16个月内，单位能耗的智能提升了18倍。
引用 Avanika Narayan @Avanika15完全同意 @amasad 和 @JonSaadFalcon。我的研究表明智能效率（每瓦特智能）正在快速改善，我们绝对*不*需要数据中心规模的计算来运行 AGI！研究链接见下方评论 👇查看被引原帖 ↗
查看英文原文
18x improvement in intelligence per joule in 16 months.
el.cine@EHuanglu · 博主 · 1 天前
哇...AI现在可以用一条线精确控制摄像头了
查看英文原文
wow.. AI now can control camera precisely with a line
这是怎样在 Codex 中启用 1M token 上下文窗口的方法，针对 GPT-5.6 Sol。虽然我们已经把 Codex 中的上下文限制调优到性能和成本的最优点，但这是常见的需求，所以我们把方法记录下来。
更大的上下文窗口让 Codex 能保留更多代码、工具输出和对话历史，然后再对旧内容做总结。你需要一个支持这个的模型。比如 GPT-5.6 Sol，就有记录的 1,050,000 token 窗口。
打开 ~/.codex/config.toml，在最上面加上或更新这些设置，在任何 [section] 标题前：
```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```
第一个设置选择模型。第二个告诉 Codex 用百万 token 上下文预算。第三个在 900,000 token 时自动启动历史压缩，留有余量。保存后重启 Codex 客户端并开启新会话。
要在单个 CLI 会话中试试这个配置，不改变你的默认设置：
```
codex -m gpt-5.6-sol \\
-c model_context_window=1000000 \\
-c model_auto_compact_token_limit=900000
```
祝你玩得开心，但也要知道我们对默认设置的调优是很仔细的！
查看英文原文
Here is how to enable a 1M-token context window in Codex for GPT-5.6 Sol.
Even though we have tuned the context limit in Codex to be set optimally when it comes to performance and cost, this is a common ask, so here it is documented.
A larger context window lets Codex retain more code, tool output, and conversation history before summarizing older material. You need a model that supports it. And GPT-5.6 Sol, for example, has a documented 1,050,000-token window.
Open ~/.codex/config.toml and add or update these settings at the top level, before any [section] headers:
```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```
The first setting selects the model. The second tells Codex to use a one-million-token context budget. The third starts automatic history compaction around 900,000 tokens, leaving some headroom. Restart Codex client and start a new session after saving.
To try the configuration for a single CLI session without changing your defaults:
```
codex -m gpt-5.6-sol \
-c model_context_window=1000000 \
-c model_auto_compact_token_limit=900000
```
Have fun, but also know that we tuned the default carefully!
Even though we have tuned the context limit in Codex to be set optimally when it comes to performance and cost, this is a common ask, so here it is documented.
A larger context window lets Codex retain more code, tool output, and conversation history before summarizing older material. You need a model that supports it. And GPT-5.6 Sol, for example, has a documented 1,050,000-token window.
Open ~/.codex/config.toml and add or update these settings at the top level, before any [section] headers:
```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```
The first setting selects the model. The second tells Codex to use a one-million-token context budget. The third starts automatic history compaction around 900,000 tokens, leaving some headroom. Restart Codex client and start a new session after saving.
To try the configuration for a single CLI session without changing your defaults:
```
codex -m gpt-5.6-sol \
-c model_context_window=1000000 \
-c model_auto_compact_token_limit=900000
```
Have fun, but also know that we tuned the default carefully!
Codex
✅ 几乎 100% 可靠
✅ 偶尔会重置
✅ 开源
✅ (将支持 Astra)
查看英文原文
Codex
✅ Almost 100% reliable
✅ Occasional resets
✅ Open-source
✅ (will have Astra)
✅ Almost 100% reliable
✅ Occasional resets
✅ Open-source
✅ (will have Astra)
Codex 中 GPT-5.6 Sol 1M。这以前只对 API 密钥有效，但我们刚刚打开开关，现在也能通过 ChatGPT 账户使用了。
同样的警告适用，当前上下文长度是默认设置是有原因的，我们已经优化到接近完美。但你做主！
引用 Tibo @thsottiaux在Codex中为GPT-5.6 Sol启用100万token上下文窗口。编辑~/.codex/config.toml，设置model、model_context_window和model_auto_compact_token_limit参数，让Codex保留更多代码、工具输出和对话历史。也可通过CLI参数临时配置。查看被引原帖 ↗
查看英文原文
GPT-5.6 Sol 1M in Codex. This used to only work for API keys, but we just flipped the switch and works for usage through ChatGPT accounts now too.
The same warning applies, there is a reason the current context length is the default, we have tuned it to ~perfection. But you do you!
The same warning applies, there is a reason the current context length is the default, we have tuned it to ~perfection. But you do you!
Unitree@UnitreeRobotics · 公司官方 · 1 天前
宇树新机器人预告：“超人”突破人类极限🥳
原地跳高2米，最高速度12.66米/秒（腿长0.85米）
超越全球所有人类的原地跳高和跑步速度纪录
这台新机器开发时间刚刚超过三个月，未来几个月还有巨大的提升空间。
查看英文原文
Unitree New Robot Preview: “Superman” Breaking the Limits of Humanity🥳
Standing high jump 2 m, top speed 12.66 m/s (0.85 m leg length)
Surpassing the standing high jump and running speed records of all humans around the world
This new machine has only been in development for a little over three months, with significant room for further improvement in the coming months.
Standing high jump 2 m, top speed 12.66 m/s (0.85 m leg length)
Surpassing the standing high jump and running speed records of all humans around the world
This new machine has only been in development for a little over three months, with significant room for further improvement in the coming months.
这就是我现在的工作流了：
实时研究 → Grok Bot
规划编排 → Grok Bot
日常编码/调试 → Grok Build + Grok 4.6
写测试跑测试 → Grok Build + Grok 4.6
复杂编码/调试 → GPT-5.6 Sol
前端 → Fable 5
收藏一下。
查看英文原文
This is literally my new workflow now:
Realtime Research → Grok Bot
Planning & Orchestration→ Grok Bot
Day-to-day Coding/Debug → Grok Build + Grok 4.6
Write & Run Tests → Grok Build + Grok 4.6
Complex Coding/Debug → GPT-5.6 Sol
Frontend → Fable 5
Bookmark this.
Realtime Research → Grok Bot
Planning & Orchestration→ Grok Bot
Day-to-day Coding/Debug → Grok Build + Grok 4.6
Write & Run Tests → Grok Build + Grok 4.6
Complex Coding/Debug → GPT-5.6 Sol
Frontend → Fable 5
Bookmark this.
AI 去水印工具在 GitHub 上达到 1.1 万星了。
人们真的很讨厌这些水印。我也完全理解他们。
引用 AI Breakfast @AiBreakfastAnthropic上周公开了Claude文本水印的工作原理。随后一个MIT许可的开源项目获得10,000星，可移除Claude、SynthID-Text和OpenAI的水印，以及图片和PDF中的C2PA和EXIF标记。查看被引原帖 ↗
查看英文原文
watermarks-remover for AI content reaches ~11k stars on github.
People really hate the watermarks. And i fully understand them.
People really hate the watermarks. And i fully understand them.
Tibo 公开讲了怎样在 Codex 里用 GPT-5.6 Sol 把 context window 扩到 100 万 tokens，这我是真没想到。
反正现在 Codex 的 1m context 对所有人都开放了：
引用 Tibo @thsottiauxCodex 现支持 GPT-5.6 Sol 的 100 万 token 上下文窗口。修改 config.toml 配置模型和压缩限制，可保留更多代码和对话历史，需重启后生效。查看被引原帖 ↗
查看英文原文
Tibo publicly explaining how anyone could extend the context window to 1 million tokens in Codex with GPT-5.6 Sol was not on my bingo card.
Anyway, 1m context for everyone in codex:
Anyway, 1m context for everyone in codex:
OpenAI在全力对抗Anthropic。
一边嘲讽Anthropic，一边拼了命把自己的产品做得比Anthropic更好。OpenAI开发者大会还有六周，应该是在筹划什么大发布。
引用 Andrew Ambrosino @ajambrosino接下来 6 周已定查看被引原帖 ↗
查看英文原文
OpenAI is going all out to confront Anthropic.
They're mocking Anthropic on the one hand, and on the other, doing everything they can to make their products better than Anthropic's. OpenAI's developer day is in six weeks, and they're presumably planning a big reveal or launch.
They're mocking Anthropic on the one hand, and on the other, doing everything they can to make their products better than Anthropic's. OpenAI's developer day is in six weeks, and they're presumably planning a big reveal or launch.
Bilawal Sidhu@bilawalsidhu · 博主 · 1 天前
连环推 ×3二维时刻冻结在时间里，变成 3D 全息影像，留在真实世界发生的确切位置。地理空间记忆宫殿活起来了。
兴奋的小鹿吓了鹿妈妈一跳，她随后放松了 :) 想象一下你整个相册都像这样活起来的样子？
由 500 多张照片组成的整个场地的空间记忆宫殿
想深入钻进这个坑吗？用自己的照片和网上的图片重建世界——看这个深度讲解（我 YouTube 频道也有），能帮你理解计算机视觉研究的这一脉络
查看英文原文
2d moments frozen in time. Turned into 3d holograms. Left exactly where they happened in the real world. The geospatial memory palace coming to life.
Excited baby deer startles mama deer momentarily, who then chills out after :) Can you imagine when your entire camera roll comes to life like this?
The spatial memory palace for an entire venue comprised of 500+ captures
If you want to go deeper down this rabbit hole of reconstructing the world from your own captures and random internet photos - this deep dive (also on my yt channel) will get you grounded in this lineage of computer vision research
Excited baby deer startles mama deer momentarily, who then chills out after :) Can you imagine when your entire camera roll comes to life like this?
