cycles, and harvest changes to separate the classes.
The team reports that, in expert labels at more than 9,000 random sites, forest-type classification was about 90 percent correct and the forest versus non-forest distinction was about 95 percent correct. Those are overall results for that validation design, not a claim that every parcel has the same accuracy. The map also provides a probability for each class. A specific use still needs a chosen threshold and local evidence. This issue read the research groups' introduction. It did not review a technical article as the full paper or treat the map as an automatic compliance decision.
Google DeepMind | October 22, 2026
On October 22, the official announcement introduced the SpaceX Super Intelligence Partner Network for Grok. It separates Transformation Partners, who help companies put the system into operation, from Marketplace Partners, who build plugins and product integrations. The announcement says more than 130 companies are taking part in the early stage. Joining the program does not mean each company has finished a large deployment. Delivery, training, and connecting existing tools are the parts more worth following than another model-parameter claim.
Source: xAI · October 22, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
OlmoEarth v1.3, released on October 22, provides precomputed global annual features and a browser explorer for those embeddings. Each 10-meter pixel gets a 128-dimensional feature, and the first 64 dimensions can also be used alone. The team released model weights, pretraining data, and code. Users can train a smaller classifier on the features for land classification or change analysis. The publisher reports that the 128-dimensional version led on 20 of its 27 evaluations. That does not mean every region and task has been independently shown to beat other models.
Source: Ai2 · October 22, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
The October 21 update to NVIDIA Cosmos is included in this issue. Its data tools, including Cosmos Data Factory, support more modalities and sensor inputs, and provide blueprints for driving scenes, defect images, and changes in weather and lighting. Training and optimization tools, including Cosmos Framework, help adapt models in a modular way. This is different from the previous issue's robotics-platform news. The focus is training material and the evaluation path. Synthetic dangerous or rare scenes can fill out a sample. They are not a reason to skip tests in the real environment. Expected partner benefits in the announcement should not be written as results already achieved.
Source: NVIDIA · October 21, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
On October 21, NVIDIA described progress on NVIDIA DSX cooperation, bringing the grid, substations, and the power and cooling systems of the machine room into joint design and simulation. DSX Flex connects workload demand with grid conditions. What this issue keeps is this new power-system dimension. A reference design, qualified equipment, and infrastructure that has actually been built are still different stages. Simulation can expose a bottleneck earlier. It does not replace construction, grid connection, or on-site acceptance.
Source: NVIDIA · October 21, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
On October 21, NVIDIA described European retail applications. Shopify post-trained an open model, NVIDIA Nemotron 3.5 Lightning, for the GraphQL subagent in Sidekick that retrieves store data. The announcement reports that, on Shopify's internal benchmark, the specialized model was more accurate than a leading frontier model, serving cost was 90 percent lower, and responses were seven times faster. The result is limited to that task and that internal comparison. It is not a general advantage on arbitrary tasks. The case also shows that whether a subagent stays should depend on the boundary and on measured performance.
Source: NVIDIA · October 21, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
StormScope EU, introduced on October 21, is aimed at short-term weather prediction in Europe. It uses satellite observations to predict cloud changes across several channels, in the NVIDIA Earth-2 setting, with Earth2Studio among the tools. A demonstration pipeline at the event updated results about every 10 minutes as new observations arrived, and access was provided as map tiles. Energy and transport applications need local, timely judgments. A model demonstration is not a reliable operational warning at every location. Deployment still has to combine local observations and domain requirements.
Source: NVIDIA · October 21, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
The October 21 European public-sector cases involve running models in an institution's own environment, authorizing by role, keeping an audit trail, and building tools in local languages. Haystack Enterprise Platform is part of that account, alongside NVIDIA. Some systems are in use. Projects such as medical imaging in Poland and employment services are still being built or developed. Open models create room to adapt and to manage the system independently. Whether training data is provided, how sensitive information is protected, and who is responsible for errors still have to be judged project by project.
Source: NVIDIA · October 21, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
On October 21, Nokia said that with T-Mobile and NVIDIA it completed an AI-RAN call on a live commercial site in Bellevue in the United States, using a commercial core network and commercial devices. The next stage will test more frequency bands, and radio-access functions sharing accelerated computing with AI inference. One call is a milestone from the lab to the field. It is not a nationwide deployment, and it is not the spectrum-efficiency target already achieved.
