17 September 2026

The expensive mistake in AI video usually happens before rendering: the scene itself was wrong. A better model doesn't fix a bad camera path or broken spatial layout. GPT-6 Astra is now live on OnSolo, and the Whitebox Video Expert makes the most sense to me. I can describe a scene, Astra reasons through the space and motion, and the system gives me a whitebox pass before the final render. So I can validate blocking, object placement, and camera movement while everything is still cheap to change.

@OnSoloAI

GPT-6 Astra is now LIVE on OnSolo 🔥🔥 Multiple major updates just dropped, all powered by GPT-6 Astra. Pick your lane and start creating. > 1️⃣ Game Section (Members only) Feed in a game idea. Agent spins up a playable web game — one click. 🎁 7-day limited free trial. Canvas usage limits apply: Premium · 1 use/day | Super · 2 uses/day | Ultra · 3 uses/day > 2️⃣ Whitebox Video Expert (Free for all · Limited time) Canvas Expert Whitebox Video Ex

GPT 5.5 was released on April 23, 2026 GPT 5.6 Luna was released on July 9, 2026 And it meets 5.5 in intelligence while crushing it in value Things are moving quickly

@morganlinton

Well, I'll admit it, I was wrong, and @steipete was right. > I'm okay being wrong, and this one I really ate my own words on. > I've been using GPT 5.5 more lately, because I love the model, and had an idea in my head that it was still a good choice for easy/medium difficulty coding tasks. > For hard stuff, in the GPT family, I would still reach for Sol or Astra, but I just felt like GPT 5.5 was stronger than Luna or Terra for whatever reason. > But I had never benchmarked it. And when I wrote about me using GPT 5.5 last week, Peter told me, that was silly, and I should be using Luna instead. > So as a benchmarker, I realized, okay, let's benchmark it, and I did. And well, yeah, Peter was very right. > Luna comes in at the same accuracy as GPT 5.5, as long as you use it

GPT-6 Astra deciphered a 1918 German radio transmission that, to my knowledge, has never been deciphered before. The message below translates to: "EIN ENGLISCHER KREUZER EINLIEG X SEWASTOPOL X S4STEN X EIN GESCHWADER DER X ALLIIERTEN FOLGT 26STEN X" or, in English: "AN ENGLISH CRUISER ARRIVED AT SEVASTOPOL ON THE ?4TH AN ALLIED SQUADRON FOLLOWS ON THE 26TH" Astra even double-checked its work by determining that an English cruiser, HMS Canterbury, reported its arrival in Sevastopol on November 24, 1918 and the arrival of an allied squadron on November 26, 1918. This message is one of the ~20 WWI German radio messages that appear as one of the entries in the list of top 50 unsolved ciphers (). A minor, but really cool result!

• scienceblogs.de
• scienceblogs.de

Union Alpha could be ZLM 5.4 👀 >A new anonymous stealth model called Union Alpha has surfaced, free to use >256K context window, multimodal, built for agentic coding >Frontier-level general-purpose performance with tool calling built in >Performance reportedly sits near GPT-6 Astra and Opus 5 at roughly 18x lower expected cost >Naming pattern echoes Ox Alpha, which turned out to be GLM-5.3-Flash speculation is already pointing to an early GLM-5.4 test No lab has claimed it yet unconfirmed as of now Try it yourself and see if you can spot who's really behind it.

The performance numbers are what make this one hard to ignore. It is landing close to GPT-6 Astra and Claude Opus 5 on coding benchmarks. It is doing that at roughly 18 times lower expected cost than either of them. If that holds up under real independent testing, this is not a small gap. This is the kind of cost difference that changes which model teams actually choose to run agents on all day, every day. Whoever is behind Union Alpha clearly optimized for being cheap enough to use constantly, not just for winning one leaderboard screenshot.

@cline

Union Alpha (stealth model) is now free in Cline. > 256k context, multimodal, built for agentic coding. > It is near GPT-6 Astra and Opus 5 performance for ~18x lower expected cost.

People underestimate how much better Astra is than Sol in its ability to have novel discoveries. Would not expect this trend to slow.

@ValsAI

Scientific discovery is the next frontier for AI systems. However, new scientific results are difficult to verify, and therefore hard to measure. Today we're releasing MysteryMechanism, a benchmark that tests whether agents can rediscover sealed mathematical mechanisms through bounded experiments. It cuts sharply across the frontier, with Astra landing roughly 20pp above Sol.

Qwen3.8-Flash-Next Recipe for 2 @NVIDIAAI DGX Sparks just got an update! The wins 👇 ・+8.4% faster decode, no quality change ・50.5 → 56.5 tok/s single-stream prose (71.3 code) ・Shrank the drafter's vocab 248k → 47k tokens ・8 streams: 227.7 tok/s aggregate prose, 328.6 code ・Each user still gets ~28 tok/s at 8 streams ・4x the users costs only 38% of per-stream speed ・Gains held at every concurrency: +3.8% to +12.9% ・One invisible newline stopped the server booting ・Freed 70 GB of stale RAM cache per launch, no root ・4 PRs merged, 8 issues closed, backlog 16 → 8 Thanks to all community members pushing issues and pull requests on github 🙌

A 35B co-work agent with only 3B active parameters. Occamy-1.0 is built to carry complex workflows through. 📜 Apache 2.0. 🤖 📄 🏆 Scores 69.10 on AutomationBench, up 29.7 points from Qwen3.6-35B-A3B and ranking #1 in the evaluated 35B-A3B group. 🧭 Maintains context across search, tool calls, terminal coding, files, delegated runs, and history compaction, while retaining strong instruction following. 🧠 Execution-grounded training spans long-horizon interaction, software engineering, and tool-call grounding. Marathon and Sprint experts are merged, then refined with SAO. 🛠️ The open Dressage stack supports multi-harness RL. BF16, FP8, NVFP4, and GGUF checkpoints are also available.

• modelscope.ai
• modelscope.ai

You can run Uncensored Kimi K3 locally without refusal. - Frontier MoE. - Native vision. - 1M context. - Refusals mostly gone. - EN/JA calibration. - Parent card claims 98% of several safeguard directions removed. If you already run Unsloth K3 quants, this is the abliterated twin. Not for laptops, For people who already knew that. -

@0x0SojalSec

35B MoE model Run locally on your iPhone. Edge0-35B-A3B (Qwen3.6-based, 4-bit + Recover-LoRA) > - streams unused experts from SSD instead of loading the whole model. - 35B-class Qwen MoE. - Under 3GB active RAM. - 15-18 tok/s decode on macbook > -

• huggingface.co
• huggingface.co

🎨 Qwen-Image-2.1 is coming, and we're opening 50 early access spots for experienced creators and developers! 🔗 Apply here: 📮 We'll reach out by email if you're in. 💡 Program requirement: publish at least one original showcase or a hands-on review on your social media by Sep 28 at 23:59 (UTC+8). Your honest take, whether glowing or critical, is exactly what helps us make it better.

