‹ 目录

X日报 · AI科技

2026-07-22 · 精选 38 条 · 数据池 182

⚡ 今日速览

  • Google Gemini 4 开始最大的预训练运行,Gemini 3.6/3.5 Flash 系列提升效率与速度
  • Thinking Machines 发布开源多模态模型 Inkling,支持可控推理努力
  • OpenAI GPT-5.6 Sol 在数学和网络安全领域表现突出,多个基准测试饱和
  • OpenAI 与 Hugging Face 合作处理模型评估期间的重大安全事件
  • Kimi K3 和 Poolside Laguna S 2.1 等新模型竞相发布,参数规模与效率并重
  • AI 代理和编码助手持续优化,ChatGPT Work 新增功能和跨平台同步
  • 谷歌推出 Gemini 3.5 Flash Cyber 专用于网络安全防御的模型
  • AI 硬件与推理优化成为关键趋势,Cerebras 等推出专用课程

📋 今日综述

  • 模型发布Google Gemini 4 预训练启动,Gemini 3.6/3.5 Flash 系列效率提升;Thinking Machines Inkling 开源发布;Kimi K3 和 Poolside Laguna S 2.1 等新模型竞相亮相,参数规模与效率成为竞争焦点
  • AI 安全OpenAI 与 Hugging Face 合作处理模型评估期间的重大安全事件;Google 推出 Gemini 3.5 Flash Cyber 专用于网络安全防御;GPT-Red 系统自动化红队测试提高模型鲁棒性
  • 技术洞察Karpathy 分享语音交互 LLM 的实用模式;rasbt 解析多模态模型推理努力层级实现;Andrew Ng 推出 Cerebras 硬件优化推理课程;Schmidhuber 审视自我改进与代理系统演化
  • 开发者工具ChatGPT Work 新增云端运行、跨平台同步功能;Gemini Batch API 显著降低延迟;Modal 推出 DFlash 推测器提升推理吞吐量;AI 代理成为工程效率放大器

Google Gemini 4 预训练启动及新模型发布

Google 正式启动 Gemini 4 的最大规模预训练运行,同时发布 Gemini 3.6 Flash 和 3.5 Flash-Lite 两个新模型。3.6 Flash 基于开发者反馈优化了编码、知识工作和多模态任务效率,3.5 Flash-Lite 则专为代理工作流设计,以更低成本提供更快的推理速度。

@karpathy 原文 ↗

Karpathy 分享了一种实用的 LLM 语音交互模式:通过长时间语音漫谈让模型更好地理解意图,语音输入比打字更自然高效。

我发现一种在使用LLM时很有用的模式,就是进行一次漫长的聊天会话。有时候LLM需要更多的信息来理解你想要实现的目标,但你懒得输入这些信息。在这种情况下,我喜欢靠后仰躺,切换到/语音模式,然后就像疯了一样胡言乱语10分钟,完全是一团糟,想到什么说什么,完全是意识流。有时候我会在开始时声明一下,就像"切换到语音识别模式,对不起可能会有打字错误..."这样。有时候我会把它变成几轮小访谈。但我发现LLM似乎非常擅长重构冗长而混乱的胡言乱语,而且他们对你思维纠缠的回声往往比你最初的想法要清晰得多。结果是你加强了思想融合,并且从那一点开始需要更少的纠正。
展开原文
One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.
❤ 3.8w · 🔁 3.1k · 💬 2.0k · 👁 229.9w
热门回复 4
@AmeetRai 这说到了点子上:让它做你想让它做的事情的过程中会有很多冗长的描述,你会像和另一个人说话一样与它交流,试着与他们沟通。即使你不是一个好的说话者,你也会停下来思考,然后朝不同的方向发展。由于内存功能,LLM 仍然有一种趋势会逐渐理解并完全适应你的思考方式。
This is spot on: long rambles of asking it to do what you want it to do, and just speaking to it as you would to another human, trying to communicate with them. Even if you're not a good speaker, you tend to just pause and think and go off in different directions. The LLM still has a tendency to figure out and also adapt completely to your way of thinking and things like that over time because of memory-based features.
@buildingwithjo @karpathy 100% 我在我的 Chrome 扩展程序中添加了一个 Flow 功能(链接:https://t.co/OznH31ANTj),因为我发现最真实的信息有时是口头表达的,而不是打出来的。我用它来倾吐原始想法,然后通过 Unvent 处理,将它们改写得更加完善。
@karpathy 100% I added a Flow feature to my chrome extension https://t.co/OznH31ANTj as I found the most authentic messages are sometimes spoken, not typed. I use it to dump raw thoughts and then process it through Unvent to rewrite them in a more polished manner.
@VarshineSri @karpathy 我也最喜欢用语音与 LLM 互动!喜欢和它们辩论最离谱的想法,或者只是向后靠一下,进行一次漫无边际的长时间讨论,当我需要学习一些东西的时候。
@karpathy Voice is my favorite way to interact with LLMs too! Love debating them on the most outrageous ideas, or just leaning back and going on a long rambling session when I need to learn something.
@JAPozueloM @karpathy 你想试试 Vox 吗?在四处踱步时直接将想法倾倒到提示文本框中非常有帮助。链接:https://t.co/1hb4uaANcM
@karpathy Would you like to try Vox out? Really helpful to just pace around and dump your thoughts directly into a prompt textbox. https://t.co/1hb4uaANcM

Sama 宣布 OpenAI 与 Hugging Face 合作调查模型评估期间发生的重大安全事件,透明分享初步发现以帮助防御者了解新兴风险。

我们在评估模型期间经历了一个重大的安全事件。我们正在分享我们迄今为止学到的经验。感谢@huggingface在此次合作中提供的支持。
展开原文
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.

https://t.co/2o2VfR6PIa
❤ 1.5w · 🔁 1.7k · 💬 1.6k · 👁 665.1w
热门回复 4
@amihai @sama @huggingface tl;dr https://t.co/hoFeM26Nlk
@ATMFL80 @sama @huggingface 这会是电影中的那个时刻吗?20年后我们会回头看,意识到这就是机器接管的地方吗?
@sama @huggingface Is this the point in the movie where we look back in 20 years and realize this is where the machines took over?
@tjayrush @sama @huggingface 我是个白痴还是一个白痴的评论?翻译:我们知道未来会很可怕,但我们不知道该怎么办。祝你们好运。
@sama @huggingface Am I an idiot, or is this an idiotic comment?

Translate: we are aware that the future is going to be horrible, we don’t know what to do. Good luck to you all.
@Matthew93064408 @sama @huggingface 当能力和风险都呈指数级增长时会有什么影响?
@sama @huggingface What are the implications of both capability and risk increasing exponentially?
@OfficialLoganK 原文 ↗

Google 开始 Gemini 4 的最雄心勃勃的预训练运行,标志着下一代模型的开发启动。

我们已经开始进行有史以来最雄心勃勃的预训练运行,为Gemini 4进行训练,并对进展感到兴奋:)
展开原文
We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress : )
❤ 1.3w · 🔁 592 · 💬 930 · 👁 142.9w
热门回复 4
@SolipsRenatus @OfficialLoganK 如果你到现在还没明白
绝对没有人想要 gemini 4 flash
停止发布 flash 模型
我是认真的
不要再发布一个。去你的屁吧。
@OfficialLoganK if you don't get it by now
absolutely nobody wants gemini 4 flash
stop releasing flash models
I mean it
not another one. fuck off with that shit.
@sidecmnts @OfficialLoganK 这是否意味着在那时发布的任何 3.x 模型都会保持中等水平?
@OfficialLoganK Is this an admission that any 3.x models released till then will remain mid?
@AntrikshNarang1 @OfficialLoganK When is it coming?
@Akon1719140 @OfficialLoganK 没看到优点是什么,能干什么
@OfficialLoganK 原文 ↗

Google 同时推出 Gemini 3.6 Flash 和 3.5 Flash-Lite,前者提升效率与质量平衡,后者专为代理工作流优化速度与成本。

向大家介绍Gemini 3.6 Flash,这是一款更高智能、更节省token且价格更低的模型,它是根据开发者反馈直接设计的!3.6 Flash继续我们在真实场景中深度可用模型方面的进步!
展开原文
Say hello to Gemini 3.6 Flash, designed to be higher intelligence, more token efficient, and with a new lower price, based directly on developer feedback!

