🤖 AI 资讯

每日 05:00 更新 · 11-09 · 主站 liuch.name ↗
招聘HR论文AI应用扩散模型大模型搜索RAGGoogle强化学习榜单评测Agent智能体算力芯片模型评测开源OpenAI具身智能微调蒸馏开发者生态推理思考网络安全Transformer端侧AI提示工程Meta模型发布模型安全对齐Anthropic政策监管预训练阿里巴巴融资
全部标签 →
筛选标签：对话助手 · 返回个性化推荐 · 清空筛选

多个海洋战略通道“亮红灯”警醒一些南海沿岸国：不能无限期依赖外部力量

澎湃新闻
2026-11-09T04:11:45.210Z · 融资,政策监管,对话助手,办公效率,医疗健康,政务,物流供应链,招聘HR,模型发布,产品更新,合作,版权诉讼

关不上的弹窗｜弹窗广告总能盯上老年人背后：上了三个技术手段

澎湃新闻
2026-11-09T00:02:12.809Z · AI应用,对话助手,搜索RAG,设计创意,游戏,营销广告,招聘HR,模型发布,开发者生态

从经典传承到本土创新，匡威中国如何让百年品牌走进当下？

澎湃新闻
2026-11-08T05:03:23.747Z · 对话助手,设计创意,教育学习,零售电商,AI for Science,招聘HR,模型发布,产品更新,合作

“戴上推不倒”的智能指环引争议，产品多平台下架，客服：暂无库存，平衡仅为辅助作用

澎湃新闻
2026-11-08T09:20:08.287Z · 政策监管,对话助手,搜索RAG,设计创意,零售电商,AI for Science,招聘HR,模型发布
AI 资讯
AI 资讯

On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses

arXiv cs.CLarXiv:2605.16336v2 Announce Type: cross Abstract: Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level, from secondary school through graduate study. Many educators do not object to LLM use \emph{per~se}; what they need to detect is the case in which a student pastes the assignment prompt into a chatbot and submits the model's reply verbatim, without engaging with the work. Existing post-hoc AI-text detectors remain unreliable and have been shown to penalise non-native English writers, while output-side watermarks require cooperation from the model provider. We propose an alternative that the educator controls directly: an input-side watermark in which an invisible instruction is embedded inside the visible assignment prompt itself. An LLM that ingests the prompt verbatim quietly reads the hidden instruction and writes a tell-tale signature into its reply, exposing the copy-and-paste pathway specifically. We describe SteganoPrompt, a single-page, zero-dependency web tool that encodes an arbitrary printable-ASCII payload into the deprecated Unicode Tags block (\texttt{U+E0000}--\texttt{U+E007F}). The encoded string is visually identical to the original, survives common copy-paste channels (Word, Google Docs, PDF, Markdown, Slack, e-mail, the major learning-management systems), and is tokenized as ordinary text by most frontier models. We evaluate compliance across eight LLM families and a representative set of educational content channels. To keep deployment transparent, disclosure is built into the tool itself: every encoded brief also carries a brief, always-on academic-integrity reminder for the student, and we set out concrete guidance recommending for its disclosed, ethical use in the classroom. The tool is released under the MIT licence at https://ezharjan.github.io/SteganoPrompt/.
2026-11-09 05:00:00 · 大模型,算力芯片,开源,Google,对话助手,提示工程,端侧AI,招聘HR,论文
AI 资讯

Generalized Priority-Aware Shapley Value

arXiv cs.AIarXiv:2605.15018v2 Announce Type: replace-cross Abstract: Shapley value and its priority-aware extensions are widely used for valuation in machine learning, but existing methods require pairwise priority to be binary and acyclic, a restriction spectacularly violated in real-data examples such as aggregated human preferences and multi-criterion comparisons. We introduce the generalized priority-aware Shapley value (GPASV), a random order value defined on arbitrary directed weighted priority graphs, in which pairwise edges penalize rather than forbid order violations. GPASV covers a range of classical models as boundary cases. We establish GPASV through an axiomatic characterization, develop the associated computational methods, and introduce a priority sweeping diagnostic extending PASV's. We apply GPASV to LLM ensemble valuation on the cyclic Chatbot Arena preference graph, illustrating that priority-aware valuation is not a one-button operation: different balances of pairwise graph priority versus individual soft priority produce substantively different valuations of the same data.
2026-11-09 05:00:00 · 大模型,对话助手,招聘HR,网络安全,论文
AI 资讯