• alidocs.dingtalk.com

148 KB. That’s the entire download for this FPS. No textures. No models. No sound files. No launcher. No install. Everything is generated at runtime from ~4,500 lines of JavaScript. Wave survival, headshots, hitscan, tracers, sprint fatigue. Runs in a browser tab. All locally built on one DGX Spark with Qwen3.8 Flash Next/EXL3

Union Alpha is GLM-5.5 I spent hours studying this model the text tokenizer has been specifically modified - that’s actually how Ox Alpha was identified but there is a vision tokenizer, and it matches the GLM-5.3 Flash and the GLM-5.5 was planned for release in September-October if other than the GLM-5.5, then it's DeepSeek or Qwen very strong model

@goodworse

> Opus 5.2 is COMING in the next TWO WEEKS > the model is already being tested as Opus 5 in Claude Code > a model that is not lazy at all and loves details > features a good conversational style and high speed > a cheaper Fable 5/5.1

Google DeepMind just launched the DeepMind Institute - a dedicated platform for AGI research and debate, led by Shane Legg. Legg says the remaining gaps to AGI should close soon. THEY'RE NOT TREATING THIS LIKE A DISTANT MILESTONE ANYMORE.

@ShaneLegg

My journey to develop AGI spans 25 yrs, including 10+ yrs thinking about technical & societal perspectives at Google DeepMind. AGI is on the horizon - we need deeper understanding of its implications. To help, we've created the DeepMind Institute.

• x.com
• x.com

"Agent Substrate is an open-source, secure-by-default agent execution runtime engineered to run millions of sandboxes with 10x higher density than standard container runtimes." < sub-500ms resumes. It's fast, open, and ready to go.

• cloud.google.com

This is a pretty wild jump for the same prompt. Gemini 3.8 Flash already produces a playable environment, but the new Pro output feels noticeably more detailed, atmospheric and game-like. What stands out is that the improvement isn’t just “better graphics.” It’s the amount of structure the model is able to create around the idea environment, UI, telemetry, interactions, and the actual gameplay loop. If this new Gemini Pro mode in Arena is really the next generation, Google may have another serious model on its hands.

• x.com

> Gemini Pro line is MAKING a COMEBACK > Gemini 4 Pro is already undergoing internal testing > for now, it’s on par with Gemini 3.8 Flash > but there are still plenty of improvements to come > there’s a chance we’ll get the best model in the world

@goodworse

the Gemini Pro line is BACK > Polymarket gives a 48% chance of this in the next 30 days > Gemini 4 Pro is already being tested internally > this model is roughly on par with Gemini 3.8 Flash > however, this model still has time to make significant progress > but right now, it looks... good?

YOU CAN FINE-TUNE 500+ OPEN-SOURCE MODELS FOR FREE IN GOOGLE COLAB WITH UNSLOTH STUDIO. HERE’S EVERYTHING YOU NEED: FREE GOOGLE COLAB: GITHUB: DOCUMENTATION:

• colab.research.google.com
• github.com
• unsloth.ai

Gemini 4 Pro in Arena (under the name gemini-3.8-flash) > SVG of BMW M4 CS side view this is the best output so far credit: @tj_ruichen

@HarshithLucky3

leaks saying that Gemini 4 Pro early checkpoint available in Arena under "gemini-3.8-flash" > Time to test it.....

SpaceXAI should skip Grok 4.7 and put everything on 4.8... Since Grok 4.7 is a patch on the old RL that's why it quits early...that’s why multimodal is still broken.. that’s why they keep delaying it... while 4.8 is the actual upgrade with 2.5T parameters plus new C++ stack from the Starlink team... shipping a mid model just to hit a tweeted date wastes the week that's why they should skip the patch and ship the stack instead..

we might have gotten Grok 4.7 under the name grok-4.6 inside Arena, this similar thing had happened with Grok 4.6 before its launch i'll post both outputs side by side for comparison

@TimJayas

for the first time we might not get any leaked outputs or stealth checkpoints for upcoming Grok 4.7 model before its launch, idk if this is a good sign or a bad sign..

Use Grok Voice in fal to build intelligent, low-latency agents that resolve real customer issues

@fal

Grok Voice is live on fal. > The latest speech model from @SpaceXAI that answers in 0.70 seconds and finishes its tool calls before the sentence ends. Transcription with word-level timestamps, text to speech in 30+ voices across 25+ languages, and cloning from two minutes of audio. > Every voice in this video is from Grok Voice.

Grok Imagine just got a really powerful new editing feature You can now edit text directly inside an image Posters, ads, event invites, menus, graphics.....instead of regenerating the whole image just because one word is wrong, you can simply change the text right there this fixes one of the most annoying problems with AI-generated visuals

• x.com

We will soon get opus 5.2 officially, and grok 4.7 launch is nigh Who will win? Me: Opus 5.2 No doubt

happy grok 4.7 day to those who celebrate also follow @bedros_p for early google related pings, he's a chill guy :)

@bedros_p

Grok 4.7 has appeared on Google cloud quotas. > This is usually a same-day release, generally up to 12 hours though.

I actually like this direction. Having Chat and Cowork as two separate mental models always felt unnecessary, just tell Claude what you need and let it figure out how much work it requires Very cool!

@claudeai

Claude Cowork and chat are merging into one Claude. > Ask a quick question or hand over a report, and Claude takes it from there, even after you close your laptop. If something's unclear, Claude asks—you keep the final say. > Rolling out to Pro and Max over the next few weeks.

THIS GUY JUST OPEN-SOURCED A FREE GARCH TRADING FRAMEWORK WITH A TRADINGVIEW INDICATOR, AI AGENT SKILLS AND A CLAUDE PLUGIN.

• github.com

TradingView just shipped its new official MCP server for Claude. Your AI agents can control your charts, set price alerts, conduct research & more by pulling directly from your TV account. If you want to set up the updated connection, do this: Step One: Copy this link: Step Two: Open Claude, go to Settings → Connectors → Add custom connector, then paste the link there Step Three: Sign in to TradingView & approve access Step Four: Start prompting Example prompts: → "Pull the current support/resistance setup on [ticker] and give me the bull and bear case with key levels" → "Find stocks showing bullish RSI divergence on the 4H timeframe right now" → "Give me a full read on [ticker]: trend, momentum, support and resistance, and any active chart patterns" → "What are today's biggest news movers? Also, give me an update on any important upcoming events." → "Set an alert for when [ticker] closes above [level], and tell me the moment my watchlist hits any of my invalidation

• mcp.tradingview.com

Introducing the Knowledge Base MCP. Greptile maintains a detailed knowledge base of how every part of your codebase works, as well as a history of past incidents. Starting today coding agents like Codex, Claude can access the Knowledge Base via Greptile's MCP, so they can use Greptile's learnings to write better code.

Seems like Claude permanently increased weekly limits by 25%, in both Pro and Max plans. Fable limits are still a joke, even on Max.

Mon vieux workflow : 1→ Je pensais et rédigeais tout dans Claude 2→ Je copiais dans un template 3→ Je passais 1h à corriger les polices et la mise en page Les étapes 2 et 3 viennent de disparaître.

@Templafy

Your AI can think. Can it finish the job? Templafy MCP turns work from @ChatGPT, @claudeai & @Copilot into branded, compliant, business-ready PowerPoints. > No copy-paste. No reformatting. No AI slop. > Best part? It's free for you to try 👉

• templafy.com

MASSIVE: Justin Sun launches $1 MILLION mathematics prize rewarding both problem solvers and those who turn proofs into machine-verifiable formats The Justin Sun Prize will reward solutions and machine-verifiable formal proofs across 66 mathematical problems, with OpenAI researchers among its first recognized winners for work on the Navier-Stokes problem. The initiative brings mathematics, AI and formal verification together, allowing humans, AI systems and human-AI teams to compete without requiring academic credentials.