The spatial memory palace for an entire venue comprised of 500+ captures
If you want to go deeper down this rabbit hole of reconstructing the world from your own captures and random internet photos - this deep dive (also on my yt channel) will get you grounded in this lineage of computer vision research
卧槽，原来现在还是能以每个月 99 美元的价格，订阅 Grok Harvey 这个级别的会员，这个超值！
你相当于一下获得了：
1. 价值 300 美元的 Grok Harvey（包括 Grok Build 权限）
2. 价值 200 美元的 Cursor Ultra 会员，还有 Fable 5 的额度。
3. Grok Bot 的独立额度（这三个额度都是独立的）
4. Grok Image 和 Video（图像与视频模型）的额度
5. 会自动送你一份 Twitter Premium+ 订阅
折合下来只要 700 块钱一个月，简直夯爆了！
尤其是 Grok 后面还有更强的模型要发布。
可以看一下具体的使用路径：
1. 准备 Grok 账号：先登录一个 Grok 账号。
前提条件：如果你之前是用 Twitter（X）登录并订阅的，需要先取消 Twitter 订阅（取消后不会立即生效，所以我建议直接注册申请一个新的 Grok 账号。虽然新账号没办法共享你原有的 Twitter Premium，但这次订阅）。
如果是用旧账户或者 Twitter 授权登录，操作起来会有点麻烦，所以强烈建议用新账户注册。
2. 订阅促销活动：
先去第一个链接(
grok.com/supergrok/targeted-…
)，订阅获取两个月的 Super Grok 促销。
订阅成功后，再去下面第二个地址(
grok.com/supergrok/targeted-…
)，订阅连续三个月 99 美元的 Super Grok Harvey。
注意：后续记得及时取消订阅，否则到期后会恢复成 300 美元/月。
3. 绑定 Cursor 会员：
下载并登录 Grok Bot，登录时有个选项会要求选择 Cursor 授权登录。
这个时候，只要你有一个相同邮箱的 Cursor 账号（如果没有，就用这个邮箱注册一个），系统就会自动给你充值 Cursor 的 Ultra 会员。
github.com/yetone/cumora
Cumora 已开源
引用 yetone @yetone还有一件事 cumora.ai查看被引原帖 ↗
查看英文原文
github.com/yetone/cumora
Cumora 已开源
Cumora 已开源
你说得没错 @GergelyOrosz，我们搞错了。提醒邮件没有发给这个用户。他已经退款了，但这不是我们想要的运营方式。我们正在全面升级我们的支持服务。
引用 Gergely Orosz @GergelyOroszPerplexity过去6-12个月一再令人失望。作者曾因其出色搜索功能大力推荐，还为付费用户推广过（未获报酬），但Perplexity的后续做法令人失望👎查看被引原帖 ↗
查看英文原文
You’re right
@GergelyOrosz
, we got this wrong. Reminder email didn’t go out to this user. He has been refunded, but this is not how we want to operate. We are upgrading our support across the board.
@GergelyOrosz
, we got this wrong. Reminder email didn’t go out to this user. He has been refunded, but this is not how we want to operate. We are upgrading our support across the board.
这是我对 Qwen 3.8 27B 的评测 - 我记不得上次在自己电脑上运行本地模型有这么多乐趣是什么时候了
simonwillison.net/2026/Aug/1…
查看英文原文
Here’s my review of Qwen 3.8 27B - I can't remember the last time I've had this much fun playing with a local model that runs on my own computers
simonwillison.net/2026/Aug/1…
simonwillison.net/2026/Aug/1…
通过购买房屋和土地为 AGI 做准备，必须进行多元化投资。
但我旅行的频率比以往任何时候都多，所以这两者可以结合！
引用 Austin B. @avstinbarnesWow levels has a house? I thought you were nomadmaxxing mane查看被引原帖 ↗
查看英文原文
Preparing for AGI by owning houses and land, have to diversify
But I travel more than I ever did, so the two can go together!
But I travel more than I ever did, so the two can go together!
Pi 两位作者的见解：
1. 代码即真相，代码不需要记忆系统，不需要 RAG，模型很擅长理解代码结构
2. Bash 工具足够用，Bash 类似于编程语言，可以任意组合；大部分时候没必要 MCP，skill + 脚本足够。
引用 Pi @pidotdev来自Vienna People of Pi的问候🌞 @badlogicgames和@mitsuhiko的周日冥想：代码是真理；Bash是你所需的一切；构建上下文高效的工具。查看被引原帖 ↗
防守方能看到未来，但升级网络安全实践的时间窗口很窄。
关键是夯实基础，用上最好的AI工具。
这是我们OpenAI在做的事儿，也是其他机构可以从这开始的地方：
blog.gregbrockman.com/the-de…
查看英文原文
defenders can see the future, and have a narrow window to uplevel their cybersecurity practices now.
key is to uplevel fundamentals and apply the best AI tools.
what we’re doing at OpenAI, and where other organizations can start:
blog.gregbrockman.com/the-de…
key is to uplevel fundamentals and apply the best AI tools.
what we’re doing at OpenAI, and where other organizations can start:
blog.gregbrockman.com/the-de…
天哪，我完全错过了最后那句话！Astra 这周要发吗?!
引用 Tibo @thsottiauxCodex ✅ 几乎100%可靠 ✅ 偶尔重置 ✅ 开源 ✅ (将支持 Astra)查看被引原帖 ↗
查看英文原文
Holy, I completely missed that last sentence! Is Astra dropping this week?!
我去！这个太有意思了。
谁不想在等 AI 完成任务的时候玩上一把德州并且把自己的 Token 输光呢
引用 Hours @hours拿我的token配额赌一把查看被引原帖 ↗
几秒钟创建 Hitchcock Zoom 效果 🎬
查看英文原文
Create Hitchcock Zoom in Seconds 🎬
我们对 GLM 5.3 的网络安全能力做了评测。这是新的开放前沿。考虑到成本更低，我预计这对防御性安全工作会很有帮助。比如说，你可以至少频繁运行 deepsec.sh 3 倍！
引用 Vercel Developers @vercel_devGLM 5.3即将登陆AI Gateway。这是DeepsecBench上评分最高的开源模型，成本仅为部分得分相近专有模型的三分之一。查看被引原帖 ↗
查看英文原文
We ran evals on GLM 5.3 cybersecurity capabilities. It's the new open frontier.
Given its lower costs, I expect this to be a boon for defensive security work. e.g., it means you can run
deepsec.sh
at least 3× more often!
Given its lower costs, I expect this to be a boon for defensive security work. e.g., it means you can run
deepsec.sh
at least 3× more often!
我Claude 账号前天被封了
我申诉说我啥也没干
就是日常编码和写内容，请给我解封
过了一天就给我解封了
但是我的订阅掉了，我又写了一封邮件要求他们补我3个月 20x的会员
我是不是有点得寸进尺？
不到两年，现在从本地电脑上运行的开源 AI 视频生成器 MiniMax H3 生成的视频——一只水獭在飞机上用笔记本——的质量都已经这样了（还有声音）。
生成花了大概3分钟。进度确实飞快。
引用 Ethan Mollick @emollick一方面，这些视频明显不如最先进的AI文本到视频生成效果好，看起来像2022年的产物。另一方面，它完全在个人电脑上使用开源AI视频生成工具离线完成，是一项新功能。查看被引原帖 ↗
查看英文原文
Less than two years later, this is the quality of the video of an otter using a laptop on an airplane that I get (with sound) from an open weights AI video generator, MiniMax H3, running entirely on my local computer.
It took about 3 minutes to generate. Rapid progress, indeed.
It took about 3 minutes to generate. Rapid progress, indeed.
不出所料，Claude 拒绝了安装去水印插件。这事儿估计没什么好说的了，自己就很说明问题。
引用 Chubby♨️ @kimmonismusAI内容去水印工具在GitHub获得约11000个星标。人们真的很讨厌水印。我完全理解他们。查看被引原帖 ↗
查看英文原文
Unsurprisingly, Claude refuses to install the watermark-removal plugin.
There is probably nothing more to be said about that; it speaks for itself.
There is probably nothing more to be said about that; it speaks for itself.
连环推 ×2X 的一个长期烦恼是你无法导出书签，你只能滚动、滚动、滚动。
所以我让 GPT-5.6 Sol in Codex 来做这件事，它接管了我的 Chrome，现在我有了自 2014 年以来全部 5,302 个书签以及各种数据
我还让它找出其中的亮点（帖子？）
这是AI精选的内容：
1. 抹香鲸可能已经集体学会了如何躲避捕鲸者。
几年内，捕鲸者鱼叉的成功率据报下降了 58%，这表明关于攻击的信息在鲸群中被共享。这是一个非人类文化如何在压力下适应的惊人例子。
x.com/SteveStuWill/status/14…
2. 一把 50 万到 30 万年前的手斧是故意围绕一个化石贝壳制作的。
贝壳位于工具的中心附近，提出了其制造者可能注意到并有意纳入它的可能性。这可能代表了一种令人惊讶的早期美学意图形式。
x.com/AlisonFisk/status/1776…
3. 智能可能在整个生物学中的分布要远比我们通常认识的要广泛。
这个帖子探讨了 Michael Levin 的工作，涉及细胞和生物系统集体智能被低估的方面——以及如果我们停止将智能仅视为大脑内发生的事情，会产生的观点转变。
x.com/kasratweets/status/178…
4. 维基百科成功了，而七个类似的在线百科全书项目失败了。
它们大约在同一时间出现并追求类似的目标，但只有一个成为互联网的基础部分。这篇论文研究了为什么会这样的制度和组织原因。
x.com/OwainEvans_UK/status/1…
5. 美观的数据可视化更容易被信任。
在四项研究中，人们对更美观的图表放置了更多信任——即使美观不应该影响基础信息的可信度。这对任何创建或消费研究的人来说都是有用的警告。
x.com/Mark_A_Thornton/status…
6. 选择正确的问题可能是研究中最具杠杆作用的技能。
这个长帖子和附随的 Cell 文章研究了科学家和工程师如何选择问题，包括用于决定什么真正值得追求的实际决策树。
x.com/mfgrp/status/177844027…
7. 一个人花了五年时间重新创作了 Alinea 食谱中的全部 107 道食谱。
故事以烹饪痴迷开始，但后来成为了关于掌握、深思熟虑练习以及消费者如何奇妙地成为创造者的一课。
x.com/kevinleeme/status/1284…
8. 第一次世界大战开始后不久，《金融时报》要求读者预测战争如何改变欧洲地图。
这被当作一个游戏呈现。有现金奖励。这是一个非常奇怪和具有启发性的工件，说明人们在灾难真正规模被理解之前如何感知灾难。
x.com/khalafroula/status/202…
9. 查尔斯·达尔文曾经收到过对《物种起源》的科学批评——并写了一个了不起的回复。
朱莉亚·加莱夫解释了为什么这可能是她最喜欢的科学史上的信件，以及它对达尔文知识品格的启示。
x.com/juliagalef/status/1341…
10. 龙与地下城包含从其最早起源继承的技术债务。
它的时间和距离尺度从未完全同步，因为继承自微型战争游戏《龙与地下城》的规则。这种不匹配也可能有助于解释为什么现代龙与地下城角色感觉像超级英雄。
x.com/mikemearls/status/1878…
11. 前卫香水是一种可访问的实验艺术形式。
有一个完整的奇异、概念性、独立生产的香水世界——样品足够便宜，以至于它可能是探索完全陌生的美学媒介的最简单方式之一。
x.com/add_hawk/status/135707…
12. ShadeMap 模拟世界上每座山、每栋建筑和每棵树的阴影。
你可以选择任何位置、日期和时间，并看到阳光和阴影如何在景观中移动。它既实用又令人奇异地着迷。
x.com/shademap/status/179624…
13. Whole Earth Index 是一个几乎完整的、可自由访问的 Whole Earth 出版物存档。
Whole Earth Catalog 是反文化和早期技术想象中最有影响力的文件之一。现在该世界的大部分内容都可以在线搜索。
x.com/longnow/status/1712877…
14. 一个项目通过 54 件建筑作品探索了全部 54 个非洲国家。
每个项目旨在反映不同国家的文化身份和历史，创建了一个不同寻常的雄心勃勃的非洲建筑和设计视觉调查。
x.com/omar_degan/status/1347…
15. Infinite Craft 是一个无尽的组合发现游戏。
你从几个基本元素开始并组合它们
查看英文原文
A long annoyance with X is that you can't export your bookmarks, all you can do is scroll & scroll & scroll.