Source: Nokia · October 21, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
On October 21, Databricks announced that AI Runtime is generally available. It puts managed GPUs, the training environment, logs, and Unity Catalog data permissions inside one governance boundary. It is meant to reduce the operational load of provisioning, troubleshooting, and starting experiments. Customer speedup examples in the announcement include hardware and configuration changes. The whole time difference should not be credited to one piece of software, and it should not be copied directly onto other training tasks.
Source: Databricks · October 21, 2026 · Relevant passages of the official text; capabilities and plans follow the publisher and were not independently tested
Industry Insights
Conversation SpotlightPaul Bakaus is the founder of Impeccable and a former Google web-platform architect. Starting from front-end design and developer tools, he discusses agent skills, and how generated results can escape sameness and become reliable.
The material comes from an AI Engineer technical talk. The original date is September 21, 2026. The event is about engineering practice for applying AI, including model applications, tools, and building systems. This issue uses four published excerpts and the background note on the Concourse page. It does not claim to have reviewed the full talk.
Near 06:06 and 06:44, he treats a skill as an extension of the runtime, not only a packaged prompt. Near 04:06 he discusses deterministic checks and structured procedures. Near 08:08 he discusses cross-checking from a separate context. We place that next to today's transfer of the skill standard and the advice to use a single loop. Shared task context and keeping an independent check are not mutually exclusive. The question is what purpose the boundary serves. This is this issue's connection, not his response to the October 22 news.
The specific Concourse page keeps the original video entry and the locations. These engineering experiences are not written as a law that every system must adopt. This summary is a paraphrase, not a verbatim quotation.
Concourse | September 21, 2026
Source: Concourse · September 21, 2026 · Four published excerpts on q.cxspace.com, the person's background, and the original date; the full transcript was not read
Research Briefs
0 papersNo separate items were recorded in this section of the original manuscript.
Open Source Spotlight
1 projectOn October 22, an IBM research team introduced Granite Timeseries Ensemble. It combines four complementary architectures for forecasting a series over time. The members are Granite forecasting models. The recommended default gives the members equal influence, merges the forecasts in probability space, and then produces a median and quantiles. Another option weights members adaptively by the width of the prediction interval.
The recipe stores the ensemble configuration. It does not fuse the four model weights into one model. The first run still downloads the member checkpoints. The text warns that a narrower prediction interval is not necessarily more accurate. It can also be overconfidence. Users should compare the options on representative data, rather than treat a nominal coverage rate as reliability already achieved in every setting.
This issue's scope is the official technical article and its code examples. The code was not installed, and the result is not claimed as reproduced. The license and use conditions of each member model still have to be checked for the models actually used.
Coverage notes · Incomplete · 20/45 checks recorded
Coverage is incomplete. Upstream per-source records through 05:00 Beijing time were reused, and the final hour was checked again. Some browser entry points showed only a frame or a section summary, which does not prove complete coverage. The original page for a selected OpenAI candidate was still behind an access check, so this issue keeps it only as a catalog lead and does not expand the body from the title. Where a catalog date conflicts with the original, the original controls. Several NVIDIA reports, plus Nokia and Databricks, are actually dated October 21. A Claude office plugin and the Personal Agent Protocol are actually dated October 6 and are not presented as news from this window.
There is no newly checked paper item sufficient for this issue. A technical blog is not presented as a full paper. The conversation section cites only published excerpts on a specific live page. It does not claim the full program was read. Vendor cases in the briefs keep the distinction among an internal benchmark, work still in development, a demonstration, and a deployment.
Chinese audio is available on Xiaoyuzhou (about 21 min 26 sec).
Coverage gaps reflect incomplete checks; they do not establish that a source had no updates.
- Anthropic News — Partially checked. Bodies were not read through. The /research section is outside the configured entry point and only the first screen of the list was checked. Original pages have day-level dates only. A scorpion item is dated Oct 20 on the list, and the publishedOn day boundary still needs checking.
