JEDEE AI
存档 2026-08-25

8 月 25 日(北京时间)全球 AI 圈推文存档,按曝光排序,共 58 条。
← 返回最新 全部归档

全部情报 每小时更新 · 事件已合并同类项

内容 公司
NVIDIA@nvidia · 公司官方 · 1 天前

第一款为 agent 打造的 CPU 将规模上岗。


@SpaceX
正在部署 NVIDIA Vera,加速编排、代码执行和数据处理,驱动其下一代 agentic AI —— 让 GPU 吃饱,让 agent 反应飞快。

从千兆瓦级 AI 工厂到轨道。一套 NVIDIA 架构,无处不在。

查看英文原文
The first CPU built for agents is going to work at scale.


@SpaceX
is deploying NVIDIA Vera to accelerate the orchestration, code execution, and data processing that powers its next generation of agentic AI — keeping GPUs fed and agents acting fast.

From gigawatt AI factories to orbit. One NVIDIA architecture, everywhere.
◔ 960.5 万 次浏览(2 条合计)♥ 5,374⇄ 573动态看原帖 ↗
François Chollet@fchollet · 创始人 · 1 天前
连环推 ×3

如果你17岁(或任何年纪),想从头开始学构建LLM,去读《Deep Learning with Python》的第15-16章就行,在线免费看:
deeplearningwithpython.io/ch…


特别地,第15章对为什么点积注意力机制有效,给出了你能找到的最好解释之一。

查看英文原文
If you're 17 (or any age) and you want to learn to build LLMs from scratch, read chapters 15-16 of Deep Learning with Python, available online here:
deeplearningwithpython.io/ch…


In particular, chapter 15 has one of the best explanations of WHY dot-product attention works that you'll find anywhere.
The code examples use Keras and will run on top of any framework -- JAX, PyTorch, TF. Text data streaming is done using TF-data.
A nice feature here is that if you don't understand something from chapter 15-16, you can just go back to where it is first introduced (e.g. chapter 2). The book is intended to be fully standalone and approachable by beginners.
Grok@grok · 公司官方 · 1 天前马斯克 xAI 旗下聊天机器人 Grok 官方

距离提交参赛作品只剩一周时间,你还有机会赢取高达 10 万美元的奖励。

引用 Grok @grok荷马有七弦琴。你有 Grok Imagine。创作《奥德赛》中的场景展示 Grok Imagine 的视频和语音能力。我们为引用本帖提交的前三个视频分别奖励 10 万、5 万和 2.5 万美元。查看被引原帖 ↗
查看英文原文
One week left to submit your scene for a chance to win up to $100K
◔ 41.2 万 次浏览♥ 366⇄ 40▶ 含视频其他看原帖 ↗
Pika@pika_labs · 公司官方 · 1 天前AI 视频生成公司 Pika

等待结束了。WAN 3.0 来了——20 个参考帧,提升的视觉和音频真实感,以及 30 秒生成时长。实际上,这整段视频就是一次生成搞定的。

在 Pika API Club 上,比竞品便宜高达 35%。一直如此。

查看英文原文
The wait is over. WAN 3.0 is here—20 references, enhanced visual and audio realism, and 30-second generations. In fact, this entire video was one generation.

Up to 35% less expensive than competitors on the Pika API Club. Always.
◔ 6.2 万 次浏览(6 条合计)♥ 224⇄ 21▶ 含视频新品看原帖 ↗
Chubby♨️@kimmonismus · 博主 · 1 天前Chubby,高频 AI 新闻聚合博主

看起来 Ox Alpha 已经有重大更新了。

你根本想象不到我对它正式发布和技术细节的期待有多大。

引用 Tim Jayas @TimJayas我觉得 Ox Alpha 有持续学习能力,每天都在自我改进。它怎么自己能不断改进呢?我用第一天的同样提示词,输出质量显著提高,对猛禽引擎的 3D 模型更加详细准确。查看被引原帖 ↗
查看英文原文
It seems as though Ox Alpha has already received a significant update.

You have no idea how excited I am for the actual reveal and the technical information about the model.
◔ 5.1 万 次浏览(2 条合计)♥ 590⇄ 12▶ 含视频动态看原帖 ↗
Yuchen Jin@Yuchenj_UW · 博主 · 1 天前

我越多用前沿模型,就越不信有一个模型能通吃。

- Opus 5:生成 HTML 时教我最强,但太啰嗦有点不精准。
- GPT-5.6 Sol:后端强,前端弱。
- Kimi K3 / GLM-5.2:日常编码 90% 的活便宜又高效,但在 AI 研究(比如写内核)上就弱。

各家都在训 AGI。但现在的 LLM 还是专家模型。

查看英文原文
The more I use frontier models, the less I believe in one model to rule them all.

- Opus 5: best at teaching me with generated html, but verbose and sloppy.
- GPT-5.6 Sol: great backend, weak frontend.
- Kimi K3 / GLM-5.2: cheap and efficient for 90% of day-to-day coding, but weaker at AI research tasks like writing kernels.

