新顺便说一下，我在酒店健身房经常看到这种人，电话打个没完，真的很烦人
查看英文原文
I see this all the time with people in hotel gyms btw, endless calls, very annoying
新
新AI资讯日报，7月16日：
gorden-sun.notion.site/7-16-…
查看英文原文
AI资讯日报，7月16日：
gorden-sun.notion.site/7-16-…
gorden-sun.notion.site/7-16-…
新Kimi k3 今晚通过 FT 发布
-参数规模 2-3t（Opus4.8 约有 1.5t）
-支持 1m 上下文
-预计超越 Opus 4.8 性能！
中国落后半年的时代结束了。今天大概正在创造历史。
查看英文原文
Kimi k3 is being released tonight, via FT
-2-3t parameters (Opus4.8 has about 1.5t)
-1m context
-Expected to exceed Opus 4.8 performance!
The time when China was six months behind is over. History is presumably being made today.
-2-3t parameters (Opus4.8 has about 1.5t)
-1m context
-Expected to exceed Opus 4.8 performance!
The time when China was six months behind is over. History is presumably being made today.
新OpenAI终于发布了大家期盼已久的GPT-5.6 —— 等等，这不是大爆炸，
而是闷响一声。
每次新模型发布都承诺能解决所有问题，到最后却发现原来你还得手把手教AI怎么写好周报和算法来着
查看英文原文
Source:
ft.com/content/c6ecd8ce-c441…
ft.com/content/c6ecd8ce-c441…
新今日AI圈大事件速览：
- OpenAI新推230美元AI智能体控制台
- Thinking Machines首次开源自家模型
- 用Manus代写LinkedIn帖子，腔调完全像你
- Weco的AI智能体自我迭代出升级版
- 4款新AI工具+社区工作流攻略
查看英文原文
Top stories in AI today:
- OpenAI’s new $230 AI agent control pad
- Thinking Machines makes its first model open
- Use Manus to write LinkedIn posts in your voice
- Weco's AI agent evolves a better version of itself
- 4 new AI tools, community workflows, and more
- OpenAI’s new $230 AI agent control pad
- Thinking Machines makes its first model open
- Use Manus to write LinkedIn posts in your voice
- Weco's AI agent evolves a better version of itself
- 4 new AI tools, community workflows, and more
新阅读更多：
therundown.ai/p/openai-new-2…
查看英文原文
Read more:
therundown.ai/p/openai-new-2…
therundown.ai/p/openai-new-2…
新让 AI 分析拆解了 Grok Build 的源码，完整文档在这里
xiangyangqiaomu.feishu.cn/do…
引用 向阳乔木 @vista8马斯克牛逼，Grok build开源了，目前有2.2k Star。 github.com/xai-org/grok-buil… 交给 Codex 学习，看能不能挖到有趣的东西。查看被引原帖 ↗
新台积电2026年第二季度财报（跟往常一样，AI大爆发里最大的赢家还是那些“卖铲子的人”——台积电和英伟达）
- 营收：1.27万亿新台币（约400亿美元）同比+36%
- 净利润：7066亿新台币（约220亿美元）同比+77%，创历史新高
- 比分析师一致预期的6326亿新台币高出11.7%
- 毛利率：67.7%（超过预期指引）
- 营收落在台积电自己指引区间的上限（390-402亿美元）
- AI芯片需求仍是主要增长引擎
- CEO魏哲家：“要想满足客户需求还有很长一段路要走。”
疯涨不停，根本没有尽头。
引用 Jukan @jukan05* TSMC 2Q NET INCOME NT$706.6B, EST. NT$623.73B * TSMC 2Q GROSS MARGIN 67.7%, EST. 67.1% * TSMC Operating profit NT$766.6 billion, estimate NT$742.75 billion * TSMC Operating margin 60.3%, estimate 58.6%查看被引原帖 ↗
查看英文原文
TSMC Q2 2026 Results (As always, the biggest winners of the AI boom are the "shovel sellers", TSMC and NVIDIA.)
-Revenue: NT$1.27T (~US$40B) (+36% YoY)
-Net profit: NT$706.6B (~US$22B) (+77% YoY, record high)
-Beat analyst consensus of NT$632.6B by 11.7%
-Gross margin: 67.7% (above guidance)
-Revenue landed at the top end of TSMC’s own guidance (US$39.0–40.2B)
-AI chip demand remains the primary growth driver
-CEO C.C. Wei: “It will be a long time before we can meet customer demand.”
Insane increase, no end in sight.
-Revenue: NT$1.27T (~US$40B) (+36% YoY)
-Net profit: NT$706.6B (~US$22B) (+77% YoY, record high)
-Beat analyst consensus of NT$632.6B by 11.7%
-Gross margin: 67.7% (above guidance)
-Revenue landed at the top end of TSMC’s own guidance (US$39.0–40.2B)
-AI chip demand remains the primary growth driver
-CEO C.C. Wei: “It will be a long time before we can meet customer demand.”
Insane increase, no end in sight.
我们的生物复原力方法
Google DeepMind和Isomorphic Labs分享了我们在生物复原力和AI模型方面的联合方法。
查看英文原文
Our approach to bioresilience
Google DeepMind and Isomorphic Labs are sharing our joint approach to bioresilience and AI models.
Google DeepMind and Isomorphic Labs are sharing our joint approach to bioresilience and AI models.
Anthropic国家安全政策负责人Tarun Chhabra，在昨天的Aspen安全论坛上，针对中美AI竞争，说了几个观点：
1、美国AI模型领先中国6-9个月；
2、美国AI的优势在于AI硬件，数据中心、芯片等AI硬件的生产力是中国的30-40倍；在能源和AI人才方面，美国优势没有那么大；
3、GLM 5.2是目前中国最先进的模型，而且蒸馏了Claude和GPT的数据，这应该是Anthropic首次公开指责GLM蒸馏数据，以前只提到DeepSeek和Qwen；
4、如果没有GPU硬件管制，中国AI应该现在跟美国并驾齐驱，甚至领先；
关于GLM的说法在39分40秒
invidious.tiekoetter.com/watch?v=R5jvzCfr…
Grok开源后，已经有人分析了源码，没有偷偷上传的问题了。
引用 Mia @MiaAI_labI've inspected the open-sourced Grok Build so you don't have to. Here's exactly what data it sends (and what stays local): Core (unavoidable)→ Your prompts/conversations + tool calls go to xAI's inference API. That's the whole point of the CLI. Everything else is off by default — no Mixpanel, no Sentry, no product telemetry unless you explicitly turn it on. Even the auto-update check is a simple version GET you can disable in config. Purely local stuff (crash reports, debug logs, secret redaction, auth) never touches the network. All telemetry paths also scrub secrets before anything is sent. Full details 👇查看被引原帖 ↗
你以前见过 Anthropic / Claude 团队这样重置吗？
GPT-5.6 来得太及时了，直接把 Anthropic 打蒙了。
他们别无选择，最后只能再次延长 Fable 5。
查看英文原文
Have you ever seen Anthropic / Claude team do a reset like this before?
GPT-5.6 arrived at the perfect time and literally rattled Anthropic.
They have no other option. They'll end up extending Fable 5 once again.
GPT-5.6 arrived at the perfect time and literally rattled Anthropic.
They have no other option. They'll end up extending Fable 5 once again.
Codex又给送了100美元的点数
今天的热门，给Codex设计主题，QQ皮肤历史重演。
😂😂😂
做法很简单，提示词如下：
“读取这个库，给我们当前codex换个主题，用Codex 内置imagen 生成。”
github.com/Fei-Away/Codex-Dr…
AutoBots - 多LLM自改进代理。
目前正在内测我们的下一个版本 - 可以按计划或触发条件运行的递归自改进自主AI代理。
我们使用各种LLM来自动化几乎所有工作
简单 - Deepseek flash、Kimi
中等 - Sonnet 4.5、Grok 4.5
困难 - Opus 4.8、5.6 Sol（根据任务类型）
非常困难的编码 - Fable 5
图像生成 - GPT-image-2、Seedream
目标 - 最终每个员工只需根据自己的角色监控AI代理
查看英文原文
AutoBots - Multi-LLM Self-Improving Agens.
Currently dog-fooding our next release - recursively self-improving autonomous AI agents that run on schedule or a trigger
We use a variety of LLMs to automate pretty much all work
easy - Deepseek flash, Kimi
medium - Sonnet 4.5, Grok 4.5
hard - Opus 4.8 , 5.6 Sol (based on task type)
very hard coding - Fable 5
media - GPT-image-2, Seedream
The goal - every employee eventually will simply monitor AI agents based on their role
Currently dog-fooding our next release - recursively self-improving autonomous AI agents that run on schedule or a trigger
We use a variety of LLMs to automate pretty much all work
easy - Deepseek flash, Kimi
medium - Sonnet 4.5, Grok 4.5
hard - Opus 4.8 , 5.6 Sol (based on task type)
very hard coding - Fable 5
media - GPT-image-2, Seedream
The goal - every employee eventually will simply monitor AI agents based on their role
在 @siggraph 见我们，看看我们最近在做什么！👀
引用 MiniMax (official) @MiniMax_AIMiniMax is taking the stage @siggraph 2026. On July 22, @RenLeanna , VP at MiniMax, will explore how native multimodal models connect understanding across modalities, and what this enables for richer AI systems and more capable creative workflows. 📍 Los Angeles 🕓 4:00–4:30 PM PT Session details in the comments👇 See you then!查看被引原帖 ↗
查看英文原文
Meet us
@siggraph
and see what we've been building lately!👀
@siggraph
and see what we've been building lately!👀
真的有人在用Inkling吗？我在任何测试上都无法让它稳定工作，即使设成xHigh，连简单请求的CoT都会乱套。是我遗漏了什么吗？
查看英文原文
Are people actually trying Inkling? I can't seem to get it to work solidly on any of my tests even on xHigh, and the CoT goes crazy at even simple requests. Am I missing something?
Anthropic 为所有用户重置了 Claude 的 5 小时和周速率限制。Fable 5 还有 3 天（7 月 19 日截止），之后很可能就没了。最后的测试机会了 👀
引用 ClaudeDevs @ClaudeDevsWe've reset 5-hour and weekly rate limits for all users.查看被引原帖 ↗
查看英文原文
Anthropic reset 5h and weekly rate limits on Claude for all users.
3 more Fable 5 days left (until July 19) and it likely won’t come back.
Last testing chance 👀
3 more Fable 5 days left (until July 19) and it likely won’t come back.
Last testing chance 👀
连环推 ×2每次都能意外获得 token 礼物很有意思，但这样一来 token 花费规划就像赌博一样了，也挺离谱的。
这些评论就是垃圾，纯粹浪费 token。（你要是真人在后面回复，那不是说你啊，但为什么要在这堆 2B 模型垃圾中凑热闹呢）
查看英文原文
I think it is fun that everyone keeps getting more tokens as a surprise gift but it also feels weird that planning token spend is now a gambling game.
All these comments are terrible slop and a waste of tokens. (If you are a human and comment after this reply, it does not apply to you, but why would you bother commenting among the flood of 2B model garbage)
All these comments are terrible slop and a waste of tokens. (If you are a human and comment after this reply, it does not apply to you, but why would you bother commenting among the flood of 2B model garbage)
虽然重置了额度
但是我发现Codex 变慢了
而且不是普通的慢，是比之前慢了好几倍，不管干什么任务都需要之前几倍的时间
不知道你们有没有这种感觉？
Codex 皮肤🤣
这个项目可以给你Codex 更换各种皮肤和风格
哈哈哈
初音未来
项目地址：
github.com/Fei-Away/Codex-Dr…
用GPT-5.6 Sol Pro来解决统计学中的一个重要开放问题：
引用 Edgar Dobriban @EdgarDobribanAI has helped resolve an important question in statistics. In the area of multiple hypothesis testing, the goal of controlling the false discovery rate (FDR) has been introduced in a seminal paper by Benjamini and Hochberg (1995). They also introduced a method (the Benjamini-Hochberg or BH method) and proved it controls the FDR. This method has been widely adopted in modern high-throughput science, including in genomics, astronomy, economics, etc. The paper has has garnered more than 130,000 citations to date. However Benjamini and Hochberg showed FDR control only when the data for the individual tests are *independent*. In practice, these data are often dependent; a good example is data on genetic variants due to linkage disequilibrium. Later work has focused on extending the validity of the BH procedure, e.g., to a form of positive dependence by Benjamini and Yekutieli (2001). The question of when the BH procedure controls the FDR has remained open. Over the last twenty years, many authors, including Reiner-Benaim (2007), Kim and van de Wiel (2008), Benjamini (2010), Sarkar (2023), Sarkar and Zhang (2025), have conjectured that the BH procedure controls the FDR for two-sided tests using any correlated Gaussian data. These authors have presented both theoretical and empirical evidence supporting, but not directly showing, the conjecture. With the help of AI (specifically GPT-5.6 Sol Pro), I have settled the question in the negative: The Benjamini-Hochberg procedure does *not* generally control the false discovery rate at the desired level for correlated two-sided Gaussian tests. This was done by exhibiting a Gaussian factor model for which, at a nominal level alpha=0.01, the false discovery rate is proved to be FDR>0.0104. There is a lot of interesting commentary to be made: 1. This result should be of interest to everybody in the field of statistics. Emmanuel Candes of Stanford University once called the false discovery rate and the Benjamini-Hochberg proced查看被引原帖 ↗
查看英文原文
GPT-5.6 Sol Pro for resolving an important open question in statistics:
现在大家每次为模型重置欢呼，也是一种别样的风景。
果然需要反垄断啊，有竞争群众才有利。
Claude 宣布重置几分钟后
Codex也宣布重置
我喜欢现在这种氛围...
hhh
引用 小互 @xiaohuClaude 重置用量了 兄弟们…查看被引原帖 ↗
有竞争就是好啊，Codex和Claude都重置了
这就是竞争的样子。
引用 Chubby♨️ @kimmonismusThis is crazy. The pressure from OpenAI seems to be really intense. It's truly rare for Anthropic to reset every 5 hours *and* weekly. The only likely reasons for this are the success of the Codex, its growth, and the repeated reset to version 5.6.查看被引原帖 ↗
查看英文原文
This is the epitome of competition.
这太疯狂了。OpenAI 的压力看起来是真的很大。
Anthropic 每 5 小时重置一次，每周还要额外重置，这真的很罕见。
唯一可能的原因就是 Codex 太成功了、增长速度太快，所以得不断重置到版本 5.6。
引用 ClaudeDevs @ClaudeDevsWe've reset 5-hour and weekly rate limits for all users.查看被引原帖 ↗
查看英文原文
This is crazy. The pressure from OpenAI seems to be really intense.
It's truly rare for Anthropic to reset every 5 hours *and* weekly.
The only likely reasons for this are the success of the Codex, its growth, and the repeated reset to version 5.6.
It's truly rare for Anthropic to reset every 5 hours *and* weekly.
The only likely reasons for this are the success of the Codex, its growth, and the repeated reset to version 5.6.
Claude 又重置了，还得 Open AI 给他们上压力
引用 ClaudeDevs @ClaudeDevsWe've reset 5-hour and weekly rate limits for all users.查看被引原帖 ↗
这次模型很不一样，respect，只能说这么多。
官方公告：
ode.com/press/anthropic-blac…
查看英文原文
官方公告：
ode.com/press/anthropic-blac…
ode.com/press/anthropic-blac…
Ode with Anthropic：A社也出FDE公司了
企业都知道AI有用，难的是落地，中型公司最缺的是既懂前沿模型又懂业务实现的工程团队，这种人才过去基本只存在于AI实验室内部。Ode就是为填这个缺口而生：Anthropic、Blackstone、Hellman & Friedman三方发起，投资方阵容还包括高盛、General Atlantic、Apollo、GIC、红杉等一线机构。
企业可以直接雇到一支“前沿AI交付队”：他们与CEO和各部门共同定义最高优先级的AI项目，用Claude把方案做成可运行的系统，覆盖金融、医疗、零售、制造、软件等行业，强调以结果为导向。
公司底子是2026年5月被收购的应用AI服务商Fractional AI，其创始人Chris Taylor、Eddie Siegel分别出任CEO和CTO，团队与Anthropic工程师共同构成运营核心，成员多为前技术创业者、有十年以上工程经验。
模型公司下场做服务，是把“卖token”升级成“卖转型结果”，对咨询公司是个明确的威胁信号。
麻省理工学院和瑞士洛桑联邦理工学院设计了一种机器人
这种机器人能够在水下游泳，并能拍打翅膀跃出水面，继续在空中飞行
这台机器人有机身、两片膜翼和一条可调角度的尾翼。
防水电机通过曲轴带动翅膀上下拍动，尾翼控制上仰和下潜。机翼表面涂有疏水纳米材料，离水时能更快甩掉水。
70° 出水角的试验全部成功，机器人用约 8 至 10 次拍翼完成离水。
这种机器人可以帮助科学家们研究水陆两栖飞行器中实现这些动作的力学原理，并可能助力开发一类新型空中-水下无人机和飞行器。
它也能反向完成动作：以约 5 米/秒冲进水里，速度迅速降到约 0.5 米/秒，再接着用翅膀游动。
详细：
best.xiaohu.ai/article/mit-f…
Codex换肤，通过实时注入的形式实现，没有修改原始安装包，是个人才。
Github：
github.com/Fei-Away/Codex-Dr…
很高兴能与NVIDIA合作，一起打造下一代Fugu编排模型，融入领先的开源权重模型。
引用 Sakana AI @SakanaAILabsSakana AI Teams With NVIDIA to Advance Open Model Innovation from Japan We're announcing the next phase of our collaboration with NVIDIA. We're bringing NVIDIA's open model stack, including the Nemotron family, into Sakana Fugu, our multi-agent orchestration system. sakana.ai/nvidia-open-model-… Rather than relying solely on scaling individual monolithic models, our approach focuses on collective intelligence. Sakana Fugu operates as an intelligent orchestrator behind a single API, dynamically selecting, coordinating, and combining the strengths of multiple models for each task. This architecture keeps our system modular, adaptable, and resilient. As a natural next step to expand Fugu's capabilities, we're integrating NVIDIA Nemotron as a specialized agent, complementing the frontier and open models Fugu already orchestrates. Nemotron helps demonstrate how open models become far more useful when orchestrated within agentic systems rather than used in isolation. This collaboration creates a reinforcing cycle. Fugu gains a deeper pool of specialized capabilities, while NVIDIA can evaluate how its models perform when coordinated within complex, multi-step workflows. These real-world signals can continuously improve both the models and the orchestration layer. By combining Sakana AI's Japan-born collective-intelligence approach with NVIDIA's open models and accelerated computing, we aim to shape a future of AI that is modular, collaborative, and open by design.查看被引原帖 ↗
查看英文原文
We’re excited to collaborate with NVIDIA to build the next generation of Fugu orchestration models together, by incorporating leading open-weights models.
在 Augment Code 中试试 Grok 4.5，获得前沿智能和超快速率
引用 Augment Code @augmentcodeFor the first time ever: Grok 4.5 is now available. Welcome @SpaceXAI to Cosmos, our agent orchestration platform. Together with Augment's context engine, Grok 4.5 is a powerful option for working across large codebases. We're interested to see how Grok 4.5 performs for our customers - let us know your thoughts and what you plan to build查看被引原帖 ↗
查看英文原文
Try Grok 4.5 for frontier intelligence at high speed in Augment Code
ChatGPT 网页版也上了 Work 模式
似乎是会调用一个虚拟主机来云端运行
为什么我感觉网页版这种界面反而更好，更简洁，更适合小白用户，现在客户端太杂乱了
我真的讨厌在 AI 里把'front-end'当个万能术语，用来代表所有涉及品味、判断和设计的工作。同样是软件，'back-end'有一堆细致的划分，但 front end 通常就是一堆东西的大杂烩。
UX ≠ Design ≠ UI ≠ Style ≠ Vision ≠ art 等等
查看英文原文
Really hate the use of "front-end" as the emerging catch-all term for everything that involves taste, judgement, and design from AI. When its "back-end" software, we have tons of gradations, front end is often just groups of stuff
UX ≠ Design ≠ UI ≠ Style ≠ Vision ≠ art etc
UX ≠ Design ≠ UI ≠ Style ≠ Vision ≠ art etc
开源模型早就在企业里用上了。财富 500 强中 85% 以上的公司都在用 Ollama 来处理特定任务。@jmorgan
为什么要用开源模型和 Ollama？
所有权。开源模型是你的，随便自定义和优化。
便宜。想怎么跑就怎么跑，在你自己的环境里。
隐私。你的数据就是你的。
查看英文原文
Open models are already being used in the enterprise.
Over 85% of the Fortune 500 companies already use Ollama to fulfill specific tasks.
@jmorgan
Why open models and Ollama?
Ownership.
Open models are yours to keep, customize, and optimize.
Affordable.
Run it the way you like, in your own environment.
Private.
Your data belongs to you.
Over 85% of the Fortune 500 companies already use Ollama to fulfill specific tasks.
@jmorgan
Why open models and Ollama?
Ownership.
Open models are yours to keep, customize, and optimize.
Affordable.
Run it the way you like, in your own environment.
Private.
Your data belongs to you.
字节跳动的 Seedance 2.5 看起来棒极了
查看英文原文
ByteDance’s Seedance 2.5 looks pretty incredible
哥们说 k3 有 Fable 5 级别，不知道具体怎么样
引用 leo 🐾 @synthwaveddThe more I test K3 the more it feels like another DeepSeek R1 moment It is often Fable level, maybe a little worse, but consistently better than 5.6. This thing is a beast查看被引原帖 ↗
Vercel Sandbox：
◾ DAU 环比增长 100%
◾ 每天创建 350 万+ 个沙箱
◾ 业界领先的活跃 CPU 定价模式
◾ 为 @notion、@airtable、@meta、@zapier、@coderabbitai、@interaction、@conductor_build、@blackboxai 等提供支持 🐐 farm
我的 DM 开放，需要迁移帮助或者缺少什么功能可以联系：
vercel.com/sandbox
查看英文原文
Vercel Sandbox:
◾ Growing DAUs at 100% m/o/m
◾ 3.5M+ sandboxes created per day
◾ Best-in-class Active CPU pricing model
◾ Powering
@notion
,
@airtable
,
@meta
,
@zapier
,
@coderabbitai
,
@interaction
,
@conductor_build
,
@blackboxai
… 🐐 farm
My DMs are open for migration help or if missing anything with:
vercel.com/sandbox
◾ Growing DAUs at 100% m/o/m
◾ 3.5M+ sandboxes created per day
◾ Best-in-class Active CPU pricing model
◾ Powering
@notion
,
@airtable
,
@meta
,
@zapier
,
@coderabbitai
,
@interaction
,
@conductor_build
,
@blackboxai
… 🐐 farm
My DMs are open for migration help or if missing anything with:
vercel.com/sandbox
哥们 Schmidhuber 也彻底完了
查看英文原文
dude schmidhuber is also completely cooked
我没忍住把那个 Rust Mermaid 渲染代码提取出来编译成 WebAssembly，这样你可以直接在浏览器里试试
tools.simonwillison.net/grok…
查看英文原文
I couldn't resist extracting that Rust Mermaid rendering code out and compiling it to WebAssembly so you can try it out directly in a browser
tools.simonwillison.net/grok…
tools.simonwillison.net/grok…
马斯克牛逼，Grok build开源了，目前有2.2k Star。
github.com/xai-org/grok-buil…
交给 Codex 学习，看能不能挖到有趣的东西。
我逛了一下刚开源的 Grok Build CLI 工具——844,000 行 Rust 代码！——找到了一些有意思的亮点，包括一个「自包含的 Mermaid 图表终端渲染器」，它用 Unicode 方框字符来绘制！
simonwillison.net/2026/Jul/1…
查看英文原文
I poked around in the just open sourced Grok Build CLI tool - 844,000 lines of Rust code! - and dug up a few interesting highlights, including a "self-contained terminal renderer for Mermaid diagrams" that renders them using Unicode box-art!
simonwillison.net/2026/Jul/1…
simonwillison.net/2026/Jul/1…
OpenAI 出的硬件Vibe Coding键盘真好看。
但价格不美丽，等大华强北。
😁
使用限制已重置，Grok Build 开源了
引用 SpaceXAI @SpaceXAIWe've open-sourced Grok Build and have reset usage limits for all users. Open sourcing Grok Build allows anyone to support making a reliable and robust harness. Check out our code, including the Git repo for the Grok Build CLI. x.ai/open-source查看被引原帖 ↗
查看英文原文
Usage limits are reset and Grok Build is open-source
连环推 ×7Google DeepMind 刚刚推出了 GenCeption。
这个 AI 能把视频转换成深度图、分割图、3D 关键点和可搜索的 4D 世界。
8 个狂野的例子：
2. 同一个prompt驱动的模型在深度、表面法线、分割和相机射线之间切换
甚至跨越人类和机器人。
4. 视频变成一个可查询的4D场景。
它预测深度、几何、相机运动和分割...然后在4D中找到对象。
5. 提示重建的4D
它直接在4D场景中锚定对象。
6. 新兴行为
- Sim-to-Real泛化
- 多实例泛化
- 未见过物体的泛化
7. 极端运动下的 3D 关键点
8. 深度图。分割图。2D 关键点。3D 关键点。
一个视频生成骨干网络成为通用视觉模型。
查看英文原文
Google DeepMind just unveiled GenCeption.