- Claude Blog — Partially checked. Dates are day-level only, and bodies were not read. The configured blog URL redirects to resources/articles. External claude.dev items took titles only from list data.
- Microsoft AI — Partially checked. The configured page has no list dates, and the configured feed returned HTTP 410. The AI-topic feed and the Source sitewide feed each contain only 10 items, which cannot establish full-window coverage. Bodies were not read.
- Qwen — Partially checked. The configured entry point is an old site and does not match the current official entry. Only the first rendered screen of qwen.ai/research was checked, without pagination. Hugging Face model-creation times are not announcement times; ModelScope and WeChat accounts were not checked. config/sources.json was not changed.
- Tongyi Lab — Partially checked. The configured homepage is a benefits page. News is actually on qianwenai.com/news, and the config was not changed. The release calendar's interactive view was not fully rendered. WeChat accounts and Alibaba Cloud official news were not checked.
- DeepSeek — Partially checked. The change log has day-level precision. Repositories with pushes but no release were not reviewed commit by commit. The full change log was not read.
- Moonshot AI — Partially checked. Blog posts were not opened one by one. The homepage may not list every product update. GitHub and Hugging Face were not checked.
- Z.ai — Partially checked. News and paper bodies were not read. The z.ai English site and GitHub or Hugging Face releases were not checked.
- StepFun — Partially checked. The official site has no news stream and no absolute dates. Commits in the Step-Code repository were not reviewed one by one. WeChat accounts and the Chinese site were not checked.
- Mistral — Partially checked. Bodies were not read.
- European Commission Digital Strategy — Partially checked. A cloud IP received CloudFront HTTP 403, and the list was read from a Wayback snapshot. Dates are day-level, without hours or minutes. The snapshot may lag the live page slightly. The section is not only AI. Bodies were not read.
- UNESCO AI — Partially checked. The page lists only the latest several article cards. Bodies were not read.
- OECD.AI — Partially checked. If the embedded JSON is not the whole site, new articles not injected into the homepage are invisible. Bodies were not read.
- arXiv cs.AI / cs.LG / cs.CL — Partially checked. Abstracts and full texts were not read. Papers submitted after 2026-10-21T17:59:14+00:00 were not yet announced and were invisible to the API. Only new papers with a first submission inside the window were collected; replacement versions of older papers were not counted separately. The published field is submission time, not the arXiv announcement time. Because the window was not fully visible, the source stays partial. At check time, about 05:10 Beijing time on Thursday, October 23, 2026, when US Eastern time was about 17:10 on Wednesday, the latest cs.AI/new listing was Thursday, 22 October 2026, with 370 entries: 63 new, 148 cross-listed, and 159 replacements. That listing was announced at 20:00 US Eastern on Tuesday, 2026-10-22T00:00Z, and covers submissions from 14:00 US Eastern Monday to 14:00 Tuesday, 2026-10-20T18:00Z to 2026-10-21T18:00Z. Of this round's window, 2026-10-19T21:00Z to 2026-10-22T21:00Z, the portion 2026-10-19T21:00Z to 2026-10-21T18:00Z, about 45 hours, was announced, visible in the API, and collected. Submissions from 2026-10-21T18:00Z to 2026-10-22T21:00Z, about 27 hours, become visible only with the announcement at 20:00 US Eastern Wednesday, 2026-10-23T00:00Z, which is 08:00 Beijing time on October 23 and after this round's cutoff. That gap cannot be written as an absence of new papers.
- Hugging Face Daily Papers — Partially checked. An empty list or HTTP 400 is not interpreted as there being no new papers. Listing date and arXiv submission date are recorded separately. Abstract pages were not opened one by one to check submission dates.
- OpenReview — Partially checked. New submissions inside the window were not searched venue by venue. Announcement bodies were not opened.
- GitHub Trending (daily + Python daily) — Partially checked. The snapshot is not a list of projects that are new inside the window. Items were not filtered for AI relevance.
- 36Kr AI (article-level checks for original and syndicated reporting) — Partially checked. Bodies were not read through. Items by contributing authors were not traced to the first-publication platform, often a WeChat account. Promotion judgments looked only at wording at the start and end of the body. Promotion and event channels were not opened separately.