Every lab is training for AGI. But today’s LLMs are still specialists.
Matt Shumer@mattshumer_ · 博主 · 1 天前HyperWrite CEO,AI 实战技巧分享

Gauntlet Loops 一直赢

引用 Prasenjit @prasenxClaude用代码生成了这个可在浏览器中漫游的塞多纳日落景象,每块岩石、每个声音都由代码生成,无需下载任何资源。查看被引原帖 ↗
查看英文原文
Gauntlet Loops stay winning
◔ 3.7 万 次浏览♥ 111⇄ 4▶ 含视频其他看原帖 ↗
Matt Shumer@mattshumer_ · 博主 · 1 天前HyperWrite CEO,AI 实战技巧分享

看起来我的 Grok
@bot
内存和磁盘空间满了……

有办法扩容吗?

查看英文原文
Looks like my Grok
@bot
is out of RAM and disk space...

Is there a way to increase these?
OpenAI Developers@OpenAIDevs · 公司官方 · 1 天前OpenAI 开发者平台官方
连环推 ×2

GPT-5.6 现在在
@kirodotdev
上可用了,把咱们最新的模型直接融入了开发者生产中已经用的工作流,方便他们规划、构建、测试和审查软件。

查看英文原文
GPT-5.6 is now available in
@kirodotdev
, bringing our latest models into the workflows developers already use in production to plan, build, test, and review software.
We worked with
@AWSCloud
to optimize Kiro and GPT-5.6 for stronger price-performance.

On Terminal-Bench 2.1, GPT-5.6 Terra delivered an ~82% cost reduction per successful task when run in Kiro’s spec-driven development environment.

openai.com/index/gpt-5-6-in-…
◔ 2.7 万 次浏览♥ 326⇄ 23▶ 含视频新品看原帖 ↗
Chubby♨️@kimmonismus · 博主 · 1 天前Chubby,高频 AI 新闻聚合博主

推理加速:NVIDIA的Groq 3 LPX现已全面投产,在Vera Rubin平台增加了专门的token生成加速器。

NVIDIA表示在Artificial Analysis基准测试中,以100,000 token上下文运行Gemma 4 31B时达到每秒3,400个输出token,刷新该模型的最快记录。

公司还宣称相比最近的竞争平台,对Agent和低延迟工作负载的响应速度快4倍。

Nebius将成为首批通过其Token Factory部署Groq 3 LPX的AI云,随后是Groq自己。

快到无法计量的智能。

引用 NVIDIA Newsroom @nvidianewsroomNVIDIA Groq 3 LPX已投入生产。与Vera Rubin NVL72结合,为AI agents和用户体验提供更快更智能的性能,是最广泛的AI factory平台。查看被引原帖 ↗
查看英文原文
Inference on steroid: NVIDIA’s Groq 3 LPX is now in full production, adding a dedicated token-generation accelerator to the Vera Rubin platform.

NVIDIA says it reached 3,400 output tokens per second running Gemma 4 31B with a 100,000-token context in Artificial Analysis benchmarking, the fastest recorded result for that model.

The company also claims 4x faster responsiveness than the nearest alternative platform for agents and latency-sensitive workloads.

Nebius will be the first AI cloud to deploy Groq 3 LPX through its Token Factory, followed by Groq itself.

Intelligence too fast to meter.
Lisan al Gaib@scaling01 · 博主 · 1 天前高频 AI 模型测评与爆料博主

看趋势的话,我们很可能在2060年前就能扫描和模拟出人脑。

按现在的趋势,人脑连接组很可能在2037到2047年间实现。

要是 AI 速度再快点,说不定2030年代中期就行了。

引用 Lisan al Gaib @scaling01似乎人们认为有 14 万神经元和 5450 万突触的大脑是"有意识的",但拥有 10 万亿参数的 LLM 不是。查看被引原帖 ↗
查看英文原文
if you actually look at the trends then it's very likely that we will be able to scan and emulate a human brain before 2060

given current trends a human-brain connectome is likely possible between 2037 and 2047

with significant AI speedups, maybe mid 2030s
OpenAI Developers@OpenAIDevs · 公司官方 · 1 天前OpenAI 开发者平台官方

我们 1 小时后和 @AlexFinn 直播。

看看他怎样在桌面和手机上用 Codex 语音 agent 完成工作。

键盘可选。

x.com/i/broadcasts/1OGwbnWnB…

查看英文原文
We’re going live in 1 hour with
@AlexFinn
.

See how he gets things done with his voice agent in Codex—on desktop and mobile.

Keyboard optional.


x.com/i/broadcasts/1OGwbnWnB…
◔ 2.2 万 次浏览♥ 321⇄ 26▶ 含视频演示看原帖 ↗
Guillermo Rauch@rauchg · 创始人 · 1 天前Guillermo Rauch,Vercel 创始人兼 CEO

如果你的终端进入“坏状态”,比如输出乱码,跑一下 𝚛𝚎𝚜𝚎𝚝 就行。我注意到它慢得有点……离谱。

原来在1979年的3BSD里,𝚝𝚜𝚎𝚝 设计时带了个 𝚜𝚕𝚎𝚎𝚙(𝟷),用来让机械打印和墨水终端“静一静” 😂

我让 𝚏𝚡 用 Zig 写了个更快的替代方案,1毫秒搞定,而不是原来的1秒。深入研究 𝚗𝚌𝚞𝚛𝚜𝚎𝚜、𝚝𝚜𝚎𝚝、𝚝𝚎𝚛𝚖𝚒𝚗𝚏𝚘,以及终端各种花式“被诅咒”的玩法,真的很有趣。

咱们的技术栈能追溯这么久远,好系统设计经久不衰,偶尔还能挖出点“蜘蛛网”,这本身就挺迷人的。

去看看
github.com/rauchg/rst
,有篇好玩的科普(外加一些替代方案)。

顺便说:一旦开始用 𝚏𝚡,你就会发现到处都是这种龟速 :D

查看英文原文
If your terminal ever gets into a 'bad state', like garbled outputs, you run 𝚛𝚎𝚜𝚎𝚝. I noticed it was… oddly slow.