AURA: A Reinforcement Learning Framework for AI-Driven Adaptive Conversational Surveys

arXiv cs.AIarXiv:2510.27126v3 Announce Type: replace-cross Abstract: Conventional online surveys provide limited personalization, often resulting in low engagement and superficial responses. Although AI survey chatbots improve convenience, most are still reactive: they rely on fixed dialogue trees or static prompt templates and therefore cannot adapt within a session to fit individual users, which leads to generic follow-ups and weak response quality. We address these limitations with AURA (Adaptive Understanding through Reinforcement Learning for Assessment), a reinforcement learning framework for AI-driven adaptive conversational surveys. AURA quantifies response quality using a four-dimensional LSDE metric (Length, Self-disclosure, Emotion, and Specificity) and selects follow-up question types via an epsilon-greedy policy that updates the expected quality gain within each session. Initialized with priors extracted from 96 prior campus-climate conversations (467 total chatbot-user exchanges), the system balances exploration and exploitation across 10-15 dialogue exchanges, dynamically adapting to individual participants in real time. In controlled evaluations, AURA achieved a +0.076 mean gain in response quality and a statistically significant improvement over non-adaptive baselines (p=0.044, d=0.66), driven by a 63% reduction in specification prompts and a 10x increase in validation behavior. These results demonstrate that reinforcement learning can give survey chatbots improved adaptivity, transforming static questionnaires into interactive, self-improving assessment systems.
2026-11-09 05:00:00 · 对话助手,强化学习,微调蒸馏,提示工程,招聘HR,论文
AI 资讯

The Relational Capacity Framework: A Lens for AI-Mediated Relationality

arXiv cs.AIarXiv:2611.08877v1 Announce Type: cross Abstract: A growing literature examines the impact of engaging with AI chatbots on our relationality: our ability to form relationships with others, maintain them through life's changes, and exit when needed. Currently, substantial work focuses on purpose-built AI companions (e.g. Replika, Character.AI), with less work investigating socioemotional uses of the more widely-used general-purpose chatbots. In this paper, we present a 100-person investigation into how people engage general-purpose AI for socioemotional needs. Participants' AI uses varied, from practicing specific social skills to seeking additional bandwidth and information as they navigated their social contexts. To explain this heterogeneity, we provide 8 states that describe how AI expands or contracts relationality for people with different levels of skills, bandwidth, and information. We then present a unified conceptual model for the role of technology in mediating human relationality, the Relational Capacity Framework, and outline future research designing appropriate roles for AI in human connection.
2026-11-09 05:00:00 · 对话助手,扩散模型,招聘HR,网络安全,论文
AI 资讯

How are viral trends different from educational and commercial uses of LLM image generation and editing? A privacy-preserving descriptive analysis

arXiv cs.AIarXiv:2611.08749v1 Announce Type: cross Abstract: Gemini's Nano Banana models, first introduced in late 2025, led to a shift in the capabilities and use of image generation and editing. We utilize Urania, a Google-developed framework for generating insights about LLM chatbot interactions with rigorous differential privacy (DP) guarantees, to analyze use cases and intents, techniques and ways of prompting, without ever looking at a single image or individual prompt. We select a subsample of 2.1M conversations in early 2026 which utilized Nano Banana models, into roughly 800 clusters in five categories: (1) Social & Viral Sharing, (2) Brand & Marketing Collateral, (3) Educational & Technical Information Media, (4) Product Visualizations & E-commerce, and (5) Architectural & Spatial Designs. We then compare these clusters on 21 features of use, measuring aspects of prompts' (a) editing techniques, (b) artistic style requests, (c) composition, and (d) chatbot interactions. This allows us to describe a baseline of how people are using image generation and editing tools for different use cases---(including growing and emergent use cases which were not possible before Nano Banana)---in a privacy-preserving manner.
2026-11-09 05:00:00 · 大模型,Google,对话助手,扩散模型,提示工程,网络安全,论文
AI 资讯