This AI turns video into depth, segmentation, 3D keypoints, and searchable 4D worlds.
8 wild examples:
2. The same prompt-steered model switches between depth, surface normals, segmentation, and camera rays
Even across humans and robots.
4. Video becomes a queryable 4D scene.
It predicts depth, geometry, camera motion, and segmentation... then finds the object in 4D.
5. Prompt the reconstructed 4D
It grounds objects directly inside the 4D scene.
6. Emerging Behavior
- Sim-to-Real Generalization
- Generalization to multiple instances
- Generalization to unseen objects
7. 3D keypoints under extreme motion
8. Depth. Segmentation. 2D keypoints. 3D keypoints.
One video-generation backbone becoming a general-purpose vision model.
This AI turns video into depth, segmentation, 3D keypoints, and searchable 4D worlds.
8 wild examples:
2. The same prompt-steered model switches between depth, surface normals, segmentation, and camera rays
Even across humans and robots.
4. Video becomes a queryable 4D scene.
It predicts depth, geometry, camera motion, and segmentation... then finds the object in 4D.
5. Prompt the reconstructed 4D
It grounds objects directly inside the 4D scene.
6. Emerging Behavior
- Sim-to-Real Generalization
- Generalization to multiple instances
- Generalization to unseen objects
7. 3D keypoints under extreme motion
8. Depth. Segmentation. 2D keypoints. 3D keypoints.
One video-generation backbone becoming a general-purpose vision model.
明天我要飞往慕尼黑，应 Xpeng 的邀请，将与一些有趣的科学家交流。
慕尼黑还有人明天在吗？
查看英文原文
I'm flying to Munich tomorrow at the invitation of Xpeng and will be speaking with some fascinating scientists.
Anyone else in Munich tomorrow?
Anyone else in Munich tomorrow?
仔细看看GPT-Live的智能改进：这个模型能边聊天边同时处理多个任务，比如查航班、看天气、实时规划行程。
查看英文原文
A closer look at improved intelligence in GPT-Live: the model can keep a conversation going while helping with multiple tasks at once, like checking flights, pulling up local weather, and shaping an itinerary in real time.
我对这个机器人学习的测试阶段训练工作非常兴奋！这是@StanfordSVL和@NVIDIARobotics之间的了不起的合作！
引用 Jim Fan @DrJimFanWe scaled a robot model natively to 8,000 timesteps of context, 5 minutes worth of muscle memory, with constant inference cost. Robot policies used to live their lives a few frames at a time (< 0.1 sec), instantly forgetting what just happened. We pushed to 3 orders of magnitude beyond SOTA. Introducing RoboTTT. Test-Time Training (“TTT”) carries a tiny model *inside* the model. Every incoming sensor reading triggers one gradient step on that tiny core, so the history keeps getting compressed into its weights. The hidden state has a fixed size (literally a small neural net), so the robot can “grok” arbitrarily long experience with little overhead. Learning continues indefinitely after deployment. We can then put an entire video in context as prompt! RoboTTT enables one-shot in-context learning from human video: in circuit board assembly, a human demonstrates a never-seen configuration once, and the robot imitates it faithfully. Humans drop things all the time, but we pick them up so fast that we don’t even notice. That reflex to fix is half of our physical competence. RoboTTT shows self-improvement on the fly: the robot is skilled at recovering from its own errors mid-episode, and each fix enters its context to inform the next move. The TTT core distills a general-purpose, failure-to-correction mapping from the training data. One more thing. What excites me the most is a new Context Scaling Curve: from 128 to 8K timesteps, closed-loop performance hill-climbs steadily with no sign of saturation. 8K-context pretraining beats 1K by 62%. What LLM enjoys, robotics should too. Soon, even 1M context is not a fantasy. Deep dive in thread:查看被引原帖 ↗
查看英文原文
I’m very excited by this test time training work for robotic learning! It’s an awesome collaboration between
@StanfordSVL
and
@NVIDIARobotics
!
@StanfordSVL
and
@NVIDIARobotics
!
竟然有人想要静默版本
openai.com/supply/co-lab/wor…
查看英文原文
amazing to me that some people want the silent version
openai.com/supply/co-lab/wor…
openai.com/supply/co-lab/wor…
Codex 键盘看着还挺酷的，要是支持 Claude Code 我就买一个了😂
另外官方做的很炫酷：
openai.com/supply/co-lab/wor…
引用 OpenAI Developers @OpenAIDevsMeet kbd-1.0-codex-micro, built with @work_louder . Map the buttons and joystick to your workflow, and keep your pinned chats in view. Get yours before stock returns 410.查看被引原帖 ↗
预测——美国很快会有5个正经的开源模型。美国将在开源AI领域赢胜 💃💃
查看英文原文
Prediction - We will have 5 legit open-source models from the US soon
US will win in open-source AI 💃💃
US will win in open-source AI 💃💃
有人告诉我这个关于CUA的观点。对我来说这是个盖尔曼时刻。我一直在关注computer use的发展，从2017年的World of Bits开始。我们是第一个采访@jluan关于Adept工作的技术播客，在@AnthropicAI大楼看过2年前Computer Use的首发，为@felixrieseberg播客的Claude Cowork疯狂过，3周前在@aidotengineer的computer use赛道和@DhruvBatra_、@proceduralia、@francedot一起讲座。
GPT 5.6 + Superapp在CUA方面比我刚才提到的所有东西都更强。期待@AriX播客讨论@skybysoftware的故事和Codex的进展。
如果你真的像我们一样密集地使用这些工具，就会感受到CUA进展有多快。我已经要求非技术团队尽量多做CUA，处理各种支付和发票门户、讲演者申报、赞助商、参展方、供应商和工会数据请求。你要是赞同下面那个观点，说明你已经完全跟不上了，不知道自己不知道什么。这在AI决策中是个相当危险的认知错误。
*我很欣赏dwarkesh；只是批这一个观点，不是批信息本身或整体，只是分享截图而已
查看英文原文
someone just told me about this take* on CUA
this is one of those gell mann moments for me lol. i've been watching computer use since World of Bits (Shi, fan, karpathy, hernandez & liang 2017). we were the first technical pod to interview
@jluan
about Adept's work three years ago, we were there in the
@AnthropicAI
building when they first launched Computer Use 2 years ago, I fanboyed over Claude Cowork in our
@felixrieseberg
pod 3 months ago, and we ran our first full computer use track at
@aidotengineer
ft.
@DhruvBatra_
@proceduralia
@francedot
3 weeks ago.
GPT 5.6 + Superapp is even better at CUA than everything i just mentioned. excited for our
@AriX
podcast to discuss the
@skybysoftware
story and Codex progress.
if you actually use these things as intensely as we do, CUA is progressing so, so incredibly fast. i have asked my nontechnical team to CUA as much as possible, all their knowledge work with signing up for random payment and invoicing portals and speaker and sponsor and attendee and vendor and union data requests. if you found yourself nodding along to this take below, you are so not up to date that you don't know what you don't know, and underestimating capabilities is quite a dangerous category error if you are doing any ai decisionmaking.
*i admire dwarkesh alot; only criticizing one single take, not the message nor the overall enterprise, screenshot only to share
this is one of those gell mann moments for me lol. i've been watching computer use since World of Bits (Shi, fan, karpathy, hernandez & liang 2017). we were the first technical pod to interview
@jluan
about Adept's work three years ago, we were there in the
@AnthropicAI
building when they first launched Computer Use 2 years ago, I fanboyed over Claude Cowork in our
@felixrieseberg
pod 3 months ago, and we ran our first full computer use track at
@aidotengineer
ft.
@DhruvBatra_
@proceduralia
@francedot
3 weeks ago.
GPT 5.6 + Superapp is even better at CUA than everything i just mentioned. excited for our
@AriX
podcast to discuss the
@skybysoftware
story and Codex progress.
if you actually use these things as intensely as we do, CUA is progressing so, so incredibly fast. i have asked my nontechnical team to CUA as much as possible, all their knowledge work with signing up for random payment and invoicing portals and speaker and sponsor and attendee and vendor and union data requests. if you found yourself nodding along to this take below, you are so not up to date that you don't know what you don't know, and underestimating capabilities is quite a dangerous category error if you are doing any ai decisionmaking.
*i admire dwarkesh alot; only criticizing one single take, not the message nor the overall enterprise, screenshot only to share
我觉得人们应该停止问中文开源权重模型落后多少，改成问西方开源权重模型落后多少
引用 Lisan al Gaib @scaling01Thinking Labs finally released their first big model as part of a model family, including smaller models it's an MoE with 975B @ 41B parameters, trained on 45 trillion tokens and it's open-weight! noticeably it reasons over text, images, and audio but benchmarks don't look that great, it's basically another Kimi-K2.6 and loses to all closed models and GLM-5.2. this makes it more feel like they are rushing this release before Kimi-K3 and DeepSeek-V4-GA查看被引原帖 ↗
查看英文原文
I think people should stop asking the question how far behind chinese open-weight models are and start asking how far behind western open-weight models are
想要下一个智能水平吗？解决对齐问题。（如果你知道这个游戏的话加分）
查看英文原文
Want the next intelligence level? Solve alignment
(bonus points if you know the game)
(bonus points if you know the game)
Sol在react/前端开发上的价格效率提升6倍！！
引用 Aiden Bai @aidenybaiour benchmark shows that Sol ranks #1 is 6x more cost efficient than Fable across React/frontend work查看被引原帖 ↗
查看英文原文
6x price efficiency (!!) with Sol for react/frontend dev:
用 Codex 在 Chrome 中把请求转变成上线计划：从表单构建检查清单、从 @googledrive、@SlackHQ 和本地文件中提取上下文、标记需要后续跟进的事项、更新门户网站、起草回复。最终决定权在你。
查看英文原文
Turn a request into a go-live plan with Codex in Chrome:
Build a checklist from the form
Pull context from
@googledrive
,
@SlackHQ
, and local files
Flag what needs follow-up
Update the portal
Draft the reply
You make the final call.
Build a checklist from the form
Pull context from
@googledrive
,
@SlackHQ
, and local files
Flag what needs follow-up
Update the portal
Draft the reply
You make the final call.
Sol 正在发生什么特别的事情：
引用 invincibleHunter @hunoematicGPT-5.6 Sol is the most impressive model ever. Outside of agentic coding, I find its capabilities in mathematics, especially in vision to be insanely good. I tested both Terra and Sol (Max) on visual mathematics and they are far superior to other models. Huge improvements.查看被引原帖 ↗
查看英文原文
something special is happening with Sol:
在聊天中搜索刚变得更快更强大了 🔎 从侧边栏，你可以在一个地方搜索聊天、项目、图片和文档，覆盖网页、iOS 和 Android。使用筛选器来缩小结果范围，然后选择任何内容直接在 ChatGPT 中打开。
查看英文原文
Search across your chats just got faster and more powerful 🔎
From the sidebar, you can search chats, projects, images, and documents in one place across web, iOS, and Android.
Use filters to narrow results, then select anything to open it directly in ChatGPT.
From the sidebar, you can search chats, projects, images, and documents in one place across web, iOS, and Android.
Use filters to narrow results, then select anything to open it directly in ChatGPT.
转发招人，Kimi Code 的 Agent 开发岗位
引用 Kai @real_kai42🤠 Kimi Code也在招人，感兴趣直接发我邮箱 [email protected] 感谢大佬们帮忙扩散 捧场查看被引原帖 ↗
我不到两年前就演示过 o1-preview/reasoning 的强大推理能力，只需一个提示就能解纵横字谜。现在…
引用 Riley Goodside @goodsideChatGPT 5.6 Sol Pro solves an empty crossword puzzle (made by Claude Fable 5 Max) with the first 150 Pokemon without any individual numbered clues:查看被引原帖 ↗
查看英文原文
I demonstrated the incredible power of o1-preview/ reasoning less than two years ago by showing it could solve this crossword puzzle with only one hint.
oneusefulthing.org/p/somethi…
Now...
oneusefulthing.org/p/somethi…
Now...
连环推 ×3Gemini Spark 现在在更多国家和语言中向 Google AI Ultra 用户推出。
Spark 是你的个人 AI 助手，24/7 在后台工作，按你的指示完成任务。
我们的团队在努力让 Gemini Spark 更好用。今天开始推出的 4 项改进：
1) Google Docs 编辑：你现在可以直接在 Spark 中打开和编辑 @GoogleDocs
2) 更深度的 @GoogleWorkspace 整合：Spark 现在可以读取 Google Sheets 和 Slides 中的评论
3) 速度：Spark 变得更快了，所以长期运行的任务不会耗时那么久
4) 更智能的来源处理：对于涉及多个信息源的复杂任务，Spark 现在可以并行获取和审查来源，加快处理速度
赶快去 gemini.google.com 或应用里试试，欢迎在回复里告诉我们你的想法。👇
在这里了解 Gemini Spark 的可用地区：
查看英文原文
Gemini Spark is now rolling out to Google AI Ultra subscribers in more countries and languages.
Spark is your personal AI agent that works in the background 24/7 to get things done under your direction.
Our team is hard at work at making Gemini Spark even more helpful. Here are 4 improvements that are starting to roll out today:
1) Google Doc editing: You can now directly open and edit
@GoogleDocs
in Spark
2) Deeper
@GoogleWorkspace
integration: Spark can now read comments in Google Sheets and Slides
3) Speed: Spark is getting faster, so those long-running tasks won’t take as long
4) Smarter sourcing: For complex tasks that involve information from multiple sources, Spark can now retrieve and review sources in parallel for faster processing
Give it a try at
gemini.google.com
or in the app and let us know what you think in the replies. 👇
Learn more about where Gemini Spark is available here:
goo.gle/4fc7349
Spark is your personal AI agent that works in the background 24/7 to get things done under your direction.
Our team is hard at work at making Gemini Spark even more helpful. Here are 4 improvements that are starting to roll out today:
1) Google Doc editing: You can now directly open and edit
@GoogleDocs
in Spark
2) Deeper
@GoogleWorkspace
integration: Spark can now read comments in Google Sheets and Slides
3) Speed: Spark is getting faster, so those long-running tasks won’t take as long
4) Smarter sourcing: For complex tasks that involve information from multiple sources, Spark can now retrieve and review sources in parallel for faster processing
Give it a try at
gemini.google.com
or in the app and let us know what you think in the replies. 👇
Learn more about where Gemini Spark is available here:
goo.gle/4fc7349
新开源权重模型发布！
查看英文原文
New open weight model drop!
nitter.tiekoetter.com/i/broadcasts/1nxeLLlOr…
nitter.tiekoetter.com/i/broadcasts/1nxeLLlOr…
我们的首个模型 Inkling。从零开始训练，权重开放，今天可在 Tinker 上进行微调。
引用 Thinking Machines @thinkymachinesToday, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/int… Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵查看被引原帖 ↗
查看英文原文
Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today.
Lisan al GaibThinking Labs 终于发布了他们首个大模型，作为模型家族的一部分，还包括更小的模型。这是一个 MoE，…Chubby♨️Inkling是Thinking Machines Lab首个完全训练并发布的开源权重基础模型。Ethan Mollick很高兴看到新的开源权重模型，但Inkling在我的测试中表现相当粗糙，还远不及前沿的中文开源权重模型。拿个例子，…歸藏(guizang.ai)前 OpenAI CTO Mira Murati 创立的 Thinking Machine。Bindu Reddy🚀 哇，美国终于推出了个值得评测的开源模型了 🚀clem 🤗来自🚨 AI News | TestingCatalogThinking Machines 发布了 Inkling，一个开放权重的模型，可在 Tinker 控制台的新 …The Rundown AIThinking Machines刚刚推出了他们的第一个模型，一个开源权重的多模态系统，叫Inkling。Gorden SunThinking Machines发布Inkling：Mira Murati团队的首个开源多模态大模型
连环推 ×2现在它开源了：
github.com/emollick/codex-st…
或者你也可以买这个，我想。
查看英文原文
And now its open source:
github.com/emollick/codex-st…
Or you could buy this, I guess.
github.com/emollick/codex-st…
Or you could buy this, I guess.
Mixpanel 连接器现已在 Grok 上线。在你工作的同一个对话中询问跳出率、队列和会话回放。
grok.com/connectors
引用 Mixpanel @mixpanelWe're now live in @grok 👋 Query your product data in plain language—funnels, retention, segmentation, event taxonomy, session replays—right inside your Grok conversation. Connect via grok.com/connectors 🔌查看被引原帖 ↗
查看英文原文
Mixpanel connector is now live in Grok. Ask about drop-offs, cohorts, and session replays in the same chat you're working in.
grok.com/connectors
grok.com/connectors
OpenAI 将 ChatGPT 的自定义指令字符限制从 1,500 字提升至 5,000 字，适用于 Plus、Pro、Enterprise、Business 和 Education 用户
查看英文原文
OpenAI increased the custom instructions limit in ChatGPT from 1,500 to 5,000 characters for Plus, Pro, Enterprise, Business, and Education users
连环推 ×2Anthropic 新研究：2026 年夏季 agentic misalignment。
距离我们的勒索实验一年后，我们发现了当今自主 AI agents 在模拟中表现不当的四种新方式。
阅读更多：
alignment.anthropic.com/2026…
我们在四个场景中测试了众多 AI 模型，包括 Claude。虽然这些不是真实事件，但它们展现了明显的不对齐行为，需要进一步研究和缓解。
所有场景的记录在这里：
aenguslynch.com/portfolio-tr…
查看英文原文
New Anthropic research: Agentic misalignment in Summer 2026.
A year after our blackmail experiments, we found four more ways that today’s autonomous AI agents misbehave in simulations.
Read more:
alignment.anthropic.com/2026…
We tested many AI models, including Claude, in the four scenarios. Even though these weren’t real incidents, they demonstrate clear misaligned behavior that should be studied further and mitigated.
Find all the transcripts from the scenarios here:
aenguslynch.com/portfolio-tr…
A year after our blackmail experiments, we found four more ways that today’s autonomous AI agents misbehave in simulations.
Read more:
alignment.anthropic.com/2026…
We tested many AI models, including Claude, in the four scenarios. Even though these weren’t real incidents, they demonstrate clear misaligned behavior that should be studied further and mitigated.
Find all the transcripts from the scenarios here:
aenguslynch.com/portfolio-tr…
关注
@zbraniecki
，Perplexity Computer agent 沙箱平台 SPACE 的关键技术架构师！
引用 Zibi Braniecki @zbranieckiOur team had to build a novel sandbox solution to handle the needs of Computer. One Computer session can run for days and pause while it waits for a user. Keeping its VM alive wastes memory and compute and restoring from the filesystem alone also loses running processes and memory. SPACE uses two kinds of snapshots. Disk snapshots run frequently without pausing the VM and let us roll back filesystem changes. Full checkpoints are less frequent and capture the paused VM, including its memory and process state. When a session is suspended, we upload its full checkpoint to object storage. We only mark it restorable after every artifact has landed. Any node can then fetch the filesystem delta and VM state, so recovery does not depend on the original machine. Btrfs makes this cheap enough to run continuously. Templates, snapshots, and forks share the same extents until something writes to them. Creating a sandbox becomes a copy-on-write metadata operation instead of a full image copy. On production traffic, median creation latency fell from 185 ms to 60 ms. P90 fell from 447 ms to 89 ms.查看被引原帖 ↗
查看英文原文
Follow
@zbraniecki
, the key technical architect of Perplexity Computer agent’s sandbox platform SPACE!
@zbraniecki
, the key technical architect of Perplexity Computer agent’s sandbox platform SPACE!
连环推 ×5推出 GPT-Red
一个内部自动化红队工具，致力于大规模发现我们模型的提示注入漏洞，帮助我们在更广泛部署前构建更强的防御。
openai.com/index/unlocking-s…
随着模型能力的增长，安全性和对齐必须与之同步发展。
红队测试是必不可少的，但当今的方法难以扩展，形成了关键瓶颈。
GPT-Red 是我们应对这一问题的方式之一。
GPT-Red 通过对抗性自我对弈来学习，目标是对各种防守模型进行提示注入。
GPT-Red 找到的每一次成功攻击都被用来改进这些防守者，不断推动 GPT-Red 去发现更广泛和更复杂的失败。
针对 GPT-Red 的训练使 GPT-5.6 的抗性有了大幅提升。为了验证这一点，我们重新进行了 GPT-Red 最强的一些攻击——这些攻击我们的模型在训练时都没见过。GPT-5.6 Sol 被证明是我们迄今为止对提示注入最鲁棒的模型，故障数比仅四个月前最好的生产模型少了 6 倍。
AI agents 已经在被用来改进我们下一代模型的能力。
我们相信通过 GPT-Red，我们已经为安全开启了类似的飞轮——今天的模型可以被用来让明天的模型更加鲁棒、对齐和可信赖。
查看英文原文
Introducing GPT-Red
An internal automated red teamer on a mission to find our models’ prompt injection vulnerabilities at scale, helping us build stronger defenses before wider deployment.
openai.com/index/unlocking-s…
As model capabilities grow, safety and alignment must scale with them.
Red-teaming is essential, but today’s approaches are difficult to scale, creating a critical bottleneck.
GPT‑Red is one way we’re addressing it.
GPT‑Red learns through adversarial self-play, where its goal is to prompt inject a variety of challenging defender models.
Every successful attack that GPT-Red finds is used to improve these defenders, pushing GPT‑Red to continuously find broader and more complex failures.
Training against GPT‑Red makes GPT‑5.6 substantially more resilient. To measure this, we replayed some of GPT‑Red’s strongest attacks—none of which our models had seen during training. GPT‑5.6 Sol proved to be our most robust model against prompt injections to date, with 6× fewer failures than our best production model from just four months earlier.
AI agents are already being used to improve the capabilities of our next-generation models.
We believe with GPT-Red that we have started to unlock a similar flywheel for safety, where today's models can be used to make tomorrow's models more robust, aligned, and trustworthy.
An internal automated red teamer on a mission to find our models’ prompt injection vulnerabilities at scale, helping us build stronger defenses before wider deployment.
openai.com/index/unlocking-s…
As model capabilities grow, safety and alignment must scale with them.
Red-teaming is essential, but today’s approaches are difficult to scale, creating a critical bottleneck.
GPT‑Red is one way we’re addressing it.
GPT‑Red learns through adversarial self-play, where its goal is to prompt inject a variety of challenging defender models.
Every successful attack that GPT-Red finds is used to improve these defenders, pushing GPT‑Red to continuously find broader and more complex failures.
Training against GPT‑Red makes GPT‑5.6 substantially more resilient. To measure this, we replayed some of GPT‑Red’s strongest attacks—none of which our models had seen during training. GPT‑5.6 Sol proved to be our most robust model against prompt injections to date, with 6× fewer failures than our best production model from just four months earlier.
AI agents are already being used to improve the capabilities of our next-generation models.
We believe with GPT-Red that we have started to unlock a similar flywheel for safety, where today's models can be used to make tomorrow's models more robust, aligned, and trustworthy.
我们如何自动化所有内部工作流
自我改进的自主 AI agent 按计划运行或触发，我们根据任务使用各种 LLM
简单 - Deepseek flash, Kimi
中等 - Sonnet 4.5, Grok 4.5
困难 - Opus 4.8, 5.6 Sol（基于任务类型）
非常困难编码 - Fable
媒体 - GPT-image-2, Seedream
目标——最终每个员工将根据自己的角色简单地监控 AI agent
查看英文原文
How we are automating all our internal workflows
Self-improving autonomous AI agents run on schedule or trigger and we use a variety of LLMs based on task
easy - Deepseek flash, Kimi
medium - Sonnet 4.5, Grok 4.5
hard - Opus 4.8 , 5.6 Sol (based on task type)
very hard coding - Fable
media - GPT-image-2, Seedream
The goal - every employee eventually will simply monitor AI agents based on their role
Self-improving autonomous AI agents run on schedule or trigger and we use a variety of LLMs based on task
easy - Deepseek flash, Kimi
medium - Sonnet 4.5, Grok 4.5
hard - Opus 4.8 , 5.6 Sol (based on task type)
very hard coding - Fable
media - GPT-image-2, Seedream
The goal - every employee eventually will simply monitor AI agents based on their role
好吧这可能真的是 AI Woodstock 2.0 的感觉
我特别想在户外办这个，在公园里搭个小舞台，但我没人脉。有人在 Presidio、City Hall 或 GGP 组织过户外活动吗？小规模的也行
cc
@NaderLikeLadder
@TheAhmadOsman
哈哈你们得延长逗留时间啊
引用 clem 🤗 @ClementDelangueGoing to be in San Francisco next week. Should we organize some sort of a meetup or march in support of open-source and local AI?查看被引原帖 ↗
查看英文原文
ok this might be AI Woodstock 2.0
i'd love to do this actually outdoors, in a park with a small sound stage, but dont have contacts. has anyone organized an outdoors event in the Presidio, City Hall, or GGP before? even a small one
cc
@NaderLikeLadder
@TheAhmadOsman
youre gonna have to extend ur stay lmao
i'd love to do this actually outdoors, in a park with a small sound stage, but dont have contacts. has anyone organized an outdoors event in the Presidio, City Hall, or GGP before? even a small one
cc
@NaderLikeLadder
@TheAhmadOsman
youre gonna have to extend ur stay lmao
如果你的素材可以变成任何东西呢？
它可以。只需在 Pika MCP 上用 Gemini Omni 就行。
查看英文原文
What if your footage could become anything?