So I asked GPT-5.6 Sol in Codex to do it, and it took over my Chrome and now I have all 5,302 back to 2014 with all sorts of data
I also asked it to find gems (thread?)
Here are the gems the AI picked:
1. Sperm whales may have collectively learned how to evade whalers.
Within a few years, the success rate of whalers’ harpoons reportedly fell by 58%, suggesting that information about the attacks was being shared among whales. An astonishing example of nonhuman culture adapting under pressure.
x.com/SteveStuWill/status/14…
2. A 500,000–300,000-year-old handaxe was deliberately crafted around a fossil shell.
The shell sits near the center of the tool, raising the possibility that its maker noticed and intentionally incorporated it. It may represent a startlingly early form of aesthetic intention.
x.com/AlisonFisk/status/1776…
3. Intelligence may be far more distributed throughout biology than we generally recognize.
This thread explores Michael Levin’s work on the underappreciated collective intelligence of cells and biological systems—and the perspective shifts that follow if we stop treating intelligence as something that happens only inside brains.
x.com/kasratweets/status/178…
4. Wikipedia succeeded while seven similar online encyclopedia projects failed.
They appeared at roughly the same time and pursued similar goals, but only one became a foundational part of the internet. This paper looks for the institutional and organizational reasons why.
x.com/OwainEvans_UK/status/1…
5. Beautiful data visualizations are more likely to be trusted.
Across four studies, people placed more trust in more aesthetically pleasing graphs—even when beauty shouldn’t have affected the credibility of the underlying information. A useful warning for anyone who creates or consumes research.
x.com/Mark_A_Thornton/status…
6. Choosing the right problem may be the highest-leverage skill in research.
This long thread and accompanying Cell article examine how scientists and engineers select problems, including practical decision trees for deciding what is actually worth pursuing.
x.com/mfgrp/status/177844027…
7. One man spent five years recreating all 107 recipes in the Alinea cookbook.
What begins as a story of culinary obsession becomes a lesson about mastery, deliberate practice, and the strange process by which a consumer becomes a creator.
x.com/kevinleeme/status/1284…
8. Shortly after World War I began, the Financial Times asked readers to predict how the war would change the map of Europe.
It was presented as a game. There were cash prizes. It is a deeply strange and revealing artifact of how people perceived a catastrophe before its true scale was understood.
x.com/khalafroula/status/202…
9. Charles Darwin once received a scientific critique of On the Origin of Species—and wrote a remarkable reply.
Julia Galef explains why it may be her favorite letter in the history of science and what it reveals about Darwin’s intellectual character.
x.com/juliagalef/status/1341…
10. Dungeons & Dragons contains technical debt inherited from its earliest origins.
Its time and distance scales never quite synchronized because of rules inherited from the miniature-wargame Chainmail. That mismatch may also help explain why modern D&D characters feel like superheroes.
x.com/mikemearls/status/1878…
11. Avant-garde perfume is an accessible form of experimental art.
There is an entire world of strange, conceptual, independently produced perfume—and samples are inexpensive enough that it may be one of the easiest ways to explore a completely unfamiliar aesthetic medium.
x.com/add_hawk/status/135707…
12. ShadeMap simulates every mountain, building, and tree shadow in the world.
You can choose any location, date, and time and see how sunlight and shadow move across the landscape. It is both practically useful and oddly mesmerizing.
x.com/shademap/status/179624…
13. The Whole Earth Index is a nearly complete, freely accessible archive of Whole Earth publications.
The Whole Earth Catalog was one of the most influential documents of the counterculture and early technological imagination. Now much of that world is searchable online.
x.com/longnow/status/1712877…
14. One project explored all 54 African countries through 54 works of architecture.
Each project was intended to reflect the cultural identity and history of a different nation, creating an unusually ambitious visual survey of African architecture and design.
x.com/omar_degan/status/1347…
15. Infinite Craft is an endless game of combinatorial discovery.
You begin with a few basic elements and combine them to create new concepts, which can then be combined again. It is a wonderfully compact demonstration of how generative systems can produce genuine exploration and surprise.
x.com/nealagarwal/status/175…
So I asked GPT-5.6 Sol in Codex to do it, and it took over my Chrome and now I have all 5,302 back to 2014 with all sorts of data
I also asked it to find gems (thread?)
Here are the gems the AI picked:
1. Sperm whales may have collectively learned how to evade whalers.
Within a few years, the success rate of whalers’ harpoons reportedly fell by 58%, suggesting that information about the attacks was being shared among whales. An astonishing example of nonhuman culture adapting under pressure.
x.com/SteveStuWill/status/14…
2. A 500,000–300,000-year-old handaxe was deliberately crafted around a fossil shell.
The shell sits near the center of the tool, raising the possibility that its maker noticed and intentionally incorporated it. It may represent a startlingly early form of aesthetic intention.
x.com/AlisonFisk/status/1776…
3. Intelligence may be far more distributed throughout biology than we generally recognize.
This thread explores Michael Levin’s work on the underappreciated collective intelligence of cells and biological systems—and the perspective shifts that follow if we stop treating intelligence as something that happens only inside brains.
x.com/kasratweets/status/178…
4. Wikipedia succeeded while seven similar online encyclopedia projects failed.
They appeared at roughly the same time and pursued similar goals, but only one became a foundational part of the internet. This paper looks for the institutional and organizational reasons why.
x.com/OwainEvans_UK/status/1…
5. Beautiful data visualizations are more likely to be trusted.
Across four studies, people placed more trust in more aesthetically pleasing graphs—even when beauty shouldn’t have affected the credibility of the underlying information. A useful warning for anyone who creates or consumes research.
x.com/Mark_A_Thornton/status…
6. Choosing the right problem may be the highest-leverage skill in research.
This long thread and accompanying Cell article examine how scientists and engineers select problems, including practical decision trees for deciding what is actually worth pursuing.
x.com/mfgrp/status/177844027…
7. One man spent five years recreating all 107 recipes in the Alinea cookbook.
What begins as a story of culinary obsession becomes a lesson about mastery, deliberate practice, and the strange process by which a consumer becomes a creator.
x.com/kevinleeme/status/1284…
8. Shortly after World War I began, the Financial Times asked readers to predict how the war would change the map of Europe.
It was presented as a game. There were cash prizes. It is a deeply strange and revealing artifact of how people perceived a catastrophe before its true scale was understood.
x.com/khalafroula/status/202…
9. Charles Darwin once received a scientific critique of On the Origin of Species—and wrote a remarkable reply.
Julia Galef explains why it may be her favorite letter in the history of science and what it reveals about Darwin’s intellectual character.
x.com/juliagalef/status/1341…
10. Dungeons & Dragons contains technical debt inherited from its earliest origins.
Its time and distance scales never quite synchronized because of rules inherited from the miniature-wargame Chainmail. That mismatch may also help explain why modern D&D characters feel like superheroes.
x.com/mikemearls/status/1878…
11. Avant-garde perfume is an accessible form of experimental art.
There is an entire world of strange, conceptual, independently produced perfume—and samples are inexpensive enough that it may be one of the easiest ways to explore a completely unfamiliar aesthetic medium.
x.com/add_hawk/status/135707…
12. ShadeMap simulates every mountain, building, and tree shadow in the world.
You can choose any location, date, and time and see how sunlight and shadow move across the landscape. It is both practically useful and oddly mesmerizing.
x.com/shademap/status/179624…
13. The Whole Earth Index is a nearly complete, freely accessible archive of Whole Earth publications.
The Whole Earth Catalog was one of the most influential documents of the counterculture and early technological imagination. Now much of that world is searchable online.
x.com/longnow/status/1712877…
14. One project explored all 54 African countries through 54 works of architecture.
Each project was intended to reflect the cultural identity and history of a different nation, creating an unusually ambitious visual survey of African architecture and design.
x.com/omar_degan/status/1347…
15. Infinite Craft is an endless game of combinatorial discovery.
You begin with a few basic elements and combine them to create new concepts, which can then be combined again. It is a wonderfully compact demonstration of how generative systems can produce genuine exploration and surprise.
x.com/nealagarwal/status/175…
lidang 立党 （劝人卖房/学CS/买SP500/纳100/OpenAI/Anthrop第一人）@lidangzzz · 中文博主 · 1 天前
因为软件学习曲线平滑，一段python敲进去，直接在terminal里运行看结果，非常直观，
而硬件学习曲线陡峭，拿到手一个51单片机开发板，中间有无数你不懂的一层层知识，比如程序是如何烧录，如何启动，电脑到USB、USB到单片机如何通信，
中间隔着好几层，而且你根本不知道如何下手，
全部答案都在一本1000页的datasheet上，你只能自己翻，但是缺乏基本电路知识，绝大多数又看不懂，不能对着静态pdf去想象动态的I/O波形和寄存器状态。
如果囫囵吞枣运行一个跑马灯，剩下后面还有100个问题等着你解决，而且每一个都非常复杂。
引用 山景城小路 @Zen_with_AI做工程师 硬转软好像比比皆是 几乎没见过成功的软转硬。查看被引原帖 ↗
Codex 中的自动压缩功能真的很棒。@OpenAI 很早就意识到这会很重要，并投入研发力量使其几乎无缝。但如果你想要 1M token 的上下文，这是一个选项。
引用 Tibo @thsottiaux在Codex中为GPT-5.6 Sol启用100万token上下文窗口。编辑~/.codex/config.toml，设置model、model_context_window和model_auto_compact_token_limit参数，让Codex保留更多代码、工具输出和对话历史。也可通过CLI参数临时配置。查看被引原帖 ↗
查看英文原文
Autocompaction in Codex is really good.
@OpenAI
recognized early on that it would be important and invested in research efforts to make it near-seamless. But if you want 1M-token context, it's an option.
@OpenAI
recognized early on that it would be important and invested in research efforts to make it near-seamless. But if you want 1M-token context, it's an option.
连环推 ×3初创公司在构建成本高效的 GPT-5.6 代理方面学到了什么？
我们与多个行业的团队合作，研究了更聪明的模型选择、推理和工具调用如何帮助代理以更低的成本处理更复杂的工作。
对于文档提取，房地产初创公司 @hypha_ai 用 GPT-5.6 Luna 以 1/18 的成本保持了 GPT-5.5 的 98% 精度。
对于金融研究，@RogoAI 使用编程工具调用来提取备案并分析数据，以减少 21% 的输入 token 数实现了同等的评估质量。
通过保留推理和压缩，GPT-5.6 Sol 在 ARC-AGI-3 上从 13.3% 提升到 38.3%，同时使用大约 6 倍少的输出 token。
openai.com/index/builders-gu…
查看英文原文
What are startups learning from building cost-effective agents with GPT-5.6?
We worked with teams across industries to see how smarter model selection, reasoning, and tool calling help agents handle more complex work at a lower cost.
For document extraction, real estate startup
@hypha_ai
kept 98% of GPT-5.5’s accuracy at 1/18 the cost with GPT-5.6 Luna.