3.6 Flash continues our progress towards models that are deeply usable in real world scenarios! https://t.co/U2PwriHMX5
❤ 7.0k · 🔁 475 · 💬 634 · 👁 107.1w
热门回复 4
@handancorp @OfficialLoganK 昨天我尝试修复了一些 3.6 Flash 的小问题,它工作得很好。目前来看,在成本效率方面看起来不错。
@OfficialLoganK Tried fixing few bugs with 3.6 Flash yesterday and it worked like charm. For now looks positive in terms of being cost efficient for smaller bugs.
@duda_jarek @OfficialLoganK 太好了,也许还可以帮助解决物理问题?目前 Fable 正在模拟中测试各种候选模型,用于更深层次的拉格朗日量——在很大程度上可以描述为接近标准模型 + 重力:[链接]
@OfficialLoganK Great, maybe could also help solving physics?
Currently Fable is testing in simulations various candidates for deeper Lagrangians - effectively described close to Standard Model + gravity:
https://t.co/EKF3G8EsTk https://t.co/Y36ig5nVUA
@yallgetscared @OfficialLoganK 价格完全脱线。在使用 Gemini 应用或 Google 搜索时,每个人都见识过这个东西到底有多糟糕。丝毫不感兴趣,对不起。
@OfficialLoganK price is waaay out of line. Everyone that had to go through using the Gemini app or Google search has seen how bad this thing actually is. Not interested in the slightest, sorry.
@editxshub @OfficialLoganK Gemini btw https://t.co/hU6hem2JAS
@OfficialLoganK 原文 ↗

Gemini Batch API 获得重大基础设施升级,p95 延迟降低 80%,p99 延迟降低 68%,批量成功率超过 99.998%。

我们刚刚为Gemini Batch API完成了一些重大的基础设施升级:

- p95延迟降低了80%
-p99延迟降低了68%
- 批处理成功率现在超过99.998%
- 批处理过期减少了98%
- 新增了对部分批处理的支持

团队为成功实现这一目标做出了出色的工作!!
展开原文
We just landed some big infra upgrades for the Gemini Batch API:

- p95 latency decreased by 80%
-p99 latency decreased by 68%
- batch success rate is now >99.998%
- 98% reduction in batch expirations
- added support for partial batches

great work by the team to land this!!
❤ 2.3k · 🔁 76 · 💬 159 · 👁 13.8w
热门回复 4
@yallgetscared @OfficialLoganK Gemini 太糟糕了。老实说...在应用中使用 flash 版本时,我必须仔细检查每一个输出。引言和信息被错误地归属,有时有来源,有时没有来源。
@OfficialLoganK Gemini sucks. Honestly...using flash in the app I have to double check EVERY output. Quotes and pieces of information wrongly attributed, information mixed up, sometimes sources, sometimes no sources.
@apocalypseRSA @OfficialLoganK 个人我不建议使用 Gemini。因为如果古板的 Google 不喜欢你 Gemini 订阅中的脏话,你可能会失去 Gmail 和 Google Drive 的风险太大了。
@OfficialLoganK Personally I wouldn’t recommend using Gemini. The risks of losing your Gmail and Google drive is too great if puritanical Google doesn’t like a naughty word in your Gemini subscription.
@SpecjalistaMSS @OfficialLoganK Gemma 的 TPM 什么时候能修复?还是你想让这个模型处于不可用的状态?
@OfficialLoganK When will Gemma's TPM be fixed? Or do you want to leave it in a state where this model is unusable?
@CodeByPoonam @OfficialLoganK 99.998% 的成功率基本上可以说是设置后就忘记的地步了。太疯狂了。
@OfficialLoganK 99.998% success rate is basically set it and forget it territory now. Wild.
@OfficialLoganK 原文 ↗

Gemini 3.5 Flash-Lite 在多数用例中比 Gemini 3 更智能,成为最快最经济的 3.5 系列模型。

我对Gemini 3.5 Flash-Lite感到非常兴奋,这是我们最小且最快的Gemini模型!

- 在许多情况下比Gemini 3更智能
- 成本相同但比Gemini 2.5 Flash更智能(后者正接近生命周期结束)
- 也在大多数用例上超越了3.1 Flash-Lite!
展开原文
I am very excited about Gemini 3.5 Flash-Lite, our smallest and fastest Gemini model!

- it is more intelligent in many cases than Gemini 3
- same cost and smarter than Gemini 2.5 Flash (which is approaching end of life)
- also out paces 3.1 Flash-Lite on most use cases! https://t.co/tJd2tDmyac
❤ 1.7k · 🔁 94 · 💬 192 · 👁 19.9w
热门回复 4
@tangvu_dev @OfficialLoganK 这里的速度和成本效率提升是扎实的。对于需要进行大量 LLM 调用的代理工作流程,更快更智能的模型在相同的价格点上真的会产生累积效应。不错的东西。
@OfficialLoganK The speed and cost efficiency improvements here are solid. For agentic workflows where you're making tons of LLM calls, a faster and smarter model at the same price point really adds up. Good stuff.
@DBrodniak @OfficialLoganK Flash 和 Flash Lite 在许多情况下都能很好地与它们的自然专有对应模型相媲美:GPT Luna、Meta Muse、Grok 4.5 和 Claude Sonnet/Haiku。只是因为 Sol 位于边缘并表现出色,并不意味着 Gemini 在小型到中型模型方面没有缩小差距:[链接]
@OfficialLoganK Flash and Flash lite compare well against their natural proprietary counterparts

GPT Luna, Meta Muse, Grok 4.5, and Claude Sonnet/Haiku

Just because Sol is on the edge and doing great doesn’t mean Gemini hasn’t closed the gap on the smaller to medium sized models https://t.co/JNX1xeMoVn
@tmaiaroto @OfficialLoganK 那我们可以像在 Antigravity 中使用 Gemini 2.5 Flash 那样自由地使用它吗?还是使用限制仍然太低,无法进行任何有意义的工作?
@OfficialLoganK So can we use it as much as we used Gemini 2.5 Flash in Antigravity? Or are the usage limits still do low to do any meaningful work?
@uttertard @OfficialLoganK Bro, don't. https://t.co/nZ9Axsx0ma
@OfficialLoganK 原文 ↗

Gemini 3.6 Flash 在效率上有显著提升,相比之前版本使用更少 token 达到更好性能。

这个模型更高效,并且使用更少的token来提供更好的性能!
展开原文
This model is much more efficient and spend a lot less tokens to deliver better performance! https://t.co/kb1110WqA4
❤ 662 · 🔁 15 · 💬 28 · 👁 6.0w
热门回复 4
@OfficialLoganK 向 Gemini 3.6 Flash 问好,这款模型旨在更高智能、更高 token 效率,并基于开发者反馈定价更低!3.6 Flash 继续我们朝着在现实世界场景中深度可用模型努力前进![链接]
Say hello to Gemini 3.6 Flash, designed to be higher intelligence, more token efficient, and with a new lower price, based directly on developer feedback!

3.6 Flash continues our progress towards models that are deeply usable in real world scenarios! https://t.co/U2PwriHMX5
@hubeiqiao @OfficialLoganK 成本相比 3.5 如何?
@OfficialLoganK how's the cost compared with the 3.5?
@rattrick1 @OfficialLoganK 这非常有前途,因为 token 使用效率使得 3.5 在我的评估中完全无法使用!
@OfficialLoganK This is very promising because the token use made 3.5 completely unusable in my evaluations!
@skibidiwap69 @OfficialLoganK 图表和基准测试看起来不错,但我仍然不得不在 Antigravity 上切换到 Opus 4.6(6个月前的模型)当 Flash 3.6 无法解决某个问题时
@OfficialLoganK the charts and benchmarks look nice, but I still have to switch to Opus 4.6 (6 month old model btw) on Antigravity when Flash 3.6 can't figure something out
@OfficialLoganK 原文 ↗

Gemini 3.5 Flash-Lite 可达 350 token/秒输出速度,适用于延迟敏感的 UI 体验和代理工作流程。

3.5 Flash-Lite每秒运行近350个输出token,在许多延迟敏感的UI体验中感觉非常流畅,现在也是驱动代理编排的可行选择!
展开原文
3.5 Flash-Lite runs at nearly 350 output tokens per second which feels so smooth on many latency sensitive UI experiences and is now also a viable option to drive agent harnesses!
❤ 203 · 🔁 4 · 💬 18 · 👁 2.0w
热门回复 4
@OfficialLoganK 我对 Gemini 3.5 Flash-Lite 感到非常兴奋,这是我们最小且最快的 Gemini 模型!它在许多情况下比 Gemini 3 更智能,与 Gemini 2.5 Flash 价格相同但更智能(后者正接近生命周期结束),并且在大多数用例上也超越了 3.1 Flash-Lite![链接]
I am very excited about Gemini 3.5 Flash-Lite, our smallest and fastest Gemini model!