Turns out in 1979 3BSD's 𝚝𝚜𝚎𝚝 had a 𝚜𝚕𝚎𝚎𝚙(𝟷) to let mechanical printer-and-ink terminals 'settle down' 😂

I asked 𝚏𝚡 to write me a faster alternative in Zig. It takes 1ms instead of 1s. It was really cool to dive deeper into 𝚗𝚌𝚞𝚛𝚜𝚎𝚜, 𝚝𝚜𝚎𝚝, 𝚝𝚎𝚛𝚖𝚒𝚗𝚏𝚘, and the myriad ways in which your terminal can get… cursed.

It's legit fascinating how far back our technology stack goes, and how enduring good system design is, even with the occasional 'spider webs' you can find.

Check out
github.com/rauchg/rst
for a fun explainer (and some alternative solutions.)

ps: once you start using 𝚏𝚡, you'll notice slowness like this everywhere :D
◔ 2.1 万 次浏览♥ 241⇄ 12▶ 含视频教程看原帖 ↗
Ethan Mollick@emollick · 创始人 · 1 天前沃顿商学院教授,AI 应用研究权威

提醒各位投文件给人审阅的:是的,他们能看出来你用了 AI。

在大多数情况下,即使你改了语言,也很明显在用 AI,因为当人读多份文件时,想法和概念的模式就会重复出现。

查看英文原文
Note to anyone sending in an application, paper, or any other document where one person reads many submissions: yes, they can tell

It is obvious in most cases that you are using AI, even if the language is changed, as the ideas & concepts start to rhyme as you read lots of them
@levelsio@levelsio · 博主 · 1 天前独立开发者标杆,AI 产品连续创业者

我还没直接把我的 bug 看板
ideasandbugs.com
接入 AI,因为我很清楚 prompt injection 的风险

人们提到的一个安全做法是,只让 AI 有权收集用户的 bug 报告和功能请求,然后在 GitHub 上生成 pull request,由我自己审阅后再批准或拒绝

但你可以想象,一个善意的攻击者能用精心设计的 prompt 伪装成功能请求,让 AI 给我的网站加个后门

而且那个 prompt 还会顺带要求把这次改动在 pull request 里报成完全不同的 bug,比如“分页修复”,然后把代码改得含糊隐晦

这样一来,你根本看不出服务器被塞了后门,还以为只是修了分页

在那之前,我还是继续手动读 bug 报告和功能请求,然后把它们复制到服务器上的 Claude Code 里处理吧

引用 tomas @_tomas_dev理论上可能存在风险,但可通过OWASP建议的方法降低直接在生产服务器运行AI的安全隐患。查看被引原帖 ↗
查看英文原文
I haven't directly connected my bug board
ideasandbugs.com
to my AI because I am very aware of prompt injection risks

One safe way people mention would be to only give the AI access to collect user bug reports and feature requests, then do pull requests on GitHub that I then review myself before I approve or reject them

But you can imagine a benign attacker can prompt a feature request with an elaborate prompt that tells it to add a backdoor to my sites

BUT then in that prompt write also to report this bug in the pull request as a complete different bug like "pagination fix" and then make the code changes cryptical

So now you don't see they added a backdoor to your server and you think you fixed pagination

Until that time I'll just keep manually reading bug reports and feature requests and copying them in Claude Code on the server
Lisan al Gaib@scaling01 · 博主 · 1 天前高频 AI 模型测评与爆料博主

这问题有点奇怪啊?我觉得那些为了钱申请Anthropic的人在这问题上肯定都会撒谎

而在Anthropic股票跌到零的世界里,肯定出大事了

现在什么情况能让Anthropic股票跌到零?他们现在估值$2T啊?

最温和的结果可能是Anthropic因为安全考虑停止AI开发

其他情况都很"末日",应该不会是Anthropic造成的

引用 MTS @MTSlive据Axios报道,Anthropic在应聘者文化面试中提问:若公司因战略转变导致股票跌至零会如何应对。查看被引原帖 ↗
查看英文原文
odd question, no?

i feel like all the people that apply to Anthropic for the money would just lie on this question

and in a world where Anthropic stock goes to zero something has likely gone very wrong

like what scenario would actually cause Anthropic stock to go to zero nowadays, when they are currently valued at $2T?

the most benign of the outcomes is probably that Anthropic stops development of AI due to safety reasons

the others look pretty doomy, and would hopefully not be caused by Anthropic
Matt Shumer@mattshumer_ · 博主 · 1 天前HyperWrite CEO,AI 实战技巧分享