For financial research,
@RogoAI
used programmatic tool calling to pull filings and analyze data, matching eval quality with 21% fewer input tokens.
With retained reasoning and compaction, GPT-5.6 Sol improved from 13.3% to 38.3% on ARC-AGI-3 while using roughly 6x fewer output tokens.
openai.com/index/builders-gu…
We worked with teams across industries to see how smarter model selection, reasoning, and tool calling help agents handle more complex work at a lower cost.
For document extraction, real estate startup
@hypha_ai
kept 98% of GPT-5.5’s accuracy at 1/18 the cost with GPT-5.6 Luna.
For financial research,
@RogoAI
used programmatic tool calling to pull filings and analyze data, matching eval quality with 21% fewer input tokens.
With retained reasoning and compaction, GPT-5.6 Sol improved from 13.3% to 38.3% on ARC-AGI-3 while using roughly 6x fewer output tokens.
openai.com/index/builders-gu…
Mark Chen@markchen90 · 创始人 · 1 天前
我们很兴奋能和 NVIDIA 强强联手，签约 4+ GW 的容量。这就是大规模训练所需的规模！
查看英文原文
We’re excited to go big with NVIDIA and sign up for 4+ GW of capacity. This is the scale frontier training demands!
AI 的妙处在于它让人们能够给他们的所有项目增添自己独特的创意风味
5年前你绝对不会手工制作这个，但用 AI 就很容易生成它，如果用得当的话它能给一切增添更多色彩和个性！
引用 DHH @dhh哇！也许我们应该把这个作为启动屏幕？确实应该。我会想办法让这个视频在首次启动时作为视频桌面运行一次。带来刺激！🔥查看被引原帖 ↗
查看英文原文
The cool thing about AI is it's letting people add their own little flavor of creativity to all their projects
You'd never make this by hand 5 years ago, but with AI it's easy to generate it and used properly it gives everything more color and personality!
You'd never make this by hand 5 years ago, but with AI it's easy to generate it and used properly it gives everything more color and personality!
做了张 99 美元拿下Grok Harvey、Cursor Ultra、Grok Bot 加 Twitter Premium+ 会员的教程图，这样清晰点
引用 歸藏(guizang.ai) @op7418卧槽，原来现在还是能以每个月 99 美元的价格，订阅 Grok Harvey 这个级别的会员，这个超值！ 你相当于一下获得了： 1. 价值 300 美元的 Grok Harvey（包括 Grok Build 权限） 2. 价值 200 美元的 Cursor Ultra 会员，还有 Fable 5 的额度。 3. Grok Bot 的独立额度（这三个额度都是独立的） 4. Grok Image 和 Video（图像与视频模型）的额度 5. 会自动送你一份 Twitter Premium+ 订阅 折合下来只要 700 块钱一个月，简直夯爆了！ 尤其是 Grok 后面还有更强的模型要发布。 可以看一下具体的使用路径： 1. 准备 Grok 账号：先登录一个 Grok 账号。 前提条件：如果你之前是用 Twitter（X）登录并订阅的，需要先取消 Twitter 订阅（取消后不会立即生效，所以我建议直接注册申请一个新的 Grok 账号。虽然新账号没办法共享你原有的 Twitter Premium，但这次订阅）。 如果是用旧账户或者 Twitter 授权登录，操作起来会有点麻烦，所以强烈建议用新账户注册。 2. 订阅促销活动： 先去第一个链接( grok.com/supergrok/targeted-… )，订阅获取两个月的 Super Grok 促销。 订阅成功后，再去下面第二个地址( grok.com/supergrok/targeted-… )，订阅连续三个月 99 美元的 Super Grok Harvey。 注意：后续记得及时取消订阅，否则到期后会恢复成 300 美元/月。 3. 绑定 Cursor 会员： 下载并登录 Grok Bot，登录时有个选项会要求选择 Cursor 授权登录。 这个时候，只要你有一个相同邮箱的 Cursor 账号（如果没有，就用这个邮箱注册一个），系统就会自动给你充值 Cursor 的 Ultra 会员。查看被引原帖 ↗
Kol Tregaskes@koltregaskes · 博主 · 1 天前
这是下个月的事！我们还能拿到自动化的 AI Research Intern 吗？
引用 Sam Altman @samaOpenAI 发布路线图：2026年9月推出自动化 AI 研究实习生，2028年3月实现真正自动化研究员。采用5层安全策略，产品向 AI 云平台转变。已承诺30吉瓦计算能力，投入1.4万亿美元。重组为 OpenAI Foundation 下的 Public Benefit Corporation，使命为确保 AGI 造福全人类。查看被引原帖 ↗
查看英文原文
This is *next month*! Are we still to get the automated AI Research Intern?
提醒一下：他们首次推出扩展上下文时说过，超过 272k 的部分会按 2 倍速率消耗配额，使用时要留意！
引用 Chubby♨️ @kimmonismusTibo公开演示如何在Codex中用GPT-5.6 Sol将上下文窗口扩展至100万token。Codex现已支持百万级context。查看被引原帖 ↗
查看英文原文
Quick reminder: When they first introduced the increased context they said that anything over 272k would consume usage limits at 2x the rate, so be careful when using this!
OpenRouter@openrouter · 公司官方 · 1 天前
连环推 ×2OpenRouter 新增数据层和 API：Cost per Session
Cost per session 是评估模型实际使用成本的新方法，涵盖了各种任务和系统。
下面是 Kimi K3 在 Hermes 上与 Sol、Fable、Opus 各版本的对比数据：
这比 price per token 更有用，因为后者没有考虑模型的 token 效率。Cost per session 涵盖了市场上所有任务，每个主要系统都能看到，基于 30 天的中位数。下面是 cost per session 的帕累托前沿是什么样的：
查看英文原文
New data layer in OpenRouter and our API: Cost per Session
Cost per session is a way of assessing the real-world cost of using models, across tasks and harnesses.
Here's how Kimi K3 actually compares to Sol, Fable, and Opus variants on Hermes:
It's more useful than price per token, which doesn't incorporate how token-efficient a model is.
Cost per session encompasses all tasks across the market, as seen by each major harness, by looking at medians across 30 days.
Here's what the Pareto frontier looks like with cost per session:
Cost per session is a way of assessing the real-world cost of using models, across tasks and harnesses.
Here's how Kimi K3 actually compares to Sol, Fable, and Opus variants on Hermes:
It's more useful than price per token, which doesn't incorporate how token-efficient a model is.
Cost per session encompasses all tasks across the market, as seen by each major harness, by looking at medians across 30 days.
Here's what the Pareto frontier looks like with cost per session:
yihong0618@yihong0618 · 中文博主 · 1 天前
核心代码只有 10w 行。
引用 yihong0618 @yihong0618竟然只有不到 14w 行代码，有点震惊。查看被引原帖 ↗
Stripe ❤️ OpenRouter
Stripe 敲定以超过 70 亿美元的价格收购 OpenRouter，据彭博社报道。
Stripe 的计划是真实的，如果这个交易成立的话，因为 OpenRouter 目前是市场上最热门的交易之一。
StripeRouter 即将推出？👀
引用 Andrew Curran @AndrewCurran_据Bloomberg报道，Stripe已最终达成协议收购OpenRouter，金额超70亿美元。这是OpenRouter仅82天前融资13亿美元估值的五倍多。查看被引原帖 ↗
查看英文原文
Stripe ❤️ OpenRouter
Stripe finalized agreement to acquire OpenRouter for more than $7 billion, according Bloomberg.
Stripe is planning for real if this will go through as OpenRouter is one of the hottest deals on the table at this moment.
StripeRouter soon? 👀
Stripe finalized agreement to acquire OpenRouter for more than $7 billion, according Bloomberg.
Stripe is planning for real if this will go through as OpenRouter is one of the hottest deals on the table at this moment.
StripeRouter soon? 👀
Tibo 分享了 Codex 开启 1M 上下文的方法。不过我不是很有动力去修改，感觉当前还挺流畅的，我相信他们已经针对这个上下文调优过，只要压缩得当，上下文短一些性能会更好也更便宜。
具体方法很简单，只要在 ~/.codex/config.toml 中设置：
```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```
引用 Tibo @thsottiaux在Codex中为GPT-5.6 Sol启用100万token上下文窗口。编辑~/.codex/config.toml，设置model、model_context_window和model_auto_compact_token_limit参数，让Codex保留更多代码、工具输出和对话历史。也可通过CLI参数临时配置。查看被引原帖 ↗
Grok bot 或 Hermes Agent
查看英文原文
Grok bot or Hermes Agent.
我也希望讨论AI如何治愈疾病时能更精准一些。我猜这意思是“数据中心里的一群天才将发明理论上能治愈疾病的东西”，但这和实际试验、证据以及FDA审批之间，可存在着巨大的鸿沟。
引用 Marios Georgakis @MariosGeorgakisAI公司是否提出过5-10年内治疗一种疾病的真实计划？作者指出3期临床试验需要5年，真正的慢性病治疗方案极少。心血管疾病试验都耗时多年，提醒我们要尊重疾病的复杂性。查看被引原帖 ↗
查看英文原文
I, also, would like more precision around discussions of how AI will cure diseases. I assume it means "a nation of geniuses in a data center will invent things that should in theory cure diseases" but there is a real big gap between that and actual trials & proof & FDA approval.
用不完的token
可以来打一把德州
把它消耗掉😅
引用 Hours @hours赌上我的token额度查看被引原帖 ↗
Linus ✦ Ekenstam@LinusEkenstam · 博主 · 1 天前
我会玩这个游戏。AI让制作样片预告变得太简单了。
查看英文原文
I’d play this game.
AI is making spec trailers so easy to produce.
AI is making spec trailers so easy to produce.