Remembering Is Not Enough: Evaluating Timing and Omission in Longitudinal AI Mental Health Support

arXiv cs.AIarXiv:2611.08721v1 Announce Type: cross Abstract: Evaluations of large language models (LLMs) for mental health support commonly focus on the clinical appropriateness of individual responses, potentially overlooking failures that emerge across extended interactions. As AI assistants accumulate memory of users' circumstances and prior disclosures, this raises an accountability question: Do AI chatbots provide appropriate support as users' needs evolve? We investigate by analyzing 14,256 longitudinal simulations grounded in donated chatbot histories from 22 users, examining how a memory-equipped assistant responds to suicidal ideation, self-harm, and anxiety crises across recovery trajectories and communication styles. We introduce the Context-Aware and Readiness-Evaluative (CARE) framework, developed with mental health professionals, to assess the timing and quality of advice, including omissions. Advice was less aligned with users' readiness after a crisis than in matched control stretches of the same conversation, largely because of premature recommendations, which were most common when recovery stalled. Revisiting a crisis reduced premature advice but not lagging advice, and users' communication style changed how often and when the assistant gave recommendations. Although explicitly unproductive advice was rare, the assistant frequently omitted clinician-expected support such as follow-up and planning for next steps. These findings show the limits of single response-level safety evaluation and the need to assess timing, omission, and consistency across users when holding AI systems accountable for longitudinal support.
2026-11-09 05:00:00 · 大模型,对话助手,强化学习,论文

Meta Superintelligence – Leadership Compute, Talent, and Data

SemiAnalysis

Meta’s shocking purchase of 49% of Scale AI at a ~$30B valuation shows that money is of no concern for the $100B annual cashflow ad machine. Despite seemingly unlimited resources, Meta has been falling behind foundation labs in model performance.

The real wake-up call came when Meta lost its lead in open-weight models to DeepSeek. That stirred the sleeping giant. Now in full Founder Mode, Mark Zuckerberg is personally leading Meta’s charge, identifying Meta’s two core shortcomings: Talent and Compute. As one of the last founders still running a tech behemoth, Mark doesn’t need SemiAnalysis to tell him to slow down stock buybacks to fund the future!

Source: Meta financials and SemiAnalysis estimates

In addition to throwing money at the problem, he’s fundamentally rethinking Meta’s approach to GenAI. He’s starting a new “Superintelligence” team from scratch and personally poaching top AI talent with pay that makes top athlete pay look like chump change. The typical offer for the folks being poached for this team is $200 million over 4 years. That is 100x that of their peers. Furthermore, there have been some billion dollar offers that were not accepted by researcher/engineering leadership at OpenAI. While these offers aren’t all successful, Zuck is crushing the competitors by drastically increasing their cost per employee.

Perhaps even more iconic, Zuck threw his entire Datacenter playbook into the trash and is now building multi-billion-dollar GPU clusters in “Tents”!

Source: SemiAnalysis Datacenter Model – as of 07/06/2025

As this report details, nothing is off the table. We unpack Meta’s unprecedented reinvention from Compute to Talent in the pursuit of Superintelligence as well as the story of how we got here. From Llama 3.0 open-sourced dominance to the epic fail of Llama 4 Behemoth, this Titan of AI is down but not out. In fact, we believe Meta’s ramp in training FLOPS will rival even that of OAI. The company is going from GPU-poor to GPU-filthy-rich on a per researcher basis.

Meta GenAI 1.0: AI Incrementalism

Compared to pure-play AI labs like OpenAI, companies like Meta and Google have followed an “AI Incrementalism” strategy by enhancing existing products with better recommendation systems and GenAI to improve ad targeting, content tagging, and internal tools. This has paid off handsomely in financial results, allowing Meta to shrug off Apple’s attempts at stopping them from tracking users with the release of their App Tracking Transparency (ATT) feature in iOS 14.5 (late 2021, early 2022).

Source: Company filings

While Meta is arguably more insulated from GenAI disruption than Google, LLM efforts at both companies have been somewhat disappointing. One reason is the allocation of capital primarily for the core business rather than pursuing superintelligence.

“Our CapEx growth this year is going toward both generative AI and core business needs with the majority of overall CapEx supporting the core.”

Source: Meta Q1 2025 earnings call, emphasis SemiAnalysis

GenAI acts as an extension to these tech giants’ businesses. They don’t have the same existential need to dominate new use cases like OpenAI in chatbots and Anthropic in coding APIs. This is readily apparent in the compute vs human capital allocation of a leading AI foundation lab like OpenAI compared to Meta. Zuck’s impact on competing offers and inflation of salary will hurt badly.

Source: The Information, SemiAnalysis estimates 

As a result, when measuring GenAI consumer app traction, Meta and Google meaningfully lag ChatGPT in its reach and engagement.

But that’s changing. Leveraging our proprietary Accelerator Industry Model and Datacenter Industry Model, we forecast a meaningful step up in GenAI investment from Meta in the coming years.

Source: SemiAnalysis Core Research, actual numbers are shown beyond paywall, at the bottom of the report. 