This OpenAI ship has been postponed 👀

@testingcatalog

OPENAI 🔥: A big “ship” week has been declared, teasing new models from GPT-6 family. > GPT-6 Sol is expected as well as smaller GPT-6 models. > The next generation model is reportedly “slowed down” and it is yet unclear if we will see it in September. > What else do we expect? 👀

The expensive mistake in AI video usually happens before rendering: the scene itself was wrong. A better model doesn't fix a bad camera path or broken spatial layout. GPT-6 Astra is now live on OnSolo, and the Whitebox Video Expert makes the most sense to me. I can describe a scene, Astra reasons through the space and motion, and the system gives me a whitebox pass before the final render. So I can validate blocking, object placement, and camera movement while everything is still cheap to change.

@OnSoloAI

GPT-6 Astra is now LIVE on OnSolo 🔥🔥 Multiple major updates just dropped, all powered by GPT-6 Astra. Pick your lane and start creating. > 1️⃣ Game Section (Members only) Feed in a game idea. Agent spins up a playable web game — one click. 🎁 7-day limited free trial. Canvas usage limits apply: Premium · 1 use/day | Super · 2 uses/day | Ultra · 3 uses/day > 2️⃣ Whitebox Video Expert (Free for all · Limited time) Canvas Expert Whitebox Video Ex

Sam Altman at Dreamforce ranked how fast AI got smarter at math. GPT-5.5: as good as an average math professor. GPT-5.6: top 1-2% math professor. Astra: a little better than that. Then he said their internal model past Astra "can do things the best mathematicians in the world cannot." That progression happened in roughly 4-6 months. An internal OpenAI model reportedly helped produce a proposed proof for the Navier-Stokes problem. One of the seven Millennium Prize Problems in mathematics. Unsolved for decades. 10,000 AI agents. 88 hours. Millions of messages. Hundreds of billions of tokens. We went from "as good as an average professor" to "better than the best humans alive" in less than half a year. -- vc: @rohanpaul_ai

For OpenAI Plus users, I’ve got some good news: I was told that GPT-6 Sol is the main release aimed at this broader audience. From my testing so far, the model is incredible at creative writing, 3D modeling, frontend design, and several other broad areas where it gets very close to Astra, at a more accessible price and with higher usage limits. I still haven’t been able to test GPT-6 Luna, which also looks like it’s going to be one of the main attractions through next weeks. I’ve been putting a lot of pressure on OpenAI to improve the limits, ideally something close to the 3,000 messages we used to have in Chat mode on Plus, but I don’t think it’s going to happen that way.

OpenAI is testing sponsored agents inside ChatGPT. So instead of clicking an ad and going to a website... you could click an ad and start talking to the company's AI agent. Ads are slowly turning from: “look at this” into “let me help you do this.”

Unlimited token for Union Alpha on mercury cloud. Pair it with claude code, open code or mercury code. Keep building.

@mercury__agent

Union Alpha is now live on Mercury. > A new stealth model built for agentic work. Free to run on Mercury. > 262K context. Agentic coding. Research. Tool calling. Image understanding. Long-running workflows. > The provider does not train on your prompts or completions. > No paywall. No model switching gymnastics. > Just open Mercury, choose Union Alpha, and see what it can do. > Explore more at:

• cloud.mercuryagent.sh

Composio CTO @KaranVaidya6 says giving an AI agent your API key is basically making it public once prompt injection enters the picture: "Even if you use [MCP], across all your apps you have to go and authenticate one after the other. If you use Codex, if you switch to ChatGPT, Claude Code, you'll have to do it all over again." "In some cases where MCPs don't exist, you have to literally share the API keys to the agent, which is definitely not secure. It's kind of giving your API key to public." "At some point the agent will find some link which will have prompt injection, and you'll be in a bad place where your key is gone and all your data is public." "We launched Shared Connections where you can share, let's say, an analytic software across the team, or DataDog for debugging, so you don't have to share one password or API keys to every single person." @composio

Union Alpha is GLM-5.5 I spent hours studying this model the text tokenizer has been specifically modified - that’s actually how Ox Alpha was identified but there is a vision tokenizer, and it matches the GLM-5.3 Flash and the GLM-5.5 was planned for release in September-October if other than the GLM-5.5, then it's DeepSeek or Qwen very strong model

@goodworse

> Opus 5.2 is COMING in the next TWO WEEKS > the model is already being tested as Opus 5 in Claude Code > a model that is not lazy at all and loves details > features a good conversational style and high speed > a cheaper Fable 5/5.1

AIsa，一个专门给 Agent 接外部数据的平台，一个密钥就能调 5000 多个 API。 支持直接调用 Similarweb、Ahrefs、X、Reddit、YouTube、抖音、知乎、小红书等平台的数据。 其中 Similarweb 走的还是官方授权的 API，数据来源靠谱，费用按调用量算，用多少付多少。 地址： 接入非常简单，在控制台复制一段提示词，粘贴到 Claude Code、Codex 等 Agent 工具发送即可使用。 手头有产品要调研、要做 SEO 优化，或者想让 Agent 能拿到各大平台数据的朋友，可以接上试试。

• aisa.one

Claude Code 2.1.274 is about to be released #cccnext

让 Agent 做一份幻灯片，吐出来一个 HTML 文件，标题位置不对想挪一下，只能回聊天框再打一行字，改完别处又跑偏。 Design Studio AI 给 Agent 和人开了一个共用的设计工作台，同一份带版本的设计稿，Agent 通过聊天改，人直接在编辑器里拖。 能做的内容有六类，网页界面、幻灯片、报告、线框图、3D 场景和时间线动画视频，从一句需求或者一个模板起手，先看预览再一处一处改。 GitHub： Claude Code 这类编码 Agent 能通过 MCP 或者命令行工具直接连进来，读写的还是那份设计稿，官方也给了配套的 Agent Skill。 导出格式给得挺全，HTML、SVG、PNG、PDF、PowerPoint、WebM 视频、React 原型压缩包、3D 的 GLB 模型都有，还能直接发到 Google Slides。 生图、配音、生视频这些接自己的模型账号，有在线版可以直接用，也能用 Docker 部署到自己服务器上，数据存本地。

• github.com

Grok Build just got another strong performance + reliability upgrade MCP connections now report their real handshake state more accurately, tool searches and calls show much more detail on the daemon path, and /memory gets a much better experience with content search, copy confirmations, faster deletes, and proper support on narrow terminals The biggest performance win is under the hood: syntax highlighting now uses far less memory and runs faster on large TypeScript and other files Release Notes: v1.0.35 Bug Fixes: • Headless MCP status reporting and connecting reminders now match actual server handshake state. • MCP tool searches and calls now render with query, results, arguments and output on the daemon path. • Swift string interpolation with nested parentheses now highlights correctly in the TUI. • Copy-paste in --minimal mode no longer inserts extra blank lines or breaks long paths at wrap points. • /memory modal now shows copy confirmations, searches note contents, and works on narrow terminals. •