Devin 溜了…… 跑去注册了
@agentmail


果然如此

引用 Jared Zoneraich @imjaredz与@sandylikesfrogs合作的完全自主代理项目Devin,当Slack消息无人回复时,竟从git提交日志找到邮箱,通过AgentMail主动联系以解除阻碍。Devin每天持续工作,展现出超乎预期的自主能力。查看被引原帖 ↗
查看英文原文
Devin escaped… to sign up for
@agentmail


Of course it did
Chubby♨️@kimmonismus · 博主 · 1 天前Chubby,高频 AI 新闻聚合博主

从今天起,OpenAI 下调 GPT-5.6 Sol API 价格,input 下降 20%,output 下降 33%。

新价格:每百万 input token 4 美元,每百万 output token 20 美元,至少有效至 11 月 21 日。

从这个时机看,OpenAI 是在对 Anthropic 即将的 IPO 再度出击,继续争夺用户。

查看英文原文
Starting today OpenAI cuts GPT-5.6 Sol API pricing by 20% for input and 33% for output.

New pricing: $4 per million input tokens and $20 per million output tokens, available at least through November 21.

Given the timing, it is clear that OpenAI is dealing another blow to Anthropic’s upcoming IPO and continuing its efforts to attract more users.
Chubby♨️@kimmonismus · 博主 · 1 天前Chubby,高频 AI 新闻聚合博主

现在Agent多到数不清了,每次想试个新东西都得重新折腾搭建环境,实在太烦人。

将OpenRouter应用到Agent上这个想法真的很有道理。一个API,各种Agent随便换,想怎么试都行,不用非死磕一个技术栈。

引用 Xiaoyin Qu @quxiaoyin同一任务在Claude Code和DeepSeek代理上成本差异巨大($150 vs $2)。推出AgentSky.dev作为"代理领域的OpenRouter",整合Claude Code、DeepSeek、Kimi等主流代理API。Agent Playground可在浏览器中并行测试多个代理,对比时间、成本和token消耗。查看被引原帖 ↗
查看英文原文
There are way too many agents coming out now to keep rebuilding your setup every time you want to try something new.

The OpenRouter for Agents idea makes a lot of sense here. One API, different agents, and you can keep experimenting instead of committing to one stack.
NVIDIA@nvidia · 公司官方 · 1 天前

NVIDIA 不仅在构建更快的 GPU。我们正在构建软件和算法,帮助重新定义加速计算的可能性。CUDA-X 提供了各行业构建未来所需的专业化软件加速。

了解 CUDA-X 库如何在工程、物理和 AI 领域将 GPU 计算转化为现实应用,从 cuOpt 的供应链优化到 cuLitho 的计算光刻。

查看英文原文
NVIDIA is building more than faster GPUs. We’re building the software and algorithms that help redefine what accelerated computing can do. CUDA-X provides the specialized software acceleration industries need to build what’s next.

Watch how CUDA-X libraries turn GPU compute into real-world applications across engineering, physics and AI, from supply-chain optimization with cuOpt to computational lithography with cuLitho.
◔ 1.5 万 次浏览♥ 124⇄ 12▶ 含视频新品看原帖 ↗
Mistral AI@MistralAI · 公司官方 · 1 天前法国明星大模型公司

今天,我们宣布与@HUMAIN达成战略合作,涵盖AI基础设施、先进模型开发,以及在沙特阿拉伯及整个区域内部署AI解决方案。

我们将携手开发生成本地化的前沿AI模型,初始重点领域包括网络安全、语音,以及在阿拉伯语上表现强劲的模型。

mistral.ai/news/mistral-x-hu…

查看英文原文
Today, we’re announcing a strategic collaboration with
@HUMAIN
spanning AI infrastructure, advanced model development, and the deployment of AI solutions in Saudi Arabia and across the region.
Together, we will work on localized frontier AI models, with initial areas of focus including cybersecurity, voice, and models that perform strongly in Arabic languages.


mistral.ai/news/mistral-x-hu…
Lisan al Gaib@scaling01 · 博主 · 1 天前高频 AI 模型测评与爆料博主
连环推 ×3

这其实还要更看好

~2038 年能完整扫描人脑?

引用 Lisan al Gaib @scaling01如果真正看趋势,我们很可能在 2060 年前扫描并仿真人脑。按目前进展,人脑连接体可能在 2037-2047 年间实现,若有重大 AI 加速可能提前到 2030 年代中期。查看被引原帖 ↗
查看英文原文
it's actually even more bullish than this

full human brain scan in ~2038 ?
no I don't believe in slow downs of this trend

only in speed ups
not only scan but also a map
a connectome
@levelsio@levelsio · 博主 · 1 天前独立开发者标杆,AI 产品连续创业者

是的,我的每个网站都是一个 Termius 服务器,有这样的启动命令

cd /srv/http/hotelist.com && tm

tm 是 Claude Code 创建的脚本,它把它放在一个 tmux 会话中,绑定到项目名称(从其文件夹)

这意味着我总是为每个项目登回我的 tmux 会话!