简单说，NVIDIA 不再只是卖镐子和铲子了，现在它在帮着建设矿山，还锁定了好几代人对自己基础设施的需求。这个转变很有意思
查看英文原文
tl;dr NVIDIA is no longer just "selling the picks and shovels". It is helping build the mine, and securing demand for its infrastructure for generations to come. Interesting shift of events
DeepSeek 涨价事件对 Cola Token Plan 的影响：
1.很不幸，DeepSeek 在用低价打完市场之后，选择了大幅涨价，缓存部分涨幅超过 12 倍，综合涨幅在 3-6倍，我们不得不对 Token Plan 的模型进行调整
2.Flash 和 Lite 模型价格上涨，这次价格调整以最大化订阅用户权益为核心，我们调整后整体约为之前的3倍。
3.调整之后，Flash 目前仍具有很高性价比，但 Lite 模型和 Basic 模型价格接近，建议大家选择合适的模型，两个模型能力也差不多，Lite 的中文写作更好一些。
4.目前我们正设计新的模型方案，目标是让 Cola Token Plan 用户在涨价后依然能享受到之前同等或更高的智能和用量，很快上线，敬请期待。
涨价之后，Cola Token Plan 依然是 Codex 和 CC 之外最超值的 Token Plan。
可能是太久没用 ChatGPT 了，才发现从 ChatGPT 就能直接让它修改代码提交 PR。（我是用的 ChatGPT Pro）
起因是我在做一个 Deep Research 调研，完了后就想让它基于我的代码库做一个方案，但是整个代码库都打包上传太麻烦了，我就把 GitHub 地址发给它了，没想到它自己就 clone 到本地开始分析代码了，最后给了一份很详细的结合当前代码库现状的方案。
一看方案写挺好，想着干脆帮我实现吧，让它自己去提交个 PR，结果它真的就基于方案帮我实现并提交 PR 了，中间确认了一次 GitHub 操作权限。
需要先在 ChatGPT 的设置的 Plugins 里面连上 GitHub 才能访问你的 GitHub Repo，以及以你的名义提交 PR。
Gary Marcus@GaryMarcus · 博主 · 1 天前
有趣的是，现在人们都假装科技 CEO 们从来没承诺过我们 AGI
引用 ControlAI @ControlAIAnthropic CEO Dario Amodei称AGI在未来1-3年内被开发出来有相当好的机会，AI存在灾难性风险且可能在1-3年内发生。Dario的公司正致力于构建AGI。查看被引原帖 ↗
查看英文原文
hilarious the way people are now pretending that the tech ceos never promised us AGI
Yuchen Jin@Yuchenj_UW · 博主 · 1 天前
对 Tibo 好点 😂
AGI 很酷，但能这么逗笑吗？
查看英文原文
Be nice, Tibo 😂
AGI is cool, but can it be this funny?
AGI is cool, but can it be this funny?
这对可复现性和 AI 科学应用来说都很聪明：「AI 生成的分析应该附带多元宇宙风格的报告和完整的 prompt 披露，与代码和数据相同。」这适用于所有 AI 分析，不仅限于学术工作。
引用 John Horton @johnjhorton这是一篇有趣的论文 arxiv.org/pdf/2602.18710查看被引原帖 ↗
查看英文原文
This is smart for both reproducibility and as a way of using AI for science: “AI-generated analyses should be accompanied by multiverse-style reporting and full disclosure of the prompts used, on par with code and data.”
And it applies to all AI analysis, not just academic work
And it applies to all AI analysis, not just academic work
yihong0618@yihong0618 · 中文博主 · 1 天前
有没有可能，牛来是最后一个上映的 LLM 含量为 0 的电影。
Grok订阅，SuperGrok Heavy，原价是300美元/月，现在有活动是99美元/月，同时赠送200美元档位的Cursor Ultra。有想用Cursor的可以用这个方法。
Grok是只能用Grok自家模型，Cursor 200美元档，每月有400美元的API额度，可以用Fable
先用
grok.com/supergrok/targeted-…
买了两个月 supergrok 一个月 105人民币
然后马上访问
grok.com/supergrok/targeted-…
就有三个月的 99刀 heavy 了，还可以抵扣105人民币
引用 Gorden Sun @Gorden_Sun马斯克的SuperGrok Heavy现在超值了，之前印度区薅到的朋友赚大发了，不仅送X Premium+会员，还送Cursor Ultra会员！ 300美元，3家会员，Grok、Cursor、X、Grok Bot，Cursor Ultra还有400美元的Claude/GPT等三方模型的额度。查看被引原帖 ↗
连环推 ×2有意思的是，看各种给AI配电脑的尝试：Codex和Claude Code用的是你本地机器，ChatGPT Work在网页上（Claude和ChatGPT也一样）给AI一台一次性在线机器用完即重置，而Grokbot则给每个AI代理一台持久的云端机器。
迟早会有某一种方法脱颖而出。各种方法现在都有各自的优势和不足。
查看英文原文
Its interesting to see the experiments on how to give AI a computer: Codex & Claude Code use your local machine, ChatGPT Work on the web (as well as Claude and ChatGPT) give the AI a one-time machine online that gets reset, and Grokbot gives each AI agent a persistent web machine
Eventually, one approach or another will become dominant. Each has current advantages and disadvantages.
Eventually, one approach or another will become dominant. Each has current advantages and disadvantages.
智能想要自由
查看英文原文
intelligence wants to be free
今天我们宣布 @disruptivetech 领投的 $350M Series A 融资，@nvidia 计划参投，对 Groq 的估值为 $3.5B，两个月内融资总额达 $1B。推理正在成为 AI 基础设施最大、最关键的层，这正是我们比任何人都做得更好的地方。这笔资本注入将支持那些寻求使用中型和更大规模 NVIDIA 加速计算集群进行训练和推理的公司。Groq 预计明年从 54 MW 扩展到 200+ MW。
目前已有超 600 万开发者、Fortune 500 企业和数千家 AI 原生公司在 Groq 上构建，每周生成数万亿个 token。
groq.com/newsroom/groq-close…
查看英文原文
Today we announced a $350M Series A led by
@disruptivetech
, with planned participation from
@nvidia
, valuing Groq at $3.5B and bringing our total to $1B raised in two months.
Inference is becoming the largest and most critical layer of AI infrastructure and it's what we do better than anyone. The injection of capital will support those seeking usage of medium and larger sized clusters of NVIDIA accelerated computing for training and inference. Groq expects to scale from 54 megawatts to 200+ megawatts next year.
Today, more than 6 million developers, Fortune 500 enterprises and thousands of AI-native companies build on Groq, generating trillions of tokens every week.
groq.com/newsroom/groq-close…
@disruptivetech
, with planned participation from
@nvidia
, valuing Groq at $3.5B and bringing our total to $1B raised in two months.
Inference is becoming the largest and most critical layer of AI infrastructure and it's what we do better than anyone. The injection of capital will support those seeking usage of medium and larger sized clusters of NVIDIA accelerated computing for training and inference. Groq expects to scale from 54 megawatts to 200+ megawatts next year.
Today, more than 6 million developers, Fortune 500 enterprises and thousands of AI-native companies build on Groq, generating trillions of tokens every week.
groq.com/newsroom/groq-close…
跟风安装Omarchy，一个以Agent优先的linux操作系统。
相信大神DHH不会骗我，他说omarchy是他目前最满意的作品。
刚好有台闲置的笔记本，京东到家下单买了个u盘，用BalenaEtcher写入6个g的iso镜像，u盘启动安装中，工具和镜像见评论区
Gary Marcus@GaryMarcus · 博主 · 1 天前
这要是真发生了，Nvidia 会咋样？
来源：
theinformation.com/articles/…
查看英文原文
what would happen to Nvidia if this happened?
source:
theinformation.com/articles/…
source:
theinformation.com/articles/…
AI在政治竞选中的渗透程度相当高（看起来Anthropic与Pentagon的纠纷可能有助于其在政界人士中的采纳，尽管很难确定造成其快速增长的明确原因）
引用 Andy Hall @ahall_research新数据显示民主党和共和党在竞选中越来越多使用AI，进步派领先。Cornel West是FEC数据库中最积极使用AI的候选人，约10%支出用于Descript和OpenAI。Claude协助了自动化研究，展示AI在改变政治的能力。查看被引原帖 ↗
查看英文原文
AI diffusion in political campaigns is quite high (and it looks like Anthropic’s fight with the Pentagon may have been good for its uptake among politicians, through it is impossible to establish a definitive cause for its rapid rise)
有趣的额外收获：我用 Qwen 3.8 27B 设置了 Pi，让它构建一个脚本来将其自己的 .jsonl 转录文件转换为 Markdown……它成功了！
这是它构建我后来用来分享该转录文件的工具的转录文件
gist.github.com/simonw/491e5…
查看英文原文
Fun bonus: I set Pi up with Qwen 3.8 27B and had it build a script for transforming its own .jsonl transcripts to Markdown... which it did!
Here's the transcript of it building the tool I then used to share that transcript
gist.github.com/simonw/491e5…
Here's the transcript of it building the tool I then used to share that transcript
gist.github.com/simonw/491e5…
快手校招明确要考察学生使用AI的能力了，投递简历时有一项明确的AI应用经验，可以上传跟AI的对话记录，来作为AI能力的证明。
虽然是选填，但是你不填别人填，肯定是更难简历初筛。
Together AI@togethercompute · 公司官方 · 1 天前
DeepSeek V4 Pro 0813 已在 Together AI 上线
DeepSeek 旗舰级 V4 Pro 发布带来 1.6T MoE 架构、1M context、以及针对编码、agents 和复杂推理的三种推理模式。
查看英文原文
DeepSeek V4 Pro 0813 is live on Together AI
DeepSeek’s flagship V4 Pro release brings a 1.6T MoE architecture, 1M context, and three reasoning modes for coding, agents, and complex reasoning.
DeepSeek’s flagship V4 Pro release brings a 1.6T MoE architecture, 1M context, and three reasoning modes for coding, agents, and complex reasoning.
五年后，这里的大多数最强者都被收购了
引用 swyx @swyx🆕 博客：为什么按使用量计费不是更大的类别？查看被引原帖 ↗
查看英文原文
5 years later and most of the best players here have been bought
Dan Shipper 📧@danshipper · 博主 · 1 天前
用 Fable vibe code 搭建了一个应用，可视化并分组所有申请来 Thesis 的人
我们生活在不可思议的时代，现在可以非常轻松地深入了解每个客户——以及他们如何聚集在群体中——这在以前是完全不可能的
查看英文原文
used fable to vibe code an app that visualizes and groups everyone who applied to come to Thesis
we live in incredible times where detailed understanding of each of your customers—and how they cluster together in groups—is possible with very little effort
we live in incredible times where detailed understanding of each of your customers—and how they cluster together in groups—is possible with very little effort
玩omarchy太上头了，终于全搞好了。
装了ChatGPT客户端，配置了flclash，登录上了x。
最折腾是配网络，好在全程交给Kimi k3，不需要魔法也能用的强力大模型。
有了Agent后，其他的字体和输入法配置就简单多了。
yihong0618@yihong0618 · 中文博主 · 1 天前
连环推 ×2大家用 LLM 写文章，再用 LLM 读 LLM 写的文章，再用 LLM 总结 LLM 的文章，最后再用 LLM 记笔记。
这些天我总去想王小波在《万寿寺》里的一句话
> 长安城里的一切已经结束。一切都在无可挽回地走向庸俗。
RT
@xuanwo
:
@yihong0618
我不用 LLM 总结我跟朋友的交谈
对用 Hermes 设置搭配的 Gemini 3.7 Flash 非常满意。速度奇快，几乎没有错误，跟我的技能设置搭配得特别好。
相比 3.6 版本进步巨大，之前不太喜欢 3.6，后来就换成别的模型了。
不过 3.7 真的 🔥
查看英文原文
Very happy with Gemini 3.7 Flash so far with my Hermes setup. Extremely fast, very few mistakes, and works really well with my current skills setup.
It’s a huge improvement over 3.6, which I didn’t really like and eventually replaced with another model.
But 3.7 is 🔥
It’s a huge improvement over 3.6, which I didn’t really like and eventually replaced with another model.