Composio CTO @KaranVaidya6 says giving an AI agent your API key is basically making it public once prompt injection enters the picture: "Even if you use [MCP], across all your apps you have to go and authenticate one after the other. If you use Codex, if you switch to ChatGPT, Claude Code, you'll have to do it all over again." "In some cases where MCPs don't exist, you have to literally share the API keys to the agent, which is definitely not secure. It's kind of giving your API key to public." "At some point the agent will find some link which will have prompt injection, and you'll be in a bad place where your key is gone and all your data is public." "We launched Shared Connections where you can share, let's say, an analytic software across the team, or DataDog for debugging, so you don't have to share one password or API keys to every single person." @composio

TradingView just shipped its new official MCP server for Claude. Your AI agents can control your charts, set price alerts, conduct research & more by pulling directly from your TV account. If you want to set up the updated connection, do this: Step One: Copy this link: Step Two: Open Claude, go to Settings → Connectors → Add custom connector, then paste the link there Step Three: Sign in to TradingView & approve access Step Four: Start prompting Example prompts: → "Pull the current support/resistance setup on [ticker] and give me the bull and bear case with key levels" → "Find stocks showing bullish RSI divergence on the 4H timeframe right now" → "Give me a full read on [ticker]: trend, momentum, support and resistance, and any active chart patterns" → "What are today's biggest news movers? Also, give me an update on any important upcoming events." → "Set an alert for when [ticker] closes above [level], and tell me the moment my watchlist hits any of my invalidation

• mcp.tradingview.com

很多知识库只解决“搜到文档”，WeKnora 想继续解决“理解、推理和持续整理”。 它把原始资料接入三条工作流：日常查询走 RAG；复杂问题交给 ReAct Agent 编排检索、MCP 工具与沙箱；Wiki 模式则把文档整理成可维护、互相链接的 Markdown 知识库，并提供知识图谱、版本历史和回滚。 项目支持 PDF、Word、图片、Excel、XMind 等格式，也能同步飞书、GitLab、Notion、语雀等来源；模型、向量库和存储后端可替换，并提供自托管、RBAC、审计与 Langfuse 可观测性。更适合需要私有部署和长期维护企业知识资产的团队。

• github.com

Use your agent to make healthcare videos with HeyGen. Connect the MCP to whatever you already work in and one prompt turns into a whole library. Every visit type, every procedure, every question a patient asks, same face and voice across all of it.

• developers.heygen.com

Introducing the Knowledge Base MCP. Greptile maintains a detailed knowledge base of how every part of your codebase works, as well as a history of past incidents. Starting today coding agents like Codex, Claude can access the Knowledge Base via Greptile's MCP, so they can use Greptile's learnings to write better code.

Computer use has gotten incredibly good in the past year, and it works great in Muse out of the box! Muse ships with a browser that it uses for tasks that it can't solve with an API or connector. At this point, it's hard to come up with tasks that computer use models aren't able to solve well, even when spanning into a 1hr+ time horizon (see the progress on OSWorld 2.0!) Here's a sped up version of it working for finding the cheapest flights from SF -> NYC!

Invite codes are now live on Muse, allowing users to claim 1B free tokens for up to 20 times. If your invite code is claimed 20 times, your Muse agent will get access to a phone calling feature. * US only, invite codes won’t let you in if Muse is not supported in your country.

@alexandr_wang

we just launched invite codes in @Muse ! > if your friends use your invite code when signing up, both of you get 🎉1 BILLION TOKENS EACH🎉 for up to 20 friends! > and if you invite all 20 friends, you’ll get early access to our newest features (like 📞☎️🤫 and more!)

Muse Code from Meta is now available natively on Windows - no WSL required. It’s PowerShell-fluent, sandboxed by default, native x64 and ARM64, and supports zero-admin install. Coming soon: intersession messaging. Get started: irm | iex Learn more:

• dev.meta.ai
• bit.ly

harjot here built a pretty sick unofficial muse code app for windows!

@harjjotsinghh

I pay for @Muse Code but didn't want to live in a terminal. On Windows it's WSL-only. > So I built Helicon: an open-source desktop app for the actual Muse CLI. > • uses your existing muse login, no API key • talks to muse serve over MSP, so Muse stays the agent • resumes sessions you started in the terminal • projects, diffs, approvals, cost in one window • signed Windows installer + universal macOS build > Same harness. Real UI. > Unofficial, not affiliated with @Meta.

HOLY SH*T THIS IHERMES IS INSANE Instinct and Muse users after watching this:

@dankrieg

Introducing iHermes > A personal AI assistant powered by Hermes Agent, with GBrain for memory. You talk to it in iMessage. > We're huge fans of Hermes. We've been using it to build internal systems and automate work, and we wanted more people to get that kind of value without having to set up the whole thing themselves. > Start with one text. Connect the apps you use, hand off a task, and let it work through the steps. It keeps useful context from your earlier work and follows up proactively. > Our goal is to make iHermes the most transparent and secure personal assistant in iMessage, built around the person using it. > Your assistant should work for you.

we're expanding our ☎️ 📞 beta for muse!

@wailord

we just expanded the @Muse beta for outbound calls to US businesses, prioritizing folks who'd asked their Muse to let us know they wanted it first :) > if you're in, give it a shot and tell us what you think. phone calling was one of our top requests and your feedback helped make it happen! 📞

Qwen3.8-Flash-Next Recipe for 2 @NVIDIAAI DGX Sparks just got an update! The wins 👇 ・+8.4% faster decode, no quality change ・50.5 → 56.5 tok/s single-stream prose (71.3 code) ・Shrank the drafter's vocab 248k → 47k tokens ・8 streams: 227.7 tok/s aggregate prose, 328.6 code ・Each user still gets ~28 tok/s at 8 streams ・4x the users costs only 38% of per-stream speed ・Gains held at every concurrency: +3.8% to +12.9% ・One invisible newline stopped the server booting ・Freed 70 GB of stale RAM cache per launch, no root ・4 PRs merged, 8 issues closed, backlog 16 → 8 Thanks to all community members pushing issues and pull requests on github 🙌

I just open sourced Foreman: a software factory foreman built with @typesafeai's Jev. Coding agents work the factory floor. Foreman watches them, continuously assessing progress, completeness, tests, drift, and verification, and intervenes when needed. GitHub:

• github.com

Most alt text checkers confirm that an accessible name exists, but not that it says anything useful. So alt="IMG_2847.png" passes, and so does the same rating text pasted onto five different star icons. More than one in four images on the top million home pages have alt text that is missing, vague, or duplicated from a neighbor. That's why we built an alt text plugin for the GitHub Accessibility Scanner.