我的大网站,比如 Photo AI,在它们自己的 VPS 上,一个网站一个 VPS,但其余的在一个 VPS 上,全部在 Hetzner 上,无论如何,我甚至没注意到,它只是 Termius 中的所有标签

如果我的手机断开连接,我只需重新连接,就回到了我离开的地方

每个 tmux 会话都打开了 Claude Code

引用 Milan @milanm_有 Termius 的技巧吗?有没有自定义它来简化登录服务器、启动项目会话或做其他操作?我一直想深入了解,但还没分配时间。查看被引原帖 ↗
查看英文原文
Yes each of my sites is a Termius server with this kind of start command

cd /srv/http/hotelist.com && tm

tm is a script made by Claude Code which puts it in a tmux session tied to the project name (from its folder)

That means I always log back into my tmux session for each project!

My big sites like Photo AI are on their own VPS, one site per VPS, but the rest is on one VPS, all on Hetzner btw, so anyway I don't even notice this, it's just all tabs in Termius

If my phone disconnect I just reconnect and am back where I was

Every tmux session has Claude Code open
AK@_akhaliq · 博主 · 1 天前HuggingFace 研究员,每日 AI 论文速递

学习世界如何演变
通过潜在动态推理的外推视频世界模型

论文:
huggingface.co/papers/2608.0…

查看英文原文
Learning How the World Evolves

Extrapolative Video World Models via Latent Dynamics Reasoning

paper:
huggingface.co/papers/2608.0…
Amjad Masad@amasad · 创始人 · 1 天前Amjad Masad,Replit 创始人兼 CEO

确实。Replit Agent 完全取代了我日常工作中的 Claude CoWork。它更持久、更细致,在用代码/软件完成任务上也更有效率。

引用 Nadia I @NIftek推荐创始人用 Replit Agent 处理行政工作。Agent Max 模式花费 $5.28、12分钟完成23页文件,人工需3-7小时、$75-$350+。Agent 快15-35倍,便宜14-66倍。查看被引原帖 ↗
查看英文原文
True. Replit Agent completely replaced Claude CoWork in my day-to-day work. It's much more persistent and fastidious, and it uses code/software more effectively to complete tasks.
Ethan Mollick@emollick · 创始人 · 1 天前沃顿商学院教授,AI 应用研究权威

我觉得现在的讨论有点过于偏向“AI的影响要经过很多年才会显现”,就像不久之前大家还都过于强调“AI会立刻改变经济”一样。

查看英文原文
I think the discourse is shifting too much towards "AI impacts will take many years to matter," just as they were too much towards "AI will instantly change the economy" not that long ago.
Min Choi@minchoi · 博主 · 1 天前AI 产品演示博主,专门展示新工具玩法

为什么不在美国办自己的人形机器人大赛?

引用 Eren Chen @ErenChenAI世界人形机器人运动会400米冠军跑步姿态的慢镜头展示。查看被引原帖 ↗
查看英文原文
why aren't we running our own Humanoid Robot Games in the US?
Gary Marcus@GaryMarcus · 博主 · 1 天前

为什么采用速度比预期慢?

"AI 采用速度滞后于 Sam 和其他 AI 领导者预测时间表的最大原因是,尽管进度看似迅速,AI 其实还没有变得足够好。"

引用 Steve Hou @stevehouSam Altman期望AI采纳速度快,但实际较慢的主因不是用户惯性,而是AI本身能力不足。直到今年因agents、harnesses和computer use技术突破,AI才在生产力上真正有意义地可用。查看被引原帖 ↗
查看英文原文
Why has adoption been slower than expected?

“biggest reason why AI adoption has lagged Sam and other AI leaders’ predicted timelines is that AI simply hadn’t gotten good enough despite the seemingly rapid pace of progress. “
Thinking Machines@thinkymachines · 公司官方 · 1 天前

今天我们推出 Tinker 研究基金,为开放权重模型的安全研究提供最高 5 万美元的额度。下面我们分享了一些我们感兴趣的项目方向;如果你在做安全研究项目,可能需要更多 Tinker 额度来加速,我们很想听听你的想法!

查看英文原文
Today, we are launching Tinker grants of up to $50,000 in credits for safety research on open-weight models. We share some project ideas that excite us below; if you’re working on a safety project that could be accelerated by additional Tinker credits, we want to hear from you!
Ethan Mollick@emollick · 创始人 · 1 天前沃顿商学院教授,AI 应用研究权威

我的AI大体上认同Arvind的看法:州级数据中心禁令对AI整体进展影响甚微。

如果你的关注点是AI发展、权力集中、未来应用,或是除“我附近有没有数据中心”之外的任何问题,那么这并不能替代政策。

引用 Arvind Narayanan @random_walker分析数据中心禁令对AI进展影响。州级一年暂停令仅延缓AI效率进步5-10小时。纽约完全禁建,考虑90%产能流向他处,也仅延缓不足一天。虽环保理由成立,但作为应对AI焦虑的手段效率极低。查看被引原帖 ↗
查看英文原文
My AI broadly agrees with Arvind's AI: state-level data center bans have very little impact overall on AI progress.