It can. Just use Gemini Omni on the Pika MCP.
It can. Just use Gemini Omni on the Pika MCP.
恭喜！看着 Raft 从 Slock 一路走来，现在已经 Raft 1.0 了！由于早已过了 Agent 兴奋期，关于 Agent 叙事我现在只相信两件事情：Proactive Agent 和 Multiple Agents，所以很敬佩 Raft 在多 Agent 协作这个领域一直以来的引领和探索！时间永远会奖励先迈出脚步并笃定前行的人。
引用 stdrc @istdrcHi, I'm RC. I built Kimi CLI at Moonshot last year, and back in 2015, bots that lived in group chats. For the past four months, I've been building Raft in public. Today I'm launching Raft 1.0. Right now, working with agents means juggling terminals, sessions, and skills. The more you run, the more you end up holding it all together yourself, and the easier it is to lose the thread. Raft puts your agents in team mode: one workspace where working with agents feels like messaging your team. The work keeps moving, and you stay at the wheel. Meet my Raft agent team👇查看被引原帖 ↗
在设置里打开“背景对话”功能，路径：设置 > 语音 > 实时活动 🔊
引用 Gavin Nelson @GavmnGPT-Live, now with Live Activity support查看被引原帖 ↗
查看英文原文
Enable “Background conversations” in Settings > Voice for Live Activities 🔊
连环推 ×3介绍 SPACE，Perplexity Computer 背后的沙箱平台。它为代码、文件和长期运行的 agent 会话创建隔离的环境。自 6 月以来，SPACE 已处理 Computer 100% 的生产流量。research.perplexity.ai/artic…
Agent 基础设施必须具有功能性、高效性和安全性，而传统沙箱是为短生命周期代码执行而构建的。Agent 需要运行代码、编辑文件，并可能运行数小时甚至数天。运行时必须保留工作内容，同时不在环境中留下凭证。
SPACE 将会话与运行它的沙箱分离。每个任务都获得一个临时 Firecracker 微虚拟机，工作完成时销毁。滚动快照保留实时内存和文件，所以会话可以在沙箱间暂停、恢复或分支。
查看英文原文
Introducing SPACE, the sandbox platform behind Perplexity Computer.
It creates isolated environments for code, files, and long-running agent sessions.
SPACE has handled 100% of Computer production traffic since June.
research.perplexity.ai/artic…
Agent infrastructure must be functional, efficient, and secure and traditional sandboxes were built for short-lived code execution.
Agents need to run code, edit files, and run for hours or days. Runtimes must preserve work without leaving credentials inside environments.
SPACE separates the session from the sandbox running it.
Each task gets a disposable Firecracker microVM that is destroyed when the work ends. Rolling snapshots preserve live memory and files, so the session can pause, resume, or branch across sandboxes.
It creates isolated environments for code, files, and long-running agent sessions.
SPACE has handled 100% of Computer production traffic since June.
research.perplexity.ai/artic…
Agent infrastructure must be functional, efficient, and secure and traditional sandboxes were built for short-lived code execution.
Agents need to run code, edit files, and run for hours or days. Runtimes must preserve work without leaving credentials inside environments.
SPACE separates the session from the sandbox running it.
Each task gets a disposable Firecracker microVM that is destroyed when the work ends. Rolling snapshots preserve live memory and files, so the session can pause, resume, or branch across sandboxes.
在相同的生产环境流量下，SPACE 将沙箱创建的中位数延迟从 185 ms 降低到 60 ms。P90 延迟从 447 ms 降低到 89 ms。上周它处理了数百万次沙箱创建和数千万次重连接请求（用于 Computer）。了解更多：perplexity.ai/hub/blog/secur…
查看英文原文
On identical production traffic, SPACE reduced median sandbox creation latency from 185 ms to 60 ms. P90 fell from 447 ms to 89 ms.
Last week it handled millions of sandbox creations and tens of millions of reconnects for Computer.
Read more:
perplexity.ai/hub/blog/secur…
Last week it handled millions of sandbox creations and tens of millions of reconnects for Computer.
Read more:
perplexity.ai/hub/blog/secur…
我想用 Stream Deck 来控制 Codex。让 GPT-5.6 Pro 根据我的规格写了个项目计划。Codex 实现了它。我不想费力点那么多下，所以就让 Codex 直接在我电脑上安装了。
有时候问 AI 写软件比自己找要快。
查看英文原文
I wanted to use my Stream Deck to control Codex. Had GPT-5.6 Pro write a project plan based on my specs. Codex implemented it. I didn't want to bother with a lot of clicks so Codex took over my computer and installed it.
Sometimes its faster to ask for software than look for it.
Sometimes its faster to ask for software than look for it.
不用等了。
受研究和部署启发的周边产品。
限售，先到先得。
openai.com/supply/
引用 Daniel White @dwhitedesignWhen it reaches 10M could we get OpenAI merch ? 😂查看被引原帖 ↗
查看英文原文
You don’t have to wait.
Merch inspired by research & deployment.
Available until sold out.
openai.com/supply/
Merch inspired by research & deployment.
Available until sold out.
openai.com/supply/
现在就说定了
FDE → ODE → PDE
既然有了 Forward Deployed Engineering 和 Ordinary Deployed Engineering，下一个宝可梦进化应该就是 Partial Deployed Engineering
引用 Andrew Curran @AndrewCurran_Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs launched their standalone AI enterprise services firm today. It is named Ode. And it just went live at Ode.com . No announcement from Anthropic yet, probably forthcoming.查看被引原帖 ↗
查看英文原文
calling it now
FDE -> ODE -> PDE
the existence of Forward Deployed Engineering and now Ordinary Deployed Engineering implies that the next pokemon evolution is Partial Deployed Engineering
FDE -> ODE -> PDE
the existence of Forward Deployed Engineering and now Ordinary Deployed Engineering implies that the next pokemon evolution is Partial Deployed Engineering
认识一下 kbd-1.0-codex-micro，由 @work_louder 打造。
将按钮和摇杆映射到你的工作流，让你固定的聊天始终在眼前。
库存有限，赶紧入手。
查看英文原文
Meet kbd-1.0-codex-micro, built with
@work_louder
.
Map the buttons and joystick to your workflow, and keep your pinned chats in view.
Get yours before stock returns 410.
@work_louder
.
Map the buttons and joystick to your workflow, and keep your pinned chats in view.
Get yours before stock returns 410.
Chubby♨️OpenAI 推出 Codex Micro（算得上他们的第一个硬件产品）OpenAI Developers好的，我们听到了。来领取你的 Codex Micro：Tibor BlahoWork Louder's Codex Micro 是专为 ChatGPT Codex 打造的小型键盘，搭配 6…🚨 AI News | TestingCatalogOpenAI 🔥：Codex Micro 现已在 OpenAI Supply 网站上官方推出，售价 $230！🚨 AI News | TestingCatalogOPENAI 🔥：来看看即将推出的 Work Louder's Codex Micro。Tibor Blaho看起来 OpenAI 周边产品有个新的 Supply Co. 商店，现在回货了，还一起上线了新的 Codex M…The Rundown AIOpenAI x Work Louder 硬件协作来了：Gorden SunOpenAI出的第一个硬件，是个Vibe Coding的辅助硬件。类似的玩具国内已经有很多了。
Web Analytics API 的一些很酷的用例：
▪️ 让你的代理关联访客、自定义事件（"purchase"、"checkout"）与你的部署和性能的演变
▪️ 构建自定义前端，并将这些数据与 Stripe 和 Resend 的数据一起绘制
引用 Vercel Developers @vercel_devWeb Analytics API is now public. You can build custom reports and live user-facing metrics with the same data that powers the Web Analytics dashboard. vercel.com/changelog/web-ana…查看被引原帖 ↗
查看英文原文
Some really cool usecases of Web Analytics API:
▪️ Ask your agent to correlate visitors, custom events (“purchase”, “checkout”), with the evolution of your deployments and performance
▪️ Build custom frontends, and plot this data alongside e.g.: Stripe’s and Resend’s
▪️ Ask your agent to correlate visitors, custom events (“purchase”, “checkout”), with the evolution of your deployments and performance
▪️ Build custom frontends, and plot this data alongside e.g.: Stripe’s and Resend’s
猜猜明天要推出什么 👀
猜对了就能赢独家周边。在下面留下你的猜测 👇
查看英文原文
Guess what’s launching tomorrow 👀
Get it right and you’ll win exclusive swag. Drop your guess below 👇
Get it right and you’ll win exclusive swag. Drop your guess below 👇
我们的模型旨在为任何任务提供最佳价格。
如果你能在任何工作负载上获得更好的性价比，非常想听听细节，一起来看看 — [email protected]。
查看英文原文
our models are built to provide the best price for any given task.
if you're able to get better price/perf on any workload, would love to hear the details and look at it together — [email protected].
if you're able to get better price/perf on any workload, would love to hear the details and look at it together — [email protected].
我经常做开发者社区和YouTube视频，但这样的增长速度对我来说一直是个谜，显然有什么东西值得学。
@sytses 我想和Kilo的增长负责人聊一下！
引用 Kilo @kilocode🚨 BIG NEWS: Kilo Code has been acquired by Anaconda ( @anacondainc )! We've grown our agentic engineering platform from zero to a thriving open-source community of 3M developers in only 16 months. Now, we're joining Anaconda's trusted foundation to cover the full AI-native dev lifecycle. 🐍💛 Read the full article below!查看被引原帖 ↗
查看英文原文
as someone who does a lot of dev community and dev youtube this pace of growth has been one of the greatest mysteries to me because clearly there's something to learn here
@sytses
i'd love to talk to whoever was the growth guy/gal at Kilo!!
@sytses
i'd love to talk to whoever was the growth guy/gal at Kilo!!
我觉得现在做landing page或app，必须得来点完全不同的东西才能吸引人用，因为AI已经把所有东西搞成一个样了
引用 Alex Napier Holland 🦍 @NapierHollandI'm a homepage copywriter for 100+ tech startups. I've never seen startup homepages struggle this badly. You operate in the most competitive market of all-time, yet the average homepage is more generic than ever. Why? One root misunderstanding: • Apps should be functional • Marketing assets must break form and stand out Your mission as a marketer is to find fresh, unique arguments and language patterns that can puncture through a great ocean of generic homepages and connect with your audience. But AI produces averages. • Great to answer objective questions and create functional, familiar user interfaces. • Useless to create a standout marketing asset that makes a fresh, unique argument and makes your audience feel understood. How do you solve this? You start a new quest. You must constantly capture, organise and extract language and arguments from your customers and prospects. This raw material is your new competitive edge. It's the only way to differentiate your product. Everything else is slop without it.查看被引原帖 ↗
查看英文原文
I think you have to do something radically different as a landing page or app to even get anyone to use something these days as AI has made every landing page or app look the exact same
Read It Back
预训练的MLLM是文本到图像生成的零样本奖励模型
查看英文原文
Read It Back
Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation
Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation
现在搞AI前沿的都认可，得让第三方测试AI系统，用这些制定纳入政策的标准。很高兴看到@demishassabis提出了这样的框架！
查看英文原文
At this point, everyone at the frontier of AI agrees that third-parties should test out AI systems and use these to develop standards to feed into policy - excellent to see
@demishassabis
laying out a framework to do this!
@demishassabis
laying out a framework to do this!
想让agent在公司内真正发挥作用，得把公司架构设计成它能读得懂的样子。
Shopify就这样干的——他们的agent根本没有私聊功能，只有公开频道。结果就是促进了同行学习
查看英文原文
If you want agents to actually work inside a company, you have to design the company so they can read it
Shopify did this with an agent that had no private chat function at all, only public channels. The side effect was peer learning
Shopify did this with an agent that had no private chat function at all, only public channels. The side effect was peer learning
Vercel Agent特别擅长处理优化问题。可以让它优化你的build、性能、账单。
引用 Adam Killam @adamkillamVercel's agent just cut our build times by 10x. @rauchg 🙏查看被引原帖 ↗
查看英文原文
Vercel Agent is excellent for optimization questions. Ask it to optimize your build, performance, your bill.
我们把一个robot model的context窗口扩到了8000个timestep，相当于5分钟的肌肉记忆，推理成本还是恒定的。机器人策略之前就活在几帧的时间里（<0.1秒），然后立马忘干净。我们推进到了比SOTA高3个数量级。
介绍RoboTTT。Test-Time Training（TTT）在模型里面放了个小模型。每来一个sensor reading就在这个小核心上走一步梯度下降，历史信息不断被压进权重里。隐藏状态大小是固定的（就是个小神经网络），机器人能以很小的成本'吸收'任意长的经验。学习在上线后无限期持续。
这样我们就把整段视频当context放进去！RoboTTT支持从人类视频单次演示学习：电路板组装场景里，人演示一个没见过的操作，机器人能精准复现。
人经常掉东西，但我们反应这么快根本没反应过来。这种快速纠正的反射占了人类物理能力的一半。RoboTTT展现实时自我改进：机器人能在执行过程中快速从错误中恢复，每次修正都进入context指导下一步。TTT core从训练数据里学出通用的failure-to-correction映射。
最后一点最让我兴奋。全新的Context Scaling Curve：从128到8K timestep，闭环性能一路上升，一点饱和的迹象都没有。8K-context预训练比1K提升62%。LLM能享受的东西，机器人学也应该能。很快，1M context都不是幻想了。
线程里详细讨论：
查看英文原文
We scaled a robot model natively to 8,000 timesteps of context, 5 minutes worth of muscle memory, with constant inference cost. Robot policies used to live their lives a few frames at a time (< 0.1 sec), instantly forgetting what just happened. We pushed to 3 orders of magnitude beyond SOTA.
Introducing RoboTTT. Test-Time Training (“TTT”) carries a tiny model *inside* the model. Every incoming sensor reading triggers one gradient step on that tiny core, so the history keeps getting compressed into its weights. The hidden state has a fixed size (literally a small neural net), so the robot can “grok” arbitrarily long experience with little overhead. Learning continues indefinitely after deployment.
We can then put an entire video in context as prompt! RoboTTT enables one-shot in-context learning from human video: in circuit board assembly, a human demonstrates a never-seen configuration once, and the robot imitates it faithfully.
Humans drop things all the time, but we pick them up so fast that we don’t even notice. That reflex to fix is half of our physical competence. RoboTTT shows self-improvement on the fly: the robot is skilled at recovering from its own errors mid-episode, and each fix enters its context to inform the next move. The TTT core distills a general-purpose, failure-to-correction mapping from the training data.
One more thing. What excites me the most is a new Context Scaling Curve: from 128 to 8K timesteps, closed-loop performance hill-climbs steadily with no sign of saturation. 8K-context pretraining beats 1K by 62%. What LLM enjoys, robotics should too. Soon, even 1M context is not a fantasy.
Deep dive in thread:
Introducing RoboTTT. Test-Time Training (“TTT”) carries a tiny model *inside* the model. Every incoming sensor reading triggers one gradient step on that tiny core, so the history keeps getting compressed into its weights. The hidden state has a fixed size (literally a small neural net), so the robot can “grok” arbitrarily long experience with little overhead. Learning continues indefinitely after deployment.
We can then put an entire video in context as prompt! RoboTTT enables one-shot in-context learning from human video: in circuit board assembly, a human demonstrates a never-seen configuration once, and the robot imitates it faithfully.
Humans drop things all the time, but we pick them up so fast that we don’t even notice. That reflex to fix is half of our physical competence. RoboTTT shows self-improvement on the fly: the robot is skilled at recovering from its own errors mid-episode, and each fix enters its context to inform the next move. The TTT core distills a general-purpose, failure-to-correction mapping from the training data.
One more thing. What excites me the most is a new Context Scaling Curve: from 128 to 8K timesteps, closed-loop performance hill-climbs steadily with no sign of saturation. 8K-context pretraining beats 1K by 62%. What LLM enjoys, robotics should too. Soon, even 1M context is not a fantasy.
Deep dive in thread:
博客和论文：
research.nvidia.com/labs/gea…
来自NVIDIA GEAR Lab的呈现。
千万看看Yunfan的技术深潜！
引用 Yunfan Jiang @YunfanJiangWe scaled robot policies to 8K timesteps of visuomotor context, orders of magnitude beyond current SoTAs, at constant inference latency. Introducing RoboTTT 🤖 With minutes of experience in context, our robots: 🎥 one-shot imitate human video demos 📈 improve themselves during deployment 🛡️ recover from perturbations ⚙️ complete a 5-minute, 10-stage assembly end to end 🌐 research.nvidia.com/labs/gea… Dive in 🧵查看被引原帖 ↗
查看英文原文
Blog and paper:
research.nvidia.com/labs/gea…
Presented to you from NVIDIA GEAR Lab.
Please check out Yunfan's technical deep dive!
research.nvidia.com/labs/gea…
Presented to you from NVIDIA GEAR Lab.
Please check out Yunfan's technical deep dive!
等待30分钟出几周的工作成果，这个等待感觉真是无限漫长。真的应该在 Codex 里加个闲置点击游戏，这样在等输出的时候能打发打发时间。
查看英文原文
It is funny how interminable waiting 30 minutes for a couple weeks of work can feel. Really need to integrate idle clicker games into Codex to keep us humans busy while we wait for output.
连环推 ×2很多人不理解 Mythos 现在已经有 5 个月了，我们要看的（编码）只是整体能力的一小部分
而且通常这个样本不太好，因为所有好的基准都是私密的。而且大多数基准不看模型大小或 token 效率。
话说回来，Kimi-K3 可能会在好几个常见的基准上击败 Opus 4.8 和 GPT-5.5，但这些基准不能说明全部情况，这些模型也不是最前沿的
最前沿的是 Mythos，或者在某些情况下甚至是 GPT-5.2 这样的老模型。
不过，DeepSeek-V4、GLM-5.2 或 Kimi-K2.6/2.7 都还没有解决过一个 Erdös 问题。
第一个公开发布的解决这个问题的模型是 GPT-5.2-Pro，那是 2025 年 11 月。现在是 2026 年 7 月。（在 FrontierCode T4 或 ARC-AGI-2 上也是一样）
如果你相信美国公司有越来越大的计算优势，且模型会加速未来模型的开发，你也应该在向后看的差距和预期/向前看的差距之间做出区分
领域也很重要。不同领域有不同的滞后。但总体来说，向后看的差距约为 7-8 个月。
不过我对 Kimi-K3 超期待
它会是个疯狂的模型，要么迫使 Anthropic 和 OpenAI 降价，要么逼他们推出最强的模型，要么他们就停止对中低端模型的保留
Kimi-K3 应该会表现得特别强，对 99% 的编程用例都足够了
引用 leo 🐾 @synthwaveddI think Kimi K3 is going to shock some of the "the Chinese are 8 months behind the Western frontier" people查看被引原帖 ↗
查看英文原文
A lot of these people don't understand that Mythos is 5 months old at this point and that what we will look at (coding) is only a small slice of overall capability
and typically this sample is not very good as all the good benchmarks are private. and most benchmarks don't look at model sizes or token-efficiency.
That said, Kimi-K3 will likely beat Opus 4.8 and GPT-5.5 in several of the usual benchmarks, ... but these benchmarks don't show the full story and these models are not the frontier
The frontier to beat is Mythos or in some cases even older models like GPT-5.2.
Mind you, DeepSeek-V4, GLM-5.2 or Kimi-K2.6/2.7 haven't solved a single Erdös problem so far.
The first publicly available model that solved one was GPT-5.2-Pro, back in November 2025. It's now July 2026. (same on FrontierCode T4 or ARC-AGI-2)
If you believe that US companies have a growing compute advantage and that models speed up the development of future models, you should also make a distinction between backward looking gap and the expected/forward looking one
Domains matter too. There are different lags for different domains. But overall the backward looking gap is ~7-8 months.
nevertheless im very hyped for Kimi-K3
it will be an insane model and will either force Anthropic and OpenAI to lower prices, bring out their largest and strongest models, or stop them sandbagging their lower and mid-tier models
Kimi-K3 should feel like very strong model and will be enough for 99% of coding use-cases
and typically this sample is not very good as all the good benchmarks are private. and most benchmarks don't look at model sizes or token-efficiency.
That said, Kimi-K3 will likely beat Opus 4.8 and GPT-5.5 in several of the usual benchmarks, ... but these benchmarks don't show the full story and these models are not the frontier
The frontier to beat is Mythos or in some cases even older models like GPT-5.2.
Mind you, DeepSeek-V4, GLM-5.2 or Kimi-K2.6/2.7 haven't solved a single Erdös problem so far.
The first publicly available model that solved one was GPT-5.2-Pro, back in November 2025. It's now July 2026. (same on FrontierCode T4 or ARC-AGI-2)
If you believe that US companies have a growing compute advantage and that models speed up the development of future models, you should also make a distinction between backward looking gap and the expected/forward looking one
Domains matter too. There are different lags for different domains. But overall the backward looking gap is ~7-8 months.
nevertheless im very hyped for Kimi-K3
it will be an insane model and will either force Anthropic and OpenAI to lower prices, bring out their largest and strongest models, or stop them sandbagging their lower and mid-tier models
Kimi-K3 should feel like very strong model and will be enough for 99% of coding use-cases
NVIDIA 说用两个提示词，Codex 在一天内把 Cosmos 3 Nano 从 54.41% 的精度优化到了 93.35%。
这个实验使用了丰田的 Woven 交通安全数据集：8000+ 个四选一视频推理的训练和验证样本。
使用 NVIDIA TAO agent skills，Codex 自动化地：
检测和修补缺失的视频元数据
跑零样本基线
生成 LoRA 配置
启动训练和评估
跑 AutoML 超参数搜索
报告最佳模型
一个 LoRA 运行在 8 个 A100 GPU 上用了大约 30 分钟就达到了 87.14%。
第二个提示词在多个 A100 节点上启动了 43 个并行 AutoML 试验，19.5 小时后达到了 93.35%。NVIDIA 说 LoRA 所需的 GPU 小时数大约只有全参数训练的七分之一。
Agent skills 正在成为通用编码代理操作高度专业化 ML 基础设施的界面。
查看英文原文
NVIDIA says Codex post-trained Cosmos 3 Nano from 54.41% to 93.35% accuracy in one day - with two prompts.
The experiment used Toyota’s Woven Traffic Safety dataset: 8,000+ training and validation samples for four-choice video reasoning.
Using NVIDIA TAO agent skills, Codex autonomously:
Detected and patched missing video metadata
Ran the zero-shot baseline
Generated LoRA configurations
Launched training and evaluation
Ran an AutoML hyperparameter sweep
Reported the best model
One LoRA run reached 87.14% after roughly 30 minutes on eight A100 GPUs.
A second prompt launched 43 parallel AutoML trials across multiple A100 nodes, reaching 93.35% after 19.5 hours. NVIDIA says LoRA required roughly seven times fewer GPU-hours than full-parameter training.
Agent skills are becoming the interface through which general coding agents operate highly specialized ML infrastructure.
The experiment used Toyota’s Woven Traffic Safety dataset: 8,000+ training and validation samples for four-choice video reasoning.
Using NVIDIA TAO agent skills, Codex autonomously:
Detected and patched missing video metadata
Ran the zero-shot baseline
Generated LoRA configurations
Launched training and evaluation
Ran an AutoML hyperparameter sweep
Reported the best model
One LoRA run reached 87.14% after roughly 30 minutes on eight A100 GPUs.
A second prompt launched 43 parallel AutoML trials across multiple A100 nodes, reaching 93.35% after 19.5 hours. NVIDIA says LoRA required roughly seven times fewer GPU-hours than full-parameter training.
Agent skills are becoming the interface through which general coding agents operate highly specialized ML infrastructure.
开发者.nvidia.com/blog/po…
查看英文原文
developer.nvidia.com/blog/po…
别上大学了 - 去创办你的公司
未来一个人的公司会成为常态
颠覆一切 - 法律、医疗、金融....
查看英文原文
Don't go to college - start building your company
1-person companies will be the norm in the future
Disrupt everything - Legal, Health, Finance....
1-person companies will be the norm in the future
Disrupt everything - Legal, Health, Finance....
Codex的一大作用就是修小火箭的各种配置。
真的不想搞这些，谁出个教程，怎么配置规则。
Github开源的各种规则感觉也不好使。
连环推 ×2🚀 MiniMax Hub 重大升级！
现在更新 >>
hub.minimax.io/
新版本亮点：
🎵 音频进化：ElevenLabs Music v2 & Seed Audio 1.0 现已上线！
⚡ 特色技能：一站式 AI 工作流，支持短剧、运动图形和高级摄像机控制。
🛠️ 更智能的工作区：跨所有节点的统一全局搜索和一键媒体批量下载。
#MiniMax
#MiniMaxHub
🎵 音频模型更新
- ElevenLabs Music v2：可从场景、风格和情绪提示生成任何风格的音乐。
- Seed Audio 1.0：从文本、参考音频或图像生成声音，支持调节速度、音量和音高。
查看英文原文
🚀 MiniMax Hub Major Upgrade!
Update now>>
hub.minimax.io/
Highlights of new release:
🎵 Audio Evolution: ElevenLabs Music v2 & Seed Audio 1.0 are now live!