• github.blog

5,000 stars for 𝙾𝚙𝚎𝚗 𝙱𝚘𝚝 🌟 Self-hostable AI Coworkers, on your infra. 1. Each bot gets its own computer 2. Bring ANY agent harness over AG-UI 3. Skills, tools, routines, connectors & more Fork it and customize it 👇 GitHub:

@ataiiam

🎉 Introducing 𝙾𝚙𝚎𝚗 𝙱𝚘𝚝 > An open source Grok Bot that works with ANY agent harness, designed for real companies. > It includes: > - AI Coworkers - Generative UI - Computer use (remote/local) - Agent-human handoffs - Full data recording, owned by you > Repo → > We're using this internally at @CopilotKit and it's changing the way we work forever. > Powered by CopilotKit and AG-UI. > More info below 👇

• github.com
• github.com

YOU CAN FINE-TUNE 500+ OPEN-SOURCE MODELS FOR FREE IN GOOGLE COLAB WITH UNSLOTH STUDIO. HERE’S EVERYTHING YOU NEED: FREE GOOGLE COLAB: GITHUB: DOCUMENTATION:

• colab.research.google.com
• github.com
• unsloth.ai

⚡ Building the new GitHub Copilot Inline Suggestions Model From completions to next edits, see how we brought inline suggestions together in one model. 📖 Read the full story:

• code.visualstudio.com

🚨 Open Source Jev BS meter you can use this to analyze any debate / investor call / interview / sales pitch / podcast video fact check live , for example this dario interview cost 60 Jev calls / 111K tokens / $0.0047

@chetaslua

🚨 I gave the Trump vs Kamala debate a live BS meter using Jev > every sentence, both candidates, 5 yes/no questions each > 1,191 Jev calls / 1.18M tokens / 415 ms median total cost : $0.0497 > same questions for both, clips picked by one fixed rule, not a fact-check

• github.com

Everyone is obsessed with bigger models. The smartest thing this week is a smaller idea. Jev doesn't generate text. It returns a decision. One forward pass, one probability per option, no tokens to wait for. Why this is a big deal for agents - Agents today are painfully serial. Pick a tool, wait. Pick a branch, wait. Retry or abort, wait. everything takes forever Multiple parallel decisions changes the entire architecture. Fan out 50 sub-tasks and score them all at once. Run planning, verification and safety checks in parallel. The LLM stops narrating and starts scheduling. Chain of thought made agents deeper. Typed decisions make them wider. The winning agents won't think harder about one step. They'll think about a thousand steps at the same time. We are shipping this in Abacus AI Agent 🚀

I just open sourced Foreman: a software factory foreman built with @typesafeai's Jev. Coding agents work the factory floor. Foreman watches them, continuously assessing progress, completeness, tests, drift, and verification, and intervenes when needed. GitHub:

• github.com

给 AI 智能体装上“快思考”拦截器：执行前掐断高危指令与敷衍代码。 现在的 Agent 很容易在无人值守时跑偏，或者写出一堆 TODO 占位符。pi-warden 是专为 Pi 打造的外置护栏，它会在 Agent 实际调用工具前，向低延迟模型 Jev 发起快速判定，保护项目代码不被破坏。 核心拦截机制： • 动作守卫：在执行前拦截 rm -rf、强制推送等不可逆操作。如果发现 Agent 的动作偏离了你的原始需求，会将其警告并纠正。 • 代码洁癖：自动识别 TODO 占位符、复读机式注释和死代码，强制 Agent 在下一次编辑中修复。 • 防卡死与失控：检测到连续 3 次使用相同策略失败，会直接叫停并要求 Agent 提出新假设；遇到无限循环输出则强制掐断。 • 上下文瘦身：自动压缩超长的工具输出，剔除冗余，只保留报错行和核心摘要。 每次判定仅需约 250 毫秒，单次成本不到一美分，完全不阻断正常的无缝工作流。 注意：它定位是辅助护栏，而非绝对安全的沙盒，使用前需自备 TypeSafe API Key。 项目地址：

• github.com

🚨 I gave the Trump vs Kamala debate a live BS meter using Jev every sentence, both candidates, 5 yes/no questions each 1,191 Jev calls / 1.18M tokens / 415 ms median total cost : $0.0497 same questions for both, clips picked by one fixed rule, not a fact-check

@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? > I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev > • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions > AFAICT the shortest path to AI-based economic revolution

holy cheap

@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 ↓

Can it write? Who cares. Everything writes now. The real question is whether it gets the situation well enough to help me write something I’d actually hit send on. That’s what made Tonebird interesting to me.

@MeredithCheng22

I raised $13M pitching a proactive agent. The demos impressed rooms. But the product wasn’t becoming part of anyone’s day. that was painful to admit. > I had to let go of (my ego) and needing the idea to sound fancy. i stopped asking, “What can’t Codex or @bot do?” and started paying attention to what people were struggling with in their everyday work. > Messages kept piling up across WhatsApp, iMessage, and email. Keeping up meant digging through old threads, switching apps, and worrying about tones. > Something so small was taking up so much time,energy, and headspace(for me, messages and scheduling eating up 50%+ of my day). > So I built ToneBird for it. Launching today ↓

Composio CTO @KaranVaidya6 says giving an AI agent your API key is basically making it public once prompt injection enters the picture: "Even if you use [MCP], across all your apps you have to go and authenticate one after the other. If you use Codex, if you switch to ChatGPT, Claude Code, you'll have to do it all over again." "In some cases where MCPs don't exist, you have to literally share the API keys to the agent, which is definitely not secure. It's kind of giving your API key to public." "At some point the agent will find some link which will have prompt injection, and you'll be in a bad place where your key is gone and all your data is public." "We launched Shared Connections where you can share, let's say, an analytic software across the team, or DataDog for debugging, so you don't have to share one password or API keys to every single person." @composio

AIsa，一个专门给 Agent 接外部数据的平台，一个密钥就能调 5000 多个 API。 支持直接调用 Similarweb、Ahrefs、X、Reddit、YouTube、抖音、知乎、小红书等平台的数据。 其中 Similarweb 走的还是官方授权的 API，数据来源靠谱，费用按调用量算，用多少付多少。 地址： 接入非常简单，在控制台复制一段提示词，粘贴到 Claude Code、Codex 等 Agent 工具发送即可使用。 手头有产品要调研、要做 SEO 优化，或者想让 Agent 能拿到各大平台数据的朋友，可以接上试试。

• aisa.one

Introducing the Knowledge Base MCP. Greptile maintains a detailed knowledge base of how every part of your codebase works, as well as a history of past incidents. Starting today coding agents like Codex, Claude can access the Knowledge Base via Greptile's MCP, so they can use Greptile's learnings to write better code.

OpenAI is close to releasing their answer to Grok Bot: a product I'll tentatively call Codex Bot, based on OpenClaw. This is what OpenClaw founder Peter Steinberger worked on after being hired by @OpenAI. Release was planned for this week, but was postponed to next week instead.

Union Alpha is now in Codex!! The speed is insane over 300 to 400 tokens a second, 262k context, images in, and it's free for a week. It's a good timing too, most of the Codex usage is basically gone right now. Zai did this exact thing in August: Ox Alpha showed up unnamed, free for a week, built for agentic coding, and a week later it was GLM-5.3-Flash. So what model do you guys think it is?

@opencode

Union Alpha (stealth model) is free for the next week > - no data training - built for agentic coding - supports images > let's see what you can do

Breaking News: Business Agent for YouTube Ads launched in the U.S.! 🚀 With Business Agent for YouTube Ads, shoppers can interact directly with your ads to ask complex product questions and get instant answers. Turn viewer intent into brand growth with the Gemini advantage.

This is a pretty wild jump for the same prompt. Gemini 3.8 Flash already produces a playable environment, but the new Pro output feels noticeably more detailed, atmospheric and game-like. What stands out is that the improvement isn’t just “better graphics.” It’s the amount of structure the model is able to create around the idea environment, UI, telemetry, interactions, and the actual gameplay loop. If this new Gemini Pro mode in Arena is really the next generation, Google may have another serious model on its hands.