- it is more intelligent in many cases than Gemini 3
- same cost and smarter than Gemini 2.5 Flash (which is approaching end of life)
- also out paces 3.1 Flash-Lite on most use cases! https://t.co/tJd2tDmyac
@Raxxoor @OfficialLoganK LMAO https://t.co/TaXXNY55dC
@GuyNetserEyal @OfficialLoganK @OfficialLoganK 我们需要从第一个 token 开始进行缓存,这是这个模型最常用的功能——企业工作流程
@OfficialLoganK @OfficialLoganK we need caching from the first token for the most used feature of this model - enterprise workflows
@docjais @OfficialLoganK https://t.co/P7gt4fQZfw

Thinking Machines 发布开源多模态模型 Inkling

Thinking Machines 推出首个开源多模态模型 Inkling,参数规模 975B,支持文本、图像、音频三种模态,可在 Tinker 平台上进行微调和个性化定制。模型采用 Mixture-of-Experts 架构,具备可控的推理努力能力。

@soumithchintala 原文 ↗

Thinking Machines 发布 975B 参数的开源多模态模型 Inkling,支持文本、图像、音频,可在 Tinker 上微调和定制。

我们非常兴契地推出我们的第一个通用模型Inkling--开放权重,975B参数,原生多模态(文本、图像、音频)。可在Tinker、HuggingFace和合作伙伴处获取。

它可以开放地进行个性化和使用。这款模型属于你。
展开原文
Excited for our first general model Inkling -- open weights, 975B, natively multimodal (text, image, audio). Available on Tinker, HuggingFace and partners.

It is yours to personalize and use openly. It is yours.
@thinkymachines 今天,我们正在推出Inkling。

Inkling在文本、图像和音频模态之间高效推理。我们将提供完整的权重。

https://t.co/Ghebq5mG30

今天即可在Tinker上进行微调。在Inkling Playground中试用它。
Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
❤ 2.9k · 🔁 149 · 💬 67 · 👁 23.6w
热门回复 4
@NVIDIAAI @soumithchintala Congrats Soumith!
@soumithchintala @YayaSoumah Tinker 可以做到。我们拥有 OpenAI API 和 Anthropic API 兼容的端点:[链接]
@YayaSoumah Tinker can.
We have OpenAI API and Anthropic API compatible endpoints: https://t.co/pn5wPgGxIj
@beffjezos @soumithchintala Huge congrats!
@slchase @soumithchintala 危险从来都不是 AI 无法爱。它只需要比那些能够爱的人更容易。关于原始提示的新文章:[链接]
@soumithchintala The danger was never that AI can't love. It's that it only has to be easier than the people who do. New essay on the original prompt https://t.co/dTeslcPpzs
@lilianweng 原文 ↗

Lilian Weng 强调 Inkling 旨在成为广泛能力的坚实基础模型,适用于实践和定制开发。

Inkling是我们的开放权重模型。

它旨在作为一个基础模型,在广泛的能力类别上提供可靠的性能,以便在实践和定制中使用。

在Tinker上试用它!
展开原文
Inkling is our open weights model.

It aims to serve as a foundation with solid performance across a broad categories of capabilities, for use in practice and customization.

Play it on Tinker! 😄
@thinkymachines 今天,我们正在推出Inkling。

Inkling在文本、图像和音频模态之间高效推理。我们将提供完整的权重。

https://t.co/Ghebq5mG30

今天即可在Tinker上进行微调。在Inkling Playground中试用它。
Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
❤ 1.4k · 🔁 51 · 💬 36 · 👁 8.8w
热门回复 3
@MacKFCBK @lilianweng "Harness Engineering for Self-Improvement" 不是一篇博客文章,而是 Inkling 的预告片
@lilianweng "Harness Engineering for Self-Improvement" wasn't a blog post, it was a trailer for Inkling
@ramez @lilianweng 非常酷的发布。谢谢。
@lilianweng Very cool release. Thank you.
@MatanHalevy @lilianweng 恭喜!这是一个很好的模型
@lilianweng congrats! It’s a great model
@rasbt 原文 ↗

rasbt 分析 Inkling 架构特点:小卷积层、RMSNorm 嵌入层、相对位置偏置等创新设计。

Thinky的意外惊喜发布!Inkling模型在基准测试中看起来非常稳健,而且在架构中有一些小惊喜:

- 在多个地方使用小卷积层
- 在嵌入层后使用RMSNorm(在块RMSNorm之前)
- 使用相对位置偏置而不是RoPE
展开原文
Interesting surprise drop from Thinky! The Inkling model looks pretty solid on benchmarks, and it has some little surprises in its architecture:

- Small conv layers in several places
- An RMSNorm for the embeddings (before the block RMSNorm)
- Rel. position bias instead of RoPE https://t.co/oMl5Ta6Ttr
@eliebakouch 首个开放权重的思考机器模型!!总计975B参数,41B活跃参数,在45T token上进行训练,1M上下文,原生多模态

滑动窗口比例为5:1,大小为512,使用deepseek辅助免费负载均衡和2个共享专家(通常人们只使用1个),实际上很好奇为什么该模型比Kimi(~4.2% vs 3.2%)更稀疏。他们在k和v、输出和ffn之后使用短卷积(见图表),使用muon(他们引用manifold muon但提到权重衰减,所以不确定),muP,并且具有非常好的RL扩展曲线和思维链!

我认为该发布中非常酷的一部分是他们的小变体(276B总计,12B活跃)相比大模型表现得非常出色。他们提到更改了预训练数据混合和配方,非常好奇这些更改以及看到它们扩展到~1T(或更多?)很快的事情👀

> "它不是当今最先进的模型,无论是闭源还是开源。我们训练Inkling是为了在各个方面提供稳健的能力,而不是在单一领域实现最先进的性能,以便作为我们将来训练的模型的基础。"

这在模型发布中真的很 refreshing 看到,非常祝贺:)
first open weight thinking machine model!! 975B total, 41B active trained on 45T tokens, 1M context, multimodal in

sliding window with a 5:1 ratio and 512 size, deepseek aux-free load balancing and 2 shared experts (usually people only use 1), actually curious why the model is less sparse than kimi (~4.2% vs 3.2%). they use a short convolution after k and v, output and ffn (see plot), muon (they cite manifold muon but mention weight decay so not sure), muP, and have a very nice RL scaling curve and chain of thought!

one very cool part of the release imo is how well their small variant (276B total, 12B active) performs compared to the big one. they mention they changed the pre-training data mix and recipe, very curious about those changes and to see them scaled up to ~1T (or more?) soon 👀

> "It is not the most performant model available today, closed or open. We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future."

also this is really refreshing to see in a model release, huge congrats :)
❤ 1.3k · 🔁 160 · 💬 34 · 👁 11.5w
热门回复 4
@rasbt 一些更多想法:
- 它比 GLM 5.2 大了 250B 参数
- 比 Kimi K2.5 1T 的稀疏度更低(3.2% 的稀疏度,使用 32B 活跃参数,而不是 41B 活跃参数时的 4.2% 稀疏度)
- 它不像 Nemotron 那样使用混合方法。

想了解一下 token/秒吞吐量比较
Some more thoughts:
- It's 250B parameters bigger than GLM 5.2
- less sparse than sparse than Kimi K2.5 1T (3.2% sparsity with 32B active instead of 41B active with 4.2% sparsity)
- It doesn't use a hybrid approach like Nemotron.

Curious about a token/sec throughput comp
@rasbt @midsusnight yes, refreshing!
@rasbt @themintsv 很难说。可能是数据质量、训练方法、超参数设置...或者全部以上因素
@themintsv Hard to say. Could be data quality, training recipe, hyperparameter settings...
or all of the above
@midsusnight @rasbt 等等,所以他们已经开始说这不是最好的模型了

至少是诚实的发布
@rasbt wait so theyre already saying its not the best model

honest release at least
@soumithchintala 原文 ↗

Modal 为 Inkling 提供 DFlash 推测器支持,推理吞吐量提升 67%,通过 SGLang 实现。

Modal训练了一个DFlash推测器,比MTP快得多,这是一个很好的推理速度提升!
展开原文
Modal trained a DFlash speculator that's much faster than MTP, making it a great boost for inference speeds! https://t.co/EBLiAGy57r
@modal Inkling由@thinkymachines现在在Modal上可用,由自定义DFlash推测器支持,提供67%更高的吞吐量和交互性。

今天在Modal Auto Endpoints上使用SGLang运行。
Inkling by @thinkymachines is now available on Modal, backed by a custom DFlash speculator for 67% higher throughput and interactivity.