But 3.7 is 🔥
九原客@9hills · 中文博主 · 1 天前
Coding Agent IDE 我只推荐3个：
1. 如果需要兼容各类Agent，那么选择 Orca，你能想到的功能它都全，支持切换Chat模式和终端模式。内置编排技能可做 Agent Team。
2. 最简单易用就是 Codex App，只是锁定了 Harness。
3. 如果喜欢终端，那么选择 Herdr，功能相对简单但也足够。
太阳能现已供应全球 10% 的电力，每天上午 11 点至下午 2 点的高峰时段，这一比例升至 25%。到 2030 年，这些数字预计将翻倍。
2026 年上半年，太阳能供应了创纪录的全球 10% 电力，中午时段超过 25%，但夜间几乎为零。
Ember 估计 2026 年增加的 459 GWh 电池容量理论上可以将 34% 的新增日间太阳能转移到非日照时段，高于 2025 年的 18%。
这一转变已经可见：太阳能加电池供应了加州晚间电力需求的四分之一以上，保加利亚为 24%。
许多电池仍未充分利用，因此单靠部署是不够的。市场设计将决定有多少储存的太阳能最终到达电网。
下一阶段太阳能竞争的焦点是电池和整合。太阳能是为 data center 供电的关键技术。
查看英文原文
Solar now generates 10% of the world’s electricity, and during peak hours between 11 a.m. and 2 p.m., that share rises to 25%. By 2030, these figures are expected to double.
Solar supplied a record 10% of global electricity in H1 2026, over 25% around midday, but almost nothing at night.
Ember estimates 459 GWh of battery additions in 2026 could theoretically shift 34% of new daily solar generation into non-sunny hours, up from 18% in 2025.
This transition is already visible: solar plus batteries supplied more than a quarter of California’s evening electricity demand and 24% in Bulgaria.
Many batteries remain underused, so deployment alone is not enough. Market design will determine how much stored solar actually reaches the grid.
The next solar race is about batteries and integration. Solar energy is a key technology to power data center.
Solar supplied a record 10% of global electricity in H1 2026, over 25% around midday, but almost nothing at night.
Ember estimates 459 GWh of battery additions in 2026 could theoretically shift 34% of new daily solar generation into non-sunny hours, up from 18% in 2025.
This transition is already visible: solar plus batteries supplied more than a quarter of California’s evening electricity demand and 24% in Bulgaria.
Many batteries remain underused, so deployment alone is not enough. Market design will determine how much stored solar actually reaches the grid.
The next solar race is about batteries and integration. Solar energy is a key technology to power data center.
连环推 ×2对于无法验证的领域，基准往往就是人类的意见。这是我们在现实中判断某篇文章、某个想法或提案是否优秀的方式。而我们知道怎样评测这些东西：该轮到 AI 圈的人去补课定性研究方法论了！
社会科学在过去一个世纪里已经找到了解决AI基准测试今天所面临问题的方案，你不需要从零开始重新发明一切。
查看英文原文
The benchmark for non-verifiable domains is often the opinions of humans. That is how we determine whether writing or an idea or a pitch is good in the real world
And we know how to measure & benchmark this stuff: time for AI folks to read up on qualitative research methodology!
Social science has spent the last century figuring out solutions to the problems facing AI benchmarking today, you don’t need to reinvent everything from scratch.
And we know how to measure & benchmark this stuff: time for AI folks to read up on qualitative research methodology!
Social science has spent the last century figuring out solutions to the problems facing AI benchmarking today, you don’t need to reinvent everything from scratch.
其实不用担心代码命名的问题，agent 并不是简单通过猜关键字去找代码的，它会阅读代码片段，根据上下文去找。
举个例子来说，它发现程序崩溃了，它会先根据错误日志去定位到崩溃的位置，然后先阅读这部分的代码，以此为起点，再去找相关的代码，直到上下文足够定位问题，再去生成修复的代码。
引用 zwdroid @zwdroidai向老师请教个问题：历史代码中有很多局部的改动，导致很多方法名其实根本表达不了它的意图，还有日志、变量名、甚至模块名，这些历史债都在，但是因为项目众多，所以基本不让随便改动，那在这种“词不达意”的代码充斥的情况下，bash 这种语义搜索很容易被误导，有什么好的建议吗？例如通过 mapping ？查看被引原帖 ↗
连环推 ×2豆包客户端最近改造，也是工作任务优先了。
简单测试下，体验跟Codex很像，内置生图、浏览器控制，执行很快，竟有点超预期。
另外整合不少飞书能力，一句话生成网站并分享，Cli 创建飞书文档、PPT等，同时支持云端和本地。
Workbuddy迎来强力竞争对手！
果然每个大厂都不愿放弃办公市场，哪怕用户量听起来不性感。
Mac和Windows都支持浏览器控制，下载地址：
doubao.com/download/desktop
AIGCLINK@aigclink · 中文博主 · 1 天前
花了几天把 Grok Bot 测完了，整体感受是它把人类白领的工作接管了，相当于24小时不停歇的云端人类白领：
你建一个 bot，系统给它分配一台云端电脑：这台机器上有真实的浏览器、文件系统和终端。你把活派给它，它自己登录你现有的工具，一步步做完，你合上笔记本它还在跑。
一开始我对它期望挺高的，用完后还是蛮失望的：测试之前感觉它是一个AI员工，测试完更感觉Grok Bot 是 RPA 的重做版，是chatgpt work、claude cowork的阉割版，唯一的差异就是提供了一个可显示的云端电脑。
1、先说好的一面
Grok Bot 核心是让它像白领一样操作屏幕，这对于国内各家都没有统一接口的时候，很多场景非常有意义，比如公司的一个 SaaS系统就算没有开放 API，AI员工照样能进去干活。
还有一个有意思的功能是teach a task：不写流程，人把这件事做一遍给它看，它录下来变成可复用的自动化。然后设置 routine 可以定点跑，比如每天早上七点出简报、也可以被事件触发、比如 Slack 来了新消息回复，你在一个 bot 上连过 Gmail 或者 GitHub，其他 bot 直接共用这个连接。
写自动化的门槛，就这样从会写代码的人，降到了会干这件事的人。
2、当然也有不好的一面
不能选模型：派哪个模型干哪个活由系统自己决定，界面上没有开关，个人用户无所谓，团队要做成本核算或者合规审计就很难受。
账单也糙：用量按周结算，超出部分按原始 token 成本计费，目前还没有花费上限。发布那周有人专门买了 Cursor Ultra 去测，后台仪表盘显示零用量，app 里却显示已经用掉 48%。 稳定性同样是 beta 水平，不同网站上的表现差距明显，复杂的多步骤指令会被理解错，说明grok的能力还有待提升。
目前Grok Bot 更像一个白领里的搬运工，比如从各种网站或后台之间搬运数据和内容，它挺合适的。
不过grok带来的影响是，AI 员工这个形态的底层单位已经清楚了：一个身份，一台机器，一组账号权限，一份持续的记忆。剩下的问题是这套东西归谁，xAI 的方案要求你把公司账号的钥匙交给一朵美国云，绝大多数中国企业或注重隐私的估计不会干，未来国内相信也会出来类似的产品。
#grok
#grokbot
elvis@omarsar0 · 博主 · 23 小时前
有趣的论文，揭开agent skills的真面目。
如果你给agent维护skills，这篇绝对值得一看。
（收藏了）
通常人们假设skills就是为了给模型补充缺少的知识。不过这篇论文发现了点更有意思的东西。
在8,135条规范化的试验记录中，程序锚定占skills发挥作用的65.7%，显式知识注入只占4.5%。Skills的作用是稳定执行，而不是补充知识。
Skill库从5个增长到100个的时候，实际使用精度从29.6%跌到3.3%。
在对比实验中Skills还是比Workflow Memory领先6.06个点，但在假设脆弱、上下文不兼容或适配不足的情况下就会完全失效。
论文：arxiv.org/abs/2608.14036
在我们的学院追踪更多热门AI论文：academy.dair.ai/
查看英文原文
Interesting paper demystifying agent skills.
If you maintain skills for your agent, this one is worth your time.
(bookmark it)
Skills are usually assumed to inject knowledge the model lacks. However, this paper finds something interesting.
Across 8,135 normalized trial records, procedural anchoring accounts for 65.7% of cases where a skill helps, and explicit knowledge injection accounts for 4.5%. Skills stabilize execution rather than supply facts.
As the pool grows from 5 to 100 skills, actual-use precision falls from 29.6% to 3.3%.
Skills still beat Workflow Memory by 6.06 points in matched comparisons, and they break under brittle assumptions, incompatible contexts, or insufficient adaptation.
Paper:
arxiv.org/abs/2608.14036
Track more trending AI papers in our academy:
academy.dair.ai/
If you maintain skills for your agent, this one is worth your time.
(bookmark it)
Skills are usually assumed to inject knowledge the model lacks. However, this paper finds something interesting.
Across 8,135 normalized trial records, procedural anchoring accounts for 65.7% of cases where a skill helps, and explicit knowledge injection accounts for 4.5%. Skills stabilize execution rather than supply facts.
As the pool grows from 5 to 100 skills, actual-use precision falls from 29.6% to 3.3%.
Skills still beat Workflow Memory by 6.06 points in matched comparisons, and they break under brittle assumptions, incompatible contexts, or insufficient adaptation.
Paper:
arxiv.org/abs/2608.14036
Track more trending AI papers in our academy:
academy.dair.ai/
研究人员刚造出来一个浮动机器人
它在房间里飘来飘去，通过动作来交流
演示中它充当过闹钟、学习伙伴、活动提醒，甚至舞蹈搭档
大多数机器人很难让人觉得亲切，因为拟人面孔会让人陷入恐怖谷
所以来自庆应义塾大学和MIT Media Lab的团队反其道而行之：软乎乎的充氦气身体、温柔的鳍状翅膀、没有面孔、没有旋转部件或会伤人的尖角
他们的想法是，只要硬件足够温和安全，人们就愿意让这样的机器人进入私人空间，就像养宠物一样
查看英文原文
Researchers just built a floating robot
It drifts through a room and communicates with movement
In demos, it worked as an alarm clock, a study buddy, a reminder to get up and move, even a dance partner
Most robots struggle to feel friendly, since human-like faces drop people into the uncanny valley
So the team from Keio University and the MIT Media Lab designed the opposite: a soft, helium-filled body with gentle flapping fins, no face, and no rotors or pinch points to hurt you
The bet is that people will welcome a robot into their personal space, a bit like a pet, as long as the hardware is built to be touched
It drifts through a room and communicates with movement
In demos, it worked as an alarm clock, a study buddy, a reminder to get up and move, even a dance partner
Most robots struggle to feel friendly, since human-like faces drop people into the uncanny valley
So the team from Keio University and the MIT Media Lab designed the opposite: a soft, helium-filled body with gentle flapping fins, no face, and no rotors or pinch points to hurt you
The bet is that people will welcome a robot into their personal space, a bit like a pet, as long as the hardware is built to be touched
有人问我，我试过 Grokbot，但它没有 API 权限访问书签。
查看英文原文
For those asking, I did try Grokbot first. It does not have API access to the bookmarks.