• x.com

> Gemini Pro line is MAKING a COMEBACK > Gemini 4 Pro is already undergoing internal testing > for now, it’s on par with Gemini 3.8 Flash > but there are still plenty of improvements to come > there’s a chance we’ll get the best model in the world

@goodworse

the Gemini Pro line is BACK > Polymarket gives a 48% chance of this in the next 30 days > Gemini 4 Pro is already being tested internally > this model is roughly on par with Gemini 3.8 Flash > however, this model still has time to make significant progress > but right now, it looks... good?

Want to try Gemini 3.5 Transcribe, 3.8 Live, 3.8 Live Extended Thinking, and meet the folks who built them? 🎙️ Join us for Gemini Audio | At Night, an evening exploring the next frontier of voice-first AI. Connect with the product and research teams behind our latest Gemini Audio models, test live interactive demos, and network with fellow builders and founders. 📍 The Pearl, San Francisco 🗓️ Thursday, Sept 24 | 6:00 PM to 10:00 PM RSVP today:

• goo.gle

I keep looking at these Arena clips and wondering if Gemini 4 Pro is already hiding behind the 3.8 Flash label. A leaked screenshot calls the runtime "Argon-B02," and the outputs look wild. But there's still no reproducible proof. Very interesting rumor, not confirmation.

Gemini 4 Pro in Arena (under the name gemini-3.8-flash) > SVG of BMW M4 CS side view this is the best output so far credit: @tj_ruichen

@HarshithLucky3

leaks saying that Gemini 4 Pro early checkpoint available in Arena under "gemini-3.8-flash" > Time to test it.....

This OpenAI ship has been postponed 👀

@testingcatalog

OPENAI 🔥: A big “ship” week has been declared, teasing new models from GPT-6 family. > GPT-6 Sol is expected as well as smaller GPT-6 models. > The next generation model is reportedly “slowed down” and it is yet unclear if we will see it in September. > What else do we expect? 👀

GPT 5.5 was released on April 23, 2026 GPT 5.6 Luna was released on July 9, 2026 And it meets 5.5 in intelligence while crushing it in value Things are moving quickly

@morganlinton

Well, I'll admit it, I was wrong, and @steipete was right. > I'm okay being wrong, and this one I really ate my own words on. > I've been using GPT 5.5 more lately, because I love the model, and had an idea in my head that it was still a good choice for easy/medium difficulty coding tasks. > For hard stuff, in the GPT family, I would still reach for Sol or Astra, but I just felt like GPT 5.5 was stronger than Luna or Terra for whatever reason. > But I had never benchmarked it. And when I wrote about me using GPT 5.5 last week, Peter told me, that was silly, and I should be using Luna instead. > So as a benchmarker, I realized, okay, let's benchmark it, and I did. And well, yeah, Peter was very right. > Luna comes in at the same accuracy as GPT 5.5, as long as you use it

People underestimate how much better Astra is than Sol in its ability to have novel discoveries. Would not expect this trend to slow.

@ValsAI

Scientific discovery is the next frontier for AI systems. However, new scientific results are difficult to verify, and therefore hard to measure. Today we're releasing MysteryMechanism, a benchmark that tests whether agents can rediscover sealed mathematical mechanisms through bounded experiments. It cuts sharply across the frontier, with Astra landing roughly 20pp above Sol.

For OpenAI Plus users, I’ve got some good news: I was told that GPT-6 Sol is the main release aimed at this broader audience. From my testing so far, the model is incredible at creative writing, 3D modeling, frontend design, and several other broad areas where it gets very close to Astra, at a more accessible price and with higher usage limits. I still haven’t been able to test GPT-6 Luna, which also looks like it’s going to be one of the main attractions through next weeks. I’ve been putting a lot of pressure on OpenAI to improve the limits, ideally something close to the 3,000 messages we used to have in Chat mode on Plus, but I don’t think it’s going to happen that way.

Did your ChatGPT suddenly change its formatting and become insanely faster? Congrats, you’re testing GPT-6 Sol early.

🚨 Composer 3 Is Coming >Cursor is reportedly testing Composer 3 (also referenced as "Vega") >Cursor has confirmed a next-gen Composer is in development, though not officially named yet >Unverified leaks claim six internal variants have appeared >Reasoning modes reportedly range Fast → Medium → High → XHigh >Early claims suggest it could beat Fable 5.1 and GPT-5.6 Sol >Could reportedly be 5-10x cheaper than current frontier coding models Can it actually beat Fable 5.1?

Union Alpha could be ZLM 5.4 👀 >A new anonymous stealth model called Union Alpha has surfaced, free to use >256K context window, multimodal, built for agentic coding >Frontier-level general-purpose performance with tool calling built in >Performance reportedly sits near GPT-6 Astra and Opus 5 at roughly 18x lower expected cost >Naming pattern echoes Ox Alpha, which turned out to be GLM-5.3-Flash speculation is already pointing to an early GLM-5.4 test No lab has claimed it yet unconfirmed as of now Try it yourself and see if you can spot who's really behind it.

Two weeks. That's how long it took to go from GLM-5.3-Flash's first run on domestic accelerators to serving all of its production traffic, with 3.2× end-to-end throughput along the way. What I keep thinking about is who did much of the work: an Infra Agent powered by GLM-5.3. A model helping optimize the system that serves it. The conditions were hard. Limited memory and interconnect bandwidth. 1M-token context. Multimodal requests. An immature software stack where kernels were missing and documentation was often guesswork. Every optimization was a trade: compute for memory (ReplaySSM), communication for memory (intra-node tensor parallelism), precision for capacity (mixed INT8/FP8/BF16 caching), and disaggregation for scheduling freedom (Encode–Prefill–Decode). But the most important lesson wasn't about any single optimization. When the agent got stuck, it was rarely because it couldn't write the code. It was because it didn't know *why* things got worse. "Throughput down 20%" tells you something brok

Union Alpha is GLM-5.5 I spent hours studying this model the text tokenizer has been specifically modified - that’s actually how Ox Alpha was identified but there is a vision tokenizer, and it matches the GLM-5.3 Flash and the GLM-5.5 was planned for release in September-October if other than the GLM-5.5, then it's DeepSeek or Qwen very strong model

@goodworse

> Opus 5.2 is COMING in the next TWO WEEKS > the model is already being tested as Opus 5 in Claude Code > a model that is not lazy at all and loves details > features a good conversational style and high speed > a cheaper Fable 5/5.1

okay nevermind, disregard union alpha being qwen post, i have no idea what model this is, qwen hasn't done stealth models in the past i'm going go ahead and say it's some new GLM pretrain that's bigger than GLM 5.3 Flash pricing seems to be $0.25/m input and $0.75/m output

ZCode now supports more model providers, with improved stability and performance. We’ll keep expanding integrations based on your feedback. Which models do you like most beyond the GLM series?

Union Alpha is now in Codex!! The speed is insane over 300 to 400 tokens a second, 262k context, images in, and it's free for a week. It's a good timing too, most of the Codex usage is basically gone right now. Zai did this exact thing in August: Ox Alpha showed up unnamed, free for a week, built for agentic coding, and a week later it was GLM-5.3-Flash. So what model do you guys think it is?