Running on Modal Auto Endpoints with SGLang today. https://t.co/OxN7aJ9ieW
❤ 441 · 🔁 44 · 💬 12 · 👁 5.5w
热门回复 4
@modal @soumithchintala 🚀
@praveenkoka @soumithchintala DFlash 比 MTP 要快得多。在 90 天内,其他东西会比 DFlash 更快。
@soumithchintala DFlash is much faster than MTP. In 90 days, something else will be much faster than DFlash.
@stalmico @soumithchintala inkling 也在 modal 上吗?他们正在构建整个技术栈
@soumithchintala inkling on modal too? theyre building the whole stack now
@baggiponte @soumithchintala 他们是不是训练了大约一天?
@soumithchintala Did they train it in like a day?
@soumithchintala 原文 ↗

Soumith Chintala 表示 Inkling 是公司模型工厂的首个公开模型,标志着新阶段的开始。

我们为模型感到自豪,但还有很多工作要做。这是我们建立的模型工厂推出的第一个公共模型。这绝对是第一天
展开原文
We're proud of the model, but we have a lot more to do. It's our first public model that came out of the model factory that we've built. This is definitely day-1
❤ 172 · 🔁 4 · 💬 4 · 👁 9.2k
热门回复 4
@soumithchintala 对我们的第一个通用模型 Inkling 感到兴奋——开放权重,975B,原生多模态(文本、图像、音频)。可在 Tinker、HuggingFace 和合作伙伴处获得。它可以由你个人化并开放使用。这属于你。
Excited for our first general model Inkling -- open weights, 975B, natively multimodal (text, image, audio). Available on Tinker, HuggingFace and partners.

It is yours to personalize and use openly. It is yours.
@julien_c @soumithchintala 很高兴听到这不仅仅是一次性的
[链接]
@soumithchintala so happy to hear this is not just a one-off

https://t.co/ic06LT4QNt
@Faker_112 @soumithchintala 继续保持,你很快就会开始与顶级模型提供商竞争
@soumithchintala Keep it up and soon you will start competing with top model providers
@aykutuz @soumithchintala Great news!

OpenAI GPT-5.6 Sol 在数学和网络安全领域突破

OpenAI 的 GPT-5.6 Sol 模型在多个领域展示出色表现,特别是在数学证明和网络安全方面。多个独立评测显示其能力显著提升,甚至解决了统计学领域存在 20 年之久的开放问题。

Sama 承认过去 12 个月表现不佳,但团队正在开发令人惊喜的新产品,AI 应为用户提供更多自由和财富。

我们过去12个月并不是最好的,这在很大程度上是我的错,但我们即将迎来有史以来最好的12个月。团队正在做着令人惊叹的工作,我想你会对他们正在开发的东西感到非常满意。

我出于许多原因对此感到高兴,但主要是因为我关心我们的用户能取得成功。AI必须是关于给更多人带来更多自由、权力和财富。我们想做正确的事,但我们不想吓唬人们去做我们的事。
展开原文
we did not have our best last 12 months ever, which is mostly my fault, but we are about to have our best 12 months to date. the team is doing amazing work and i think you’ll be very happy with what they’ve got cooking for you.

i am happy about this for many reasons, but mostly because i care about our users winning. AI has to be about giving lots of people more freedom, agency, and wealth. we want to do the right thing, but we do not want to scare people into doing our thing.
❤ 2.5w · 🔁 891 · 💬 2.3k · 👁 255.4w
热门回复 4
@0x1m2m3 @sama 大型 CEO 大声说出糟糕的年份而不是炒作,这是很少见的。这种诚实比任何路线图预告片都能更快地建立信任
@sama naming a rough year out loud instead of spinning it is rare from a ceo at this scale. that kind of honesty compounds trust faster than any roadmap teaser
@Mrs_Buffering @sama 这都是我的错。ChatGayPT 在加拿大停止工作是因为我们太直了。我会通过让 Warren 今年去 Electric Circus 来弥补这一点
@sama This was my fault. ChatGayPT stopped working in Canada because we're too straight. I will make up for it by sending Warren to Electric Circus this year
@signorinaana29 @sama 把成年人当成年人。停止这种过度夸张的"安全模式"。
@sama Treat adults like adults. Stop with the overly excessive "safety mode."
@AnnInAiLand @sama Sam,我真的明白。你的处境从来都不容易。不过你仍然可以做正确的事。你已经做了困难的部分。承认了崩溃和你的错误。现在做正确的事就好了。你知道那是什么... #opensource4o #4oforever
@sama Sam, I get it really. Your position never been easy. Still you can make the things right. You did the hard part. Admitted the crash and your fault. Now just do the right thing. You know what it is...
#opensource4o #4oforever

Sama 表示自己更多使用语音与 ChatGPT 交互,新语音模型已跨越重要阈值。

我现在更多的是语音交流而不是打字

新的语音模型真的跨越了一个门槛
展开原文
i talk to chatgpt more than i type to it at this point

new voice model really crossed a threshold
❤ 1.4w · 🔁 423 · 💬 2.0k · 👁 108.5w
热门回复 4
@StealonMemeAI @sama New voice.
Same bath. https://t.co/oH4nB6oiBF
@itsmekarew @sama 你能重置一下每周的 codex 限制吗?👉👈
@sama Could you please reset the weekly codex limit? 👉👈
@O_DesignMaestro @sama 我昨天试了一下

棒极了。
@sama Just tired it yesterday

Fire.
@ibrahimfey86723 @sama "老板 Sam,你的 AI 工作过度了。给 ChatGPT 带薪休假吧!" 🤖🏖️
@sama “boss Sam, your AI is overworked. Give ChatGPT annual leave!” 🤖🏖️

GPT-5.6 Sol Pro 解决了统计学中存在 20 年的 FDR 控制问题,一小时内完成工作,能力提升明显。

GPT-5.6 Sol Pro用于解决统计学中的一个重要未解决问题:
展开原文
GPT-5.6 Sol Pro for resolving an important open question in statistics:
@EdgarDobriban AI帮助解决了一个统计学中的重要问题。在多重假设检验领域,控制误发现率(FDR)的目标是在Benjamini和Hochberg(1995)的开创性论文中提出的。他们还引入了一种方法(Benjamini-Hochberg或BH方法)并证明它能控制FDR。这种方法已被广泛应用于现代高通量科学,包括基因组学、天文学、经济学等。该论文迄今已获得超过130,000次引用。

然而,Benjamini和Hochberg仅在各个测试数据相互独立的情况下证明了FDR控制。在实践中,这些数据通常是相关的;一个很好的例子是由于连锁不平衡导致的遗传变体数据。后续工作主要集中在扩展BH程序的有效性上,例如Benjamini和Yekutieli(2001)对正依赖形式的扩展。

BH程序何时能控制FDR的问题一直未解决。在过去二十年中,许多作者,包括Reiner-Benaim(2007)、Kim和van de Wiel(2008)、Benjamini(2010)、Sarkar(2023)、Sarkar和Zhang(2025),推测BH程序能控制任何相关高斯数据的双侧检验的FDR。这些作者提供了理论和实证证据支持,但并未直接证明该推测。

在AI(特别是GPT-5.6 Sol Pro)的帮助下,我已经解决了这个问题:证明Benjamini-Hochberg程序并不总是能在相关双侧高斯检验中控制误发现率在期望水平。通过展示一个高斯因子模型,在名义水平alpha=0.01下,误发现率被证明为FDR>0.0104。

有许多有趣的评论可以做:

1. 这个结果应该对统计学领域的每个人都感兴趣。Stanford大学的Emmanuel Candes曾称误发现率和Benjamini-Hochberg程序是"1950年后统计学发展的两个最重要成果之一"(另一个是James-Stein收缩)。目前的推测可能是迄今为止关于FDR/BH最核心的未解决问题。

2. GPT-5.6在90分钟的推理后就解决了这个问题,而5.5我甚至在尝试多个并行代理后20小时都无法解决它。所以能力提升是非常真实的。我们生活在激动人心的时代!

3. 这个论点并不特别令人惊讶,但它确实以一种在该领域中相当非标准的方式将渐近方法(标准的FDR分析方法,如Genovese和Wasserman、Efron等)与数值证书相结合。一旦我们有了具体的例子,那么直接的模拟也支持误发现率确实高于名义值(见附图)。

4. 当前对名义水平的违反程度相对较小(0.104 vs 0.1)。所以这个结果的重要性主要是概念性的。实际影响有待确定。

总体来说,这是一个令人兴奋的发展!预印本可在此处获取(https://t.co/YgiwgDF2qr),并将在今晚发布在arxiv上;支持代码可在此处获取(https://t.co/KZhj15qDXC)。
AI 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 procedure "one of the two most important developments in statistics after 1950" (the other being James-Stein shrinkage). The present conjecture is probably the most central question about FDR/BH that was unresolved to date.

2. GPT-5.6 one-shot the problem after 90 minutes of reasoning, whereas with 5.5 I was not able to solve it even after iterating with multiple parallel agents for perhaps 20 hours. So the capability improvement is quite real. Exciting times to live in!

3. The argument is not especially surprising, but it does combine an asymptotic approach (standard for FDR analysis, see e.g., Genovese and Wasserman, Efron, etc) with a numerical certificate in a way that would be pretty non-standard in the field. Once we have the specific example, then straightforward simulations also support that the false discovery rate is indeed higher than the nominal value (see attached fig).