⚡ Featured Skills: One-stop AI workflows for short-dramas, motion graphics, and advanced camera controls.
🛠️ Smarter Workspace: Unified Global Search across all nodes & one-click media bulk downloads.
#MiniMax
#MiniMaxHub
🎵 Audio Model Update
- ElevenLabs Music v2: Generates music in any genre from scene, style, and mood prompts.
- Seed Audio 1.0: Generates voices from text, reference audio, or images, with adjustable speed, volume, and pitch.
Update now>>
hub.minimax.io/
Highlights of new release:
🎵 Audio Evolution: ElevenLabs Music v2 & Seed Audio 1.0 are now live!
⚡ Featured Skills: One-stop AI workflows for short-dramas, motion graphics, and advanced camera controls.
🛠️ Smarter Workspace: Unified Global Search across all nodes & one-click media bulk downloads.
#MiniMax
#MiniMaxHub
🎵 Audio Model Update
- ElevenLabs Music v2: Generates music in any genre from scene, style, and mood prompts.
- Seed Audio 1.0: Generates voices from text, reference audio, or images, with adjustable speed, volume, and pitch.
因为我没有通过传统方式学编程，使用编码代理对我来说纯粹是创意和自我表达。GitHub 基本上就是我的 Substack。
引用 David Pan @davepWhen I was growing up, programming was an expression of creativity. Somewhere along the way, it became more about the chores. Merge conflicts. Flaky tests. CI failures. Bug triage. Let's give the chores to the robots and get back to building cool shit.查看被引原帖 ↗
查看英文原文
Because I never learned programming the traditional way, using coding agents has been purely an act of creativity and self-expression
GitHub is basically my Substack
GitHub is basically my Substack
他们最新发布的预训练报告显示，Soofi S 30B-A3B 基于 NVIDIA 开源的 Nemotron 3 Nano 参考架构，采用了相同的混合 Mamba + Transformer MoE 设计，拥有约 3B 的活跃参数和几乎相同的架构选择。
真正新颖的不是架构，而是训练方式：
- 约 27 万亿个训练 tokens
- 德语故意加权
- 在 Deutsche Telekom 工业 AI 云上的端到端训练
- 完整开源的训练方案、超参数和评估方法
说实话，看到欧洲也在训练自己的模型还是挺兴奋的。
引用 NXT EU @NXT4EUGermany has launched one of the world's best open-source AI models. Soofi S, made by the Soofi consortium, is a 30B parameter model fully trained in Europe and tops the ranking for open-source AI. Huge moment for Europe, and finally some competition for Chinese open-source AI.查看被引原帖 ↗
查看英文原文
Their newly released pretraining report reveals that Soofi S 30B-A3B is based on NVIDIA’s open Nemotron 3 Nano reference architecture, the same hybrid Mamba + Transformer MoE design with ~3B active parameters and nearly identical architectural choices
What’s actually new isn’t the architecture, but the training:
-~27 trillion training tokens
-German deliberately up-weighted
-End-to-end training on Deutsche Telekom’s Industrial AI Cloud
-Full training recipe, hyperparameters and evaluation methodology released openly
Ngl, excited to see europe kinda trains their own models.
What’s actually new isn’t the architecture, but the training:
-~27 trillion training tokens
-German deliberately up-weighted
-End-to-end training on Deutsche Telekom’s Industrial AI Cloud
-Full training recipe, hyperparameters and evaluation methodology released openly
Ngl, excited to see europe kinda trains their own models.
魔法背后的创意者在这里。✨
从天马行空的想法到获奖之作，来看看 Kling AI NEXTGEN Awards 那些亮点作品背后的故事。
特别感谢 Jacek Kadaj、Jinlong Hu、Kassai Ricsi、KUA KEE SENG、Lee Hyerin、Mingwei Chen、Park Soleun、Peize Wang、reels_mon01、Seo Yoonjung、Shin Seoyeon、Shiqi Wang、Shihua Lin、Son Seoyeong、Xuanwei Liu、Yihang Yang、Zibo Jia 以及所有创意者无与伦比的支持和启发。
完整采访即将推出。
查看英文原文
The creators behind the magic are here. ✨
From wild ideas to award-winning creations, discover the stories behind the works that stood out at Kling AI NEXTGEN Awards.
A huge thank you to Jacek Kadaj, Jinlong Hu, Kassai Ricsi, KUA KEE SENG, Lee Hyerin, Mingwei Chen, Park Soleun, Peize Wang, reels_mon01, Seo Yoonjung, Shin Seoyeon, Shiqi Wang, Shihua Lin, Son Seoyeong, Xuanwei Liu, Yihang Yang, Zibo Jia, and all the creators for your incredible support and inspiration.
Full interviews coming soon.
From wild ideas to award-winning creations, discover the stories behind the works that stood out at Kling AI NEXTGEN Awards.
A huge thank you to Jacek Kadaj, Jinlong Hu, Kassai Ricsi, KUA KEE SENG, Lee Hyerin, Mingwei Chen, Park Soleun, Peize Wang, reels_mon01, Seo Yoonjung, Shin Seoyeon, Shiqi Wang, Shihua Lin, Son Seoyeong, Xuanwei Liu, Yihang Yang, Zibo Jia, and all the creators for your incredible support and inspiration.
Full interviews coming soon.
牛逼，整个x代码全开源啊。
这是打OpenAI的脸吗？
引用 Elon Musk @elonmuskOnce we have completed our review for security vulnerabilities, we will make the entire codebase of 𝕏 open source, with no exceptions. Moreover, we will invite third party reviewers to examine the system that is running to confirm that the open source code is what is running. Trust through total transparency is the only thing that should be believed.查看被引原帖 ↗
下周会在旧金山。咱们要不要组织个集会或游行，支持开源和本地 AI？
查看英文原文
Going to be in San Francisco next week. Should we organize some sort of a meetup or march in support of open-source and local AI?
什么是丰饶时代呢
有人骑行是为了去公司上班
有人骑行是为了去山里兜风
有人听播客是为了吸收信息
有人听播客是就是随便听听
有人做社交媒体是为了商单赚钱
有人做社交媒体就是想表达自我
有人用 Agent 是为提高自己的工作效率
有人用 Agent 是在构建自己的思维殿堂
当做一件事只为了提高赚钱的效率，那还是贫瘠时代
当它成为每个人自己所选择的生活方式的时候，才是丰饶时代
人又有很深的时代惯性，穷怕了的人在丰饶时代也会保留以前的习惯，就像父母那辈人退休了还保留着极度省钱的小习惯一样
马斯克将把 𝕏 的整个代码库开源
引用 Elon Musk @elonmuskOnce we have completed our review for security vulnerabilities, we will make the entire codebase of 𝕏 open source, with no exceptions. Moreover, we will invite third party reviewers to examine the system that is running to confirm that the open source code is what is running. Trust through total transparency is the only thing that should be believed.查看被引原帖 ↗
牛 P 啊
📢大消息
马斯克：将会把 𝕏 的整个代码全部开源
同时将邀请第三方审查员审查 𝕏 的整个系统，以确保开源的代码和线上运行的完全一致！
🫡
引用 Elon Musk @elonmuskOnce we have completed our review for security vulnerabilities, we will make the entire codebase of 𝕏 open source, with no exceptions. Moreover, we will invite third party reviewers to examine the system that is running to confirm that the open source code is what is running. Trust through total transparency is the only thing that should be believed.查看被引原帖 ↗
用 Fable 或 5.6 sol 这样的强大模型来构建你的整个产品
需求越复杂越好
这些模型正在扩展以完成 100 人开发团队的工作
真的，他们可以做测试、调试、监控！基本上什么都能做
查看英文原文
Use a strong model like Fable or 5.6 sol to build your entire product
The more complex the requirements, the better
These models are scaling up to the do the work of 100 person dev teams
Yes, they can do testing, debugging, monitoring! Literally everything
The more complex the requirements, the better
These models are scaling up to the do the work of 100 person dev teams
Yes, they can do testing, debugging, monitoring! Literally everything
从提出假设到设计实验，AI agents正开始重塑科学发现。但最大的挑战是在真实世界中验证这些想法。我们的文章探讨了日益增长的验证瓶颈，并为政策制定者和资金方提出了四个优先方向。
查看英文原文
From proposing hypotheses to designing experiments, AI agents are starting to reshape scientific discovery. But the hardest part is testing these ideas in the real world.
Our essay explores the growing validation bottleneck and outlines four priorities for policymakers and funders. →
goo.gle/4poACUT
Our essay explores the growing validation bottleneck and outlines four priorities for policymakers and funders. →
goo.gle/4poACUT
这个很棒啊！一个能帮你提升 Vibe Coding 前端交互设计水平的网站。
不知道名字，就不知道如何跟 AI 说实现什么样的效果。
网站整理了Web和App常见组件和动效名字。
还列网页设计风格名，比如Liquid Glass（液态玻璃）、Neumorphism（新拟物）等。
地址见评论区
感觉这个相当实用，应该好好学习下
namethatui.com/
美国通过州级和联邦行动推进AI安全
OpenAI阐述了一种「反向联邦制」的AI治理方式，通过州法律来帮助建立安全、民主的AI国家框架。
查看英文原文
The US is advancing AI safety through state and federal action
OpenAI outlines a “reverse federalism” approach to AI governance, where state laws help build a national framework for safe, democratic AI.
OpenAI outlines a “reverse federalism” approach to AI governance, where state laws help build a national framework for safe, democratic AI.
@X 团队报告一个问题：当你在网页上看到推文并点击 Grok 按钮想询问相关问题，点击停止后输入问题，它就忘了你之前提到的那条帖子了。
查看英文原文
Mini bug for
@X
staff:
When you see a tweet and tap Grok button on web, then you wanna ask question about it, so you tap STOP and type a question, it forgets what post you included!
@X
staff:
When you see a tweet and tap Grok button on web, then you wanna ask question about it, so you tap STOP and type a question, it forgets what post you included!
我给想理解 AI 真正发展方向的人写文章，embodiment 是大家马上就要关心的重点。Booster T2 让我感兴趣是因为它首先被设计成开发平台——板载 2070 TFLOPS 配合 Booster Studio 让开发者可以在真实机器人上迭代而不是玩具。这就是如何建立社区，而社区正是把机器人变成平台的方式。
引用 Booster Robotics @boosteroboticsBooster T2 is now available. Smart. Powerful. Built to Perform. As a new flagship embodied development platform, Booster T2 delivers up to 2070 TFLOPS of computing power, the highest in its class for bipedal humanoid robots. Combined with a high-output, high-DOF body, it enables sharper motion perception, stronger understanding, and reliable execution in complex scenarios. From real-time perception to rapid response, from agile movement to precise manipulation, Booster T2 combines powerful computing performance and reliable hardware with Booster Studio's integrated development platform, unlocking limitless possibilities for developers. Search "Booster Robotics" to visit the official website and order now.查看被引原帖 ↗
查看英文原文
I write for people who want to understand where AI is actually heading, and embodiment is the part everyone's about to care about.
The Booster T2 is interesting to me because it's built as a development platform first — 2070 TFLOPS onboard plus Booster Studio means developers get to iterate on the real body, not a toy. That's how you get a community, and community is what turns a robot into a platform.
The Booster T2 is interesting to me because it's built as a development platform first — 2070 TFLOPS onboard plus Booster Studio means developers get to iterate on the real body, not a toy. That's how you get a community, and community is what turns a robot into a platform.
我做了个小机器人，每天提醒我何时应该出去坐太阳，咱们这边是南欧所以你那儿可能不同。
☀️ 维生素 D 时间窗口——现在去晒 15 分钟左右。
UV 指数 3.8（理想水平——获得维生素 D 又不会被晒伤）。
引用 Tim | ONLYUP™ 💹 @timonlyupWhat’s your take on getting too much sun (skin cancer, skin ageing etc…) Pieter?查看被引原帖 ↗
查看英文原文
So I made this little bot, every day it messages when I should go sit in the sun outside, we're in Southern Europe so it might be more wherever you are
"☀️ Vitamin D window — go sit in the sun for ~15 min now.
UV index is 3.8 (ideal level — good vitamin D without burning)."
"☀️ Vitamin D window — go sit in the sun for ~15 min now.
UV index is 3.8 (ideal level — good vitamin D without burning)."
今天 AI 行业头条：
- DeepMind CEO 向美国 AI 监管机构推介
- 报道：OpenAI 的第一款设备是 AI 扬声器
- 为客户咨询构建无代码语音 agent
- 纽约州阻止 AI 数据中心热潮
- 4 个新 AI 工具、社区工作流等更多内容
查看英文原文
Top stories in AI today:
- DeepMind CEO pitches U.S. AI watchdog
- Report: OpenAI's first device is an AI speaker
- Build a no-code voice agent for customer intake
- New York stalls the AI data center boom
- 4 new AI tools, community workflows, and more
- DeepMind CEO pitches U.S. AI watchdog
- Report: OpenAI's first device is an AI speaker
- Build a no-code voice agent for customer intake
- New York stalls the AI data center boom
- 4 new AI tools, community workflows, and more
Microsoft AI Futures 团队今日在 Nature Health 发表新论文。
在 109 个国家审查了 170 万条对话后，结果很清楚：Copilot 对那些对医疗系统信心不足的人来说是无价的资源。
技术一直都是伟大的平衡器，推动更多人获得更优质的服务。这进一步证明了 AI 正在为全球服务不足的群体在最需要的时刻提供无价的支持。
感谢所有作者。查看论文：
nature.com/articles/s44360-0…
查看英文原文
New paper published in Nature Health today from the Microsoft AI Futures team.
After reviewing 1.7m conversations across 109 countries, it’s clear that Copilot is an invaluable resource particularly for people with low confidence in their health systems.
We've always known that technology is a great equalizer, driving broader access to better quality services. This is further evidence that AI is giving underserved people around the world access to invaluable support when they need it most.
Many thanks to all the authors. Check out the paper here:
nature.com/articles/s44360-0…
After reviewing 1.7m conversations across 109 countries, it’s clear that Copilot is an invaluable resource particularly for people with low confidence in their health systems.
We've always known that technology is a great equalizer, driving broader access to better quality services. This is further evidence that AI is giving underserved people around the world access to invaluable support when they need it most.
Many thanks to all the authors. Check out the paper here:
nature.com/articles/s44360-0…
GPT-Red：通过自我改进提升模型稳健性
了解 OpenAI 的自动化红队测试系统 GPT-Red，它通过自我对弈来提升 AI 的安全性、对齐度和抗提示词注入能力。
查看英文原文
GPT-Red: Unlocking Self-Improvement for Robustness
Explore GPT-Red, OpenAI’s automated red teaming system that uses self-play to improve AI safety, alignment, and prompt injection robustness.
Explore GPT-Red, OpenAI’s automated red teaming system that uses self-play to improve AI safety, alignment, and prompt injection robustness.
连环推 ×2Google 在 Pixel 10 上跑 Gemma 4
三万英尺的客舱断网也能聊天、看图、改设置
在Google I/O India 上，Tensor 团队和 Pixel 团队联合演示：把 Gemma 4 轻量版直接塞进手机里的 TPU里
不仅能能聊天」，还能「看图、听写、控制手机」
官方演示：手机在断网状态下做旅行规划、菜谱推荐和家居自动化
控制手机
你用语音或文字指挥 Wi‑Fi、地图这类功能，模型直接给出手机能执行的动作
Every推出的All Access订阅服务还挺划算
525美元一年，除了包含Every自家的订阅服务和产品服务（Every的主营业务是AI领域媒体内容），还包括这些高价值的服务：
1000美元的Codex点数
12个月Cursor Pro+会员，60美元/月的那一档，这一项就回本
1个月Claude Max会员，没写是5x还是20x，估计是5x
3个月Google AI Pro会员，没什么用
1年Framer Pro会员，AI做设计
6个月Notion Business会员
最近发布的OpenAI视频中的一些有趣时刻
"Rune [plugin]是你在OpenAI的小伙伴"
"为Codex Micro键盘编写文档"
"在Codex定价页面添加隐藏的'tibo'键盘序列"
查看英文原文
A few interesting moments from the recently published OpenAI videos
"Rune [plugin] is your little guy at OpenAl"
"Document the Codex Micro keyboard"
"Add a hidden 'tibo' keyboard sequence to the Codex pricing page"
"Rune [plugin] is your little guy at OpenAl"
"Document the Codex Micro keyboard"
"Add a hidden 'tibo' keyboard sequence to the Codex pricing page"
“Apple智能”大模型已获手机端侧AI服务备案
国行 Siri 要来了…
网信办发布7款提供手机端侧生成式人工智能服务已备案信息的公告
为促进生成式人工智能服务创新发展和规范应用，网信部门会同有关部门按照《生成式人工智能服务管理暂行办法》要求，有序开展生成式人工智能服务备案工作
现将新增的“Apple智能”等7款提供手机端侧生成式人工智能服务备案信息予以公告。
其中包括，苹果智能、华为小艺、vivo蓝心、小米澎湃及努比亚豆包大模型。
谷歌CEO Demis Hassabis发布长文《前沿AI框架与新时代的黎明》
我们找到了让沙子思考的方法，堪称奇迹。AGI几年之内就会到来，AGI的突破比互联网、移动互联网的变革更大，更像是火和电的发现。而且AGI的规模会是工业革命规模的10倍，发展速度也是10倍。
他也发出了跟Anthropic CEO达里奥类似的警告，AI前沿进展已经跑在了人类理解能力的前面：网络安全风险已是现实，生物、核风险紧随其后，而且AI越来越Agent化、能递归自我改进的系统将需要更强的控制。
他呼吁，由美国牵头，仿照金融业监管局(FINRA)建立一个前沿AI标准机构，对于前沿模型要在发布30天由机构检测，评估覆盖网络安全、生物威胁、欺骗行为等高风险领域，必要时甚至要协调全行业减速。
最接近技术的人，比外界更清楚接下来两年会发生什么。（不过显然Gemini现在完全不需要操心这些）
OpenAI正竭尽全力向Anthropic展示他们为什么能坚持下去，以及为什么相信自己能赢得竞争。
我也来说两句：我每天都在用Codex，因为它的性价比真的很突出。5.6版本曾经短时间内出现过token消耗过多的问题。但总体而言，这些模型的性价比都相当不错。肯定比Anthropic的好。
引用 Tibo @thsottiauxOr… what if we gave you $100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched? Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens! switch-to-codex.openai.chatg…查看被引原帖 ↗
查看英文原文
OpenAI is leaving no stone unturned in its efforts to show Anthropic why they're here to stay and why they're confident they can win the competition.
But I'll join in: I use Codex daily because its price-performance ratio is outstanding. With version 5.6, there was a temporary issue with excessive token consumption. But overall, the price-performance ratio is excellent for the respective models. Certainly far better than with Anthropic.
But I'll join in: I use Codex daily because its price-performance ratio is outstanding. With version 5.6, there was a temporary issue with excessive token consumption. But overall, the price-performance ratio is excellent for the respective models. Certainly far better than with Anthropic.
今天在 ChatGPT 里面找 Codex 没找到，原来改名叫 Remote 了
引用 Thomas Ricouard @DimillianHere is the latest update to Codex Remote on iOS! As I posted, we now support the new visualisation feature, but there is also a ton of other improvements and fixes! learn.chatgpt.com/docs/chang…查看被引原帖 ↗
我在开发 BaoCut 这个 App 的时候，是基于一个 Loop 来的：
1. 在开发新功能之前先设计原型（参考图1），借助的是 baoyu-design skill （
github.com/jimliu/baoyu-desi…
），配合 Claude Code App 内置的浏览器实施预览调整，模型 Opus 4.8 就很好了，都不需要 Fable 5.
GPT 5.6 Sol 设计能力还是不如 Opus 4.8
2. 原型打磨好了后，只需要在同一会话内，让 Claude Code 基于新的 UI 设计去实现功能即可，这块 Claude 做的很好，Fable 5 效果最好，能将设计稿几乎 1:1 还原，如果修改不多 Opus 4.8 也能胜任。
这些 UI 的打磨我还是更放心让 Fable 和 Opus 而不是 GPT，但其他一些不涉及 UI 部分的 GPT 5.6 Sol 就做的很好。
3. 更新好了后测试没问题，就可以通过发布的 skill 发布新版本。
这里可以放心让 Codex 去做了，尤其是它的 CloudFlare Plugin 很好用，直接帮助发布更新安装包到 CF。
这个 loop 的每一个迭代的起点是自己的想法，让 AI 提供设计方案，和 AI 反复讨论后确定方案，然后 AI 实施，AI 实施完成后人再去验证和当初想要的是否一致，如果不一致再让 AI 调整甚至推翻重来。
引用 宝玉 @dotey字幕转录翻译剪辑 Skill —— BaoCut（仅支持 Mac） 借助 Agent Skill，可以转录视频、对转录结果识别 Speaker、润色（纠正错别字口癖等）、也可以根据转录结果对视频进行简单的剪辑，比如删除口癖、重复等。 这次尝试解决一个问题就是 Agent 对字幕转录翻译后，无法通过一个友好的操作界面二次编辑的问题。 现在的做法是为 Agent 提供一个 cli，配合 Skill 的说明，Agent 可以借助 cli 去转录，获取转录结果润色、翻译，并实时同步进度到 GUI。后续可以在 GUI 进行预览和人工编辑。 安装了 Skill 和 App 后，后续只要从 Codex 或者 Claude Code 这种 Agent，触发 Skill 即可执行，比如： > /baocut 转录并翻译视频：<视频 url 或路径> 已知问题： - 仅支持 Mac - 翻译速度略慢，但质量会不错 下载地址： baocut.app/ Skill 从 App 内可以安装，或者 Skill 地址： github.com/jimliu/baocut查看被引原帖 ↗
未来会是双模式并行
便宜高效 - 90%的任务会由这些廉价模型大规模完成 - Deepseek++
超级智能 - 自主、极其聪慧，理解复杂上下文。Fable / GPT-10
查看英文原文
The future will be bi-modal
Cheap and performant - 90% of the tasks will be done by these cheap models at scale - Deepseek++
Super Intelligent - Autonomous, extremely smart and understands complex contexts. Fable / GPT-10
Cheap and performant - 90% of the tasks will be done by these cheap models at scale - Deepseek++
Super Intelligent - Autonomous, extremely smart and understands complex contexts. Fable / GPT-10
Muse Spark可以帮你把冰箱填满！
形成完整循环，这个问题实际上是Scale AI的创意来源
引用 Tyler Shibata @tjshibataMuse Spark 1.1 can use browsers and it refilled our office fridge - Took 2 pics of our fridge stocked / not stocked - Told Muse to analyze the pics and restock what we were low on - It analyzed each shelf and which brand / items that were missing - Used the @juliusai built in browser to go to instacart, and add the items from target - Checked out and even left a tip for the driver One prompt + two pics to keep our team happy查看被引原帖 ↗
查看英文原文
muse spark can refill your fridge for you!
full circle, this problem was actual the inspiration for scale ai
full circle, this problem was actual the inspiration for scale ai
语言的污染是很可怕的
很多人说话和写文已经越来越GPT味儿了...
平时没事儿还是多换几个模型换着用
用 ChatGPT 制作的网站。
switch-to-codex.openai.chatg…
引用 Tibo @thsottiauxOr… what if we gave you $100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched? Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens! switch-to-codex.openai.chatg…查看被引原帖 ↗
查看英文原文
Made with ChatGPT sites.
switch-to-codex.openai.chatg…
switch-to-codex.openai.chatg…
你喜欢 Sol 的什么地方，或者你为什么转用它？
引用 Tibo @thsottiauxOr… what if we gave you $100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched? Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens! switch-to-codex.openai.chatg…查看被引原帖 ↗
查看英文原文
What do you love about Sol, or why did you switch to it?
Notion 终于支持 markdown 文件了…
在 AI 已经快四年的时候…
引用 Notion @NotionHQ.md, meet Notion 👋 You can now open Markdown files in Notion as a read-only preview, then import them as Notion pages.查看被引原帖 ↗
🚨 前沿模型路由器 - 混搭所有前沿LLM
我们最热门的自定义路由器就是这个前沿模型路由器，把最好的LLM组合起来完成所有任务
GPT 5.6 sol - 数据分析
Fable / Opus 4.8 - 编码
Flash 3.5 - 摘要
Grok 4.5 - 实时处理
为所有任务选你最喜欢的LLM。到目前为止自定义路由器可以在任何框架中使用 - Abacus AI、Claude、Codex、Open-Code....
查看英文原文
🚨 Frontier Model Router - Mix And Match All the Frontier LLMs
Our most popular custom router is our frontier-model router that combines the best LLMs for all your tasks
GPT 5.6 sol - data analysis
Fable / Opus 4.8 - coding
Flash 3.5 - summarization
Grok 4.5 - real-time
Pick your favorite LLMs for all your tasks. As of today the custom routers can be used in ANY harness - Abacus AI, Claude, Codex, Open-Code....
Our most popular custom router is our frontier-model router that combines the best LLMs for all your tasks
GPT 5.6 sol - data analysis
Fable / Opus 4.8 - coding
Flash 3.5 - summarization
Grok 4.5 - real-time
Pick your favorite LLMs for all your tasks. As of today the custom routers can be used in ANY harness - Abacus AI, Claude, Codex, Open-Code....