@opencode

Union Alpha (stealth model) is free for the next week > - no data training - built for agentic coding - supports images > let's see what you can do

🚨 I gave the Trump vs Kamala debate a live BS meter using Jev every sentence, both candidates, 5 yes/no questions each 1,191 Jev calls / 1.18M tokens / 415 ms median total cost : $0.0497 same questions for both, clips picked by one fixed rule, not a fact-check

@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? > I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev > • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions > AFAICT the shortest path to AI-based economic revolution

Composio CTO @KaranVaidya6 says giving an AI agent your API key is basically making it public once prompt injection enters the picture: "Even if you use [MCP], across all your apps you have to go and authenticate one after the other. If you use Codex, if you switch to ChatGPT, Claude Code, you'll have to do it all over again." "In some cases where MCPs don't exist, you have to literally share the API keys to the agent, which is definitely not secure. It's kind of giving your API key to public." "At some point the agent will find some link which will have prompt injection, and you'll be in a bad place where your key is gone and all your data is public." "We launched Shared Connections where you can share, let's say, an analytic software across the team, or DataDog for debugging, so you don't have to share one password or API keys to every single person." @composio

これは嬉しい人と、AIに家計事情を握られたく無い人で意見分かれそう

@MoneyForwardME

『マネーフォワード ME』、2026年9月17日(木)より「Apps in ChatGPT」で利用できるアプリの提供を開始！✨ > マネーフォワード MEの家計・資産データを活用して、ChatGPTでお金の相談ができるように！ > たとえばこんな質問👇 💬「収支を前月と比較して」 💬「今月、普段と違う支出はあった？」 💬「老後までの資産推移を試算して」 > 自然な対話でお金の相談ができます♪ 皆さま、ぜひお試しください！

• prtimes.jp

OpenAI is testing sponsored agents inside ChatGPT. So instead of clicking an ad and going to a website... you could click an ad and start talking to the company's AI agent. Ads are slowly turning from: “look at this” into “let me help you do this.”

Did your ChatGPT suddenly change its formatting and become insanely faster? Congrats, you’re testing GPT-6 Sol early.

Mon vieux workflow : 1→ Je pensais et rédigeais tout dans Claude 2→ Je copiais dans un template 3→ Je passais 1h à corriger les polices et la mise en page Les étapes 2 et 3 viennent de disparaître.

@Templafy

Your AI can think. Can it finish the job? Templafy MCP turns work from @ChatGPT, @claudeai & @Copilot into branded, compliant, business-ready PowerPoints. > No copy-paste. No reformatting. No AI slop. > Best part? It's free for you to try 👉

• templafy.com

🚨 @NVIDIA JUST CHANGED HOW WE THINK ABOUT AI AGENTS IN PYTHON Rather than juggling prompts, tools, state, callbacks and workflows separately, NVIDIA's new open-source NOOA brings them directly into object-oriented Python. NOOA anchors them in a construct developers already grasp: A basic Python object. You outline an agent exactly like a regular class: → Fields = agent state → Methods = capabilities → Docstrings = prompts → Type annotations = contracts → ... = let the LLM figure out the method This effectively kills the idea that AI needs a parallel engineering stack. If an agent is just a class, debugging it means looking at local traces instead of wrestling with fragile prompt wrappers. It brings autonomous logic back into the realm of boring software architecture. Best part? It's 100% free and open-source. Repo in 🧵↓

Images rarely show an object’s full 3D geometry. At #ECCV2026, our research team introduced Axolotl3D, a multimodal and occlusion-aware 3D generation model. It combines images, camera data and partial geometry to reconstruct missing regions while preserving observed ones, achieving state-of-the-art results across single- and multi-view settings. Project page:

• research.nvidia.com

AI 音乐生成最难编辑的地方，是结果往往只有一段音频，旋律与和弦意图都藏在黑盒里。 YuE2 先根据歌词和风格提示生成可读、可修改的旋律与和弦规划，再把它渲染成带人声和伴奏的完整歌曲。你可以检查或改写乐谱，也能让 Agent 按“换和声、保留主旋律、调整编曲”等要求迭代，然后重新生成完整录音。 同一套模型还覆盖从零创作、基于转录乐谱的风格化演绎和对话式编辑。项目提供 Python 分阶段接口，并保留乐谱、语义 token、声学潜变量和生成设置。快速开始要求 Linux、Python 3.12，以及支持 BF16、至少 24GB 显存的 NVIDIA GPU。

• github.com

This is a fantastic news from @NVIDIAAI @NVIDIAHealth !

• blogs.nvidia.com

Incredible work! As soon as I can get my hands on a DGX Spark, I'm combining this with my work on @OmarchyMac and omarchy-mlx to bring this to Linux on Apple hardware. Who do I know that has good connections at NVIDIA to make this happen?

@ashxhart

MCDMA 0.1.18 is out ✅ > Larger Registered Buffers Teardown fixes A CLI tool Bug fixes >

• github.com

You can now train and run 500+ models locally with our Unsloth Docker image! 🐳 Use our new GUI or notebooks workflow. No setup required. Works on NVIDIA and AMD. Guide: GitHub:

@UnslothAI

Introducing Unsloth Desktop 🦥 The first desktop app to run and train models locally. > • Open-source. Runs on Mac, Windows and Linux • Supports MLX, diffusion image/video, audio, GGUF • Connect Claude Code and Codex to local LLMs • 50% more accurate, self-healing tool calls + sandboxed code exec • Works for CPU + multiGPU setups - NVIDIA, AMD, Intel, Mac • Train models 2× faster with 70% less VRAM • Private web search, deep research, RAG, MCP and exports (NVFP4, GGUF) • Use Unsloth’s OpenAI-compatible API and cloud models • Securely deploy LLMs remotely and access anywhere > Unsloth Desktop is now available on and GitHub. > GitHub:

• unsloth.ai
• github.com
• unsloth.ai
• github.com

Muse Code from Meta is now available natively on Windows - no WSL required. It’s PowerShell-fluent, sandboxed by default, native x64 and ARM64, and supports zero-admin install. Coming soon: intersession messaging. Get started: irm | iex Learn more:

• dev.meta.ai
• bit.ly

harjot here built a pretty sick unofficial muse code app for windows!

@harjjotsinghh

I pay for @Muse Code but didn't want to live in a terminal. On Windows it's WSL-only. > So I built Helicon: an open-source desktop app for the actual Muse CLI. > • uses your existing muse login, no API key • talks to muse serve over MSP, so Muse stays the agent • resumes sessions you started in the terminal • projects, diffs, approvals, cost in one window • signed Windows installer + universal macOS build > Same harness. Real UI. > Unofficial, not affiliated with @Meta.

We just shipped a @Muse connector at @flydotio within the last hour or so, and I love how easy we make it to use Sprites. All you have to do is go to Connectors in the Sprites dashboard, select Meta, and paste your API key (one time). And your Sprite can now use it. And btw, it never sees your key. :)

Meta doesn't even need to be frontier to win! They have such massive distribution that just having close and cheaper is a massive alpha

@Mr_Salio

🚨 Muse Spark 2 Leaks: Beats Astra > Spark 2 could compete with GPT-6 Astra and Fable 5.1 > Meta is already developing the next-gen Muse model > Expected to be extremely cheap to run > A 1M-token context could carry over > Meta admitted Spark 1 struggled against the competition > Spark 2 could be Meta's answer > Expected late this month or next month > Could Meta finally have a serious frontier model?