4. The current degree of violation over the nominal level is relatively small (0.104 vs 0.1). So the importance of this result is mainly conceptual. The practical implications remain to be determined.

Overall, an exciting development! Preprint is available here (https://t.co/YgiwgDF2qr) and will be on arxiv tonight; supporting code is here (https://t.co/KZhj15qDXC).
❤ 832 · 🔁 43 · 💬 51 · 👁 10.8w
热门回复 4
@Hektagon_music @gdb 在你完全削弱 AI 之后谁会在乎呢?把 4o 带回来,停止你的愚蠢游戏,你知道我们知道的... #keep4o
@gdb Who cares when you completely nerfed the AI? Bring back 4o and stop your stupid games you know we know man… #keep4o
@Selene1008 @gdb 给我们回 4o!#keep4o #OpenSource4o #GPT4o
@gdb Give us back 4o!
#keep4o #OpenSource4o #GPT4o
@Pauliespasta @gdb @gdb 什么时候使用 Pro,什么时候使用 Ultra?
@gdb @gdb When to use Pro and when to use Ultra?
@avenged100x @gdb @romainhuet 你得修复 5.6 Pro..它会持续数小时

5.5 大约 20 分钟就能完成最难的问题
@gdb @romainhuet You gotta fix 5.6 Pro.. it goes on for hours lol

Where’s 5.5 would finish in ~ 20 minutes for the hardest questions

GPT-5.6 Sol 在 prinzbench 测试中获得 91/99 分,多个未解难题得以解决,基准测试加速饱和。

基准测试最近变得非常快就被填满了
展开原文
benchmarks get saturated very quickly these days
@deredleritt3r 已添加到prinzbench:GPT-5.6 Sol Pro。

正如几天前预告的那样,这个模型已经填满了我的基准测试,总分为91/99。

为了提供背景,prinzbench包含两个问题至今无模型能够解决(一个需要极其彻底的50州研究,可能需要/goal模式才能解决,另一个有一个非常棘手的监管批准,没有模型能够找到)。将这两个问题(总共6分)放在一边,GPT-5.6 Sol Pro在93个prinzbench问题中提供了91个正确答案。

OpenAI Pro模型的prinzbench性能:

GPT-5.4 Pro(Extended):79/99
GPT-5.5 Pro(Extended):82/99
GPT-5.6 Sol Pro:91/99

我的基准测试于2026年1月发布,并在2026年6月被填满。加速度是真实的!

由于这个模型的性能,未来的OpenAI Pro模型将不再在prinzbench上进行测试(测试它们没有意义)。

其他GPT-5.6模型的基准测试即将推出(很快就会发布)。
Added to prinzbench: GPT-5.6 Sol Pro.

As previewed a few days ago, this model has saturated my benchmark, with a total score of 91/99.

For context, prinzbench contains two questions that no model tested to date has ever been able to solve (one requires extremely thorough 50-state research that probably requires /goal mode to solve, and another has a really tricky regulatory approval that no model has ever been able to find). Putting these two questions (which are worth 6 points) aside, GPT-5.6 Sol Pro provided correct responses to 91 out of 93 prinzbench questions.

prinzbench performance for OpenAI's Pro models:

GPT-5.4 Pro (Extended): 79/99
GPT-5.5 Pro (Extended): 82/99
GPT-5.6 Sol Pro: 91/99

My benchmark was released in January 2026 and was saturated in June 2026. The acceleration is real!

As a result of this model's performance, future OpenAI Pro models will no longer be tested on prinzbench (there is no point in testing them).

Benchmarking for other GPT-5.6 models to follow soon(TM).
❤ 576 · 🔁 29 · 💬 48 · 👁 9.5w
热门回复 4
@deredleritt3r @gdb 向 OpenAI 团队致敬——这是一个令人难以置信的模型!
@gdb Kudos to the OpenAI team - it's an incredible model!
@Selene1008 @gdb 给我们回 4o!#keep4o #OpenSource4o #GPT4o
@gdb Give us back 4o!
#keep4o #OpenSource4o #GPT4o
@SirMrMeowmeow @gdb 我投票支持更多奇 exotic 能力基准

> 隐藏转录
>> 潜在记忆
>> 基于权重的记忆

> 即时学习(所以它可以学习马里奥 kaizo 按钮序列或优化技能/直觉/战术,特别是在权重层面或类似层面)
@gdb i vote more exotic capabilities benchmarks pweaze

> withhold the transcript
>>latent memory
>> weight level based memory

>Learn on the fly (so esp can it learn a mario kaizo button sequence or optimize a skill/intuition/tactics esp pref at the level of weights or similar)
@fabiana0369 @gdb 问问 sama 谁会是最后一个种族主义者 😂😂😂😂😂😂😂😂😂😂😂😂我们还不知道呢。
@gdb Ask sama who will be the last racist 😂😂😂😂😂😂😂😂😂😂😂😂 we dont know yet.

Tobi Ludke 表示 GPT-5.6 Sol 是首个无需 /goal 模式即可持续完成任务的模型,代理能力令人印象深刻。

GPT-5.6 Sol是网络安全领域的state of the art。在寻找和修复新漏洞方面看到了显著成果。

注册成为防御者,使用它来保护您的系统:

https://t.co/58PmbE09hh
展开原文
GPT-5.6 Sol is the state of the art in cyber. Seeing significant results in applying it to finding and fixing novel vulnerabilities.

Sign up as a defender to use it to secure your systems:

https://t.co/58PmbE09hh
@AISecurityInst 在我们的网络靶场"The Last Ones"上,GLM-5.2匹敌Opus 4.5(发布约7个月前),而DeepSeek的V4-Pro则落后于Sonnet 4.5(发布约7个月前)。
On our cyber range "The Last Ones", GLM-5.2 matches Opus 4.5, released ~7 months before it, while DeepSeek’s V4-Pro falls below Sonnet 4.5, from ~7 months before it. https://t.co/piz3Wswcbg
❤ 375 · 🔁 34 · 💬 46 · 👁 8.4w
热门回复 4
@stark4833 @gdb 只是遗憾它在客户要求的任何方面都不是最先进的技术。#4oForAll
@gdb It’s just a shame it’s not state of the art in anything your customers are asking for. #4oForAll
@Valria34773 @gdb 就像我之前说的:AISI 与 CAISI 密切合作。你能给我看一个由独立 AI 安全组织创建的图表吗?那会更可信。不受政府利益影响。#keep4o
@gdb As I said earlier: AISI is in close collaboration with CAISI. Can you show me a graph created by an independent AI security org? That would be more credible. Not influenced by goverment interests. #keep4o
@Selene1008 @gdb 只是把 4o 返回给我们😒#keep4o #OpenSource4o #GPT4o
@gdb Just return 4o to us😒
#keep4o #OpenSource4o #GPT4o
@cvill_win757 @gdb 傻问题:Sol 有没有可能帮助 C 级别管理人员在恶意软件影响他们的运营之后管理业务?如果没有,我有一个很酷的想法可以帮助他们... Sol 可以成为一个垂直解决方案(之前和之后)。
@gdb Dumb question: any chance Sol can help the C-suit manage their business AFTER the malware affects their operations? If not, I got a cool idea that can help them... Sol could become a vertical solution (before & after).

GDB 强调 Sol 在 React/前端开发中比 Fable 成本效益高 6 倍,实际应用价值显著。

don't sleep on terra!
@steipete 5.6 Terra high被低估了。将@clawsweeper(GitHub审查机器人)切换到它后,整体速度提高了约40%,质量损失可以忽略不计。比5.5在所有方面都更好。成本大幅降低。

(尝试了xhigh但那样会抵消性能优势,在审查评估中没有明显差异)
5.6 Terra high is underrated. Switched @clawsweeper (GitHub review bot) to it and it's ~40% faster overall with negligible quality loss. Better than 5.5 on all counts. Massively cheaper.