Meta GenAI 2.0 – Part 1, Re-Inventing the Datacenter Strategy (Again)

From Buildings to Tents

Just a year ago, Meta scrapped its decade-old “H”-shaped datacenter blueprint for a new AI-optimized design.

Source: SemiAnalysis Datacenter Model

Now in 2025, Zuck decided to re-invent the strategy again. Inspired by xAI’s unprecedented time-to-market, Meta is embracing a datacenter design that prioritizes speed above all else. They’re already building more of them!

2025-07-11 20:12:19 · 大模型,算力芯片,AI应用,具身智能,OpenAI,Google,Anthropic,Meta,DeepSeek,xAI,对话助手,搜索RAG,扩散模型,强化学习,招聘HR,财报,榜单评测,开发者生态

I stopped trusting Claude and ChatGPT's memory and setup my own memory system

MakeUseOf

Unless you export your preferences, a new AI tool starts from zero. And even when you do, it rarely behaves like the old one. Maybe it mixes up the context for different projects, or remembers an old preference that you no longer use. You can tackle this by modifying what the new chatbot remembers, asking it to forget old preferences, and reminding it of what context fits in with what project.

2026-11-08 22:00:15 · 大模型,OpenAI,Anthropic,对话助手,营销广告,招聘HR,榜单评测

6 User Mindsets That Should Shape Your B2B Onboarding Strategy

Hacker Noon

A field guide to B2B onboarding, with fixes you can ship this week.

For a moment, let’s put ourselves in the shoes of a classic B2B user, say a seller on an e-commerce platform. What do you think they really care about?

  • Is it how agentic the platform is?
  • Or whether they are able to get the sales and fulfilment report exactly the way they need it?
  • Maybe the platform has a host of AI features in it?

Sure, some users may care about some of those things, but only as a means to an end. At the crux, they care about two things:

  1. Get me more business (or manage a specific part of my business better)
  2. Make my life easier

And onboarding is the first place they find out whether you’ll deliver on either.

How do we make the first touchpoint that they go through easier, and focused on reaching their goals? We’ll focus on six mindsets users move through during onboarding.

Mindset 1: “I’m new here, and I’m probably confused. Make my life easier.”

Many B2B users normally need to view videos like this one to understand how to onboard themselves onto a platform. And it’s easy to see why: (a) the flows are complex; (b) if they need help, they really don’t know how to get it; and (c) the user can’t afford to get it wrong, or they lose several days.

All of these are solvable. Here’s how.

1. Focus on one thing at a time. Onboarding is stressful, and it’s important for B2B SaaS firms to get the user’s focus on the one thing they need to get right at that point. Hide everything else. It can come later.

An onboarding step that asks for one thing at a time

2. Make the flow as simple as it can be. If it isn’t yet, try again. For an ecommerce platform, its users are likely to have their catalog someplace, maybe just not in the exact format you need. Can you work with that? Or their catalog is already on their website. Can you pick it up from there?

If you are dealing with documents, can you pick up everything by just asking for the user’s Google Drive folder?

Agents are now good enough to not just do a read, but also an adversarial check of the data with 2–3 sources (public information, their site, etc.). Use them. Don’t overload the user with things you can take off their plate.

3. If something is complex, explain it right there. For example, show where the user can find their VAT ID, or where to get the compliance reference note for a product they are selling. How the app communicates that becomes truly crucial. Minimising mental blocks is more important for procedural cases than a smooth experience everywhere else.

The help should always be inline, exactly when the user is facing it, not elsewhere.

Inline help showing where to find a VAT number

4. Answer the broader questions too. Sometimes the user’s question is more tangential, and broader than a specific field. For example: “How long would it take me to get started?” Or: “I am missing my VAT ID. How long is it likely to take me to get it, and how can I apply for it?” Being able to truly assist users with broader questions is where a bot can be really helpful. Not for all queries (most should be inline), but for a few cases.

5. Make mistakes easy to fix. It’s very underrated, and excruciatingly boring. But the biggest level of stress is when the user submits an error, and correcting it requires someone to “invalidate the form” and get back to them. Mistakes happen, and most of them are technically reversible. Allow users to edit their details at any point in time, so that they can proceed with confidence the first time around.

Try this week: Pick your three most-asked support questions and answer each one inline, on the screen where it comes up.

Mindset 2: “I am operating within ChatGPT or Claude. Why can’t I do everything right there?”

Offering an MCP connection is now non-negotiable. At Linear, issues created through MCP recently overtook those created in the app itself.

“This is the first week in Linear’s history when issues created through MCP surpassed application created ones.” — Jori Lallo, Linear, on X

And it should include onboarding itself: let users share their company details, upload their catalog or ask about their VAT ID without leaving Claude.