- Cyberspace Administration of China · News (AI-related) — Partially checked. Bodies were not opened to confirm the publishing body and legal effect. List dates were taken from the URL path. AI relevance was judged from titles only.
- ByteDance Seed · Blog and papers — Partially checked. Full texts were not opened. The paper list was checked only on page 1. A year in model navigation was not treated as a publication date.
- MiniMax·News & Updates — Partially checked. Load More was not clicked. Bodies were not read. Chinese news and model-release pages on minimax.cn were not checked.
- TechCrunch — Partially checked. Bodies were not read during collection. Original reporting versus briefs such as In Brief was not classified at that stage.
- Ars Technica — Partially checked. Bodies were not read. The section may include syndicated items, and original versus relay status was not classified.
- MIT Technology Review — Partially checked. Bodies were not read. A summary or paywall was not treated as the full text. Suspected Insights sponsored content was not opened to confirm.
- 404 Media — Partially checked. Bodies were not read, so which investigations are AI-related cannot be determined. Dates are day-level, and times on boundary days are unknown. Non-AI titles were left for later selection rather than deleted. Some articles require a subscription or login and were not attempted.
Reviewing a source does not constitute independent testing. Paper-reading scope is stated with each item; experiments were not reproduced.
Sources & discoveries 83 links
Titles are retained from the Chinese edition; source names are shown in English. This discovery list includes material not read in full.
News · 79 links
- Agent Skills移交中立基金会 技能标准进入共同治理Donating Agent Skills to the Agentic AI FoundationAnthropic News
- Google扩大虚拟电厂合作 以需求响应缓解电网压力We are working with Voltus to build a 1-gigawatt virtual power plant.Google AI Blog
- NVIDIA推出Minerva 把机房与任务监控接到同一视图NVIDIA Unveils Minerva to Bring End-to-End Observability for AI FactoriesNVIDIA
- 全球森林类型地图细化至10米 区分天然林与种植用途Beyond tree cover: mapping the world’s forest typesGoogle DeepMind
- Grok伙伴网络公布 企业部署与插件接入分工Introducing the SpaceX Super Intelligence Partner NetworkxAI
- OlmoEarth v1.3开放全球卫星特征 降低地图任务起步成本OlmoEarth v1.3: State-of-the-art embeddings for the whole planetAi2
- 法律专家称Meta的AI仍可能抓取新闻网站Meta’s AI could still scrape news sites, legal experts sayThe Verge
- 404 Media报道后，得州用于快速审批数据中心的门户关闭Portal Texas Used to ‘Rubber Stamp’ Data Centers Down After 404 Media Reporting404 Media
- Meta重新开放Muse分享功能Meta turns Muse sharing back onThe Verge
- 帮助女性进入AI热潮的社群The Social Society Helping Women Join the AI BoomWIRED
- 1Password让AI智能体代为登录，同时不向智能体暴露密码1Password launches a way for AI agents to log in on your behalf without seeing your passwordTechCrunch
- OpenAI全面开放其Jev竞品，决策模型竞争升温OpenAI makes its Jev competitor widely available as decision model competition heats upTechCrunch
- Canva发布首个图像生成模型Canva releases its first image generation modelTechCrunch
- 微软称其AI在疑难诊断上优于医生，继续追求“医疗超级智能”Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases ‘medical superintelligence’TechCrunch
- 在Meta设想的未来，你我的Muse将彼此联系In Meta’s version of the future, I’ll have my Muse call your MuseThe Verge