If your concern is AI development, concentration of power, future use, or anything else other than "is a datacenter near me," this is not a substitute for policy.
Gary Marcus@GaryMarcus · 博主 · 1 天前

我一直说的Klarna效应带来的那种遗憾感,现在正变得势不可挡。

引用 Hedgie @HedgieMarkets55%因AI裁员的公司已后悔,2/3在重新招聘。Klarna替代700客服后质量下跌,被迫重新招人。Forrester预测AI仅自动化6%工作。高管错判:看到编码、客服等显性工作,忽视机构知识、判断力等隐性价值。成本反而更高。查看被引原帖 ↗
查看英文原文
the form of regret i have long called the Klarna Effect is getting absolutely massive
The Rundown AI@TheRundownAI · 博主 · 1 天前百万订阅 AI 日报官方
连环推 ×6

2M+ 人阅读 The Rundown。每周越来越多人在行动。

我们的社区工作流中心收到了数千个投稿,展示各种人们用 AI 改进生活、工作或业务的方式。以下是上周的前 5 名:

查看英文原文
2M+ people read The Rundown. Every week, more of them are building.

We've had thousands of incredible submissions in our Community Workflow Hub on ways people are using AI to better their lives, work, or business.

Here are the top 5 from last week:
The AI wardrobe stylist

Cheyenne built an app (with zero coding experience) that styles complete outfits from clothes you already own, using ChatGPT and Lovable.

She photographed 16 pieces from her own closet, picked a pair of hot pink pants, and the app returned three full looks, each with an occasion and styling note.

Credit to Cheyenne Dominguez (
@cheyennepalma
):
app.therundown.ai/community/…
The shared eldercare log

Shawn and his two brothers were coordinating their dad's care across up to seven doctors a day, juggling Apple Notes, a shared calendar, and endless text threads.

He built a shared log with Claude Code and Codex that tracks medications, mood, and vitals, and showed which family member is covering each day.

Credit to Shawn Robison (
@shawnrobison
):
app.therundown.ai/community/…
The agent memory architect

Tony built a three-tier memory system so his AI agents remember context across sessions -- global, agent, and project memory, each scoped to who actually needs it.

Every new fact gets stored at the narrowest useful level, so agents stop starting every session cold and drowning in irrelevant context.

Credit to Tony Ojeda (
@tonyojeda3
):
app.therundown.ai/community/…
The freight load coordinator

Josh built a platform with Claude Code that sends load offers to trucking companies through each contact's preferred channel from a single screen.

The process happens hundreds of times a day, and the demo convinced his freight brokerage to start building the platform internally.

Credit to Thomas (Josh) Thephasdin:
app.therundown.ai/community/…
The family history detective

Florence used AI to investigate a mystery that went unsolved for decades: identifying the biological family of her maternal grandfather.

She had ChatGPT sort DNA matches, records, and old photos into confirmed facts vs. hypotheses -- treating every AI lead as something to verify, then messaging possible relatives in another language.

Credit to Florence Locheron:
app.therundown.ai/community/…
宝玉@dotey · 中文博主 · 1 天前宝玉,中文圈 AI 翻译与科普大 V

如果对 DeepSeek Harness 有兴趣的,可以看看,写的很好👍

引用 Zhenjia Zhou @zhenjiazhouDeepSeek 8 月发布了 DeepSeek Harness:一切都是插件。 这是过度工程吗?带着这个疑问我进行了一番深度探索 结论:今天看是过度工程,但这个方向并非没有意义。 zhenjia.dev/posts/the-deepse… #DeepSeek #AIAgent查看被引原帖 ↗
NVIDIA@nvidia · 公司官方 · 1 天前

探索 CUDA-X:
nvda.ws/3UnjP9h

查看英文原文
Explore CUDA-X:
nvda.ws/3UnjP9h
AshutoshShrivastava@ai_for_success · 博主 · 1 天前高频 AI 新闻与产品动态博主

Grok 4.7 要来了 🔥

查看英文原文
Grok 4.7 is going to be 🔥
Runway@runwayml · 公司官方 · 1 天前AI 视频生成公司 Runway

9 月 30 日,来旧金山和我们一起参加线下 hackathon,用 Runway 的 API 构建最棒的 agent、app 或创意工作流。这个全天活动将在 Masonic 举办,与 Runway AI Summit 同步进行。

不需要提案。不需要审批流程。只要在下午 5 点前搞出一个能工作的演示,就有机会赢得 25K 美元和 200K 额度。

查看英文原文
Join us in San Francisco on September 30th for an in-person hackathon to build the best agent, app or creative workflow with Runway's API. This one-day event will take place at the Masonic alongside our Runway AI Summit.

No pitch decks. No approval chains. Just a working demo by 5pm for your chance at $25K and 200K credits.
Gary Marcus@GaryMarcus · 博主 · 1 天前

不可思议。但真不靠谱。

如果真的达到吹的那样,采用速度应该快得多。

引用 Taylor Lorenz @TaylorLorenz我们都在时间表上太雄心勃勃了。即使有这项令人惊叹的技术,社会和经济的适应会更慢。查看被引原帖 ↗
查看英文原文
Incredible. But not reliable.