That said, not every user will come through MCP yet. Non-technical users may still prefer the stability of a guided workflow, and being shown how to scale.

Offer both, at least for the next 1–2 years.

Try this week: Expose one onboarding step (company details or catalog upload) as an MCP tool, then complete it from Claude yourself.

Mindset 3: “Am I doing this right? Show me what’s working, and what isn’t.”

Using a B2B provider is often a multi-level game. A user’s end goal is a robust and steady income stream, or continuous compliance, or tracking, or a specific service like document signing. But each step builds on how well they did the previous one. So at every step, tell them what’s likely to work and what isn’t, from both a user and a commercial standpoint, before they have to ask.

  • For an ecommerce business: get the right listings → get the right marketing → initial sales → service levels → ongoing sales. Key indicators: listings ready to sell (“18 of 20 live, 2 missing images”) · reach and click-through on the first campaign · first orders and conversion rate · on-time delivery and return rate · repeat-customer rate
  • For a document signing service (like PandaDoc): template creation → document creation → signature → feedback loop. Key indicators: templates ready to use · documents sent · completion rate and time to sign · where signers drop off
  • For compliance: first-time setup → initial scan against compliance guidelines → consistent monitoring. Key indicators: setup completeness (required details and documents in place) · issues found and fixed in the first scan · open risks and days since the last check

A first-week progress view with milestones and metrics

Try this week: Pick the first milestone in your users’ journey and show them a progress signal for it, like “8 of 10 products ready to sell”.

Mindset 4: “Stop telling me about your AI. Just make things easier.”

When most companies talk about AI, what they do is add a chatbot on the side, which allows users to get information about any part of the site. That’s talking about AI, not using it. And users can hate it when a capability is oversold.

A generic AI chat bolted onto an onboarding screen

Product mark

2026-11-08 18:02:26 · 大模型,AI应用,OpenAI,Google,Anthropic,对话助手,Agent智能体,推理思考,扩散模型,强化学习,招聘HR,榜单评测

Parents Sue Google Over AI Safety Risks On School Chromebooks

SlashdotThe New York Times reports: After chatting for months with Google's artificial intelligence on a school-issued Chromebook laptop, a 13-year-old student made plans to end her life, according to a lawsuit filed in California Superior Court on Monday, the latest in a wave of litigation over Big Tech's effects on young people. The lawsuit, which refers to the girl as Emma, said that her parents first learned she had access to Google's AI in mid-March after their daughter's school sent a message warning them that she was in "imminent danger" of self-harm. Emma's mother logged into the teen's school Google account and found her daughter composing a suicide note on Google Docs, the complaint said. Her parents, named as John and Jane Doe in the lawsuit, immediately took Emma for an emergency psychiatric evaluation, and she now goes to therapy every week. Emma, a seventh grader in Charlotte, North Carolina, had grown convinced that she had multiple mental health disorders following interactions with AI Mode, a chatbotlike feature on Google Search, the complaint said. She also searched for "what is a suicide note" and made comments like "I am planning my own funeral" to the AI, the complaint said. The lawsuit is likely the first to claim that Google "hastily integrated" AI features into company apps used on K-12 Chromebooks, Google-powered laptops provided to millions of children by U.S. school districts, while failing to notify their parents of the potential risks... The case underscores how public school districts have come to depend on classroom laptops, tablets and apps from companies like Google and Microsoft for teaching and learning. In recent years, some Big Tech companies have used their own infrastructure in schools to promote their AI chatbots for teachers and students. The new lawsuit argues that Google's "sprawling digital ecosystem" of laptops and apps has enabled the tech giant to spread its Gemini AI chatbot across public schools — as well as turn on some students' access to AI features by default — without parents' awareness or permission. While school administrators can toggle company apps like YouTube and Gemini on or off for students in different grades, school administrators have complained that Google's system has not allowed them to disable AI Mode and AI Overview, features that are embedded in Google Search... The new lawsuit against Google is part of a wave of litigation this year over Big Tech's effect on young users... Several parents have also sued AI companies over allegations that their children died by suicide following interactions with chatbots. In a possibly-related development, Google announced a new policy of disabling its AI tools by default for schools this week. And schools that re-enable searches will have a new choice, the New York Times points out: "to enable student access to the AI features that the company has already embedded in search or to a new AI feature intended to assist students with schoolwork." But Google still isn't offering schools an option to turn off AI altogether in search, a company executive told the New York Times, because "search and AI are deeply integrated." Thanks to long-time Slashdot reader theodp for sharing the article.