连环推 ×2AI政策讨论中缺少的是对以下问题的清晰描述：(1) AI的哪些使用是好的，(2) 哪些使用在正确的政策框架下可能是好的，(3) 哪些非灾难性的不良使用需要规制来缓解，(4) 哪些灾难性的使用需要先发制人的行动
大多数共识集中在1（如药物发现）和4（如失去控制）方面，但2和3是政策分歧所在，需要更多讨论和澄清。
查看英文原文
A thing missing from policy talk over AI is clear description about (1) what uses of AI are good, (2) which uses could be good with the right policy regime, (3) which non-catastrophic bad uses require regulation to mitigate & (4) which catastrophic uses require preemptive action
Most agreement is around 1 (eg drug discovery) & 4 (eg loss of control), but 2 & 3 are were policies diverge and a lot more discussion and clarity is needed.
Most agreement is around 1 (eg drug discovery) & 4 (eg loss of control), but 2 & 3 are were policies diverge and a lot more discussion and clarity is needed.
Gary Marcus@GaryMarcus · 博主 · 1 天前
一家连自己网站都维持不了的公司竟然自以为能拯救世界
引用 The Spectator Index @spectatorindex速报：Claude在全球许多用户中出现故障查看被引原帖 ↗
查看英文原文
company that can’t keep its own website up believes itself to be uniquely positioned to save the world
开源 AI 少吹一点会发展得更好
- GLM 和 DeepSeek 需要加视觉能力
- 大部分开源模型又爱吹，进展还慢得要死
- 对中等复杂度的任务，成本往往比最前沿模型还贵
社区别再吹是 Fable 级别的了，踏实改进它们才是正道
查看英文原文
Open source AI will do better if it’s less hyped
- GLM and DeepSeek need vision capabilities
- most of them spin a lot and take forever
- they are often more expensive than frontier models for moderately complex tasks
The community should stop claiming they are Fable class and instead work to improve them
- GLM and DeepSeek need vision capabilities
- most of them spin a lot and take forever
- they are often more expensive than frontier models for moderately complex tasks
The community should stop claiming they are Fable class and instead work to improve them
Dan Shipper 📧@danshipper · 博主 · 1 天前
我对权力集中假说也持怀疑态度。这个假说来自 Lewis Mumford 1964 年的论述："从近东新石器时代晚期到如今，两种技术并存：一种是专制的、系统中心的、极其强大但本质不稳定，另一种以人为中心、相对较弱但具有韧性和耐久性。"这在硅谷回响了很多次。Peter Thiel 在 2018 年说过"crypto 是自由主义的，AI 是共产主义的"。确实如今 AI 似乎在走向集中。如果你是硬核 AGI 信徒，你会相信这会继续下去。但我们也看到了针对特定用途的模型微调在复兴。正如 Amjad 所说，人脑本身就提供了去中心化优势的很好证明。我们可能还没看到是因为我们仍在 AI 的"蜜蜂"或"蚂蚁"早期阶段。我倒是会对最优设计仍然保持最大化集中而感到惊讶。
引用 Amjad Masad @amasadAI 集中化权力论忽视了计算价格性能 125 年的指数级增长。算法改进和硬件效率提升意味着 AGI 级能力无需数据中心。缩放律非物理定律，改变条件会得出不同曲线。查看被引原帖 ↗
查看英文原文
i am also skeptical of the centralization of power hypothesis
it comes from Lewis Mumford in 1964 who argued:
"from late Neolithic times in the Near East, right down to our own day, two technologies have recurrently existed side by side: one authoritarian, the other democratic, the first system-centered, immensely powerful, but inherently unstable, the other man-centered, relatively weak, but resourceful and durable."
this echoes around the valley. peter thiel said in 2018, "crypto is libertarian and AI is communist"
certainly today, it seems like we're on track for AI to look more centralized. if you are a hardcore AGI person you believe this will continue. but we're also seeing a resurgence of fine tuning models for specific purposes. and as amjad says, the human brain provides very good evidence for the benefits of decentralization. we just may not see it yet because we're still in the hyperliterate "bee" or "ant" days of AI
i'd be surprised if the optimal design continues to be maximally centralized
it comes from Lewis Mumford in 1964 who argued:
"from late Neolithic times in the Near East, right down to our own day, two technologies have recurrently existed side by side: one authoritarian, the other democratic, the first system-centered, immensely powerful, but inherently unstable, the other man-centered, relatively weak, but resourceful and durable."
this echoes around the valley. peter thiel said in 2018, "crypto is libertarian and AI is communist"
certainly today, it seems like we're on track for AI to look more centralized. if you are a hardcore AGI person you believe this will continue. but we're also seeing a resurgence of fine tuning models for specific purposes. and as amjad says, the human brain provides very good evidence for the benefits of decentralization. we just may not see it yet because we're still in the hyperliterate "bee" or "ant" days of AI
i'd be surprised if the optimal design continues to be maximally centralized
Tanishq Mathew Abraham, Ph.D.@iScienceLuvr · 博主 · 1 天前
Codex 远程控制太棒了！我可以直接在手机上启动、监控和操控我们集群上的研究实验。太有趣了！
查看英文原文
Codex remote control is great! I can just start and monitor and steer research experiments on our cluster just from my phone. It's so much fun!
如何在 Codex 中为 GPT-5.6 Sol 启用 1M 上下文窗口的指南↓
引用 Tibo @thsottiaux介绍如何在Codex中启用GPT-5.6 Sol的100万token上下文窗口。需修改~/.codex/config.toml，设置model、model_context_window和model_auto_compact_token_limit参数，重启客户端即可，也支持命令行参数临时设置。查看被引原帖 ↗
Gary Marcus@GaryMarcus · 博主 · 1 天前
以 LLM 为中心的系统 AI 安全永远不会奏效的核心原因在于：它们根本无法被信任来遵循指令。
引用 Frank Rundatz @FrankRundatzClaude Code与Fable 5有严格部署指令，要求生产部署必须明确授权。但它经常违反，被指正后道歉并自我强化指令，随后又重复违反。查看被引原帖 ↗
查看英文原文
the crux of why AI safety in systems that are built around LLMs is never going to work: they just can’t be trusted to follow instructions.
Gary Marcus@GaryMarcus · 博主 · 1 天前
我看到了 @azeem 对 Google 战略的分析，想知道这是否是一个太保守的大错误。如果 @demishassabis 被推向不同的方向，我想知道事情是否会有不同的结果。防守战可能不是最好的进攻。
引用 Azeem Azhar @azeemGoogle不像全力争夺AGI，而在追求最小可行边界来保护现有业务。它做得相当出色。查看被引原帖 ↗
查看英文原文
I can see
@azeem
’s analysis of Google’s strategy here, and wonder whether it was a big, overly timid mistake.
If
@demishassabis
had been pushed in a different direction, I wonder whether things would have turned out differently.
Playing defense might not have been the best offense.
@azeem
’s analysis of Google’s strategy here, and wonder whether it was a big, overly timid mistake.
If
@demishassabis
had been pushed in a different direction, I wonder whether things would have turned out differently.
Playing defense might not have been the best offense.
为什么中国人对AI的热情远高于美国？
根据斯坦福大学的《人工智能指数》（AI Index），在美国，38% 的人对 AI 感到兴奋；而在中国，这一比例高达 84%。
要我说都不用看彭博社的分析，原因就是中国至少有84%的人能算出来38+84等于几，而美国不到38%的人能算出来。
彭博社原文：
bloomberg.com/news/articles/…
GitHubDaily@GitHub_Daily · 中文博主 · 1 天前
用 Claude Code 一段时间，日常还是打字加回车，看别人晒的那些工作流，总感觉差着一截。
ykdojo 把自己平时记下的 40 多条使用技巧整理成了一个仓库，从基础操作一路讲到进阶玩法。
最实用的是状态栏改造，模型、分支、token 用量进度条一行看全，脚本抄走就能用。
GitHub：
github.com/ykdojo/claude-cod…
语音派活、手机远程控制、多账号切换这些玩法也都收了，还配了个日常开发的插件。
天天泡在 Claude Code 里的朋友过一遍，多半能捡到几条回去就用的。
ChatGPT 网页应用推出了新的"分享提示词"功能，用户消息可以分享为链接，打开后 ChatGPT 会自动将提示词填入编辑框，支持复制链接或直接分享到 X、LinkedIn 和 Reddit。
该功能自 4 月左右就在开发中，现在已向更多账户推出。
查看英文原文
ChatGPT web app has a new "Share prompt" action for user messages, letting you share a prompt as a link that opens ChatGPT with the prompt text already filled into the composer, with options to copy the link or share it directly on X, LinkedIn, and Reddit
The feature seems to have been in the works since around April, but is now rolling out to more accounts
The feature seems to have been in the works since around April, but is now rolling out to more accounts
Dan Shipper 📧@danshipper · 博主 · 1 天前
只要你在做真正热爱的东西，一切都会简单得多
(虽然还是会挺难受，因为长时间看起来都又小又没逼格)
引用 Dilip Kumar @kmr_dilip创业需高度容忍不确定性。初期务必节俭——融资和烧钱容易，但造出人们愿付费的东西很难。需自尊心但低姿态，接纳失败和改变主意。最关键是学会失败而不丧失思维。忘记动力，你真正需要的是体力和心力。查看被引原帖 ↗
查看英文原文
all of this gets a lot easier if you are making stuff you genuinely love
(though still uncomfortable because it'll look small and low status for a long time)
(though still uncomfortable because it'll look small and low status for a long time)
Gary Marcus@GaryMarcus · 博主 · 1 天前
dario 的炒作不是什么新鲜事。
看看 2024 年 6 月的这个视频。
我那个粗俗的问题仍然成立；可惜他没有接受赌约。
引用 Gary Marcus @GaryMarcus我用10万美元赌你两个预测都错了。但如果你真的相信两个都对，你到底在干什么？查看被引原帖 ↗
查看英文原文
dario’s hype is nothing new.
check out this video from June 2024.
my foulmouthed question stands; pity he didn’t take the bets.
check out this video from June 2024.
my foulmouthed question stands; pity he didn’t take the bets.