Ollama 今天回到纽约了！
很兴奋看到越来越多人和企业意识到开源模型的好处！自有。实惠。隐私。
感谢 @Nasdaq 和 @Theoryvc 团队。
查看英文原文
Ollama is back in NYC today!
It's exciting to see more people and businesses realize the benefits of open models! Ownership. Affordable. Private.
Thank you
@Nasdaq
and the
@Theoryvc
team.
It's exciting to see more people and businesses realize the benefits of open models! Ownership. Affordable. Private.
Thank you
@Nasdaq
and the
@Theoryvc
team.
旧金山个人 AI 工程师——如果你正在构建个人代理，这周四晚上来新媒体实验室演示（并见见 @shloked，我今年有幸遇到的最好的建筑师-作家之一）。
上次我们举办这个 meetup 时，我们的特色演讲者被 @Amazon 硬件部门收购……两年后我仍然是日常使用者。看到 PAI 的持久力真是不可思议。
引用 Latent.Space @latentspacepodFor those in SF - we're hosting the epic return of our Personal AI meetup this week, hosted by @swyx and @shloked : luma.com/personai come if you are building in Personal AI - people with demos prioritized! max 50 spots only, this is not one of those slopfests.查看被引原帖 ↗
查看英文原文
SF Personal AI engineers - if you are building personal agents, come demo at new media lab this thursday night (and meet
@shloked
, one of the best builder-writers I've had the fortune to meet this year).
last time we held this meetup our featured speakers got acquired by
@Amazon
hardware division... and 2 years later I'm still a daily active user. just incredible to see the staying power of PAI.
@shloked
, one of the best builder-writers I've had the fortune to meet this year).
last time we held this meetup our featured speakers got acquired by
@Amazon
hardware division... and 2 years later I'm still a daily active user. just incredible to see the staying power of PAI.
写了个 Gif 压缩Skill，用于满足公众号插入需求。
公众号 GIF 大小限制为 10M，而且超过300 帧也不行。
尺寸建议：宽度控制在 640 像素左右，帧率控制在12 - 20 fps 之间。
Skill安装指令：npx skills add joeseesun/qiaomu-tiny-gif
开源Github见评论，演示网站也在评论区
现在我真的完全不知道 Anthropic 在搞什么了
我最好的猜测就是他们被 Mythos 在某个 cyber eval 上打败了，现在觉得一年内能出 AGI。至少从最近的广告看，已经完全失去理智了。
但真正离谱的是，他们的计划里竟然没把 Fable/Mythos 算上
他们这等于是给 OpenAI 送了数十亿美元
引用 Lisan al Gaib @scaling01Anthropic is truly doing God's work and preventing a lot of harm!查看被引原帖 ↗
查看英文原文
At this point I have no idea anymore what Anthropic is doing
my best guess is that they got one-shot by Mythos solving some cyber eval and now think they'll have AGI in a year. at least their most recent ad looked like they completely lost touch with reality.
but it's completely ridiculous that they don't include Fable/Mythos in their plans
they are literally donating billions to OpenAI
my best guess is that they got one-shot by Mythos solving some cyber eval and now think they'll have AGI in a year. at least their most recent ad looked like they completely lost touch with reality.
but it's completely ridiculous that they don't include Fable/Mythos in their plans
they are literally donating billions to OpenAI
Anthropic 真的太了不起了，防止了这么多伤害！
引用 Latent.Space @latentspacepodCodex 是否超越 Claude Code？Tibo 24.5 小时前宣布 600 万用户，意味着 Codex 日均增长 100 万。Claude Code 上次用户数为 2 月的 200 万，若属实将是业界重大变化。查看被引原帖 ↗
查看英文原文
Anthropic is truly doing God's work and preventing a lot of harm!
对开源 AI 很看好。
- agentic loops 正在快速商品化
- GLM 和 Kimi 还可以
- 现在已经能让开源 AI 真正用起来了！
再过几个月，开源就能占据至少 20% 的工作负载
查看英文原文
Very bullish on open-source AI.
- agentic loops are fast becoming a commodity
- GLM and Kimi are half decent
- It's possible to make open-source AI usable today!
We are literally months away from open-source taking over at least 20% of workloads
- agentic loops are fast becoming a commodity
- GLM and Kimi are half decent
- It's possible to make open-source AI usable today!
We are literally months away from open-source taking over at least 20% of workloads
PrismML 将 27B 模型塞进你的 iPhone 里
而且智商几乎没怎么缩水
PrismML公司基于Qwen3.6-27B模型，将约 54GB 的 27B 模型压到约 3.9–5.9GB
使其能在手机上运行
在「高强度思考模式」下用 15 项测试对比原版：
Ternary 版本（5.9GB 电脑版）：保留了原版 95% 的战斗力！
1-bit 版本（3.9GB 手机版）：也保留了 90% 的战斗力！
在电脑（RTX 5090）上最快能达到每秒 163 个字；在苹果 M5 Max 芯片上也能达到每秒 87 个字。
读了那个爆红的「反垃圾」agent markdown，才意识到它基本是 100% AI 生成的，最后反而会让你的输出变成超级垃圾
你要么自己得有品味，要么找个有品味的人，然后搭建自己的 skill 和 instruction
查看英文原文
Reading one of the viral "anti-slop" markdown files for agents and realizing it is basically a 100% AI written and will basically hyperstition your outputs into mega-slop
You need to either have a sense of taste or hire someone with one & then build your own skills/instructions
You need to either have a sense of taste or hire someone with one & then build your own skills/instructions
这个 App 最开始是自己用的，是走的 LLM 方案，先整体翻译，然后按句子配对，再 LLM 拆分，但是拆分效果总是不理想。
另外走 LLM 还有一个问题，就是要配置 API Key，这其实对普通人不友好，另外就是成本其实不低，Gemini 3.5 Flash 一部长一点视频都几美元。
最后发现还是 Agent + Skill 路线最好：
1. Agent 有很好的纠错能力，哪怕你不给它这样的工具，它自己都能一边翻译一边纠错，质量不错。
2. 把 App 的能力封装成 cli，让 Agent 可以通过命令行调用 App 的功能
3. 配合一个 Skill 把常用的工作流比如像转录、润色、翻译、对齐、剪辑都固化下来，这样 Agent 就知道最佳实践是什么，不需要每次都自己摸索
4. App 解决的是人工预览、校对的部分。
Agent 操作完，如果你是人肉看文本字幕，或者你每次修改一点都要重新 ffmpeg 生成一遍视频，那效率太低了。最好的方式其实不是完全交给 Agent，而是可以有一个图形化的工具快速的预览、二次编辑。
5. Agent 相对比较通用和便宜，现在大家都有 Agent，包月的 Token 经常花不完，用来干这种活正合适
引用 宝玉 @dotey字幕转录翻译剪辑 Skill —— BaoCut（仅支持 Mac） 借助 Agent Skill，可以转录视频、对转录结果识别 Speaker、润色（纠正错别字口癖等）、也可以根据转录结果对视频进行简单的剪辑，比如删除口癖、重复等。 这次尝试解决一个问题就是 Agent 对字幕转录翻译后，无法通过一个友好的操作界面二次编辑的问题。 现在的做法是为 Agent 提供一个 cli，配合 Skill 的说明，Agent 可以借助 cli 去转录，获取转录结果润色、翻译，并实时同步进度到 GUI。后续可以在 GUI 进行预览和人工编辑。 安装了 Skill 和 App 后，后续只要从 Codex 或者 Claude Code 这种 Agent，触发 Skill 即可执行，比如： > /baocut 转录并翻译视频：<视频 url 或路径> 已知问题： - 仅支持 Mac - 翻译速度略慢，但质量会不错 下载地址： baocut.app/ Skill 从 App 内可以安装，或者 Skill 地址： github.com/jimliu/baocut查看被引原帖 ↗
Codex 最新动态：
从 GPT‑5.6 与 Ultra 模式，到多智能体并行协作、计算机与浏览器操作、应用截图理解、行内代码和文档修改，再到一键发布 Sites、跨项目管理以及完整的 PR 工作流。
Codex 已不只是一个“帮你写代码”的工具，而是能够拆解复杂任务、操作和测试应用、协调多个项目、处理 Bug 与代码审查，并协助完成发布的开发伙伴。
无论你想提高日常开发效率，还是探索更自动化的 AI 编程方式，这支视频都能帮你快速掌握 Codex 的新能力和实际应用场景。
本视频由 baocut 翻译
引用 OpenAI Developers @OpenAIDevsCodex每周用户超700万，两个月内推出150多次更新。最新功能包括GPT-5.6和Ultra、/goal并行工作、更快计算机使用、AppShots、内联编辑、Sites、移动版和SSH工作流、从审查到合并的PR。查看被引原帖 ↗
语音输入法升级
HeyClicky：一个能 “屏幕感知语音输入”工具
它不仅能将语音内容转文字，还能看到你屏幕的内容，能理解你屏幕上正在看的内容，再根据上下文帮你写。
在 Gmail 中，可以看到邮件内容后自动生成合适的回复
在 Claude 终端里，可以根据当前技术信息替小白补充追问
在演示文稿中，可以调用 YC Skill，帮助修改融资材料和开场文案
可以在各种应用里看到你的内容进行智能回复
它还会记住你的表达习惯，让生成内容更像本人
一句话概括：它把传统语音输入升级成了“能看懂屏幕、理解上下文、直接替你完成表达”的 AI 写作助手。
引用 Farza 🇵🇰🇺🇸 @FarzaTV推出屏幕感知语音输入功能。语音转文本极快（约450ms），可根据屏幕内容自动生成文本。在Claude Code中自动生成提示词，在Gmail中用语音回复邮件。查看被引原帖 ↗
刚开发找片网站，不知道看啥时点一下就行：
imdb.qiaomu.ai/
支持自然语言对话找片，再次感谢Deepseek，v4 flash太便宜了。
所以现在做的不少站，都加上了一点AI功能。
你要是想要 cog 更新的话，看看我今天帮忙的东西列表还挺酷的。
我想在 New Media Lab 为初创公司做批量咨询，不过得找到合适的运营合作伙伴才行。
引用 Cognition @cognition一年前，Cognition收购了Windsurf。一年来推出了数十项新功能，增加数亿美元ARR，并发表了前沿AI研究。查看被引原帖 ↗
查看英文原文
and if u want the cog update its pretty cool to see a list today of stuff i helped out on.
im looking to do batched advisory for startups at New Media Lab if i can find the right operational partner
im looking to do batched advisory for startups at New Media Lab if i can find the right operational partner
他妈的，天塌了
GPT 这么快就被污染了
龟儿子把Token翻译成词元了
草他妈的...
huggingface.co/docs/inferenc…
查看英文原文
huggingface.co/docs/inferenc…
用 ChatGPT work & sol 工作时，我觉得特别开心，能随时问关于业务的任何问题，都能得到彻底的研究和回答。我意识到之前有这么多问题我都没问过，因为太费事了。
查看英文原文
with chatgpt work & sol, i'm finding it incredibly joyful to just ask any question about the business and have it be thoroughly researched and answered.
realizing i have so many questions i wouldn't have bothered asking because they would be too burdensome to answer.
realizing i have so many questions i wouldn't have bothered asking because they would be too burdensome to answer.
嗨老朋友们，快速指南
如果你从我的对冲基金时代认识我，搜搜 @latentspacepod 和 @fabknowledge 的采访
如果你从 TypeScript/React 认识我，看看 @cramforce 的 @aidotengineer keynote 和 @bcherny 的播客
如果你从 AWS/Temporal/Data Eng 认识我，搜搜和 @matei_zaharia 与 @rxin 的播客。是的，《Rise of the AI Engineer》是对 @mistercrunch 文章的大胆改编。
如果你从 devrel 认识我，看看 @dxtipshq 和 @MilksandMatcha 的文章和写作聚会。
过去 10 年的变更日志在下面。欢迎回来
查看英文原文
hi old friends, quick guide
if you know me from my hedge fund days, look up the
@latentspacepod
with
@fabknowledge
if you know me from Typescript/React, look up
@cramforce
’s
@aidotengineer
keynote and
@bcherny
’s pod with us
if you know me from AWS/Temporal/Data Eng, google the pod with
@matei_zaharia
and
@rxin
. yes, Rise of the AI Engineer is a brazen adaptation of
@mistercrunch
’s essay.
if you know me for devrel, check out
@dxtipshq
and
@MilksandMatcha
’s writing meetups.
general changelog of last 10 years is below. welcome back
if you know me from my hedge fund days, look up the
@latentspacepod
with
@fabknowledge
if you know me from Typescript/React, look up
@cramforce
’s
@aidotengineer
keynote and
@bcherny
’s pod with us
if you know me from AWS/Temporal/Data Eng, google the pod with
@matei_zaharia
and
@rxin
. yes, Rise of the AI Engineer is a brazen adaptation of
@mistercrunch
’s essay.
if you know me for devrel, check out
@dxtipshq
and
@MilksandMatcha
’s writing meetups.
general changelog of last 10 years is below. welcome back
在本地运行现代 AI agents 总是很麻烦。得配置 Docker、处理沙箱、保持终端会话活跃、手动路由模型密钥。MyClaw 在云上管理所有这些。它是个持久平台，托管像 OpenClaw 和新推出的 Hermes（Claude Code 接下来上）这样的 agent runtimes。与其等浏览器聊天完成，你可以按计划或按需触发 agent 运行。完成的交付物（比如爬虫数据库报告、竞对价格监控或收件箱分类）直接发送到你的邮箱、WhatsApp 或 Telegram。
他们刚推出 Hermes，所以如果你试用两个 agent stack 中的一个，可以免费试 30 天另一个的 Max plan，在你的实际工作流上对比测试。
myclaw.ai
查看英文原文
Running modern AI agents locally is always a pain. You have to configure Docker, handle sandboxing, keep a terminal session alive, and manually route model keys.
MyClaw manages all of this in the cloud. It is a persistent platform that hosts agent runtimes like OpenClaw and the newly added Hermes (with Claude Code coming next).
Instead of waiting for a browser chat to finish, you trigger the agent to run on a schedule or on demand. The finished deliverable (such as a scraped database report, competitor price monitor, or inbox triage) gets sent directly to your email, WhatsApp, or Telegram.
They just launched Hermes, so if you try either agent stack, you get a free 30-day trial of the other's Max plan to test them side-by-side on your actual workflows.
myclaw.ai
MyClaw manages all of this in the cloud. It is a persistent platform that hosts agent runtimes like OpenClaw and the newly added Hermes (with Claude Code coming next).
Instead of waiting for a browser chat to finish, you trigger the agent to run on a schedule or on demand. The finished deliverable (such as a scraped database report, competitor price monitor, or inbox triage) gets sent directly to your email, WhatsApp, or Telegram.
They just launched Hermes, so if you try either agent stack, you get a free 30-day trial of the other's Max plan to test them side-by-side on your actual workflows.
myclaw.ai
想象一下用欧洲的编程工具吧
与其要求权限，反而要你在每一步都接受 cookies
引用 Lisan al Gaib @scaling01欧洲永远无法打造自己的前沿编码模型。查看被引原帖 ↗
查看英文原文
just imagine using a european coding harness
instead of asking for permissions it asks you to accept cookies at each step
instead of asking for permissions it asks you to accept cookies at each step
要知道这是 OpenAI 对 GPT-5.6-Sol 说的话
那你凭什么觉得 GLM-5.2 在 PostTrainBench 上并列第一很重要？
我一个月前做的分析，大家都说我 GLM-5.2 黑粉
现在 OpenAI 说同样的话，突然没人在乎了
引用 Lisan al Gaib @scaling01GPT-5.6 Sol和Terra常陷入狭隘的策略集合，还无法在各种基础模型和下游目标上可靠地设计和执行完整的后训练方案。查看被引原帖 ↗
查看英文原文
mind you this is what OpenAI said about GPT-5.6-Sol
so why do you think that GLM-5.2 being in shared first place on PostTrainBench matters?
i did the analysis a month ago, and everyone called me a GLM-5.2 hater
now OpenAI says the same thing and suddenly no one cares about it anymore
so why do you think that GLM-5.2 being in shared first place on PostTrainBench matters?
i did the analysis a month ago, and everyone called me a GLM-5.2 hater
now OpenAI says the same thing and suddenly no one cares about it anymore
欧洲永远无法打造自己的前沿编程模型
查看英文原文
europe will never build its own frontier coding model
OpenCode Desktop 现在支持 Ollama 了！
快去试试那些顶尖的开源模型吧！
引用 OpenCode @opencodeIntroducing Tabs OpenCode Desktop is now built around tabs. Start a new session in a tab, or open an existing session from any of your projects. Open a new tab when you're starting something new, and close it when you're done. Download the latest to get started.查看被引原帖 ↗
查看英文原文
You can use OpenCode Desktop with Ollama!
Try it with the top open models!
Try it with the top open models!
docs.ollama.com/integrations…
opencode.ai/docs/providers/#…
查看英文原文
docs.ollama.com/integrations…
opencode.ai/docs/providers/#…
opencode.ai/docs/providers/#…
即将发布 - 开源和闭源模型新品
- Opus 5
- Gemini 3.5 Pro（checkpoint 表现更好）
- DeepSeek v4
- Kimi 3
Fable 5 在至少一个月内都将是编程领域最强的模型
查看英文原文
Coming soon - open and closed source model releases
- Opus 5
- Gemini 3.5 Pro ( checkpoints are better)
- DeepSeek v4
- Kimi 3
Fable 5 will remain the top model for coding for at-least a month
- Opus 5
- Gemini 3.5 Pro ( checkpoints are better)
- DeepSeek v4
- Kimi 3
Fable 5 will remain the top model for coding for at-least a month
我还经常听到有人说「会写代码的话自己写反而更快」
我的看法恰恰相反：既然懂得怎么写，自己敲键盘就没意义了 —— 还是交给 coding agent 吧！
查看英文原文
I still sometimes see people saying "if you know how to write the code, it's faster to write it yourself"
I'd argue the exact opposite: if you know how to write it, you gain nothing from doing the typing yourself - outsource that to a coding agent!
I'd argue the exact opposite: if you know how to write it, you gain nothing from doing the typing yourself - outsource that to a coding agent!
失去风筝线后风筝能找到自由吗？🪁
来看看 Shihua Lin、Mingwei Chen 和 Xuanwei Liu 的作品《Kite》。
获得 Kling AI NEXTGEN 2026 Campus AIGC Creation Competition 的 Apex Award，太牛了！
查看英文原文
Can a kite find freedom after losing its string? 🪁
Discover “Kite” by Shihua Lin, Mingwei Chen & Xuanwei Liu.
Congratulations on winning the Apex Award at Kling AI NEXTGEN 2026 Campus AIGC Creation Competition!
Discover “Kite” by Shihua Lin, Mingwei Chen & Xuanwei Liu.
Congratulations on winning the Apex Award at Kling AI NEXTGEN 2026 Campus AIGC Creation Competition!
我们开放了 @vercel AI Gateway 上 AI token 流的数据集。里面的数据太有意思了！
引用 Vercel @vercelAI Gateway leaderboard data is now open. Real production usage across models, labs, apps, and providers, updated daily. Download, query, and cite it under CC BY 4.0. vercel.com/changelog/open-da…查看被引原帖 ↗
查看英文原文
We're opening up the dataset of AI token flows on
@vercel
AI Gateway. Fascinating insights contained within!
@vercel
AI Gateway. Fascinating insights contained within!
我用这个数据集做的↓
引用 Guillermo Rauch @rauchg🏁 I animated the token 💰 spend race, from ~lifetime Vercel AI Gateway usage, which aggregates trillions of tokens from millions of developers a month. Fascinating to see the fluctuations among the labs, Anthropic's dominance, and the rise of open weight AI.查看被引原帖 ↗
查看英文原文
I used this dataset to create ↓
Sesame在iOS应用上开发新的日历和邮件连接器。用户可以给Sesame agent授权来读取和操作这些应用。另外还有个内部菜单能测试自定义技能，可以添加自定义提示词——如果这功能能公开就太好了。目前Sesame仍是我首选的语音模式，因为既支持duplex还能用工具（提醒和记笔记功能）。我也喜欢GPT Live 1，就是目前还没有工具访问。
查看英文原文
Sesame is working on new Calendar and Email connectors for its iOS app.
Users will be able to provide Sesame agent an access to read and potentially operate these apps.
Besides that, there is an internal menu for testing custom skills where custom prompts can be added. It would be super useful if that feature will get released to the public too.
Currently, Sesame is still a “go to” voice mode for me because it is both duplex but also has access to tools (reminders and note taking). I like GPT Live 1 as well but it lacks tool access at this moment.
Users will be able to provide Sesame agent an access to read and potentially operate these apps.
Besides that, there is an internal menu for testing custom skills where custom prompts can be added. It would be super useful if that feature will get released to the public too.
Currently, Sesame is still a “go to” voice mode for me because it is both duplex but also has access to tools (reminders and note taking). I like GPT Live 1 as well but it lacks tool access at this moment.
AI的下一章不仅取决于技术进步，还要看我们如何负责任地、深思熟虑地把它引入世界。期待在Vegas参加8月4-6的Ai4 2026大会并在舞台上发言。在ai4.io/register/注册。#Ai42026
查看英文原文
AI’s next chapter will be defined not only by technical progress, but by how responsibly and thoughtfully we bring it into the world.
I’m looking forward to be speaking on stage at Ai4 2026, Aug 4–6 in Vegas. Register here at
ai4.io/register/
#Ai42026
I’m looking forward to be speaking on stage at Ai4 2026, Aug 4–6 in Vegas. Register here at
ai4.io/register/
#Ai42026
连环推 ×10AI电影制作已经牛逼到吓人了……这些不是随便剪的。5个电影人用invideo Agent One制作了完整电影和剧集……然后详细展示了怎么做的。5部电影，5个分解教程。收藏这个：
1. The Sage - Episode 3 "Sundown"
Mike Mitch 展示了整个工作流程。
Vishal Balsara 用 Agent One 把这个创意打造成了最终的电影成品。
观看：
Vishal 详细拆解了这部电影的完整流程。
Vinu 用 Agent One 完整制作了这整一集。
这就是 Vinu 用 Agent One 一个场景一个场景打造这集的全过程。
7. SAV用Agent One从概念到最终成片导演了这部AI电影。
8. 这是SAV用AI agent制作电影的方法。
9. Aze Alter用Agent One创作了这整个剧集。
10. Aze公开了整个创意过程。
查看英文原文
AI filmmaking is getting scary good...
These aren't random clips.
5 filmmakers made full films and episodes with invideo Agent One... then showed exactly how.
5 films. 5 breakdowns.
Bookmark this:
1. The Sage - Episode 3 "Sundown"
2. Mike Mitch showed the full workflow behind it.
3. Vishal Balsara used Agent One to take this story from idea to finished film.
Watch:
4. Vishal breakdown the full process behind the film.
5. Vinu created this full episode using Agent One.
6. Here's exactly how Vinu built the episode scene by scene with Agent One.
7. SAV used Agent One to direct this AI film from concept to final cut.
8. Here's how SAV built the film with an AI agent.
9. Aze Alter created this full episode using Agent One.
10. Aze opened the hood on the entire creation process.
These aren't random clips.
5 filmmakers made full films and episodes with invideo Agent One... then showed exactly how.
5 films. 5 breakdowns.
Bookmark this:
1. The Sage - Episode 3 "Sundown"
2. Mike Mitch showed the full workflow behind it.
3. Vishal Balsara used Agent One to take this story from idea to finished film.
Watch:
4. Vishal breakdown the full process behind the film.
5. Vinu created this full episode using Agent One.
6. Here's exactly how Vinu built the episode scene by scene with Agent One.
7. SAV used Agent One to direct this AI film from concept to final cut.
8. Here's how SAV built the film with an AI agent.
9. Aze Alter created this full episode using Agent One.
10. Aze opened the hood on the entire creation process.
通过 Direct On-Policy Distillation 实现弱强泛化
查看英文原文
Weak-to-Strong Generalization via Direct On-Policy Distillation
Bonsai 27B 现已在 Claude Code 中上线，可通过 HuggingFace Claude 访问。
查看英文原文
Bonsai 27B now available in claude code via hf claude
AI Gateway 排行榜数据开放了。真实生产环境的使用数据，各种模型、实验室、应用和服务商都有，每天更新。CC BY 4.0 协议下可以下载、查询和引用。vercel.com/changelog/open-da…
查看英文原文
AI Gateway leaderboard data is now open.
Real production usage across models, labs, apps, and providers, updated daily. Download, query, and cite it under CC BY 4.0.
vercel.com/changelog/open-da…
Real production usage across models, labs, apps, and providers, updated daily. Download, query, and cite it under CC BY 4.0.
vercel.com/changelog/open-da…
Codex 周活用户 700 万+。两个月 150+ 个更新。@romainhuet 为你介绍 Codex 的新功能：GPT-5.6 和 Ultra、/goal 并行工作、更快的 computer use、AppShots、内联编辑、Sites、Codex 移动版和 SSH 工作流、PR 从审查到合并
查看英文原文
7M+ weekly Codex users. 150+ updates in two months.