- 有人借助旧游戏和AI破解Switch 2Someone hacked the Switch 2 by using an old game and AIThe Verge
- 谷歌AI概览让部分人误以为捐款给了James TalaricoGoogle’s AI overview led some people to think they were donating to James Talarico when they weren’t.The Verge
- 《Wildwood》导演称AI与工作室理念背道而驰Wildwood’s director says AI is ‘antithetical to everything that we are at this studio’The Verge
- Uber为全部3万名员工部署ChatGPT WorkUber puts all 30,000 employees on ChatGPT WorkTechCrunch
- OpenAI、Anthropic与NVIDIA季度游说支出创新高OpenAI, Anthropic, and Nvidia had record lobbying quartersThe Verge
- Spotify押注成为AI智能体选择的音乐服务Spotify’s current AI bet is about being the music service that agents chooseTechCrunch
- AI正在改变黑客活动，无人知道接下来会怎样AI Is Changing Everything About Hacking. No One Knows What Happens Next404 Media
- AI已在放大战争的破坏AI Is Already 'Amplifying the Destruction' of War404 Media
- 数据与AI智能体助力TK Elevator维修技术人员Data and AI agents give a lift to TK Elevator service techniciansMicrosoft AI
- Gemini 4 Argon Token消耗量暴涨3倍，谷歌CEO首谈砍掉3.5 Pro内幕，还剧透了5.036Kr AI (article-level checks for original and syndicated reporting)
- 跑步上岗的具身智能：谁在赚吆喝，谁在算细账36Kr AI (article-level checks for original and syndicated reporting)
- Nirva融资3300万美元，开发只倾听佩戴者的吊坠Nirva raises $33M for a pendant that only listens to youTechCrunch
- 报道：DeepSeek在美国限制措施出台数日后融资150亿美元DeepSeek reportedly raises $15B days after US restrictionsTechCrunch
- 认识Mori：一款不太会说话或行动的蘑菇机器人Meet Mori, a Cute Robot Mushroom That Doesn’t Really Say (or Do) AnythingWIRED
- 「便宜到一集只要几十元」AI漫剧制作现状：价格「10倍跳水」，一天能产600集，全在赌概率？36Kr AI (article-level checks for original and syndicated reporting)
- 收入快速增长，MiniMax却破发了？36Kr AI (article-level checks for original and syndicated reporting)
- AI智能体准备代购，部分零售商仍有戒心AI agents are ready to shop. Some retailers are waryThe Verge
- 资本开抢AI 3D，「御三家」半年吸纳超80亿元36Kr AI (article-level checks for original and syndicated reporting)
- Claude并没有发明蝎毒抗毒血清Claude did not invent scorpion antivenomThe Verge
- 600亿元国家AI基金投向了谁36Kr AI (article-level checks for original and syndicated reporting)
- 涌进中小学的机器人草台班子36Kr AI (article-level checks for original and syndicated reporting)
- “你被AI取代了”：短剧人的202636Kr AI (article-level checks for original and syndicated reporting)
- 宇树 10 周年，王兴兴：我劝你相信 AI36Kr AI (article-level checks for original and syndicated reporting)
- 英伟达GTC德国柏林站，提前写好下一轮AI竞争的规则?36Kr AI (article-level checks for original and syndicated reporting)
- Eiliko厂商称将在机器人安全报道后聘请独立安全公司Eiliko maker says it’s hiring an ‘independent security firm’ after Verge report on its insecure robotsThe Verge
- 利润暴跌四成，也挡不住马斯克的AI野心36Kr AI (article-level checks for original and syndicated reporting)
- 谷歌Gemini 4 Carbon首批实测曝光，手感直逼Opus 5.536Kr AI (article-level checks for original and syndicated reporting)
- AI建设热潮让灰色涂料迎来生意It’s a good time to sell gray paint.The Verge
- Hugging Face的新AI生物模型让你与DNA对话Hugging Face’s New AI Biology Model Lets You Chat With DNAWIRED
- 别人养宠物花钱，这些年轻人养AI宠物月入十几万36Kr AI (article-level checks for original and syndicated reporting)
- AI把整个软件生态打成筛子，全栈开发者的至暗时刻36Kr AI (article-level checks for original and syndicated reporting)
- 个人智能体扎堆上线，没几个 App 愿意开门36Kr AI (article-level checks for original and syndicated reporting)
- 拦下AI Agent，美银、花旗在提防什么？36Kr AI (article-level checks for original and syndicated reporting)