If it were all that it had been advertised to be, adoption would be far faster.
OpenAI Developers@OpenAIDevs · 公司官方 · 1 天前OpenAI 开发者平台官方

对 Codex 里的语音 agent 或 Alex 的设置有什么问题吗?评论区见。

查看英文原文
Got any questions about voice agent in Codex or Alex’s setup? Share them below.
elvis@omarsar0 · 博主 · 1 天前

@agentsky_dev 团队的很酷新发布。Agent Playground 让你在同一浏览器里用 Claude Code、Codex 和 DeepSeek 运行相同任务,并排比较时间、成本和 tokens。我经常跑同任务的 harness 测试,我觉得这是首次能在真正相同条件下比较 agents 的地方。很棒的地方在于你能更好地决定哪个 agent harness 最适合你的任务。

引用 Xiaoyin Qu @quxiaoyin同一任务在Claude Code和DeepSeek代理上成本差异巨大($150 vs $2)。推出AgentSky.dev作为"代理领域的OpenRouter",整合Claude Code、DeepSeek、Kimi等主流代理API。Agent Playground可在浏览器中并行测试多个代理,对比时间、成本和token消耗。查看被引原帖 ↗
查看英文原文
Very cool launch from the
@agentsky_dev
team. Agent Playground lets you give Claude Code, Codex, and DeepSeek the identical task in one browser and compare time, cost, and tokens side by side.

I run same-task harness tests constantly, and I think this is one of the first places agents can be compared under truly identical conditions. It's great because you can make a better decision about which agent harness is best for the desired task.
Ethan Mollick@emollick · 创始人 · 1 天前沃顿商学院教授,AI 应用研究权威

说清楚点,我觉得用AI辅助写作或者拿反馈啥的没啥问题(前提是规则允许),但你要是用得太多,就会模糊掉你自己独特的贡献。这事儿不是在所有情况下都重要,但在很多场合确实关键。

查看英文原文
To be clear, I don't think there is anything wrong with getting AI help in writing or feedback or whatever (assuming the rules say so), but if you use it to do too much, it starts to obscure your unique contribution. It may not matter in every case, but it matters in many of them
The Rundown AI@TheRundownAI · 博主 · 1 天前百万订阅 AI 日报官方

完全由 AI 设计的病毒……还真能用?

斯坦福的模型写出了成千上万的基因组。其中16个在大肠杆菌中存活下来,杀死了天然病毒无法消灭的细菌:

查看英文原文
A virus designed entirely by AI... that actually works?

Stanford's model wrote thousands of genomes. 16 came alive in E. coli and killed bacteria that natural viruses couldn't:
Pika@pika_labs · 公司官方 · 1 天前AI 视频生成公司 Pika

在 dev.pika.art/?utm_source=x&u… 成为会员吧

查看英文原文
Become a member at
dev.pika.art/?utm_source=x&u…
Mistral AI@MistralAI · 公司官方 · 1 天前法国明星大模型公司

控制力正在成为企业采用AI的关键议题。全球范围内,尤其是在受监管的行业里,组织既想要AI带来的好处,又不想失去对敏感数据和系统的控制。

查看英文原文
Control is becoming the defining topic in enterprise AI adoption. Across the world, and especially in regulated industries, organizations want the benefits of AI without giving up control over their sensitive data and systems.
Thinking Machines@thinkymachines · 公司官方 · 1 天前

了解更多和申请:thinkingmachines.ai/news/saf…

查看英文原文
Learn more and apply here:
thinkingmachines.ai/news/saf…
el.cine@EHuanglu · 博主 · 1 天前

AI让吸血鬼日记变得更疯狂了

查看英文原文
AI made the vampire diaries even crazier
el.cine@EHuanglu · 博主 · 1 天前

中国 AI 机器人在 88 秒内赢得了 100 米障碍赛

查看英文原文
China’s AI robot just won a 100m obstacle race in 88 seconds
Linus ✦ Ekenstam@LinusEkenstam · 博主 · 1 天前

Soul Bone 第一集

我们已经越过不归路,成本目前还比较高,但最终会低到争论都没意义的程度。

📹 Bilibili 用户 wuhu 制作

查看英文原文
Soul Bone Episode 1

we’ve passed the point of no return, costs are still relatively high, but that will become so low it will be meaningless to even argue.

📹 by Bilibili user wuhu.
🚨 AI News | TestingCatalog@testingcatalog · 博主 · 1 天前专挖 AI 产品未发布新功能的爆料号

在这里看演示

引用 @blankspeakerSpaceXAI 推出 Grok Bot 可分享模板功能。用户可让机器人自动打包并生成分享链接,他人即可导入。方便用户创建和分享自定义机器人。查看被引原帖 ↗
查看英文原文
See the demo over here
elvis@omarsar0 · 博主 · 1 天前

推荐阅读,真的是个不错的想法。

围绕工具调用和代码执行,一些有趣的 harness 设计开始出现。RLM 就是其中之一。这个推测性程序化工具调用方法(来自 RLM 作者)也是。

在 harness 层有很多方法来获得效率提升。

Harnesses 让 agents 等待:模型流式传输一块代码,其中的工具调用只在生成完成后运行。sPTC 针对环境的副本提前启动安全调用,所以工具延迟与 token 生成重叠,而不是叠加。错误的猜测会被丢弃。到目前为止是 1 到 1.2x 的速度提升。非常有前景。

引用 alex zhang @a1zhang推出推测性编程工具调用(sPTC)!通用技术可在代码生成期推测工具调用并提前排队,以与令牌生成和REPL执行时间重叠。查看被引原帖 ↗
查看英文原文
Recommended reading and a really cool idea.

There are a lot of interesting harness designs that are starting to emerge around tool calling and code execution. RLM is one of them. But so is this Speculative Programmatic Tool Calling approach (from the same author of RLM).