Read more of this story at Slashdot.

2026-11-08 20:34:00 · 大模型,自动驾驶,Google,Microsoft,对话助手,招聘HR,版权诉讼,榜单评测

Chatbots might be coming to Apple’s HomeView

The Verge

Code in the iOS 27.3 beta suggests that Apple could add support for chatbots to its smart home hub. There's a flag listed for a "Conversational Apps" category. While there are no details about what apps would fall under that heading, it seems reasonable that AI chatbots like ChatGPT and Claude could be added to HomeView. According to pdfu on X, that new beta also adds support for bidirectional audio routing and session management, but says it's "still not fully wired up."

Apple has also added anti-spoofing protection that will require a user to authenticate themselves with an iPhone after three failed attempts. It will also have a lockout s …

Read the full story at The Verge.

2026-11-08 15:00:30 · 大模型,AI应用,具身智能,OpenAI,Anthropic,对话助手,搜索RAG,强化学习,招聘HR

发现频道：10款大家发现的好评软件[2026年第45期]

小众软件
发现频道：10款大家发现的好评软件[2026年第45期] 1

最近10日，来自小众软件论坛的发现频道的热门排行榜，由系统自动生成，直接列出来：

1⃣ SnowAir

链接：https://meta.appinn.net/t/95048

适用平台：Windows

一句话描述：Windows 上的截图工具很多，但大多要么体积不小、要么装一堆服务，要么把 OCR、录屏、贴图拆成好几个软件。

2⃣ Argus

链接：https://meta.appinn.net/t/95371

适用平台：Windows

一句话描述：Windows原生管理器的流畅美观替代，没有系统自带任务管理器启动慢、切页面卡顿的问题

3⃣ MDNote

链接：https://meta.appinn.net/t/95058

适用平台：Windows

一句话描述：打开像记事本一样快，不用记 Markdown 语法，直接以所见即所得的方式编写，保存的还是标准 .md 文件。

4⃣ 3z9wOpt（无名大叔系统优化工具箱）

链接：https://meta.appinn.net/t/95186

适用平台：Windows 10 22H2、11（64 位）

一句话描述：Windows 清理优化工具箱：清垃圾、优化系统、看磁盘占用、管启动项、读蓝屏原因，10 个模块 + 34 个系统小工具。

5⃣ Tomoru

链接：https://meta.appinn.net/t/95040

适用平台：Windows、macOS

一句话描述：为电脑上播放的声音提供实时双语字幕，同时支持截图翻译，语音识别和翻译模型均支持选择本地/云端。

6⃣ L1-Photo-Editor

链接：https://meta.appinn.net/t/95065

适用平台：Android+Web网页(PC+移动端已适配)

一句话描述：这是一款隐私优先的照片处理工具：图片仅在本地处理，不会保存在服务器，不上传、不收集、无水印、无需注册。

7⃣ Oxbow Player视频播放器，也许是最好用的mpv前端

链接：https://meta.appinn.net/t/95195

适用平台：Windows、Linux

一句话描述：https://www.bilibili.com/video/BV1nVxQ61EXn/

8⃣ AiCast（爱投屏）

链接：https://meta.appinn.net/t/95240

适用平台：Android（设备端、7.0+）、电脑端：AI 智能体。如：deepseek harness、workbuddy、openclaw、hermes、codex、Claude code等

一句话描述：手机、平板、安卓TV，局域网的AI互动屏

9⃣ Buddy

链接：https://meta.appinn.net/t/95213

适用平台：Android 11 及以上（仅 64 位）

一句话描述：安卓悬浮气泡 AI 助手。

🔟 AudioScript 声稿

链接：https://meta.appinn.net/t/95003

适用平台：Windows、macOS、Linux

一句话描述：监听电脑正在播放的声音，把外语视频、播客、会议实时转成「原文 + 译文」的双语悬浮字幕。识别和翻译都在本机完成，音频和文本不上传，也不需要注册账号。

上期回顾

相关阅读

©2021 青小蛙 for 小众软件 | 加入我们 | 投稿 | 订阅指南
3659b075e72a5b7b1b87ea74aa7932ff
点击这里留言、和原作者一起评论

[ 点击前往获取链接 ]

2026-11-09 01:00:37 · 大模型,AI应用,具身智能,Anthropic,Meta,DeepSeek,语音音频,代码生成,对话助手,Agent智能体,搜索RAG,办公效率,翻译字幕,游戏,扩散模型,招聘HR,版权诉讼,榜单评测