@romainhuet
catches you up on what’s new in Codex:
GPT‑5.6 and Ultra
Parallel work with /goal
Faster computer use
AppShots
Inline edits
Sites
Codex mobile and SSH workflows
PRs from review to merge
@romainhuet
catches you up on what’s new in Codex:
GPT‑5.6 and Ultra
Parallel work with /goal
Faster computer use
AppShots
Inline edits
Sites
Codex mobile and SSH workflows
PRs from review to merge
接招: 卷我们最擅长了
国模这波直接霸榜了 OpenRouter 上周和本周的 Top6 用量模型. 与去年7月份可以说是两极反转了, 那个时候破圈的只有DeepSeek和Qwen.
现在用量第一的是腾讯混元3的免费版 (可以配置到龙虾里).
然后就是小米MiMo-V2.5, 强烈怀疑前几个月的连续免费起效果了, MiMo 方案简单粗暴有效, 上新模型就嗷嗷免费, 培养出用户习惯了.
DeepSeek-V4-Flash 现在来看定位特别精准, 大部分用量都是Agent用户, 而且flash模型虽然定价低, 但是并不意味着赚钱少反而用户敢用了, 特别依靠长上下文, 请求量很大. 按照长上下文排序, DeepSeek-V4-Flash 就是 Top. 其余的是其它的 flash 模型.
而且很有趣的一点启发是, 如果决定要打Agent场景, 那么发布的模型请一定要叫 xxx-flash, 因为这个名字认可度很大. 小脑决定大脑了可以说.
>
#deepseek
#openrouter
#flash模型
Wide Research 现在可以在 Perplexity Agent API 里用了
引用 Perplexity Developers @perplexitydevsPerplexity认为窄搜索已解决，前沿是宽研究。推出WANDR基准（500个任务），即使对最强大模型也很困难。其Agent API的Search as Code架构在WANDR上表现出色，因为它允许模型一次设计研究方案，然后大规模确定性执行而不压倒上下文。查看被引原帖 ↗
查看英文原文
Wide Research is now available in Perplexity Agent API
发布了一些关于反向工程宠物生成机制工作原理的笔记 simonwillison.net/2026/Jul/1…
查看英文原文
Blogged a few notes on reverse-engineering how the pet generation mechanism works
simonwillison.net/2026/Jul/1…
simonwillison.net/2026/Jul/1…
现在创业你真正需要的全是免费开源软件，每月零成本
+ VPS 服务器
+ 用来处理 AI 功能的 API
+ R2 或 S3 文件存储
他们不想让我发这条推，因为这会砸了他们的生意，但我不能对你说谎啊！
引用 Lukasz @woocassh用户从Supabase迁移至SQLite，因为Supabase虽然设置便捷，但增加了明显延迟。迁移后应用前端速度明显提升。SQLite便于用Claude调试。建议保持简单架构，并定期备份。用户还在探索用Cloudflare替代Postmark进行邮件服务。查看被引原帖 ↗
查看英文原文
All you really need today to build a business is 100% free open source software that charges you $0/mo
+ a VPS server
+ an API to do some required AI stuff
+ some R2/S3 file hosting
They don't want me to tweet that because it destroys their businesses but I can't lie to you!
+ a VPS server
+ an API to do some required AI stuff
+ some R2/S3 file hosting
They don't want me to tweet that because it destroys their businesses but I can't lie to you!
@agentmail 团队一直在搞事情。告诉你的 agent 来 𝚟𝚎𝚛𝚌𝚎𝚕 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 𝚊𝚐𝚎𝚗𝚝𝚖𝚊𝚒𝚕 – 无需注册、自动配置、统一账单
引用 Vercel Developers @vercel_devAgentMail已上架Vercel Marketplace。您的agents可以从真实收件箱发送、接收和回复邮件。功能包括：完整线程记忆、结构化数据提取、可交付性处理。运行vc i agentmail或点击了解更多。查看被引原帖 ↗
查看英文原文
The
@agentmail
team has been cooking. Tell your agent to 𝚟𝚎𝚛𝚌𝚎𝚕 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 𝚊𝚐𝚎𝚗𝚝𝚖𝚊𝚒𝚕 – no signup, automatic setup and unified billing.
@agentmail
team has been cooking. Tell your agent to 𝚟𝚎𝚛𝚌𝚎𝚕 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 𝚊𝚐𝚎𝚗𝚝𝚖𝚊𝚒𝚕 – no signup, automatic setup and unified billing.
Fable：'把奥德修斯包装成管理顾问，论证他已经实现了产品市场契合，应该继续做木马生意而不是去 PowerPoint 上回伊萨卡'
我太喜欢卡桑德拉的 1 星评价了：'10000 个读者中 0 个觉得有用'
查看英文原文
Fable: "Pitch Odysseus as a management consultant, arguing that he has found product-market fit, and he should just stick with Trojan Horse making as opposed to going home to Ithaca in a powerpoint"
I like the 1 star review from Cassandra "0 of 10,000 readers found this helpful
I like the 1 star review from Cassandra "0 of 10,000 readers found this helpful
OpenAI 🔥：彭博社报道，OpenAI 的首款设备将是无屏幕智能音箱。
> OpenAI 正在开发一款移动无屏幕智能音箱，作为拟人化 AI 伙伴，能控制智能家居并支持 ChatGPT 的各项能力。
> 该设备会随着对用户理解的深化而变得越来越个性化和主动，其核心卖点是具有人格、能和用户建立类人级别的连接。
> Apple 起诉 OpenAI 盗取商业机密，这使 OpenAI 的设备计划复杂化，设备上市可能会因诉讼结果而延迟。
引用 Mark Gurman @markgurmanOpenAI首款产品是一款移动、无屏家庭智能扬声器，用户可与其建立像AI伴侣一样的联系。在Apple的商业秘密诉讼背景下，这家iPhone制造商在市场上没有类似产品。查看被引原帖 ↗
查看英文原文
OPENAI 🔥: The first device from OpenAI will be a screenless smart speaker, as reported by Bloomberg.
> OpenAI is developing a mobile, screen-free smart speaker to serve as a humanlike AI companion that can control smart-home appliances and tap into the capabilities of OpenAI's ChatGPT.
> The device is designed to become increasingly personalized and proactive as it gains a deeper understanding of its owner over time, and its defining feature will be its personality and ability to connect on a humanlike level with users.
> OpenAI's plans for the device have been complicated by a lawsuit from Apple, which accuses OpenAI of stealing trade secrets, and the company's ability to sell the device may be delayed depending on the outcome of the legal process.
> OpenAI is developing a mobile, screen-free smart speaker to serve as a humanlike AI companion that can control smart-home appliances and tap into the capabilities of OpenAI's ChatGPT.
> The device is designed to become increasingly personalized and proactive as it gains a deeper understanding of its owner over time, and its defining feature will be its personality and ability to connect on a humanlike level with users.
> OpenAI's plans for the device have been complicated by a lawsuit from Apple, which accuses OpenAI of stealing trade secrets, and the company's ability to sell the device may be delayed depending on the outcome of the legal process.
Muse Spark 在 2026 年亚洲物理奥林匹克竞赛中拿了满分！
这是我们用 AI 系统加速科学这个长期目标的一个大进展。
引用 AI at Meta @AIatMetaMeta AI模型参加亚洲物理奥林匹克理论考试，获得30/30满分，并列前三学生。展示推理和多模态能力。查看被引原帖 ↗
查看英文原文
Muse Spark achieved a perfect score on the 2026 Asian Physics Olympiad!
Great milestone against our long-term goal of building AI systems to accelerate science.
Great milestone against our long-term goal of building AI systems to accelerate science.
为了展示 Meta AI 强大的推理和多模态能力，我们提交了一个模型参加亚洲物理奥林匹克的理论竞赛。很高兴能告诉你们，我们的模型拿到了 30/30 的满分，和前三名学生并列。
感谢 APhO 委员会让我们的模型参加竞赛：
apho2026.kr/en/index.html
查看英文原文
To demonstrate Meta AI's advanced reasoning and multimodal capabilities, we submitted a model to participate in the Asian Physics Olympiad’s theoretical exam. We’re happy to share that our model achieved a perfect score of 30/30, tying with the top 3 student contestants.
We appreciate the APhO committee for letting our model participate in the competition:
apho2026.kr/en/index.html
We appreciate the APhO committee for letting our model participate in the competition:
apho2026.kr/en/index.html
等待结束了，发布季节来临。
六月相对平静之后，节奏又加快了。GPT-5.6 是我一直期待的突破。我从未这么快用完过配额。Fable 5 回归了，Meta 推出了竞争力很强的 Spark 1.1，SpaceX 则发布了 Grok 4.5。不过据 SpaceX 说，他们真正的旗舰模型还在来的路上。看来 Cursor 的交易已经开始产生效果了。
众所周知，Fable 5 的故事还在继续。它在订阅计划中的可用性已经延长了两次，而 Opus 5 已经在 Vertex 上出现了。这强烈暗示 Opus 5 很快就会发布，很可能作为 Fable 5 的替代品。
更有意思的是，GLM-5.2 对很多人来说成了真正的'啊哈'和'哇'时刻。这说明开源已经达到了一个水平，至少在关键领域，它已经是西方闭源模型真正可行的竞争替代品。
更令人惊讶的是，这家公司的创始人今天又暗示他们的下一个模型也即将推出。不久前在和埃隆·马斯克的讨论中，他也提到他们仍计划在年底前发布自己的 Mythos 级开源模型。
而且还有报道称 Kimi K3 明天就要发布。值得回顾的是，Kimi K2.6 仅仅几周前还是公认最受欢迎的开源模型。AI 时代的变化速度真的快到不可思议。那是在 GLM-5.2 吸引了所有人目光之前。Kimi K3 预计将支持 100 万 token 的上下文窗口，这是个巨大的飞跃，让它更加吸引人。
简单来说，Opus 5、GLM-5.3 或者可能是 GLM-6，还有 Kimi K3，都近在咫尺了。与此同时，各种消息表明 ChatGPT 6 可能在几周内推出，采用全新的预训练流程。
漫长的等待寒冬已经结束。加速的夏季已经开始。
查看英文原文
The wait is over. Release season has arrived.
After a relatively quiet June, the pace has picked up again. GPT-5.6 was the breakthrough I had been hoping for. I’ve never burned through my rate limits so quickly. Fable 5 is back, Meta has released another competitive model with Spark 1.1, and SpaceX followed with Grok 4.5. According to SpaceX, however, their truly flagship model is still on the way. It looks like the Cursor deal is already paying off.
As many know, the Fable 5 saga continues. Its availability in the subscription plan has already been extended twice, while Opus 5 has meanwhile become visible in Vertex. That strongly suggests Opus 5 will be released soon, likely as a replacement for Fable 5.
Even more interesting, though, is that GLM-5.2 turned out to be the real “aha” and “wow” moment for many people. It made it clear that open source has now reached a level where, at least in key areas, it stands as a genuinely viable and competitive alternative to Western closed source models.
What makes this even more surprising is that the company’s founder once again hinted today that their next model is also coming soon. Not long ago, during a discussion with Elon Musk, he also said they still plan to release their own Mythos-class open source models before the end of the year.
As if that weren’t enough, leaks suggest that Kimi K3 will be released tomorrow. It’s worth remembering that Kimi K2.6 was, only a few weeks ago, arguably the most popular open source model around. Time moves unbelievably fast in the AI era. That was until GLM-5.2 captured everyone’s attention. Kimi K3 is expected to launch with a one million token context window, a significant leap that makes it even more compelling.
In short, Opus 5, GLM-5.3 or perhaps GLM-6, and Kimi K3 all appear to be just around the corner. At the same time, the rumor mill suggests that ChatGPT 6 could arrive within a matter of weeks, featuring an entirely new pre-training pipeline.
The long winter of waiting is over. The summer of acceleration has begun.
After a relatively quiet June, the pace has picked up again. GPT-5.6 was the breakthrough I had been hoping for. I’ve never burned through my rate limits so quickly. Fable 5 is back, Meta has released another competitive model with Spark 1.1, and SpaceX followed with Grok 4.5. According to SpaceX, however, their truly flagship model is still on the way. It looks like the Cursor deal is already paying off.
As many know, the Fable 5 saga continues. Its availability in the subscription plan has already been extended twice, while Opus 5 has meanwhile become visible in Vertex. That strongly suggests Opus 5 will be released soon, likely as a replacement for Fable 5.
Even more interesting, though, is that GLM-5.2 turned out to be the real “aha” and “wow” moment for many people. It made it clear that open source has now reached a level where, at least in key areas, it stands as a genuinely viable and competitive alternative to Western closed source models.
What makes this even more surprising is that the company’s founder once again hinted today that their next model is also coming soon. Not long ago, during a discussion with Elon Musk, he also said they still plan to release their own Mythos-class open source models before the end of the year.
As if that weren’t enough, leaks suggest that Kimi K3 will be released tomorrow. It’s worth remembering that Kimi K2.6 was, only a few weeks ago, arguably the most popular open source model around. Time moves unbelievably fast in the AI era. That was until GLM-5.2 captured everyone’s attention. Kimi K3 is expected to launch with a one million token context window, a significant leap that makes it even more compelling.
In short, Opus 5, GLM-5.3 or perhaps GLM-6, and Kimi K3 all appear to be just around the corner. At the same time, the rumor mill suggests that ChatGPT 6 could arrive within a matter of weeks, featuring an entirely new pre-training pipeline.
The long winter of waiting is over. The summer of acceleration has begun.
关于本地 AI 模型的看法真是个很好的智力测试
查看英文原文
Your opinion on local AI models is a great IQ test
如果你没注意的话，现在你可以用 ChatGPT 构建并部署完整的网络应用了。@coreyching 展示了怎么做。
查看英文原文
ICYMI, you can now build and ship full web apps with ChatGPT.
@coreyching
shows you how.
@coreyching
shows you how.
是不是只有我，还是 Claude Fable 今天的配额消耗速度特别快？
查看英文原文
Is it just me or is Claude Fable eating usage limits way faster today?
又一轮 Codex 重置来了。这群人绝对是传说。
时间刚刚好。我超爱！<3
引用 Tibo @thsottiaux已达到800万活跃用户（跨Codex和ChatGPT Work）。再次重置使用限制，无5小时速率限制，用户可探索GPT-5.6 Sol的边界。查看被引原帖 ↗
查看英文原文
Another Codex reset incoming. Absolute legends.
right on time for me. I love it! <3
right on time for me. I love it! <3
Codex 最近增长可真快！800 万了。
又重置了！
Anthropic 加油！
引用 Tibo @thsottiaux已达到800万活跃用户（跨Codex和ChatGPT Work）。再次重置使用限制，无5小时速率限制，用户可探索GPT-5.6 Sol的边界。查看被引原帖 ↗
一天一个重置，远离速率限制惨案 👀
引用 Tibo @thsottiaux已达到800万活跃用户（跨Codex和ChatGPT Work）。再次重置使用限制，无5小时速率限制，用户可探索GPT-5.6 Sol的边界。查看被引原帖 ↗
查看英文原文
One reset a day keeps rate limits away 👀
现在 Claude Code 增加了内置浏览器，你终于可以原生运行 MagicPath 作为扩展画布来设计和构建了。就像 Dmitry 在这里展示的那样 👇
引用 Dmitry Chesnokov @dmitrychesnok0v在 X 上首次亮相，分享了如何设置 MagicPathAI 与 Claude AI 浏览器的教程。作为 MagicPathAI 的工程师，期待听到你的想法，看到你使用它创造美好事物的各种方式。查看被引原帖 ↗
查看英文原文
Now that Claude Code has added an in-app browser, you can finally run MagicPath natively as an extended canvas to design and build with Claude. As Dmitry shows here 👇
我分享了它用的所有 sprites 和 prompts，链接在这儿 github.com/simonw/pedalican-…
查看英文原文
I shared all of the sprites and prompts it used here
github.com/simonw/pedalican-…
github.com/simonw/pedalican-…
连环推 ×2想超越prompt，开始用生成音频实际做东西吗？
欢迎报名参加我们和@MusicHackspace合作的免费实战workshop，学习怎样用Stable Audio 3.0打造定制的音频app和工作流。
Stability AI音频研究员CJ Carr（@cortexelation）workshop要讲的内容：
🎵 在本地运行模型（甚至MacBook Pro也能跑）
🎵 训练自己的LoRA
🎵 在Ableton中使用Stable Audio
🎵 基于模型构建自己的音频工具
……以及更多
查看英文原文
Want to go beyond prompting and start building with generative audio?
Join us for a free hands-on workshop in partnership with
@MusicHackspace
, exploring how you can use Stable Audio 3.0 to create custom audio applications and workflows.
What Stability AI audio researcher CJ Carr (
@cortexelation
) is covering in the workshop:
🎵 Run the model locally (even on a MacBook Pro)
🎵 Train your own LoRA
🎵 Use Stable Audio with Ableton
🎵 Build your own audio tools on top of the model
… and more
Join us for a free hands-on workshop in partnership with
@MusicHackspace
, exploring how you can use Stable Audio 3.0 to create custom audio applications and workflows.
What Stability AI audio researcher CJ Carr (
@cortexelation
) is covering in the workshop:
🎵 Run the model locally (even on a MacBook Pro)
🎵 Train your own LoRA
🎵 Use Stable Audio with Ableton
🎵 Build your own audio tools on top of the model
… and more
5.6 sol增长太疯狂了。
inference团队为了扛住需求真的费了老劲。
我们会继续拼命扩容，不过可能很快就会有些波澜。
查看英文原文
5.6 sol growth is insane.
the inference team has done heroic work to be able to support demand.
we are going to move mountains to continue to scale, but it is possible there are some hiccups soon.
the inference team has done heroic work to be able to support demand.
we are going to move mountains to continue to scale, but it is possible there are some hiccups soon.
我在招@GoogleAIStudio的TPM lead，工作是在产品、GTM、工程等各方面全方位加快进度。如果你是AI信徒、主动性强、想在Google DeepMind推动最前沿的工作，可以考虑加入 :)
查看英文原文
I am hiring a TPM lead for
@GoogleAIStudio
, the job is to accelerate our progress in every way possible across product, GTM, engineering, and more. If you are AI pilled, high agency, and want to push the frontier in Google DeepMind, consider working with us : )
@GoogleAIStudio
, the job is to accelerate our progress in every way possible across product, GTM, engineering, and more. If you are AI pilled, high agency, and want to push the frontier in Google DeepMind, consider working with us : )
申请链接在下面，我的DM也随时开放：google.com/about/careers/app…
查看英文原文
Here is the link to apply, my DM's are also open:
google.com/about/careers/app…
google.com/about/careers/app…
你的免费自定义域名还能领取，但只有几天时间了。⏰
Replit Core 和 Pro 会员可以免费获得一年的符合条件的域名，零 DNS 设置。优惠截止至美国太平洋时间 7 月 17 日（周五）晚上 11:59。
打开 Domains 工具领取你的域名吧。
了解更多：docs.replit.com/promotions/f…
查看英文原文
Your free custom domain is still up for grabs, but only for a few more days. ⏰
Replit Core and Pro members get their first year of an eligible domain free, with zero DNS setup. Offer ends Friday, July 17 at 11:59 PM PT.
Open the Domains tool and claim yours before it's gone.
Learn more here:
docs.replit.com/promotions/f…
Replit Core and Pro members get their first year of an eligible domain free, with zero DNS setup. Offer ends Friday, July 17 at 11:59 PM PT.
Open the Domains tool and claim yours before it's gone.
Learn more here:
docs.replit.com/promotions/f…
连环推 ×2Codex 现在可以做完整的 PowerPoint，但我在使用默认演示技能时遇到了困难。
它总是加载这个令人恐怖的通用幻灯片作为示例。但演示技能也包含了大量关于工具使用的重要信息。很难修改。
如果我编辑这个技能，下一次 Codex 更新时可能就被覆盖了。而且不清楚它会如何与第二个新演示技能交互。
结果导致 Sol 生成的每个演示文稿都惊人地通用，这很遗憾，因为幻灯片本身通常很有用。
查看英文原文
Codex can now do full Powerpoints, but I am struggling with the default presentation skill.
It always seems to load this nightmarishly generic slide deck as its example. But the presentation skill also includes tons of important information on tool use. Makes it hard to modify.
If I edit the skill, it may just get overwritten in the next Codex update. And it isn't clear how it would interact with 2nd new presentation skill
As a result every presentation Sol makes is heartbreakingly generic, which is sad because the slides themselves are often useful.
It always seems to load this nightmarishly generic slide deck as its example. But the presentation skill also includes tons of important information on tool use. Makes it hard to modify.
If I edit the skill, it may just get overwritten in the next Codex update. And it isn't clear how it would interact with 2nd new presentation skill
As a result every presentation Sol makes is heartbreakingly generic, which is sad because the slides themselves are often useful.
用于评估 Notion 架构上新模型的命令行工具是从零开始用 GPT-5.6 完整构建的。
查看英文原文
Command-line tools for evaluating new models on
@NotionHQ
’s architecture were built from start to finish with GPT‑5.6.
@NotionHQ
’s architecture were built from start to finish with GPT‑5.6.
连环推 ×3我们开源了 WANDR。WANDR 是我们内部开发的一个基准，用来在 Perplexity Computer 内部建立深度和广度的研究能力。
research.perplexity.ai/artic…
广泛深入的研究需要两种能力。1) Agent 要搜索得足够广，找到所有符合条件的实体。2) Agent 要深入调查，用证据支撑每一项声明。
WANDR 用分层、独立可验证的记录来表示这些要求。它包含 500 个研究任务，需要跨三个难度等级的 170,495 条有出处的记录。
查看英文原文
We’re open sourcing WANDR.
WANDR is an internal benchmark we built and used for building deep and wide research capabilities inside Perplexity Computer.
research.perplexity.ai/artic…
Wide-and-deep research requires two capabilities.
1) Agents must search broadly enough to find all qualifying entities
2) Agents must investigate deep enough to support every claim with evidence.
WANDR represents these requirements as hierarchical, independently verifiable records.
It consists of 500 research tasks that require 170,495 source-backed records across three tiers of difficulty.
WANDR is an internal benchmark we built and used for building deep and wide research capabilities inside Perplexity Computer.
research.perplexity.ai/artic…
Wide-and-deep research requires two capabilities.
1) Agents must search broadly enough to find all qualifying entities
2) Agents must investigate deep enough to support every claim with evidence.
WANDR represents these requirements as hierarchical, independently verifiable records.
It consists of 500 research tasks that require 170,495 source-backed records across three tiers of difficulty.
一个 27B 的模型仅用 3.9GB
引用 PrismML @PrismML发布Bonsai 27B，首款可在手机上运行的27B级多模态模型，基于Qwen3.6 27B，支持多步推理、结构化工具使用、长上下文工作流和智能体循环。Ternary版本5.9GB，1-bit版本3.9GB，已开源。查看被引原帖 ↗
查看英文原文
A 27B model using a mere 3.9GBs
终于开发好了，电脑端有音效，更带感。
在线玩（支持手机）：
mastermind.qiaomu.ai/
开源地址：
github.com/joeseesun/masterm…
引用 向阳乔木 @vista8《世界游戏大全51》中的猜颜色游戏特别好玩，然后用AI Coding复刻了一个。 原来这个游戏叫Mastermind。 用Threejs 做3D效果，交互上也可以利用鼠标优势。 等测试好了开源。查看被引原帖 ↗
最新的 ChatGPT iOS 新功能： Codex 可视化现在也在 iOS 上可用，可以生成各种图表和自定义内容，实时预览效果。
引用 Thomas Ricouard @Dimillian更新到最新版 ChatGPT iOS 应用可获得新功能：Codex 可视化现已在 iOS 上可用，可实时构建各种图表和自定义内容。查看被引原帖 ↗
快来加入 @steipete 今天上午 11 点（PT）的 Build Week 直播，学习如何用 Codex 把想法变成实际可用的应用。
nitter.tiekoetter.com/i/broadcasts/1vJpPPXVR…
查看英文原文
Join
@steipete
today at 11am PT for a Build Week live session on turning an idea into a working build with Codex.
nitter.tiekoetter.com/i/broadcasts/1vJpPPXVR…
@steipete
today at 11am PT for a Build Week live session on turning an idea into a working build with Codex.
nitter.tiekoetter.com/i/broadcasts/1vJpPPXVR…
Claude Opus 5 已经上线 Vertex，意味着发布日期近在咫尺了。
我的预测：他们会在这个周日从订阅计划中下架 Fable 5，但会在此前不久发布 Opus 5 来安抚大家。
引用 can @marmaduke091Claude Opus 5 已在 Vertex 上线，即将推出。查看被引原帖 ↗
查看英文原文
Claude Opus 5 on vertex means release is kinda imminent.
My prediction: They are removing Fable 5 from the subscription plan this coming Sunday, but will release Opus 5 shortly beforehand to appease everyone.
My prediction: They are removing Fable 5 from the subscription plan this coming Sunday, but will release Opus 5 shortly beforehand to appease everyone.
GPT 5.6 Sol 在 Codex App 的 System Prompt
现在我已经不怎么关注这些 System Prompt 了，对普通人来说不用管这些，好用就好。
引用 Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius系统提示泄露：GPT 5.6 Sol 在 Codex Desktop 的完整系统提示和工具。系统提示超 42000 字，分享了核心内容和完整文件链接。Codex 是基于 GPT-5 的代理，设计为与用户在共享工作区协作。查看被引原帖 ↗
Demis 说得好！值得一读
查看英文原文
Well said Demis! Worth reading
GOOGLE 🔥：Jules 即将换新 logo，同时推出 V2 更新。
感谢 Jules 团队的礼物！❤️
> Jules 是 Google Labs 的自主 AI 编码代理，于 2025 年 5 月 20 日公开测试版发布。
>
@julesagent
预计不久将进行重大改版，以处理更复杂的编码任务和目标。
查看英文原文
GOOGLE 🔥: Jules is about to get a new logo along with an upcoming V2 update.