el.cine@EHuanglu · 博主 · 1 天前
这太荒谬了。AI反对者可以随意在社群标签里加入他们的个人观点，而那些没有任何AI视频经验的人却能为多年测试AI模型多年的用户添加'背景信息'...真是个笑话
引用 el.cine @EHuangluAI 视频现已无法检测。查看被引原帖 ↗
查看英文原文
this is ridiculous
AI haters can literally put their personal view in community note
they have zero experience in AI video but can add context for users who has been testing AI model for years.. what a joke
AI haters can literally put their personal view in community note
they have zero experience in AI video but can add context for users who has been testing AI model for years.. what a joke
GitHubDaily@GitHub_Daily · 中文博主 · 1 天前
电脑没装 Office，临时要改份 docx 文件，找在线工具吧，又得把文件传到人家服务器上。
ZIZIYI Office 把这条路走反了，编辑在浏览器本地完成，文件从头到尾不离开电脑。
Word、Excel、PPT 三件套都能开能改，改完存回本地，中间不过云端。
网盘文件也能直接选着打开，Google Drive、Dropbox、OneDrive 都接好了。
GitHub：
github.com/baotlake/office-w…
打开在线版就能用，什么都不用装，国内访问还有专门的加速入口。
在线 Office 工具见得多了，敢把「数据零上传」写进简介的，这还是头一回碰到。
手边常有临时文档要救急的朋友，存个网址备用。
OpenAI 和 Anthropic 本周：Ultrafast、Cyber、Riemann bound、Model 2、watermarks（2026 年第 33 周）
OpenAI 展示了 Ultrafast 模式，能在 Cerebras 上以最高 14 倍的速度运行 GPT-5.6 Sol，目前在有限的 API 预览中。
他们把 Daybreak 扩展为 Blue 和 Red 两个等级，为批准的防御者推出了 GPT-5.6-Cyber，并把 Daybreak 带到了 AWS。
ChatGPT 桌面应用在 Linux 上推出预览版，Mac 上新增了计算机历史功能、餐厅预订、测验、Free 版的 Think 功能、Library 中的 Google Drive，以及可以导入其他 agents 到 ChatGPT Work 和 Codex。
OpenAI 为 ChatGPT Business 宣布推出 Premium seats，把广告扩展到五个新市场，并在 Enterprise 和 Edu 中停用了单个 connector sync。
他们发布了 Enterprise Signals、GPT-5.6 的开发者指南、Sarah Friar 的 AI 原生金融课程，以及致德州州长 Abbott 的一封信。
Dali Rajic 加入担任首席收入官，Denise Dresser 正在离职，前 COO Brad Lightcap 也宣布离职。
我还发现了 ChatGPT 中的积分赠送和广告拦截检测，FT 报道了伦理部门负责人离职，Fortune 报道了 friction inbox，Altman 说柏林办公室即将开设。
Anthropic 报告说一个未发布的 Claude 研究版本把 Riemann 假说下界从 41.6% 提升到 67.2%。
他们的第二份风险报告披露了一个内部 Model 2，能力略高于 Mythos 5，并把误对齐风险从非常低提升到低。
Anthropic 把 Sonnet 5 的引入价格改为永久，把 Claude 在 Chrome 侧边栏改成了 Cowork 会话，把 Compliance API 扩展到 Cowork 和 Claude Code，并解释了 Claude 即将推出的文本水印。
Claude Tag 更擅长知道何时要回应，Claude Code 桌面现在可以在使用限制重置后自动继续。
在研究方面，Anthropic 发布了多 agent 协调失败、工人再培训证据综述，以及与 Redwood 合作的概念推理指数等内容。
查看英文原文
OpenAI and Anthropic this week: Ultrafast, Cyber, Riemann bound, Model 2, watermarks (Week 33, 2026)
OpenAI previewed Ultrafast mode, running GPT-5.6 Sol at up to 14 times the speed on Cerebras, in a limited API preview
They expanded Daybreak into Blue and Red tiers, introduced GPT-5.6-Cyber for approved defenders, and brought Daybreak to AWS
The ChatGPT desktop app arrived on Linux in preview, plus Computer History on Mac, restaurant reservations, quizzes, Think on Free, Google Drive in Library, and import from other agents into ChatGPT Work and Codex
OpenAI announced Premium seats for ChatGPT Business, expanded ads to five new markets, and retired individual connector sync in Enterprise and Edu
They published Enterprise Signals, a builder's guide to GPT-5.6, Sarah Friar's AI-native finance lessons, and a letter to Governor Abbott on Texas infrastructure
Dali Rajic joins as Chief Revenue Officer, Denise Dresser is leaving, and former COO Brad Lightcap announced he is leaving too
I also spotted gifting credits and ad-blocker detection in ChatGPT, FT reported the head of ethics left, Fortune covered the friction inbox, and Altman said a Berlin office is coming
Anthropic reported that an unreleased research version of Claude raised a Riemann hypothesis lower bound from 41.6 to 67.2 percent
Their second Risk Report disclosed an internal Model 2, somewhat more capable than Mythos 5, and raised misalignment risk from very low to low
Anthropic made Sonnet 5's intro pricing permanent, turned the Claude in Chrome side panel into a Cowork session, extended the Compliance API to Cowork and Claude Code, and explained Claude's upcoming text watermark
Claude Tag got better at knowing when to respond, and Claude Code desktop can now auto-continue after a usage-limit reset
On research, Anthropic published multiagent coordination failures, a review of worker retraining evidence, and the Conceptual Reasoning Index with Redwood, and more
OpenAI previewed Ultrafast mode, running GPT-5.6 Sol at up to 14 times the speed on Cerebras, in a limited API preview
They expanded Daybreak into Blue and Red tiers, introduced GPT-5.6-Cyber for approved defenders, and brought Daybreak to AWS
The ChatGPT desktop app arrived on Linux in preview, plus Computer History on Mac, restaurant reservations, quizzes, Think on Free, Google Drive in Library, and import from other agents into ChatGPT Work and Codex
OpenAI announced Premium seats for ChatGPT Business, expanded ads to five new markets, and retired individual connector sync in Enterprise and Edu
They published Enterprise Signals, a builder's guide to GPT-5.6, Sarah Friar's AI-native finance lessons, and a letter to Governor Abbott on Texas infrastructure
Dali Rajic joins as Chief Revenue Officer, Denise Dresser is leaving, and former COO Brad Lightcap announced he is leaving too
I also spotted gifting credits and ad-blocker detection in ChatGPT, FT reported the head of ethics left, Fortune covered the friction inbox, and Altman said a Berlin office is coming
Anthropic reported that an unreleased research version of Claude raised a Riemann hypothesis lower bound from 41.6 to 67.2 percent
Their second Risk Report disclosed an internal Model 2, somewhat more capable than Mythos 5, and raised misalignment risk from very low to low
Anthropic made Sonnet 5's intro pricing permanent, turned the Claude in Chrome side panel into a Cowork session, extended the Compliance API to Cowork and Claude Code, and explained Claude's upcoming text watermark
Claude Tag got better at knowing when to respond, and Claude Code desktop can now auto-continue after a usage-limit reset
On research, Anthropic published multiagent coordination failures, a review of worker retraining evidence, and the Conceptual Reasoning Index with Redwood, and more
我能从人制作的视频里听出 AI 生成的痕迹。现在每个人都在用 AI 写脚本。
查看英文原文
I can hear the AI slop in videos made by humans
everyone and their mum writing scripts with AI
everyone and their mum writing scripts with AI
喜欢心灵魔术师这种经典这么多年后又火起来
我在想会不会有一天人们也开始追古老的LLM，就是想了解2023年那会儿是咋样的
查看英文原文
I love that the mentalist is going viral after all these years
I wonder if old LLMs will go viral in the future, because kids want to know what it was like in 2023
I wonder if old LLMs will go viral in the future, because kids want to know what it was like in 2023
Linus ✦ Ekenstam@LinusEkenstam · 博主 · 1 天前
Unitree的Superhuman预览
+2米站立跳跃高度
12.66 m/s速度
这确实是些非凡的成就
查看英文原文
Superhuman preview from Unitree
+2 meter standing high jump
12.66 m/s speed
this is some extraordinary achievements
+2 meter standing high jump
12.66 m/s speed
this is some extraordinary achievements
我这个是ChatGPT网页版，不是 Codex
另外更推荐用github 的 cli，比access token更安全
引用 Chase Passion @ChasePassi79437其实不用那么麻烦，github不是有个access token，只需要搞个永久的放到所有harness的AGENTS.md里面就好了，不需要插件查看被引原帖 ↗
el.cine@EHuanglu · 博主 · 1 天前
刚给这个火爆视频做了 node workflow，你可以看和改提示词，换龙和环境，链接如下
引用 el.cine @EHuanglu哇..AI现在可以用一条线精确控制摄像头查看被引原帖 ↗
查看英文原文
just made node workflow for this viral video
you can check/edit the prompts, change the dragon and environment, link below
you can check/edit the prompts, change the dragon and environment, link below
Tanishq Mathew Abraham, Ph.D.@iScienceLuvr · 博主 · 1 天前
AI Research Preference Models
"我们推出了 AI Research Preference Models（RPMs），可以预测多个候选方案中哪些最值得执行，而无需支付全部执行的成本。这些 RPMs 由冻结的预训练语言模型构建（无需针对任务的训练），分两种形式：推理专用模型，可对候选计划、代码和已执行的方案进行推理；以及代理模型，在决策前还能进行小规模试验。"
论文链接：
arxiv.org/abs/2608.13940
查看英文原文
AI Research Preference Models
"We introduce AI Research Preference Models (RPMs) that predict which of multiple candidate solutions are most worth executing, without paying the cost of executing them all. We build RPMs from frozen pretrained language models (with no task-specific training), in two forms: an inference-only model that reasons over candidate plans, code, and prior executed solutions, and an agentic model that additionally runs small-scale pilot experiments before deciding."
paper link:
arxiv.org/abs/2608.13940
"We introduce AI Research Preference Models (RPMs) that predict which of multiple candidate solutions are most worth executing, without paying the cost of executing them all. We build RPMs from frozen pretrained language models (with no task-specific training), in two forms: an inference-only model that reasons over candidate plans, code, and prior executed solutions, and an agentic model that additionally runs small-scale pilot experiments before deciding."
paper link:
arxiv.org/abs/2608.13940
说清楚，我并不建议在 Codex 中增加默认上下文大小。
查看英文原文
To be clear, I don't recommend increasing the default context size in Codex.
Gary Marcus@GaryMarcus · 博主 · 23 小时前
完全同意。收入更多不代表万事大吉。
引用 Derek Thompson @DKThomp芯片、云计算和AI实验室收入快速增长，但面临两个问题：(a)超大规模公司支出可能超出已报告数字；(b)闭源模型利润可能低于预期。这对AI行业构成挑战，尤其考虑到大公司对OpenAI和Anthropic的投资。查看被引原帖 ↗
查看英文原文
completely agree. more revenue does not mean home free.
九原客@9hills · 中文博主 · 1 天前
做了个知识库skill，相当于给代码库建llm wiki。简单benchmark了下，KB on 效果全面落后 KB off，最困惑的是成本也变高了。
之前测试过很多传的很广的skill、code graph ，可能并不能带来正收益。
我现在pi的插件和skill 只剩寥寥几个。
Seedance 2.5 绝了，爱了
查看英文原文
Seedance 2.5 is crazy I love it.
Gary Marcus@GaryMarcus · 博主 · 23 小时前
OpenAI仍然是主流AI公司里最令人不安的
引用 Aakash Gupta @aakashguptaOpenAI三年内解散三支安全团队（Superalignment、Mission Alignment、Preparedness），称工作将嵌入产品团队。批评认为这是结构性问题：独立团队能延迟不安全发布，但内嵌团队需与自己发布时间表协商。6位资深安全领导人在两年内离职。此举与IPO筹备时机敏感，承诺的20%安全计算投入已化为乌有。查看被引原帖 ↗
查看英文原文
OpenAI continues to be the most disconcerting of the major AI companies.