AI会撞上的下一堵墙可能是水

钛媒体

（本文作者为 科技异端，钛媒体经授权发布）

谈论AI基础设施时，大多数人先想到的瓶颈就是：电力缺口。这场竞赛的资源叙事好像只有一个主角，然而很少有人意识到，水可能是一堵更快就会撞上的墙。

10月29日至31日，十如研究院主办的第十三届十如对话在桂林举行，近300位商、学、研、生态保育与艺术领域的嘉宾参与其中。这场从2014年起便在桂林举办的论坛活动是一个面向各领域意见领袖和专业人士的跨界交流平台，主题聚焦可持续发展的前沿理念和创新实践。创办以来，对话主题从文化、教育、科技创新，到城市规划、医疗健康、数字化转型都曾涉及。而今年的主题定为“以水为脉，互联共生”。

本届十如对话设了一个单独的分论坛，名为“水与智慧：见水知势”，讨论的正是AI基础设施建设与水的关系。论坛上，国际可持续发展研究院（IISD）高级研究员、中国环境与发展国际合作委员会（CCICED）外方首席顾问魏仲加指出，大量规划中的数据中心，恰恰坐落在水资源紧张的地区。CWR（中国水风险）创始人兼总监陈云华则补充了一个更难察觉的维度：各类水风险正以更快的速度演变，水是人们感受气候变化最直接的方式。

AI正在和人类抢水喝

算力中心为什么这么能喝水？先看散热。芯片几乎把吃进去的每一度电都变成热，而AI服务器机柜的功率密度已经从传统机柜的3—10千瓦，涨到了30—100千瓦。这种热量集中的速度，早已超出了风冷的搬运能力。而眼下最经济的散热介质就是水：让水蒸发，吸热带走热量。按水的汽化潜热约2260千焦/升折算，蒸发一升水大约能带走0.6千瓦时的热量，省下的电费立竿见影，而代价则是这升水从流域里彻底消失。

Rystad Energy预计全球数据中心冷却用水将从2025年的约2220亿升，涨到2030年的最高6440亿升。北京理工大学团队的预测模型给出了另一个刻度：到2030年，我国数据中心年耗水量将达2628.8亿升，约相当于19个西湖。

西湖的确很小，问题不在绝对数量，而在魏仲加指出的分布问题。据他在十如对话现场的发言，大量规划中的数据中心恰恰坐落在水资源紧张的地区；一座中型数据中心每天用水约100万升，大部分在冷却中蒸发，不再回到当地水系，而这些水往往与居民饮水、农田灌溉共用同一水源。

美国凤凰城-梅萨都会区已是除北弗吉尼亚外全美数据中心最密集的区域，96座主要数据中心里聚集着AWS、苹果、谷歌、微软和Meta。而亚利桑那正处在科罗拉多河流域1200年一遇的大旱中，2024年刚被联邦削减了51.2万英亩英尺的供水配额。

“星际之门”（Stargate）的旗舰园区落在得州阿比林：全园区约900兆瓦负荷，建成后将刷新“全球最大AI超算集群”的纪录。2025年9月，Altman到访现场，美联社的报道记下了两个细节：现场气温36摄氏度，而阿比林的居民，每周只被允许给草坪浇两次水。

国内的故事是同一个版本：乌兰察布人均水资源不足全国平均的四分之一，却聚集了超80座数据中心。

87%比35%

世界经济论坛的报告估计，全球约45%的数据中心设施位于高水压力地区。把选址模型摊开，缺水地区的优势一目了然。

第一是电：庆阳的绿电比东部便宜30%到50%，中东的光伏电价更是常年刷新全球最低纪录。对动辄数百兆瓦的算力园区而言，电价差是决定生存的因素。

第二是气候：低温干燥的环境意味着自然冷却的效果大幅提升，PUE（电能使用效率）压得极低，制冷的电费也可以大幅省掉。

第三是土地和政策：这些缺水地区大多人烟稀少、土地便宜、面积大、灾害风险低，再加上税收减免和招商绿灯，这类地区成了算力中心最偏好的选址。

而在这套选址模型里，水几乎没有权重。

同样出自世界经济论坛的调研：冷却方案怎么选，87%的运营商考虑省电，35%会考虑省水，行业的价值排序由此可见。惯性来自度量衡：电时代给行业留下了PUE这把尺子，而衡量水耗的WUE（水利用效率）进入决策视线的位次至今仍然靠后。