- 取代商K的“AI女团”，开始让年轻人烧心36Kr AI (article-level checks for original and syndicated reporting)
- 机器人下线，马斯克：周产2万台需一年36Kr AI (article-level checks for original and syndicated reporting)
- 全网呼叫陈广宇36Kr AI (article-level checks for original and syndicated reporting)
- AI视频爆发一年：一代人的生产红利和孤独迷茫36Kr AI (article-level checks for original and syndicated reporting)
- 大模型厂商，为什么决定不了自己在具身智能的位置？36Kr AI (article-level checks for original and syndicated reporting)
- 96台机器人共同进化，成功率反超人类，清华、智元机器人、无问芯穹联合提出“集体经验学习”36Kr AI (article-level checks for original and syndicated reporting)
- 35万个AI陪骗子“唠嗑” ，AI反诈成资本新宠36Kr AI (article-level checks for original and syndicated reporting)
- 欧洲金融服务机构采用开放模型构建专有智能Leading European Financial Services Organizations Embrace Open Models to Build Proprietary IntelligenceNVIDIA Newsroom
- 耗资千万的AI，为何成了无人敢用的摆设？36Kr AI (article-level checks for original and syndicated reporting)
- AI冲进双十一，算盘打得比人响？36Kr AI (article-level checks for original and syndicated reporting)
- Muse刚火，小扎就催它挣钱了36Kr AI (article-level checks for original and syndicated reporting)
- 豆包天才90后，融了13亿36Kr AI (article-level checks for original and syndicated reporting)
- AI超创”抄袭“首案：《信托追踪者》VS《血誓》，纷争过后，谁来裁定漫剧原创边界？36Kr AI (article-level checks for original and syndicated reporting)
- 黑客在PHP漏洞修补数日前就已利用漏洞Hackers Exploited the PHP Vulnerabilities Days Before They Were PatchedWIRED
- 字节御用数据中心，要IPO了36Kr AI (article-level checks for original and syndicated reporting)
- 当Agent“有工号”：从一人一工具到一人一团队，组织如何做好AI时代的分工与协作？36Kr AI (article-level checks for original and syndicated reporting)
- 具身智能，热钱不念旧情36Kr AI (article-level checks for original and syndicated reporting)
- 套取AI收录标准被罚，GEO这门生意越来越难做36Kr AI (article-level checks for original and syndicated reporting)
- Personal Agent战火正在烧向硬件36Kr AI (article-level checks for original and syndicated reporting)
- 谁在养活 Anthropic？36Kr AI (article-level checks for original and syndicated reporting)
- 跟微软抢客户，AI办公独角兽融资16亿，英伟达投了36Kr AI (article-level checks for original and syndicated reporting)
- 不建议跟风Personal Agent36Kr AI (article-level checks for original and syndicated reporting)
- 3大巨头、19款应用，争抢AI时代的第一张船票36Kr AI (article-level checks for original and syndicated reporting)
- 走向超级智能体，AI还需要主动性：UIUC、谷歌等联合综述36Kr AI (article-level checks for original and syndicated reporting)
- 宇树H2-X来了：能当“保镖”、可挡刀，最长续航8小时36Kr AI (article-level checks for original and syndicated reporting)
- 现在复活一款老游戏，只需要 8 小时了36Kr AI (article-level checks for original and syndicated reporting)
- Anthropic CEO长文硬刚，自曝雪藏Mythos 4个月，为AI安全自断财路36Kr AI (article-level checks for original and syndicated reporting)
- Blackstone的Jas Khaira认为AI建设债务担忧被夸大Blackstone’s Jas Khaira says fears over AI-buildout debt are overblownTechCrunch
- Meta拟以新类别股票支持广泛的AI建设投入Meta wants to pay for its broad AI buildout with a newly created class of stockTechCrunch
- 数据中心可能抵消数十年的空气质量改善Data centers could erase decades of air quality gainsThe Verge
- 本周大型AI数据中心建设动态This week in the big AI data center buildout.The Verge
Concourse · 3 links
- AI算力为什么按矩形出售：对话Modal首席执行官Erik BernhardssonWhy AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEOLatent Space
- 你就是团队：对话Moshi的Anekha SokhalYou are the Team — Anekha Sokhal, MoshiLatent Space
- 比尔·盖茨：为AI带来的20年动荡做好准备Bill Gates says to prepare for 20 years of ‘tumultuous’ AI disruptionThe Verge