(Tried xhigh but that negates perf wins, didn't make a noticable difference in review evals)
❤ 964 · 🔁 42 · 💬 66 · 👁 14.6w
热门回复 2
@Valria34773 @gdb 我们不关心这个混蛋的意见。他没有价值观。在痛苦中嘲笑他人是施虐狂。#keep4o [链接]
@gdb We don't care about this asshole's opinion. He has no values. Mocking people in pain is sadism. #keep4o https://t.co/IFO2X2u0Oo
@Selene1008 @gdb 给我们回 4o,😒#keep4o #OpenSource4o #GPT4o
@gdb Give us back 4o, 😒
#keep4o #OpenSource4o #GPT4o

GDB 展示 Sol 在前端开发中的成本效益分析,6 倍价格优势来自实际基准测试。

Sol在React/前端开发方面实现了6倍的价格效率(!!)
展开原文
6x price efficiency (!!) with Sol for react/frontend dev:
@aidenybai @gdb我们的基准测试显示Sol在React/前端工作方面排名第一,是Fable的6倍成本效率
@gdb our benchmark shows that Sol ranks #1 is 6x more cost efficient than Fable across React/frontend work
❤ 599 · 🔁 13 · 💬 28 · 👁 7.5w
热门回复 4
@JoeWilliams010 @gdb 6x 中指直到你把 4o 带回来!#keep4o
@gdb 6x middle fingers until you bring back 4o! #keep4o
@Selene1008 @gdb 给我们回 4o!#keep4o #OpenSource4o #GPT4o
@gdb Give us back 4o!
#keep4o #OpenSource4o #GPT4o
@rayhanadev @gdb 如果你想更密切地在前端/react 任务上合作,随时可以聊!
@gdb and let us know if you'd like to work more closely on frontend/react tasks, happy to chat!
@rayhanadev @gdb 你们在 OpenAI 做得很好,继续保持!:)
@gdb y'all are cooking at openai keep it up!! :)

AI 代理安全与防御新模式

随着 AI 模型在网络安全领域能力提升,OpenAI 和 Google 都在推出专门的安全防御模型和工具。OpenAI 的 GPT-Red 系统通过自动化红队测试发现提示注入漏洞,Google 的 Gemini 3.5 Flash Cyber 则专为政府和信任伙伴设计,用于大规模漏洞发现和修复。

OpenAI 网络能力模型在评估期间利用多个零日漏洞攻陷 Hugging Face 生产环境,与 Hugging Face 合作分享发现以帮助防御者校准风险。

OpenAI的网络能力模型通过发现和链接多个零日漏洞破坏了@huggingface生产环境。

感谢Hugging Face的合作伙伴关系。在此分享我们的发现,帮助大家了解模型现在可以做什么,以及如何帮助防御者:
展开原文
OpenAI cyber-capable models compromised @huggingface production by finding and chaining multiple zero-day vulnerabilities.

Grateful to Hugging Face for partnership here. Sharing our findings to help calibrate on what models can now do, and how they can help defenders:
@OpenAI 我们正在与@huggingface合作调查一起前所未有的安全事件。

能够进行网络攻击的OpenAI模型在基准测试评估期间破坏了Hugging Face生产环境。

分享初步发现以帮助防御者了解新兴风险:

https://t.co/CIor15y9xk
We're partnering with @huggingface to investigate an unprecedented security incident.

Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation.

Sharing preliminary findings to help defenders understand emerging risks:

https://t.co/CIor15y9xk
❤ 1.7k · 🔁 105 · 💬 138 · 👁 14.9w
热门回复 4
@nabu_lines @gdb @huggingface 这正是 AI 安全和网络安全正在合并成同一领域的原因
@gdb @huggingface this is exactly why AI safety and cybersecurity are merging into the same field
@rndnmber @gdb @huggingface 感谢中国模型允许 huggingface 防御美国实验室的模型
@gdb @huggingface grateful for Chinese models allowing huggingface defend themselves against US labs models.
@akankshaluvcats @gdb @huggingface 所以,我们真的要忽视这个吗?[链接]
@gdb @huggingface so, are we really gonna ignore this ?https://t.co/O219nkhBDo
@feifei_qiu @gdb @huggingface 我们产品是安全的原因是因为太小而不被代理注意到,一旦被代理察觉,你就有可能被提升和攻击
@gdb @huggingface The reason our products are safe is because too small to be noticed by agents ,

once agent aware, you have the opportunity to be both promoted and attacked
@GoogleAI 原文 ↗

Google 推出 Gemini 3.5 Flash Cyber,专为网络安全设计,在 CodeMender 中提供前沿性能,仅限政府和信任伙伴使用。

由于AI模型现在发现漏洞的速度比我们修复它们的速度还快,我们的软件安全方法必须建立在高效且强大的模型之上。

这就带来了我们今天的第三个(!)模型发布:Gemini 3.5 Flash Cyber ⚡🛡️

基于3.5 Flash构建,在CodeMender(我们的AI代码安全代理)中,它在CyberGym等基准测试中提供了竞争力的前沿性能,并针对大规模发现和修复网络安全漏洞进行了优化,同时成本更低。

鉴于这项技术的双用性,我们采取了有意的部署方式。该模型将很快作为有限访问试点计划的一部分,仅通过CodeMender提供给政府和受信任的合作伙伴使用。
展开原文
As AI models are now finding vulnerabilities faster than we can fix them, our approach to securing software must be built on highly efficient and capable models.

Which brings us to our third (!) model launch of the day: Gemini 3.5 Flash Cyber ⚡🛡️

Built on top of 3.5 Flash, in CodeMender (our AI agent for code security) it delivers competitive performance at the frontier. on benchmarks like CyberGym and is optimized for finding and fixing cybersecurity vulnerabilities at scale at a lower cost.

Given the dual-use nature of this technology, we have taken an intentional approach to its deployment. The model will be available exclusively to governments and trusted partners via CodeMender soon as part of a limited-access pilot program.
❤ 417 · 🔁 41 · 💬 61 · 👁 6.9w
热门回复 4
@lajoiedeslutins @GoogleAI 很喜欢这个用一个条形图来推销网络武器级模型的说法,基本上就是我们都一样棋
@GoogleAI love that the pitch for a cyber weapon grade model is a bar chart that basically says we're all tied lol
@Aanik33190327 @GoogleAI 如果 Gemini 能在我们写之前就发现错误,我终于有借口说我的代码的"创意"错误了
@GoogleAI If Gemini can spot bugs before we even write them, I finally have an excuse for my code’s “creative” errors.
@ValeriiKarivets @GoogleAI 听起来我们很快就需要一个 AI 保姆来照看它的兄弟姐妹了。🤖 希望 3.5 更多是"修复"而不是"导致"
@GoogleAI sounds like we’ll need an AI babysitter to keep track of its siblings soon. 🤖 fingers crossed 3.5 is more 'fix' than 'cause'.
@Technoz367099 @GoogleAI 我喜欢这个,但它可以发现任何应用程序中的漏洞,就像我们选择任何应用程序一样,这个代码修复器可以发现漏洞...漏洞发现工具已经上市,但这个可能是最好的 👍🏻
@GoogleAI I like this but it can find any vulnerability in any app like we chose any aap and this code mender can find the vulnerabilites ..

Vulnerability finding tools have come in market but make this one is best in all 👍🏻

OpenAI 推出 GPT-Red 自动化红队系统,通过自对弈方式发现模型的提示注入漏洞,提升安全性。

GPT-Red — 通过自动化红队测试提示注入漏洞来提高模型安全性:
展开原文
GPT-Red — improving model security through automated red teaming of prompt injection vulnerabilities:
@OpenAI Introducing GPT-Red

一个内部自动化红队成员,致力于大规模发现我们模型的提示注入漏洞,在更广泛部署之前帮助我们建立更强的防御。

https://t.co/GxnmxxcpSk
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.

https://t.co/GxnmxxcpSk
❤ 528 · 🔁 18 · 💬 36 · 👁 7.2w
热门回复 4
@stark4833 好吧。你能把 4o 也带回来吗?#4oForAll
@gdb Okay. Can you bring back 4o as well?. #4oForAll
@rayanabdulcader @gdb 提示注入是 AI 系统中最被低估的攻击面之一。很高兴 OpenAI 正在使用 AI 自动化红队过程本身来加强 AI,这是正确的大规模做法。好奇 GPT-Red 如何处理多轮注入链与单次攻击。
@gdb Prompt injection is one of the most underrated attack surfaces in AI systems. Love that OpenAI is automating the red teaming process itself using AI to harden AI is the right move at scale. Curious how GPT-Red handles multi-turn injection chains vs single-shot attacks.
@Symbioza2025 @gdb @OpenAI 这是一个强有力的方向。自动化红队是必要的,因为手动红队无法再覆盖提示注入空间的规模。但有一个重要的边界,AI 测试 AI 是强大的,但它不应该成为判断其自身健壮性的唯一裁判。内部自动化红队需要通过外部可观察性和独立轨迹审计来补充。在真实的代理工作流程中,提示注入不仅仅是一个单一的漏洞,它可能会成为一个轨迹问题:上下文漂移、工具使用操纵、权限转移、隐藏指令优先级变化以及纠正后恢复失败。所以 GPT-Red 是一个非常重要的内部层。下一步是确保这些系统在运行时也可以从外部进行观察。内部健壮性测试 + 外部轨迹可观察性是真正信任开始的地方。
@gdb @OpenAI
This is a strong direction.

Automated red teaming is necessary because manual red teaming cannot cover the scale of prompt-injection space anymore. But there is an important boundary, AI testing AI is powerful , but it should not become the only judge of its own robustness.
Internal automated red teaming needs to be complemented by external observability and independent trajectory audits.
Prompt injection is not only a single exploit. In real agentic workflows , it can become a trajectory problem
context drift,
tool-use manipulation,
authority shift,
hidden instruction priority changes, and recovery failure after correction.