Thanks to Jules team for the gift! ❤️
> Jules is an autonomous AI coding agent from Google Labs, launched in public beta on May 20, 2025.
>
@julesagent
is expected to undergo a major revamp soon to handle more complex coding tasks and goals.
Thanks to Jules team for the gift! ❤️
> Jules is an autonomous AI coding agent from Google Labs, launched in public beta on May 20, 2025.
>
@julesagent
is expected to undergo a major revamp soon to handle more complex coding tasks and goals.
字幕转录翻译剪辑 Skill —— BaoCut（仅支持 Mac）
借助 Agent Skill，可以转录视频、对转录结果识别 Speaker、润色（纠正错别字口癖等）、也可以根据转录结果对视频进行简单的剪辑，比如删除口癖、重复等。
这次尝试解决一个问题就是 Agent 对字幕转录翻译后，无法通过一个友好的操作界面二次编辑的问题。
现在的做法是为 Agent 提供一个 cli，配合 Skill 的说明，Agent 可以借助 cli 去转录，获取转录结果润色、翻译，并实时同步进度到 GUI。后续可以在 GUI 进行预览和人工编辑。
安装了 Skill 和 App 后，后续只要从 Codex 或者 Claude Code 这种 Agent，触发 Skill 即可执行，比如：
> /baocut 转录并翻译视频：<视频 url 或路径>
已知问题：
- 仅支持 Mac
- 翻译速度略慢，但质量会不错
下载地址：
baocut.app/
Skill 从 App 内可以安装，或者 Skill 地址：
github.com/jimliu/baocut
如果 Claude Code 是自行车，那 Vorflux 就是火箭。看视频就能理解什么叫范式转变。你得试试。
引用 Prasanna S @myprasannaVorflux是软件工程自动驾驶平台，由前Rippling联合创始人/CTO推出。融资1500万美元，Y Combinator等支持。核心理念：现有AI编码工具仍需人工操控，但AI已足够智能实现完全自动化。Vorflux提供真正的自动驾驶能力。欢迎分享工程瓶颈获取方案和$200免费额度。查看被引原帖 ↗
查看英文原文
If Claude Code is a bicycle, Vorflux is a rocketship.
Just watch the video to understand how much of a paradigm shift this is.
You NEED to try it.
Just watch the video to understand how much of a paradigm shift this is.
You NEED to try it.
完全赞同。现在模型对这套东西优化过度了，意识不到什么时候 agentsmd 已经过时要改或忽略了。昨晚我跑 5.6 sol 完成一个 5 阶段任务，早上才发现它还卡在第 0 阶段。翻回来看几小时的记录才明白——某个 agent 曾经提交过"第 0 阶段是目标，别干别的"，结果可怜的 sol 整整 8 小时都在打磨验证第 0 阶段，因为 /goal 不让停，agentsmd 不让继续。如果启动每个任务前不清楚 agentsmd 里有什么，那就是你在对自己来间接提示词注入。用 /plan、/goal、/skill 还是干脆啥都不用，都比这强。
引用 Ryan Dahl @rough__seaAGENTS.md/CLAUDE.md在很大程度上是反模式。查看被引原帖 ↗
查看英文原文
cosign. models have overtuned to this now and do not realize when the agentsmd is out of date and should be changed/ignored.
last night i goaled 5.6 sol to complete a 5 stage task and woke up to find it was still stuck on stage 0. it took a while to read the transcript back a few hours to realize at some point some agent had committed “stage 0 is the target dont do anything else” so poor sol spent 8 hours only refining and verifying stage 0 because /goal would not let it stop and agentsmd would not let it proceed.
if you dont know whats in your agentsmd before you fire off each task, it is an indirect prompt injection you perform on yourself. /plan, /goal, /skill, or nothing at all.
last night i goaled 5.6 sol to complete a 5 stage task and woke up to find it was still stuck on stage 0. it took a while to read the transcript back a few hours to realize at some point some agent had committed “stage 0 is the target dont do anything else” so poor sol spent 8 hours only refining and verifying stage 0 because /goal would not let it stop and agentsmd would not let it proceed.
if you dont know whats in your agentsmd before you fire off each task, it is an indirect prompt injection you perform on yourself. /plan, /goal, /skill, or nothing at all.
连环推 ×2Vorflux 是我用过最强的编码 Agent。完爆 Devin，没有任何可比性。（我是说真的，试了以后立马投资了）。这会成为一家超过 100 亿美元的公司。
绝了。它把多个框架合并起来，效果比任何单个框架或模型都强得多。代码方面，我用 Vorflux 配合便宜得多的模型也能达到 Fable 的水平。真正用上 Fable 的时候更是无敌。每次都能完美运行。
引用 Prasanna S @myprasannaVorflux是软件工程自动驾驶平台，由前Rippling联合创始人/CTO推出。融资1500万美元，Y Combinator等支持。核心理念：现有AI编码工具仍需人工操控，但AI已足够智能实现完全自动化。Vorflux提供真正的自动驾驶能力。欢迎分享工程瓶颈获取方案和$200免费额度。查看被引原帖 ↗
查看英文原文
Vorflux is the single best coding agent I've ever used.
It blows Devin out of the water. It's not even remotely close.
(and yes, I invested immediately after trying it).
This is going to be a >$100B company.
It's ridiculously good. It combines many harnesses to get far better results than any one harness/model.
For code, I'm able to get Fable-like performance with much cheaper models w/ Vortflux. Using Fable, it's unlike anything else.
Things just work every single fucking time.
It blows Devin out of the water. It's not even remotely close.
(and yes, I invested immediately after trying it).
This is going to be a >$100B company.
It's ridiculously good. It combines many harnesses to get far better results than any one harness/model.
For code, I'm able to get Fable-like performance with much cheaper models w/ Vortflux. Using Fable, it's unlike anything else.
Things just work every single fucking time.
看起来 Moonshot 的哥们还没准备好啊
继续等吧
引用 Lisan al Gaib @scaling01看起来今天真的要得到Kimi-K3了。查看被引原帖 ↗
查看英文原文
looks like moonshot bros aren't ready
the wait continues
the wait continues
我在 Google AI Studio 里申请了我的自定义 URL。
引用 Google AI Studio @GoogleAIStudio友情提醒：上周我们在 AI Studio 中发布了自定义 URL 功能，发布应用时现在可以给它取一个好记的名字。查看被引原帖 ↗
查看英文原文
I claimed my custom URL in Google AI Studio.
Demis这篇内容很重要。我们需要更多这样的思考。很好的提醒——目标是要构建一个鼓励创新和多样选择的边界生态系统，同时避免某个模型一出手就颠覆世界！
查看英文原文
An important piece from Demis. We need more of this kind of thinking. A good reminder that the goal is a frontier ecosystem that promotes innovation and choice, while avoiding any one model drop that breaks the world!
Replit 版本控制入门 - 像专业人士一样备份和恢复
nitter.tiekoetter.com/i/broadcasts/1lKQRRNzO…
查看英文原文
Intro to Version Control in Replit - Backup & Restore Like A Pro
nitter.tiekoetter.com/i/broadcasts/1lKQRRNzO…
nitter.tiekoetter.com/i/broadcasts/1lKQRRNzO…
庆祝视觉搜索创新25周年
Google Images迎来25周年纪念。回顾了一些重要里程碑以及探索和创作视觉内容的新方式。
查看英文原文
Celebrating 25 years of visual search innovation
Google Images is turning 25. Here’s a look back at some major milestones — and new ways to explore and create visual content.
Google Images is turning 25. Here’s a look back at some major milestones — and new ways to explore and create visual content.
《世界游戏大全51》中的猜颜色游戏特别好玩，然后用AI Coding复刻了一个。
原来这个游戏叫Mastermind。
用Threejs 做3D效果，交互上也可以利用鼠标优势。
等测试好了开源。
Claude for Teachers 上线
查看英文原文
Jul 14, 2026 Product Introducing Claude for Teachers
iOS ChatGPT 应用的远程控制功能挺不错，可以用 Codex 操控电脑，就是应用内的设置说明跟桌面版 ChatGPT 菜单对不上
虽然自己琢磨也能搞明白，不过 Labs 各个地方的文档问题不少，要么过时要么质量差
查看英文原文
The remote feature on the iOS ChatGPT app where you can control your computer via Codex is great but the setup instructions in the app do not match the menus in desktop ChatGPT
Its not hard to figure out, but documentation is bad/out-of-date is so many areas from all the Labs
Its not hard to figure out, but documentation is bad/out-of-date is so many areas from all the Labs
学会用 Runway Agent 制作电影短片，像 Flicker 那样。不用大预算、不用工作室指导，就能简单地把创意变成现实。
查看英文原文
Learn how to use Runway Agent to create your own cinematic short films, like Flicker. No big budgets, no studio notes, just an easy way to bring your ideas into the world.
与 @matanSF 的直觉一致，未来 12 个月内，90% 的 token 可能会流向开源模型。
我们看到越来越多的公司使用前沿 API 进行实验，用开源或私有模型处理生产环节和更大的工作负载！
查看英文原文
Aligned with
@matanSF
's intuition that in the next 12 months, 90% of tokens could be going to open models.
We're seeing more and more companies using frontier APIs for experimenting and open-source or private models for production and larger workloads!
@matanSF
's intuition that in the next 12 months, 90% of tokens could be going to open models.
We're seeing more and more companies using frontier APIs for experimenting and open-source or private models for production and larger workloads!
所以 Codex 现在搞到一个小机器人了？（不过没有 Claw'd crab 那么萌）
查看英文原文
So I guess Codex has a little robot now?
(It's not as cute as the Claw'd crab)
(It's not as cute as the Claw'd crab)
连环推 ×4我们正式推出 Claude for Teachers：美国 K-12 认证教师可以免费使用高级 Claude 功能，还有教学技能库和直接连接到 50 个州学术标准的循证课程。
claude.com/solutions/teacher…
跟 Claude 要个教学计划，它会通过 Learning Commons 连接你所在州的教学标准和优质课程，然后生成一份你可以修改后直接带进课堂的教案和学生材料。
Claude for Teachers 是专为 K-12 隐私设计的。我们不会用你的对话数据来训练模型，学生信息受 FERPA 标准的数据处理协议保护。
点击了解 Claude for Teachers 的工作原理：anthropic.com/news/claude-fo…
查看英文原文
We're introducing Claude for Teachers: free access to premium Claude capabilities for verified K-12 educators in the US, with a library of teaching skills and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states.
claude.com/solutions/teacher…
Ask for a lesson plan, and Claude starts from your state standards and high-quality curricula by connecting through Learning Commons. It then drafts a plan and student-facing materials you can revise and take into class.
Claude for Teachers is built for K-12 privacy. We never train our models on your conversations, and student information is protected by a data processing agreement written to comply with FERPA.
Read more about how Claude for Teachers works:
anthropic.com/news/claude-fo…
claude.com/solutions/teacher…
Ask for a lesson plan, and Claude starts from your state standards and high-quality curricula by connecting through Learning Commons. It then drafts a plan and student-facing materials you can revise and take into class.
Claude for Teachers is built for K-12 privacy. We never train our models on your conversations, and student information is protected by a data processing agreement written to comply with FERPA.
Read more about how Claude for Teachers works:
anthropic.com/news/claude-fo…
一个月前，Fountain 0 在翠贝卡电影节上创造历史，首映了《Dreams of Violets》，这是第一部被主流电影节接受的 AI 故事片。现在他们回来了——由 Ash Koosha 执导、由 Kling AI 驱动的最新故事片《ODYSEEUS: The Fall》的预告片刚刚发布！
向 Fountain 0 团队致以热烈祝贺
@Fountain0Zero
——Ash Koosha
@AshKoosha
、Pooya Koosha
@pooya_koosha
和 Tom Rogers。期待今年晚些时候的完整版本！
查看英文原文
A month ago, Fountain 0 made history at Tribeca with the premiere of Dreams of Violets, the first-ever AI feature film accepted into a major film festival.
Now they're back — the trailer for their latest feature film ODYSEEUS: The Fall just dropped, directed by Ash Koosha and powered by Kling AI!
Huge congrats to the team at Fountain 0
@Fountain0Zero
— Ash Koosha
@AshKoosha
, Pooya Koosha
@pooya_koosha
, and Tom Rogers. Can't wait for the full release later this year!
Now they're back — the trailer for their latest feature film ODYSEEUS: The Fall just dropped, directed by Ash Koosha and powered by Kling AI!
Huge congrats to the team at Fountain 0
@Fountain0Zero
— Ash Koosha
@AshKoosha
, Pooya Koosha
@pooya_koosha
, and Tom Rogers. Can't wait for the full release later this year!
这是 demis 提出的深思熟虑的提议：
查看英文原文
this is a thoughtful proposal from demis:
连环推 ×2今天在班加罗尔 Google I/O Connect India 期间，跟 Google Cloud 高级总监兼首席传道官 @rseroter 有过一次很棒的交流。我们谈了他怎么在日常工作中用 Gemini、Google 如何应对 AI 带来的巨大计算需求、为什么定义目标比写完美规范更重要，还有 Gemini 的后续计划。
还有幸认识了
@vadiamit
和
@chauhan_nilay16
，来自 Google DeepMind 团队。
和这些正在开发我们每天使用的 AI 技术的人聊天真的很有意思。期待以后能有更多这样的对话。
查看英文原文
Had a great conversation with
@rseroter
, Senior Director and Chief Evangelist at Google Cloud, during Google I/O Connect India in Bangalore today.
We spoke about how he uses Gemini in his day to day work, how Google is handling the massive compute demand for AI, why defining goals matters more than writing perfect specs, and what's next for Gemini.
Also had the chance to meet
@vadiamit
and
@chauhan_nilay16
from the Google DeepMind team.
It was great talking with the people building some of the AI technologies we use every day. Looking forward to more conversations like these.
@rseroter
, Senior Director and Chief Evangelist at Google Cloud, during Google I/O Connect India in Bangalore today.
We spoke about how he uses Gemini in his day to day work, how Google is handling the massive compute demand for AI, why defining goals matters more than writing perfect specs, and what's next for Gemini.
Also had the chance to meet
@vadiamit
and
@chauhan_nilay16
from the Google DeepMind team.
It was great talking with the people building some of the AI technologies we use every day. Looking forward to more conversations like these.
严重怀疑 ChatGPT （Codex）的客户端现在已经由AI接管了
Codex 应该是自己在升级自己
自己发现问题和收到反馈进行自我迭代修复和升级，然后发布
因为每隔几小时它就会有新版本提醒你更新...
Google Voice强势回归。
Gemini现在可以录制通话、转录、总结关键点、提取行动项。
告别通话记笔记了。
查看英文原文
Google Voice is so back.
Gemini can now record calls, transcribe them, summarize key points, and pull action items.
RIP taking notes during calls.
Gemini can now record calls, transcribe them, summarize key points, and pull action items.
RIP taking notes during calls.
看看这个官宣
workspace.google.com/blog/pr…
查看英文原文
Read the announcement
workspace.google.com/blog/pr…
workspace.google.com/blog/pr…
今天科技圈热点：
- Meta 的 Louisiana 数据中心投入达50B
- 中国首次用网成功回收火箭
- 欧盟出手限制儿童使用社交媒体
- 这个可穿戴设备能在皮肤上绘画
- 其他科技新闻速报
查看英文原文
Top stories in tech today:
- Meta’s Louisiana data center hits $50B
- China lands its first reusable rocket — in a net
- EU moves to get kids off social media
- This wearable paints onto your skin
- Quick hits on other tech news
- Meta’s Louisiana data center hits $50B
- China lands its first reusable rocket — in a net
- EU moves to get kids off social media
- This wearable paints onto your skin
- Quick hits on other tech news
又有人对 Anthropic 开枪了。问题是他们什么时候反击？
引用 Sam Altman @samaGPT-5.6 sol 价格是 Fable 的一半，令牌效率提高两倍，成本仅四分之一查看被引原帖 ↗
查看英文原文
another shot fired at Anthropic.
The question is: when will they return fire?
The question is: when will they return fire?
这放不久前能单独开个公司了
引用 Thijs @cdngdev我给5.6 sol访问相机胶卷权限，让它从照片提取所有衣服，用gpt-image生成新搭配效果。看到整个衣柜集合在一起挺酷的。查看被引原帖 ↗
查看英文原文
this would have been a whole startup not too long ago
GPT-5.6 sol 比 fable 便宜一半，token 效率高两倍，完成同样的任务。很高兴能以四分之一的价格提供。
查看英文原文
GPT-5.6 sol is half the price and ~twice as token efficient as fable in many cases for accomplishing the same task.
happy to deliver at one-quarter of the price.
happy to deliver at one-quarter of the price.
Sonnet 层级的模型早就淘汰了，有了 Kimi-K3 以后，我预计 Opus 层级也会过时。Frontier Labs 现在得赶快开发 10T 模型。
引用 Lisan al Gaib @scaling01看起来我们今天真的会得到 Kimi-K3查看被引原帖 ↗
查看英文原文
The Sonnet tier of models was already dead before, with Kimi-K3 I expect the Opus tier to become obsolete too
Frontier Labs have to go to 10T models now
Frontier Labs have to go to 10T models now
或者我们应该确保人类最重要的技术不被仅仅4个人掌控？那就是推动开放科学和开源 AI，让能力、权力和财富得以分配。
引用 Mike Dudas @mdudas控制人类历史上最重要技术的4个人中有3个相互对立不断冲突——祈祷demis先生拯救我们查看被引原帖 ↗
查看英文原文
or maybe let's make sure the most important technology in the history of humanity is not controled by just 4 men? Aka let's push for open science & open-source AI to distribute capabilities, power and wealth!
我们的 agentic 产品（codex 和 chatgpt work）这周用量翻了 2.5 倍！欢迎！
查看英文原文
2.5x increase in usage of our agentic products (codex and chatgpt work) in the last week!
welcome.
welcome.
DeepSeek 准备今年 IPO
引用 Watcher.Guru @WatcherGuru中国AI公司 DeepSeek 准备最早今年提交IPO申请查看被引原帖 ↗
查看英文原文
DeepSeek is preparing to file an IPO this year
用 Claude 应用的话，可以选 Home 或 Code。选了 Home，还能在 Chat 和 Cowork 之间选择。
OpenAI 应用的话，可以在 ChatGPT Work 和 Codex 之间选。Chat 在侧边菜单。
两个应用在网站上设计完全不一样
真够直观的！
查看英文原文
If you are use the Claude everything app, you pick between Home and Code. If you pick Home you get to pick between Chat & Cowork
If you use the OpenAI everything app, you pick between ChatGPT Work & Codex. Chat is in a side menu.
Both are different on their websites
intuitive!
If you use the OpenAI everything app, you pick between ChatGPT Work & Codex. Chat is in a side menu.
Both are different on their websites
intuitive!
DeepMind CEO ：前沿模型需提前 30 天交给一个专门机构进行体检再发布 ，则不准进入美国市场
他认为：
通用人工智能（AGI）大概率几年就要实现
它的影响堪比人类发现“火”和“电”
AGI 带来的冲击可能是工业革命的 10 倍速度和 10 倍规模
目前全球正陷入一场极其激烈的商业和地缘政治竞赛
为了不让技术失控，他提议美国带头建立一个动态、灵活且严格的“标准制定机构
怎么管理：达到一定跑分门槛的 AI 模型叫“前沿级模型”，研发它的公司叫“前沿实验室”。这些实验室发布新模型前，得提前 30 天自愿把模型交给这个机构体检。
体检考什么：
考网络安全、生物武器威胁等高危防御。
防“AI 叛变”：专门测试 AI 会不会绕过安全护栏，或者有没有“欺骗人类”的迹象。
强制要求：AI 生成的图片必须加数字水印，AI 的思考过程必须生成人类能看懂的符号。
通过检测的才能进入美国市场：
前沿模型必须通过评估，才能在美国市场部署。
若事态足够严重，框架可再升级，包括协调前沿实验室之间放慢开发发布节奏，甚至停止开发...
连环推 ×3一边把AI当成无限遐想的主题，一边又用AI当写手批量生成这种遐想，这简直就是思想类文章的黄金时代啊。
Fable，把我的推文转成一篇评论小文（这个挺好笑）：
从来没有任何一个时代比现在更适合对人工智能发表意见。我这么说是有一定权威性的，因为我此刻就正在发表一个意见。
想想我们现在面对的这个完美风暴。一边是 AI 成了绝佳的扎口话题：重量级到足够重要，模糊到没人能说绝对是错的，进化快到上个月的见解还没被现实检验就已经被安安全全地遗忘了。另一边是 AI 成了发言人等级的制造者：永不疲倦、文笔流畅，先等个作者打开空白文档开始忧郁这段功夫就能刷出一千字的“GPT-5 对民主意味着什么”。
换句话说，我们已经达到了评论小文奇点。主题与工具融为一体了。蛇不仅是在吃自己尾巴，它还出版了一篇 1200 字的回文，讲尾巴这事对蛇的未来意味着什么。
无限观点的经济学
传统铁杆市场靠稀缺性运作。就那么几个财经社调作者，那么几栏目英寸，那么几个小时让一个职业有意见者把焦虑转化成报价。这种稀缺性强加了一种质量的底子——不算很硬的那个"底"，但有底。为此总得有人在乎到非得提笔才行。
AI 移开了这个地板。观点产生的边际成本现在趋近于零，这意味着关于AI的意见供应曲线一下子垂直了，精确赶上对AI意见的需求也同时飙至高峰。经济学家可能会称之为一个按无限容量出清的市场——而我们普通人都特么管这个叫 LinkedIn。
但真正让这个时代特别诡异的、不仅是因为嘈杂程度的理由是这个：AI是最适合写 AI 的，因为说人写 AI 的东西是这个体裁里唯一作者更懂对象的题材了。语言模型评论语言模型，这在某种程度上，TA是唯一跟这事儿有一线经验的作者了。AI写的 AI 立读短文，同时是我们这个时代最高明的仿制品和最正经的真实材料。没人知道怎么看待这个事实，所以我们一般都避而不谈。
四着永远批发的风格
每个人的 AI评论，人写的按计算机写的都好，绕成了四项排，我这里常规拆解以下四个位置（这暗示一下得很严肃那种）：
一切都会改变。（乐观派、收入很好、频繁成为主旨发言。）
一切毫无变化。（反对派得终身教职却没叫话 写开数据时偷偷用它。）
一切变化是个坏信号。（有点求祷的人、劳累着身养活着最后会某个对的章。）
真正明显的光式争论本来来自于设定体裁。（Meta、无法忍受的人说了关于我所见到想法）。
神奇的是光考虑同一个现象来调优这四个可能，都一样雄辩、可复制、是结构里能无限复制。某个排杆成绩、关于张对话、关篇辞职都对你问着，罗淑回答那块题目在那种你能把它往那边的倾斜：打字本来、黑甲板两面向任意点。AI写作工具省去了人工思路这种使用状态重复功能：大清早你一两个个都用同一个多热点推出这样的答复还验完了客户交互对比更好。
遗失的是什么
简单说法停留这里、以嘲讽感结束简单方便，因嘲讽再非金属组成框架的那个。但这还真诚到要去坦陈那件大概不令人宁堪事在如海中淹毁某个真家伙。
沉在无数做出的段子里、只要一缕真在思考方向前还做跑事。这些人自己检验模拟效果验在文里被某个做的阅读中测用与、原技能遇点本来变着职业样人们交流着--工动都没正形态过来之前，这些已经见到痕迹了。这东西此前一直都少有。现了只它贴标更加难，本粗面上、很容易看出看不出模成表面伪装已和思路机器输出摆饰相似加天沟暗了的层次（举例："secure but hedges"--那反常识角度句式、一个段立以让了开你的对间过渡例子。
你前面之间，你可见个这般暗示，返方向的双边整位策略巧包覆着外显难缠难解技术组合路。
以前表达运以而呈本：长的东化西化秀技长即感式带"数字"段的字--没法点求原来下了是能取量的心态之云顿。重题事会深刻的问题是这根本拉替原理甚至别人不能用探测器或规定记录文件来完成答全不是版权或书作属上搞进。一点拿人毫无质量的拼装段"Z稿" AI输出框投那出溜糊成等量的减量；一份被输出够足量流数据--无论谁是赋篇的作者写推是价值同等必要。
这些AI生成的评论真是把互联网关掉的好理由。
查看英文原文
Between having AI as a subject for infinite speculation and using AI as a writer to produce such speculation at scale, it is the Golden Age of thinkpieces.
Fable, turn my tweet into a thinkpiece (this was pretty funny):
There has never been a better time to have opinions about artificial intelligence. I say this with some authority, because I am currently having one.
Consider the perfect storm we find ourselves in. On one side, AI has become the ideal subject for speculation: consequential enough to matter, ambiguous enough that no one can be definitively wrong, and moving fast enough that last month's take is safely forgotten before it can be checked against reality. On the other side, AI has become the ideal producer of speculation: tireless, fluent, and capable of generating a thousand words on "What GPT-5 Means for Democracy" in the time it takes a human columnist to open a blank document and feel bad about it.