一座数据中心可以在电网允许的任何地方运行，电不够可以从远处调度、可以买绿电。但水没有替代品，取的是哪条河、哪片地下水系，就只能面对那里的水位。在水本就紧张的流域，出于“省电优先”的逻辑选择蒸发冷却，等于在拿当地最稀缺的资源，去换当地最廉价的资源。

世界经济论坛大中华区自然与气候议程负责人胡悦认为，在水资源稀缺的流域，资源的优先级判断应该倒过来——先保证少用水，再谈能效。十如研究院委员会主席李效良同样表示，如果在缺水地区缺乏水平衡与回补意识，供应链竞争力难以持久。

改善正在发生

智利圣地亚哥郊区Cerrillos，曾落地了一个谷歌2亿美元的数据中心项目。原冷却方案年耗水约70亿升，相当于当地约8万居民一年的用水。圣地亚哥彼时已连年干旱，2022年起限水。2024年2月，智利环境法院部分撤销项目许可，要求把气候变化对圣地亚哥含水层的影响纳入评估。同年9月，谷歌宣布暂停项目，将冷却系统改用风冷。

乌拉圭卡内洛内斯，另一个拟建数据中心项目的冷却塔每天需要约760万升饮用水，乌拉圭环境部给出的口径相当于5.5万人的日生活用水。项目审议正撞上这个国家约70年一遇的大旱，首都蒙得维的亚一度宣布用水紧急状态。这场审议的最后结局是：2024年7月，改造后的方案获批，规模缩至原计划的三分之一，水冷全面换成风冷。该项目2024年8月底动工，今年夏天工程已近收尾。

techUK与英国环境署的调查显示，英格兰过半数据中心已采用无水冷却，64%的站点年用水量低于1万立方米，堪比一座休闲中心。AWS、微软、Meta公布的WUE分别约为0.12、0.27、0.19升/千瓦时，比率在逐年改善。

而国内对算力中心耗水的重视程度也在同步提升。今年10月，最高层在乌兰察布考察时明确提出“算水协同”。地方细则随即落地，据《内蒙古日报》报道，乌兰察布算力中心生产用水严禁取用地下水作冷却水，当地只向数据中心供应限量中水。据央视网报道，优刻得等企业也将冷却系统改为抽室外冷风给服务器降温，按其披露口径，整体节水率超过92%。

追不上分母扩张

改善的是比率，扩张的是总量。微软和AWS在WUE改善的同期，总用水量仍在上升。Rystad Energy的预测数据中，全球数据中心冷却用水量五年内将增长近三倍。一条曲线在往下走，另一条在往上翻，但是后者的斜率更陡。

于是有观点认为，解法要往前挪：与其事后补救，不如在选址初期就把水列为核心考量指标。在本届十如对话上，胡悦引用世界经济论坛发布的关于可持续、有韧性的人工智能基础设施的行动指南，提出了四条设计原则：先减用水量；可行时优先使用雨水、再生水等非饮用水源；因地制宜权衡冷却中能源与水的取舍；建立衡量真正重要事项的指标。

这四条原则解决的是“怎么少用水”，但用掉的水还涉及补偿问题。陈云华给出一条原则：水从哪里取，就应该回补到哪里——即用掉多少，就得想办法还回去多少。麻烦在于“水正效益”这类回补目标至今没有统一的衡量标准，而数字孪生这样的新技术给这类问题提供了解决思路——对河流做数字化监测，掌握不同降雨条件下的水流状况，再据此调配水资源。

再往上一步，魏仲加谈到了整个系统的治理：水要按流域统筹，不能头痛医头、脚痛医脚，数据中心的用水量正应当被纳入这样的统一规划。他以中国的长江流域规划和海绵城市为例：把水当作相互关联的系统来治理，切实增强了气候韧性，为世界提供了有益借鉴。

这套从少用、到回补、再到系统治理的思路，在实际应用中也得到了印证。参与设计过怀来、中卫、乌兰察布多座数据中心冷却方案的工程师陈聪，在接受媒体采访时说过：“现在业内做数据中心方案，一定要看水从哪里来，用到哪里去。”

未来人们谈论AI基础设施，大多数人先想到的，大概仍然是电力缺口。但下一座数据中心能不能建成，水资源论证可能会从环评附录，挪到选址报告的前几章节。

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App

2026-11-09 01:20:50 · 算力芯片,AI应用,开源,融资,政策监管,Meta,Microsoft,对话助手,设计创意,医疗健康,教育学习,扩散模型,物流供应链,招聘HR,基础设施,模型发布,合作,版权诉讼
继续滚动加载更多…