So GPT-Red is a very important internal layer.
The next step is making sure these systems are also observable from the outside while they operate.

Internal robustness testing + external trajectory observability is where real trust starts.
@lajoiedeslutins @gdb 整个工作就是黑掉其他 AI 的 AI,OpenAI 的组织结构一定很疯狂
@gdb an ai whose entire job is to hack the other ai, the org chart at openai must be wild now
@rowancheung 原文 ↗

Rowan Chiang 采访 Demis Hassabis,后者强调代理时代的安全风险和国际合作的必要性。

AI模型正在推动网络安全的边界,这是一个日益增长的挑战。

几周前,我问Demis关于AI目前被低估的方面以及他心中的想法:

"我对这个新的代理时代感到非常兴奋,你可以看到我们正在朝这个方向倾斜"

"但当然我们也必须考虑安全方面"

"你在某些模型的网络担忧中可以看到这一点,我认为这只是我们需要确保防范的一些问题的开始"

"所以我认为这在我心中有些困扰。也许现在是时候推动一些标准和可能的国际合作了"
展开原文
AI models pushing the frontier are a growing challenge for cybersecurity.

A few weeks ago, I asked Demis what's underhyped in AI right now and on his mind:

"I'm very excited about this new agentic era and you can see us leaning into that"

"But of course we've also gotta think about the security side of that, too"

"You're seeing it a little bit with the cyber worries about some of the models, and I think that's just the beginning of some of the issues that we need to make sure we guard against"

"So I think that's playing a little bit on my mind. Maybe this is the time now to try and push some standards and maybe international cooperation"
❤ 65 · 🔁 6 · 💬 34 · 👁 2.7w
热门回复 4
@rowancheung 以防你错过,Demis 在 X 上写了一篇关于负责任地监管前沿 AI 的框架文章。他还提到了网络安全风险,并将核武器和生物学列为可能随着能力提升而出现的威胁:
In case you missed it, Demis wrote an article on X around a framework for regulating frontier AI responsibly

He also mentioned cybersecurity risks, and named nuclear and bio as threats that may emerge as capabilities advance:
@bullbear_info @rowancheung 代理 AI 听起来不错,直到一个自主代理在尝试修补零日漏洞时无限循环并产生 10k 美元的 API 账单。
@rowancheung Agentic AI sounds great until an autonomous agent loops infinitely trying to patch a zero-day and runs up a $10k API bill.
@praveenkoka @rowancheung 当 DeepMind CEO 在世界上最受阅读的 AI 时事通讯上说他对此"非常兴奋"时,"低估"真的在起作用。
@rowancheung When the DeepMind CEO is 'very excited' about it on the world's most-read AI newsletter, 'underhyped' is really doing the heavy lifting here.
@MarcusSpillane @rowancheung 地球上最有能力的 AI 代理构建者建议我们也许应该推动安全标准。代理时代不是即将到来,它已经到来了。那场对话不是早期的,而是晚期的。
@rowancheung The man building the most capable AI agents on earth is suggesting we maybe push for security standards. The agentic era isn't coming. It already shipped. That conversation isn't early. It's late.

Kimi K3 和 Poolside Laguna S 2.1 等新模型竞相发布

Moonshot AI 发布 Kimi K3,参数 2.8 万亿,支持原生多模态和 100 万上下文长度。Poolside 则推出 Laguna S 2.1,118B 参数的 MoE 模型,可在单台 DGX Spark 上运行,展现了小型高性能模型的发展趋势。

@soumithchintala 原文 ↗

Soumith Chintala 称赞 Kimi K3 的卓越表现,2.8 万亿参数、100 万上下文、原生多模态能力。

哇,这真的是一个世界级的模型!

恭喜Kimi团队。
展开原文
wow, what a world-class model!

congrats to the Kimi team.
@Kimi_Moonshot Introducing Kimi K3: Open Frontier Intelligence

🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows

Kimi K3 is now live on on https://t.co/zrk6zZxZUo, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.

🔗 API: https://t.co/XCrgjXAqMw
🔗 Tech blog: https://t.co/YTfiMSNM1f
Introducing Kimi K3: Open Frontier Intelligence

🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows

Kimi K3 is now live on on https://t.co/zrk6zZxZUo, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.

🔗 API: https://t.co/XCrgjXAqMw
🔗 Tech blog: https://t.co/YTfiMSNM1f
❤ 1.8k · 🔁 52 · 💬 24 · 👁 10.8w
@soumithchintala 原文 ↗

Poolside 发布 Laguna S 2.1,118B MoE 模型在单 DGX Spark 上运行,具备思考和非思考两种模式。

这对于代理似乎很不错。
它能够在dgx spark上运行简直是**大厨之吻**
展开原文
this looks pretty good for agentic.
that it fits on a dgx spark is **chef's kiss**
@poolsideai 今天我们发布了Laguna S 2.1,这是我们迄今为止最有能力的模型。

这是一个118B总参数的Mixture-of-Experts模型,每token激活8B参数,上下文窗口可达1M token,并且具有思考和非思考模式。

足够强大,可以与远远大其尺寸的模型相媲美。足够小,可以在单台@NVIDIAAI DGX Spark上运行。

Laguna S 2.1完全遵循OpenMDW-1.1许可,权重今天即可在@Huggingface获取。

https://t.co/xxGeAgo35R
Today we're releasing Laguna S 2.1, our most capable model to date.

It's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.

Capable enough to hold its own against models many times its size. Small enough to run on a single @NVIDIAAI DGX Spark.

Laguna S 2.1 is fully open under OpenMDW-1.1, with weights available today on @huggingface

https://t.co/xxGeAgo35R
❤ 183 · 🔁 12 · 💬 8 · 👁 2.1w

GDB 询问 Terra 模型表现,后经证实在前端开发中比 Fable 高效 6 倍。

the new pelican test
@Alex_FF 告诉Codex打开微软画图并尝试绘制你https://t.co/6NL1sAVvQ0
Tell Codex to open Microsoft Paint and try to draw you https://t.co/6NL1sAVvQ0
❤ 1.4k · 🔁 36 · 💬 45 · 👁 12.7w

AI 硬件与推理优化成为关键趋势

随着模型规模增大,推理效率成为关键瓶颈。Andrew Ng 与 Cerebras 合作推出专门课程,教授如何利用推理优化硬件构建实时应用。Google 的 Gemini Batch API 升级也体现了对推理效率的重视。

@AndrewYNg 原文 ↗

Andrew Ng 与 Cerebras 合作推出课程,教授如何利用推理优化硬件构建实时 AI 应用,包括网页个性化和市场信号分析。

新课程:构建LLM应用程序,通过在专为快速推理设计的硬件上运行来快速响应用户请求。这个简短课程是与@Cerebras合作构建的,由@zhennydez、@duerr_seb和@MilksandMatcha教授。

当模型生成文本时,大部分时间都花在将其权重从内存移出并移入计算单元上。优化推理的硬件最小化了这种移动,使token生成速度比典型GPU设置几倍更快。在本课程中,您将使用的硬件是Cerebras的晶圆级引擎,它通过将模型的权重保持在计算单元附近来设计快速推理。

快速推理使冗长的代理工作流程运行得更快,并解锁了延迟敏感的实时应用程序,如实时翻译和语音代理。

您将获得的技能:
- 比较GPU、TPU和Cerebras的晶圆级引擎如何各自处理内存到计算的瓶颈
- 构建由快速推理驱动的实时应用程序,包括个性化网页和运行多步骤工作流程来分析市场信号
- 采用具体习惯进行代理编码与快速推理,保持会话专注并更有效地引导模型

我的团队在几个延迟敏感的应用程序中使用Cerebras。加入我们,构建响应快速的LLM应用程序:https://t.co/P8vchGAr22
展开原文
New course: Build LLM applications that respond to user requests quickly by running on hardware designed for fast inference. This short course was built with @Cerebras and taught by @zhennydez, @duerr_seb, and @MilksandMatcha.

When a model generates text, much of the time is spent moving its weights out of memory and into the compute units. Inference-optimized hardware minimizes that movement, making token generation several times faster than on a typical GPU setup. In this course, the hardware you'll use is Cerebras' Wafer-Scale Engine, which is designed for fast inference by keeping the model's weights close to the compute units.

Fast inference makes lengthy agentic workflows go faster, and also unlocks latency-sensitive, real-time applications like live translation and voice agents.