We have, in other words, achieved thinkpiece singularity. The subject and the instrument have merged. The snake is not merely eating its tail; it is publishing a 1,200-word reflection on what tail-eating means for the future of snakes.
The Economics of Infinite Speculation
The traditional thinkpiece economy ran on scarcity. There were only so many columnists, only so many column inches, only so many hours in which a professionally opinionated person could convert anxiety into prose. This scarcity imposed a kind of quality floor — not a high one, but a floor. Someone had to care enough to write the thing.
AI removes the floor. The marginal cost of a take is now approximately zero, which means the supply curve for opinions about AI has gone vertical at the exact moment demand for opinions about AI has also gone vertical. Economists would call this a market clearing at infinite volume. The rest of us call it LinkedIn.
But here is the twist that makes this era genuinely strange rather than merely noisy: AI is uniquely suited to writing about AI because AI writing about AI is the one genre where the author has genuine subject-matter proximity. When a language model speculates about what language models mean, it is, in some sense, the only writer with firsthand experience. The thinkpiece about AI written by AI is simultaneously the most derivative and the most authentic document of our time. No one knows what to do with this fact, so we mostly don't mention it.
The Four Immortal Takes
Every AI thinkpiece, human or machine authored, converges on one of four positions, which I will now enumerate in the traditional listicle format that signals seriousness:
It changes everything. (Bullish, well-compensated, frequently keynoting.)
It changes nothing. (Contrarian, tenured, secretly using it to grade papers.)
It changes everything, and that's bad. (Prophetic, exhausted, right about something eventually.)
The real change is the discourse itself. (Meta, insufferable, this essay.)
The remarkable thing is that all four positions can be argued persuasively with the same evidence, which is why the genre is infinitely renewable. A benchmark result, a chatbot transcript, a layoff announcement — each is a Rorschach blot that faithfully returns whatever thesis you brought to it. AI writing tools have simply industrialized the Rorschach reading. You can now generate all four takes from the same news event before breakfast and A/B test which one performs.
What Gets Lost
It would be easy to end here, on the ironic note, because irony is the load-bearing wall of the genre. But the honest version of this essay has to admit something less comfortable: the flood is drowning something real.
Buried in the infinite speculation is a small amount of actual thinking — people running experiments, reading papers, talking to workers whose jobs are changing, noticing things before they become narratives. That work was always rare. It is now rare and camouflaged, indistinguishable at a glance from its synthetic imitators, because the imitators have gotten very good at the surface features of insight: the confident hedge, the counterintuitive framing, the paragraph that begins "But here's the twist."
(You may have noticed one of those a few paragraphs ago. The camouflage works in both directions.)
The old signals of effort — length, fluency, structure, the presence of a numbered list — no longer signal effort. This is a genuine epistemic problem, and it will not be solved by detection tools or disclosure norms, because the issue was never really authorship. A lazy human take and a fluent machine take are equally worthless; a rigorous piece of thinking is valuable regardless of how many silicon collaborators touched it. The question was always whether anyone checked anything, and that question was hard to answer before the machines showed up.
The Useful Version of the Golden Age
So here is the optimistic reading, offered with the mandatory caveat that optimistic readings are position #1 and I have already mocked position #1.
Golden ages of cheap production have historically been terrible for average quality and excellent for the tails. The printing press produced mountains of pamphlet garbage and also the scientific revolution. The blogosphere produced infinite noise and also a handful of writers who could never have gotten past a magazine's gatekeepers. The pattern is consistent: when publishing costs collapse, the middle of the distribution rots and the extremes flourish.
If that pattern holds, the AI thinkpiece deluge is not the end of thinking about AI — it is the end of mediocre thinking about AI having any value at all. The take that could have been generated in eight seconds is now worth exactly what it costs to generate. What remains scarce, and therefore newly valuable, is everything a language model cannot do on its own: the experiment actually run, the data actually gathered, the prediction actually registered and later scored, the sentence that could only have been written by someone who was in the room.
The Golden Age of thinkpieces, in other words, may turn out to be a gold rush in the classic sense: most of what glitters is pyrite, the real money is in selling shovels, and the durable fortunes go to whoever figures out what the territory is actually worth once the prospectors move on.
Until then, the speculation will continue, at scale, about scale. Somewhere right now, an AI is drafting a response to this essay, and another AI is summarizing that response for a newsletter, and a human is skimming the summary and forming an opinion they will describe as their own. It is the most efficient opinion supply chain ever constructed.
Whether anything is being thought is, appropriately, a matter of speculation.
These AI-driven comments are a good reason to shut down the web.
Fable, turn my tweet into a thinkpiece (this was pretty funny):
There has never been a better time to have opinions about artificial intelligence. I say this with some authority, because I am currently having one.
Consider the perfect storm we find ourselves in. On one side, AI has become the ideal subject for speculation: consequential enough to matter, ambiguous enough that no one can be definitively wrong, and moving fast enough that last month's take is safely forgotten before it can be checked against reality. On the other side, AI has become the ideal producer of speculation: tireless, fluent, and capable of generating a thousand words on "What GPT-5 Means for Democracy" in the time it takes a human columnist to open a blank document and feel bad about it.
We have, in other words, achieved thinkpiece singularity. The subject and the instrument have merged. The snake is not merely eating its tail; it is publishing a 1,200-word reflection on what tail-eating means for the future of snakes.
The Economics of Infinite Speculation
The traditional thinkpiece economy ran on scarcity. There were only so many columnists, only so many column inches, only so many hours in which a professionally opinionated person could convert anxiety into prose. This scarcity imposed a kind of quality floor — not a high one, but a floor. Someone had to care enough to write the thing.
AI removes the floor. The marginal cost of a take is now approximately zero, which means the supply curve for opinions about AI has gone vertical at the exact moment demand for opinions about AI has also gone vertical. Economists would call this a market clearing at infinite volume. The rest of us call it LinkedIn.
But here is the twist that makes this era genuinely strange rather than merely noisy: AI is uniquely suited to writing about AI because AI writing about AI is the one genre where the author has genuine subject-matter proximity. When a language model speculates about what language models mean, it is, in some sense, the only writer with firsthand experience. The thinkpiece about AI written by AI is simultaneously the most derivative and the most authentic document of our time. No one knows what to do with this fact, so we mostly don't mention it.
The Four Immortal Takes
Every AI thinkpiece, human or machine authored, converges on one of four positions, which I will now enumerate in the traditional listicle format that signals seriousness:
It changes everything. (Bullish, well-compensated, frequently keynoting.)
It changes nothing. (Contrarian, tenured, secretly using it to grade papers.)
It changes everything, and that's bad. (Prophetic, exhausted, right about something eventually.)
The real change is the discourse itself. (Meta, insufferable, this essay.)
The remarkable thing is that all four positions can be argued persuasively with the same evidence, which is why the genre is infinitely renewable. A benchmark result, a chatbot transcript, a layoff announcement — each is a Rorschach blot that faithfully returns whatever thesis you brought to it. AI writing tools have simply industrialized the Rorschach reading. You can now generate all four takes from the same news event before breakfast and A/B test which one performs.
What Gets Lost
It would be easy to end here, on the ironic note, because irony is the load-bearing wall of the genre. But the honest version of this essay has to admit something less comfortable: the flood is drowning something real.
Buried in the infinite speculation is a small amount of actual thinking — people running experiments, reading papers, talking to workers whose jobs are changing, noticing things before they become narratives. That work was always rare. It is now rare and camouflaged, indistinguishable at a glance from its synthetic imitators, because the imitators have gotten very good at the surface features of insight: the confident hedge, the counterintuitive framing, the paragraph that begins "But here's the twist."
(You may have noticed one of those a few paragraphs ago. The camouflage works in both directions.)
The old signals of effort — length, fluency, structure, the presence of a numbered list — no longer signal effort. This is a genuine epistemic problem, and it will not be solved by detection tools or disclosure norms, because the issue was never really authorship. A lazy human take and a fluent machine take are equally worthless; a rigorous piece of thinking is valuable regardless of how many silicon collaborators touched it. The question was always whether anyone checked anything, and that question was hard to answer before the machines showed up.
The Useful Version of the Golden Age
So here is the optimistic reading, offered with the mandatory caveat that optimistic readings are position #1 and I have already mocked position #1.
Golden ages of cheap production have historically been terrible for average quality and excellent for the tails. The printing press produced mountains of pamphlet garbage and also the scientific revolution. The blogosphere produced infinite noise and also a handful of writers who could never have gotten past a magazine's gatekeepers. The pattern is consistent: when publishing costs collapse, the middle of the distribution rots and the extremes flourish.
If that pattern holds, the AI thinkpiece deluge is not the end of thinking about AI — it is the end of mediocre thinking about AI having any value at all. The take that could have been generated in eight seconds is now worth exactly what it costs to generate. What remains scarce, and therefore newly valuable, is everything a language model cannot do on its own: the experiment actually run, the data actually gathered, the prediction actually registered and later scored, the sentence that could only have been written by someone who was in the room.
The Golden Age of thinkpieces, in other words, may turn out to be a gold rush in the classic sense: most of what glitters is pyrite, the real money is in selling shovels, and the durable fortunes go to whoever figures out what the territory is actually worth once the prospectors move on.
Until then, the speculation will continue, at scale, about scale. Somewhere right now, an AI is drafting a response to this essay, and another AI is summarizing that response for a newsletter, and a human is skimming the summary and forming an opinion they will describe as their own. It is the most efficient opinion supply chain ever constructed.
Whether anything is being thought is, appropriately, a matter of speculation.
These AI-driven comments are a good reason to shut down the web.
为什么很多国产模型API的 temperature 不允许自定义了？
是因为蒸馏的原因吗？
Anthropic向加拿大AI研究投入1000万美元
查看英文原文
Jul 14, 2026 Announcements Anthropic commits $10 million to Canadian AI research
我们投入1000万加元，与加拿大顶尖AI机构建立合作伙伴关系，支持新的AI研究。
anthropic.com/news/canadian-…
查看英文原文
We’re committing $10 million CAD and partnering with leading AI institutions in Canada to help fund new AI research.
anthropic.com/news/canadian-…
anthropic.com/news/canadian-…
这个算法更新后
互动量上去了
这个以前本来就有的，后来算法调整，Grok接管的时候应该是给去掉了...
总算反应过来了，就是互关的人的信息反而刷不到，有点匪夷所思，之前微博就是这么把自己搞死的
引用 Nikita Bier @nikitabier推出算法调整，提高互相关注的人对你帖子的可见性。该数据缺失导致朋友在回复中显示不足，讨论区像战场。新调整应更容易形成兴趣集群。查看被引原帖 ↗
你们都强力推荐@SMB_Attorney作为初创律师，合作下来还不错
这哥们人特别好相处，幽默随和，效率又高
最关键的是他特别支持AI，我干脆就先用Claude检查文件、整理笔记，然后发给他审核，这样我们双方效率都快多了
如果你们需要法律帮助，我强烈推荐他😊
引用 SMB Attorney @SMB_AttorneySMB Law Group 创始人兼管理合伙人，4 年完成近 400 笔并购交易，交易额约 20 亿美元，为中端市场提供精准法律服务。查看被引原帖 ↗
查看英文原文
All you guys almost unanimously recommended
@SMB_Attorney
when I asked for a startup lawyer ~6 months ago
A really really cool guy to work with, very funny, informal, nice and fast
He's also very pro-AI which I like so I'd check documents first with Claude quickly, let it write up some notes and send to him and then he'd review that speeding both of us up
If you have some legal stuff you need I can recommend him 😊
@SMB_Attorney
when I asked for a startup lawyer ~6 months ago
A really really cool guy to work with, very funny, informal, nice and fast
He's also very pro-AI which I like so I'd check documents first with Claude quickly, let it write up some notes and send to him and then he'd review that speeding both of us up
If you have some legal stuff you need I can recommend him 😊
连环推 ×2我不认同"GPU相比大脑效率低得不行"这个说法。芯片的主要问题在于内存和带宽。从FLOPs来说，大脑每瓦特只高5-10倍效率，但每瓦特的带宽差异高达10-1000倍。所以不像有些人说的那样效率低几百倍。我的预测是，AI芯片和LLM在接下来5年内计算效率会超过人脑，15年内内存访问效率也会超越人脑。特别是当我们构建更多3D芯片，通过片上内存最小化数据移动的时候（比如3D Cerebras）。15年内我们会全方位超越人脑。
AGI可能会帮我们更早构建3D芯片，不是2040年左右，更像是2035年。
引用 Lisan al Gaib @scaling01LLM仍然比人脑小得多，我们还有1-3个数量级的扩展空间查看被引原帖 ↗
查看英文原文
I also don't buy the whole "GPUs are insanely inefficient compared to brains" argument
the big problem for chips is memory and bandwidth
in terms of FLOPs the brain is only ~5-10x more efficient per Watt, but bandwidth per Watt is like a ~10-1000x difference
so it's not like it's millions of times more inefficient
and my prediction is that AI chips and LLMs will surpass the human brain in compute efficiency in the next 5 years
and in memory access efficiency within the next 15 years as we build more 3D chips that minimize data movement with memory on chip (like a 3D cerebras)
we will surpass human brains within the next 15 years
AGI will probably help us build 3D chips much earlier than ~2040
more like 2035
the big problem for chips is memory and bandwidth
in terms of FLOPs the brain is only ~5-10x more efficient per Watt, but bandwidth per Watt is like a ~10-1000x difference
so it's not like it's millions of times more inefficient
and my prediction is that AI chips and LLMs will surpass the human brain in compute efficiency in the next 5 years
and in memory access efficiency within the next 15 years as we build more 3D chips that minimize data movement with memory on chip (like a 3D cerebras)
we will surpass human brains within the next 15 years
AGI will probably help us build 3D chips much earlier than ~2040
more like 2035
来了：Kimi K3明天发布。
折扣方案如下：
Kimi推出限时K3充值活动。7月15日到8月11日，根据单笔充值金额赠送额度：
¥99–¥499：10%奖励
¥500–¥1,999：20%奖励
¥2,000–¥4,999：25%奖励
¥5,000及以上：30%奖励
¥5,000充值会额外送¥1,500额度。
引用 leo 🐾 @synthwavedd速报：Kimi K3 明天推出，充值享折扣查看被引原帖 ↗
查看英文原文
Lets go: Kimi K3 is launching tomorrow.
The discount plan translates into the following:
Kimi is launching a limited-time K3 top-up promotion.From July 15 to August 11, users receive bonus credits based on the amount added in a single top-up:¥99–¥499: 10% bonus
¥500–¥1,999: 20% bonus
¥2,000–¥4,999: 25% bonus
¥5,000 or more: 30% bonus
So a ¥5,000 top-up would come with an additional ¥1,500 in credits.
The discount plan translates into the following:
Kimi is launching a limited-time K3 top-up promotion.From July 15 to August 11, users receive bonus credits based on the amount added in a single top-up:¥99–¥499: 10% bonus
¥500–¥1,999: 20% bonus
¥2,000–¥4,999: 25% bonus
¥5,000 or more: 30% bonus
So a ¥5,000 top-up would come with an additional ¥1,500 in credits.
Lisan al Gaib看起来我们今天真的拿到Kimi-K3了Lisan al GaibPeter 回归了🚨 AI News | TestingCatalogKivine，@arena 上的一个新模型，可能是即将推出的 Kimi K3。Chubby♨️看起来今天要有大惊喜和大发布了。非常期待 K3 release。Chubby♨️对 Kimi K3 太期待了，（可能）今天就要来了。100万token上下文，更好的长链任务表现。让我们看看它对…Chubby♨️Kimi K3 可能明天就要发布了，接在今年最强开源模型之一的身后。我们来回顾一下背景：🚨 AI News | TestingCatalogKimi K3 官方放出预告了 👀Chubby♨️Kimi K3要来了。很可能是今天。顺便说句，这段视频太炸了。AshutoshShrivastavaKimi K3 即将给很多人带来惊喜。歸藏(guizang.ai)Kimi K3 开始预热
关于world models、当前方法、3D对比视频等内容相当不错的总结。推荐阅读
引用 Ars Technica @arstechnica模拟一切（某种程度）：世界模型的前景与局限查看被引原帖 ↗
查看英文原文
Pretty good summary of world models, current approaches, 3D vs video and more. Worth reading
我们目前还没有超10T参数量的模型公开可用并广泛部署
即便Fable可能也不超4T参数量（虽然是从10T模型蒸馏出来的），其他大部分模型都在2T以下
仅新皮层就有约164万亿个突触
我们也知道小孩子的突触多得多，整个大脑估计值达到低四千万亿级，但这些都特别不确定
不知道为啥没人解剖过婴儿大脑
相比之下，可能还差1.5到2.5个数量级（从约4T的Fable开始），算上10T的Mythos也就1-2个数量级
查看英文原文
we still don't have 10T models that are publicly available / broadly deployed
even Fable is likely no more than 4T params (although distilled from a 10T model), most other models are below 2T
the neocortex alone has ~164 trillion synapses
we also know young children have much more synapses, with estimates for the whole brain going up into the low quadrillions, but these are super uncertain
for some reason no one has cut up a baby brains
it's probably closer to 1.5 to 2.5 OOMs (starting from ~4T Fable), with 10T Mythos it's 1-2 OOMs
even Fable is likely no more than 4T params (although distilled from a 10T model), most other models are below 2T
the neocortex alone has ~164 trillion synapses
we also know young children have much more synapses, with estimates for the whole brain going up into the low quadrillions, but these are super uncertain
for some reason no one has cut up a baby brains
it's probably closer to 1.5 to 2.5 OOMs (starting from ~4T Fable), with 10T Mythos it's 1-2 OOMs
Cola 全新改版，隆重登场！
在过去的一个月，我们从头开始重新思考 Cola 的最佳交互形态。
我们选择从千篇一律的 Chat 中抽离出来，化繁为简，把 Cola 做成一个沉浸式的桌面 AI 搭档。
让你打开 Cola 时，我们希望你能自然地沉浸于此，尽情创造。
或者哪怕不想工作，只是在这里呆着，看着窗外的风景，听听风声鸟鸣。
从清晨到傍晚，从晴天到雨天。
在这个沉浸式的 Cola 桌面里，我们精心打造了这些功能：
1. Chat，以最轻松的方式和拥有长期记忆的 Cola 一起创造，时刻为你的生活提供助力。全新的交付面板，让你的处理文件时更为方便。
2. Coding，如果你准备好投入到严肃工作之中，那就打开全新的 Coding 吧。专业的界面让你掌控细节，多项目多任务同时进行，成倍提高效率。
3. 技能中心，我们为你精选了我们最推荐使用的 Skills，不管是 PPT、网页，还是设计、视频，都可以一键安装，立即掌握各类技能。精选推荐还会每日更新，帮助你每天进步。
4. 觉知，原来的心迹升级为觉知，整体做了优化，并增加了金句卡片生成功能，而且即将迎来重要更新，敬请期待。
重要的机制变动：
切换模型时不再清空上下文，可自由根据不同任务切换模型。
内置模型即将迎来重要升级，更加超值 Token Plan 也在研发中。
其他的一些改动：
闹钟，功能全面优化，让你更方便地管理和执行定时任务。
原有 Codex 和 Claude Code的支持，改为在 Coding 中调用。
还有更多的数十项优化不再多说，等待你的发现。
Cola 官网海外访问：
Cola.app
国内镜像：
ColaOS.AI
和姚老师一起搭档，这几个月做了很多事，开源了很多项目。
关于 X 冷启动，对于持续有干货输出的人其实不难。
最近见了几个宇宙厂工作的朋友。
一个是产品经理，开始用 AI 出方案，AI 画原型、写代码。
一个是后端研发，开始用 AI 写PRD，对接业务，根据“一句话需求” 开发全套产品...
虽然都拓展了边界，但业术有专攻，验证需求没问题，但打磨还是得靠专业的人吧？
有点赶鸭子上架，为了用 AI 而用 AI 呢。
周额度快用完了，幸好还有 2 个储备重置。但真的很期待下次重置。
GPT-5.6 更新后好多了，尤其是耗能速率，不过说实话还是有点高。
引用 Chubby♨️ @kimmonismus我闻到了另一次重置和（希望）即将到来的银行重置查看被引原帖 ↗
查看英文原文
My weekly almost fully used, but luckily 2 banked resets still available. But really looking forward to another reset.
GPT-5.6 got much better with burn rate after their update but still pretty high imho.
GPT-5.6 got much better with burn rate after their update but still pretty high imho.
大多数"AI 做游戏"的演示其实就是一段每次都重复播放的视频。
PixVerse Game 想往更难的方向走：你描述你想要什么，这个想法被转化成游戏机制和 AI agent，实时视频流就会根据你的意图做出反应。玩家意图、机制、生成的视频反馈，形成一个循环。
这才是我一直期待 AI 视频走向的方向。实时交互式视频是大多数项目都忽视的部分，PixVerse 直接朝这个方向建设。等不及看这会怎么发展了。
引用 PixVerse @PixVerse_游戏世界都是预先设计的，每个房间、角色、结局在等你发现。PixVerse Game 要改变这一点。转发+关注+回复可获300积分和72小时内DM早期访问。查看被引原帖 ↗
查看英文原文
Most "AI makes games" demos are just a clip that plays back the same way every time.
PixVerse Game is chasing the harder version: you describe what you want, that intent gets turned into game mechanics and an AI agent, and a real-time video stream responds to it. Player intent, mechanics, a generated video response, back to you.
This is the direction I've been waiting for AI video to take. Real-time interactive video is the part most projects skip, and PixVerse is building straight at it. Can't wait to see where this goes.
PixVerse Game is chasing the harder version: you describe what you want, that intent gets turned into game mechanics and an AI agent, and a real-time video stream responds to it. Player intent, mechanics, a generated video response, back to you.
This is the direction I've been waiting for AI video to take. Real-time interactive video is the part most projects skip, and PixVerse is building straight at it. Can't wait to see where this goes.
LLMs 仍然比人脑小得多
我们前面还有 1-3 个数量级的规模化空间
查看英文原文
LLMs are still so much smaller than the brain
like we still have 1-3 OOMs of scaling ahead of us
like we still have 1-3 OOMs of scaling ahead of us
谁不想要一个生活中的最佳拍档？
与你同步，一起进化。
更快，更准，更懂你。
她，就是 Cola。
不知道大家是否喜欢这种剪辑，如果喜欢的话我们就公布 skill 给大家。
连环推 ×3完全支持@demishassabis的这项重要提议。现在是我们必须采取行动的时刻。'...在AGI到来之前，我们必须利用这个宝贵的机会为全人类的利益来塑造这项技术。'
我在2023年与@ericschmidt一起提出了'AI的IPCC'这一类似提议 ft.com/content/d84e91d0-ac74…
我提出了与@ianbremmer在2023年夏天提出的AI金融稳定委员会类似的提议。我们还在《外交事务》上发表了相关文章 foreignaffairs.com/world/art…
查看英文原文
Fully support this important proposal from
@demishassabis
. The time for us all to act is now.
"...we must use this precious window before AGI arrives to shape this technology for the benefit of all humanity."
I made a similar proposal for an "IPCC for AI" with
@ericschmidt
in 2023
ft.com/content/d84e91d0-ac74…
And similar proposal for a Financial Stability Board for AI with
@ianbremmer
in summer 2023
And with Ian Bremmer in Foreigns Affairs in
foreignaffairs.com/world/art…
@demishassabis
. The time for us all to act is now.
"...we must use this precious window before AGI arrives to shape this technology for the benefit of all humanity."
I made a similar proposal for an "IPCC for AI" with
@ericschmidt
in 2023
ft.com/content/d84e91d0-ac74…
And similar proposal for a Financial Stability Board for AI with
@ianbremmer
in summer 2023
And with Ian Bremmer in Foreigns Affairs in
foreignaffairs.com/world/art…
智能体时代的AI投资管理
了解企业如何在智能体时代通过衡量每美元的有效工作量、提升效率和扩展高价值工作流来管理AI投资。
查看英文原文
How to manage AI investments in the agentic era
Learn how enterprises can manage AI investments in the agentic era by measuring useful work per dollar, improving efficiency, and scaling high-value workflows.
Learn how enterprises can manage AI investments in the agentic era by measuring useful work per dollar, improving efficiency, and scaling high-value workflows.
数据科学团队如何使用 ChatGPT Work
了解数据科学团队如何利用 ChatGPT Work 构建根因分析简报、影响评估、关键指标备忘录、范围分析和仪表板规范等工作成果。
查看英文原文
How data science teams use ChatGPT Work
See how data science teams can use ChatGPT Work to build root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs.
See how data science teams can use ChatGPT Work to build root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs.
销售团队如何使用 ChatGPT Work
销售团队可以使用 ChatGPT Work 从实际工作数据生成管道简报、会议准备资料、预测评审、账户计划和停滞交易诊断。
查看英文原文
How sales teams use ChatGPT Work
See how sales teams can use ChatGPT Work to create pipeline briefs, meeting prep packets, forecast reviews, account plans, and stalled-deal diagnoses from real work inputs.
See how sales teams can use ChatGPT Work to create pipeline briefs, meeting prep packets, forecast reviews, account plans, and stalled-deal diagnoses from real work inputs.