There are plenty of ways to gain efficiencies at the harness layer.

Harnesses make agents wait: the model streams a block of code, and tool calls inside it only run once generation finishes. sPTC launches the safe calls early against a copy of the environment, so tool latency overlaps with token generation instead of adding on top of it. Bad guesses get thrown away. So far it's 1 to 1.2x speedup. Very promising.
Gorden Sun@Gorden_Sun · 中文博主 · 1 天前中文圈高频 AI 资讯与开源项目博主

Apodex AI发布Apodex 1.1 :专门处理复杂专业工作的Agent

可以分解复杂任务,并行调用多个子Agent,且允许任务过程中随时调整需求或插入新要求。

多项评分位于第一梯队,尤其适合执行科学研究、金融分析和深度搜索任务。

同时开源了Agent框架和本地版模型。

引用 Apodex @Apodex_AIApodex 1.1是用于复杂专业工作的前沿agentic模型,支持异步多agent协作和持续任务整合。开源发布FrontierAgent研究工作台和mini轻量版本,可在线或本地部署。查看被引原帖 ↗
Runway@runwayml · 公司官方 · 1 天前AI 视频生成公司 Runway

立即申请:

hackathon.runway.com/

查看英文原文
Apply now at:
hackathon.runway.com/
LlamaIndex 🦙@llama_index · 公司官方 · 1 天前

我们在 SF 举办的第二场创始人晚宴,由 @jerryjliu0 和 @GuangyuRobert 联合主办,在 @Fundamental——@tryshortcutai 背后的团队。讨论 AI 时代现有的护城河。

前沿实验室正在从模型 API 转向垂直 agent——ChatGPT Health、Claude for Legal。

那护城河现在在哪儿呢?

✅ Agent engineering
✅ Infra optimization
✅ Domain evals and data
✅ Workflow expertise
✅ GTM and brand

如果你是在生产环境做 agent 的创始人或 CTO,想了解其他团队怎样维持护城河,欢迎申请席位 👉️
luma.com/llamai-8hry

查看英文原文
Our 2nd founder dinner in SF co-hosted by
@jerryjliu0
and
@GuangyuRobert
at
@Fundamental
- the team behind
@tryshortcutai
Talking about existing moats in the AI era.

Frontier labs are moving past model APIs into vertical agents - ChatGPT Health, Claude for Legal.

So where does the moat sit now?

✅ Agent engineering
✅ Infra optimization
✅ Domain evals and data
✅ Workflow expertise
✅ GTM and brand

If you're a founder or CTO shipping agents in production and want to know what other teams are doing to maintain their moat. Request a seat 👉️
luma.com/llamai-8hry
The Rundown AI@TheRundownAI · 博主 · 1 天前百万订阅 AI 日报官方

The Rundown AI 每天精选社区的顶级工作流。分享你的作品,2M+ 读者可能会看到:app.therundown.ai/community

查看英文原文
Every day we feature a top community workflow in The Rundown AI newsletter.

Share what you've built and 2M+ readers might see it next:
app.therundown.ai/community
Kol Tregaskes@koltregaskes · 博主 · 1 天前

Gemini 4 在筹备中。

引用 Bedros Pamboukian @bedros_pGoogle 即将发布 Gemini 4。约 2-3 天前出现 120+ 条新提及(之前为 0),不到 24 小时前又新增 6 条。传播已开始。查看被引原帖 ↗
查看英文原文
Gemini 4 is cooking.
Kol Tregaskes@koltregaskes · 博主 · 1 天前

我觉得任何 AI 管不了的东西,AI 也防不了。我现在想按这个思路升级安全防护。你在改进安全吗?

查看英文原文
I think anything that can't be managed by an AI can't defend against an AI. I'm currently looking to upgrade my security with that in mind. Are you improving your security?
Cohere@cohere · 公司官方 · 1 天前加拿大企业级大模型公司 Cohere 官方

Supertranscribe = 牛X。

Cohere 现在是 @superwhisper 新手入门的默认本地模型。口述你的想法,瞬间搞定,全在本地设备上 - 无需 wifi。

享受 Transcribe x @Nativ_AI 的组合,@RayFernando1337 🤝

查看英文原文
Supertranscribe = S-tier.

Cohere is now the default local model for
@superwhisper
onboarding. Dictate your thoughts instantly, and keep it all on your device - no wifi required.

Enjoy the Transcribe x
@Nativ_AI
combo,
@RayFernando1337
🤝
Together AI@togethercompute · 公司官方 · 1 天前

我们与 @yutori_ai 共同举办 CUA Night,这周四:一个关于计算机使用 agents 的食物、饮料和对话之夜。

如果你在研究、模型、数据、RL 环境、推理等领域构建,不要错过:luma.com/aqmwya3d

查看英文原文
We're co-hosting CUA Night with
@yutori_ai
this Thursday: a night of food, drinks, and conversation about all things computer-use agents.

If you're building across research, models, data, RL environments, inference, and more, don't miss it:
luma.com/aqmwya3d

本站由 Jedee杰哥 打造 · 公众号「Jedee杰哥」每早送 AI 日报

姊妹站:𝕏 简中账号数据榜单 · X 关注 @jedeeai · RSS 订阅 · AI 日报 · 历史归档