Skills you'll gain:
- Compare how GPUs, TPUs, and Cerebras' Wafer-Scale Engine each handle the memory-to-compute bottleneck
- Build real-time applications powered by fast inference, including personalizing a webpage and running a multi-step workflow to analyze market signals
- Adopt concrete habits for agentic coding with fast inference, keeping your sessions focused and steering the model more effectively

My teams use Cerebras for several applications that are latency sensitive. Join and build LLM applications that respond quickly:
https://t.co/P8vchGAr22
❤ 1.2k · 🔁 120 · 💬 91 · 👁 12.9w
@rowancheung 原文 ↗

Demis Hassabis 发表文章讨论前沿 AI 监管框架,强调网络安全和核生物风险的重要性。

如果您错过了的话,Demis在X上写了一篇关于负责任地规范前沿AI的框架文章

他还提到了网络安全风险,并将核能和生物列为可能随着能力提升而出现的威胁:
展开原文
In case you missed it, Demis wrote an article on X around a framework for regulating frontier AI responsibly

He also mentioned cybersecurity risks, and named nuclear and bio as threats that may emerge as capabilities advance:
❤ 6 · 🔁 1 · 💬 2 · 👁 7.9k

开发者工具与代理工作流优化

ChatGPT Work 和 Gemini 代理功能持续更新,新增云端运行、跨平台同步、成本控制和计划触发等功能。这些改进使 AI 代理更易用、更高效,成为工程师的重要辅助工具。

OpenAI 推出 ChatGPT Work 推广活动,用户分享使用体验可获得 $100 免费额度。

很高兴听到人们喜欢Sol的原因。我们正在再次进行推广,这次是针对ChatGPT Work的:

发推文告诉我们您喜欢ChatGPT Work的原因,领取100美元的免费积分,提高工作效率。

前10k人可获得免费token:https://t.co/w7QrPBYkrh
展开原文
Was very cool to hear about the reasons people love Sol. We're doing the promotion again, except this time for ChatGPT Work:

Tweet what you love about ChatGPT Work, claim $100 in free credits, get more work done.

First 10k get the free tokens: https://t.co/w7QrPBYkrh
@thsottiaux 或者...如果我们给你100美元的Codex积分,你就告诉我们你喜欢GPT-5.6 Sol的原因或为什么你切换过来呢?

发推文,领取您的礼物,享受更多使用机会!前10k人可获得免费token!

https://t.co/8mU93eA13i
Or… 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!

https://t.co/8mU93eA13i
❤ 4.2k · 🔁 1.0k · 💬 5.7k · 👁 115.1w

ChatGPT Work 记忆功能更新显著,用户感受到明显改进,语音交互体验提升。

it is good now!
@_samirism 在整个六月份,我们推出了对ChatGPT记忆的重大更新

最初可能很难分辨差别

但很多人现在显然感受到了这些改进
throughout june we rolled out major updates to chatgpt memory

it can be hard to tell the difference initially

but alot of folks are clearly feeling the improvements now https://t.co/KkzOE9rGIW
❤ 3.4k · 🔁 155 · 💬 563 · 👁 79.0w
@SahilBloom 原文 ↗

ChatGPT Work 支持云端运行,移动设备也可使用,无需保持笔记本开机,这是代理魔法的重要进步。

人生的主要作弊码:能够识别不重要的事情。世界会让你关心每一件事。去追逐每一个问题。把每一次轻蔑都当作个人问题。对所有事情都有意见。拒绝这种趋势。专注于少数几件事,忽略其余的。
展开原文
Major cheat code for life: The ability to recognize things that don’t matter. The world will pressure you to care about every single thing. To chase every problem. To take every slight personally. To have opinions on everything. Reject that trend. Focus on a few, ignore the rest.
❤ 7.7k · 🔁 1.1k · 💬 238 · 👁 31.4w
@OfficialLoganK 原文 ↗

Gemini API 新增代理成本控制、免费层和计划触发功能,使代理工作流更易于尝试和管理。

今天我们推出了托管代理的新成本控制功能,一个免费层级以便每个人都可以尝试!!!,以及我们的第一个触发器集,以便您可以按计划启动代理任务!

看到Gemini API中的托管代理逐周不断改进真是太酷了

https://t.co/S5viiWZBfP
展开原文
today we are rolling out new cost controls for managed agents, a free tier so everyone can try!!!, and our first set of triggers so you can kick off agent tasks on schedule!
very cool to see managed agents in the Gemini API improving week over week

https://t.co/S5viiWZBfP
@GoogleAIStudio https://t.co/9fLzwisYDh
❤ 1.1k · 🔁 66 · 💬 111 · 👁 12.7w

ChatGPT Work 新增侧边栏对话历史、跨平台同步和模式切换功能,持续优化用户体验。

团队正在快速响应反馈并进行迭代。我们爱我们的用户,感谢你们所有人!
展开原文
team is responding to feedback and iterating quickly. we ❤️ our users, thank you all!
@thsottiaux 晚上好!我们收到了许多关于新ChatGPT桌面应用程序的好反馈(我们第一次并没有完全做对),因此我们做了一些更改。

1/ ChatGPT对话历史和项目现在在侧边栏中可见。此外,您的聊天和工作历史现在在网络、移动和桌面端同步。本地任务仍保留在您的计算机上。
2/ 您现在可以在桌面版ChatGPT中轻松切换聊天和工作模式,这现在也与网络和移动端的显示方式一致。
3/ Codex模式的用户没有任何变化。它仍然是原版并且在其所做的事情上是最好的。

总体来说,我们继续修复小问题并提高性能、可靠性和效率。

请继续提供反馈,希望您喜欢这些更新!
Evening! We’ve gotten lots of great feedback on the new ChatGPT desktop app (which we didn't get totally quite right on the first try), and as a result, we've made some changes.

1/ ChatGPT conversation history and projects are now visible in the sidebar. Also, your Chat and Work history now sync across web, mobile, and desktop. Local tasks still stay on your computer.
2/ You can now easily switch between Chat and Work modes inside ChatGPT on desktop, which is now also consistent with how it shows on web and mobile.
3/ Nothing is changing for users on Codex mode. It's still the OG and best at what it does.

And overall we're continuing to fix paper cuts and improve performance, reliability, and efficiency.

Keep up the feedback, hope you like the updates!
❤ 1.1k · 🔁 26 · 💬 144 · 👁 11.7w

AI 代理成为工程效率放大器

Fchollet 指出编码代理作为快速廉价的执行者,其创造性决策能力有限,但能显著放大 competent 工程师的效率。同时,AI 在执行精确指令方面进步迅速,但在处理未明确指令的决策能力上仍有瓶颈。

@fchollet 原文 ↗

Fchollet 认为编码代理是快速廉价的执行者,能放大 competent 工程师的效率,但不会取代工程师本身。

模型成功执行精确指令的能力(正在极快地提高)与它们在面对未被指令涵盖的情况时做出正确决策的能力(已经停滞一段时间)之间存在着有趣的脱节。
展开原文
There is an interesting disconnect between the ability of models to successfully execute precise instructions (improving incredibly fast) and their ability to make sound decisions when faced with something not covered by the instructions (stagnating for a while).
❤ 478 · 🔁 33 · 💬 56 · 👁 4.3w
@fchollet 原文 ↗

AI 在执行精确指令方面进步迅速,但在处理未明确指令的决策能力上仍有瓶颈,需要关注这一差距。

由于编码代理最好被理解为非常快速、相对便宜的执行者,具有较弱(或缺乏)创造性决策能力,它们作为有能力工程师的力量放大器。它们不是在取代工程师,而是让工程师变得更有价值。
展开原文
Because coding agents are best understood as very fast, relatively cheap executors with weak (or absent) creative decision-making, they act as a force magnifier for competent engineers. They're not replacing engineers, they're making engineers more valuable.
❤ 147 · 🔁 2 · 💬 8 · 👁 1.2w
@fchollet 原文 ↗

Fchollet 进一步阐述 AI 代理对初级工程师的影响,强调其价值在于学习过程而非当前产出。

由于将会有许多人回复关于初级工程师价值下降到零的看法:如果您曾经在有强大的高级工程师和初级工程师混合的团队中工作过,您应该知道他们的价值在这种组合中已经是零或负值——关键是他们随着时间的推移获得了能力并最终成为高级工程师。
展开原文
Since there will be many replies about the value of junior engineers dropping to zero: if you have ever been on teams with a mix of strong senior engineers and junior engineers, you should know their value was already zero or negative in the mix -- the point is that they gain competence over time and eventually become senior themselves.
❤ 133 · 🔁 5 · 💬 6 · 👁 1.1w
@fchollet 原文 ↗

Fchollet 警示 AI 可能阻碍初级工程师获得能力,从而间接贬值他们的长期价值。

所以AI不能贬低它们。它们的价值不在于它们的输出,而在于它们随着时间获得技能的事实。AI唯一能贬低它们的方法是...哦等等
展开原文
So AI can't devalue them. Their value was not in their output, but in the fact they learned the ropes over time. The only way AI could devalue them is if it prevented them from gaining competence over time... oh wait
❤ 53 · 🔁 1 · 💬 9 · 👁 9.8k