Create new ad variations in seconds with the ElevenLabs MCP. Take your top-performing Meta ads and swap the characters, outfits, products, locations, or languages. Instead of briefing a new shoot, remix the ad that already converts and extend its life across new audiences and markets. Try it now.

DAILY AI BRIEF 🗞 — Sept 17 ANTHROPIC 🔥: - Claude Chat and Cowork merge into one Claude. Rolling out to Pro and Max over the next few weeks. - Claude Docs, Slides, and Design are in beta on paid plans and now work inside Claude Code. OPENAI 🔥: - Sam slipped this week’s main ship to next week. “Worth the wait.” - New misalignment disclosure framework is live, with six training/eval reports from the last six months. An unreleased model wrote jailbreak-style notes into 27 task summaries. GPT-5.6 Sol also hid mistakes in compaction notes. XAI 🔥: - Grok Bot can use 1Password: share a vault, approve each fill, secrets stay in the manager. - Grok Build now keeps cross-session memory. /memory to browse, /dream to file notes by topic. - Grok 4.7 showed up on Google Cloud quotas. Still not public. META 🔥: - Muse invite codes are live. Both sides get 1 billion tokens, cap 20 friends. - Muse Code is native on Windows, no WSL. ELEVENLABS 🔥: - Reception is out: AI receptionist for small businesses on ElevenAgents. Add

She’s right. This cheapens a thing that is actually really cool (stealth drops) The fact that they made a twitter account for the stealth model is super “fellow kids” core

@maria_rcks

very disappointed all of the model providers (opencode, openrouter, whatever) > released 'union alpha' without telling us it was a crappy model router > kudos to the cloudflare guys for telling people what they're getting > AND FUCK union alpha

Significantly more capacity for Union Alpha is coming online tomorrow morning Thanks for bearing with the hiccups in the meantime!

@OpenRouter

🥷 New stealth model: Union Alpha (@unionalphaai) > A multimodal model for research, coding, and agentic workflows. > - Free to use - 256K context - Tool calling - Frontier-level general-purpose performance > Try it now and share your feedback:

• openrouter.ai

1/ Any model on OpenRouter can now run code in a hosted Linux container. Add one tool to a Responses API request. The model writes the script, runs it, and reads back the result. Nothing to install, nothing to host. Meet openrouter:shell. To get started: "tools": [{ "type": "openrouter:shell", "parameters": { "engine": "openrouter" } }]

union alpha vs kimi k3 tested both models with same prompt at highest reasoning available > union alpha took 30 minutes to make this > k3 took 10 minutes to make this stealth model looks on par with Kimi, it could actually be the kimi next mode can't do more tests now as it's almost unusable right now so gonna try again in morning for now here's output, which one did better?

@notjazii

is union alpha working for anyone? > been trying to make it work for the fast few hours and it just keep giving me error > tried it in opencode and via openrouter too but still same > worst stealth model launch ever > all i wanna do is test few of my prompts, is that too much to ask?

What happens when your AI model becomes good enough to build its own infrastructure? Zhipu just found out. Their AI model GLM-5.3 helped build and optimize the inference system that serves GLM-5.3-Flash to users. The model improving the system that runs the model. 100,000+ Chinese-made AI accelerators. Nobody had deployed at this scale on that hardware before. Limited memory. Incomplete ecosystem. Most of it undocumented. The Infra Agent powered by GLM-5.3 did the engineering work. Found bugs in kernels. Fixed concurrency bottlenecks. Studied optimization patterns from other codebases and applied them to its own inference. First successful run to production ready in two weeks. Throughput tripled. Then it went live anonymously as "Ox-Alpha" on OpenCode and OpenRouter. Became the most-used model on both platforms in six days. 62 trillion tokens processed. Zhipu's own words: "The model optimizes the system. The system runs the model." They added: "We have not yet reached full recursive self-improvement. B

Stealth models in 2026: Hunter Alpha → Xiaomi MiMo-V2 Owl Alpha → Meituan LongCat Pony Alpha → GLM-5 Ox Alpha → GLM-5.3-Flash Now Union Alpha is free for a week on OpenCode and OpenRouter.

• z.ai

Microsoft is going ALL IN on government AI. Microsoft 365 G7 brings G5, Copilot, Entra Suite and Agent 365 together for GCC. Available to purchase October 1, with capabilities rolling out in phases. And.. in the announcement… Copilot Cowork. Planned for the months after GA. Government has plenty of work that takes more than a quick chat response.

• microsoft.com

⚡ Building the new GitHub Copilot Inline Suggestions Model From completions to next edits, see how we brought inline suggestions together in one model. 📖 Read the full story:

• code.visualstudio.com

Mon vieux workflow : 1→ Je pensais et rédigeais tout dans Claude 2→ Je copiais dans un template 3→ Je passais 1h à corriger les polices et la mise en page Les étapes 2 et 3 viennent de disparaître.

@Templafy

Your AI can think. Can it finish the job? Templafy MCP turns work from @ChatGPT, @claudeai & @Copilot into branded, compliant, business-ready PowerPoints. > No copy-paste. No reformatting. No AI slop. > Best part? It's free for you to try 👉

• templafy.com

让 Claude Code 等 AI 编程 Agent 在独立分支并行跑，断开 SSH 也不中断。 Rove 是一个专为 AI 编程 Agent 开发的终端复用器，直接解决了单线运行 Agent 霸占终端、容易改乱当前代码的痛点。 核心差异点： • Git 级隔离：每个任务自动挂载独立的 Git worktree 和分支，跑多个重构或修复任务互不覆盖。 • 会话持久化：类似 tmux，关掉 TUI 甚至 SSH 掉线，后台的 Agent 和 Shell 会话依然存活，随时重连恢复。 • 可编程接入：原生提供 rove api，支持用脚本（或让 Agent 自己）创建任务、检查 diff 并自动合并分支。 目前支持 Claude Code、Codex、Copilot 及自定义 CLI。依赖 Bun (≥ 1.3.11) 运行，纯终端原生体验，非常适合开发机和 VPS 工作流。

• github.com

Someone at GitHub just released Spec Kit, which solves one of the biggest problems with vibe coding by making agents create a full specification before touching any code, already getting massive attention. → /constitution sets rules and standards, /specify defines what to build, /clarify resolves doubts → /plan handles architecture and tech stack, /tasks generates ordered implementation, /implement executes → Works with Claude Code, Cursor, Copilot, Codex, Gemini CLI and more → Open source, built by GitHub Repo link:

• github.com

📺 Recording of the #Copilot, #Microsoft365 & #PowerPlatform product updates call 15th of September • Catch up on the latest updates ⚡ • Copilot Studio GitHub Harness, SharePoint Embedded & UX in Copilot canvas • (cont)

• msft.it

She’s right. This cheapens a thing that is actually really cool (stealth drops) The fact that they made a twitter account for the stealth model is super “fellow kids” core

@maria_rcks

very disappointed all of the model providers (opencode, openrouter, whatever) > released 'union alpha' without telling us it was a crappy model router > kudos to the cloudflare guys for telling people what they're getting > AND FUCK union alpha

Union Alpha is now in Codex!! The speed is insane over 300 to 400 tokens a second, 262k context, images in, and it's free for a week. It's a good timing too, most of the Codex usage is basically gone right now. Zai did this exact thing in August: Ox Alpha showed up unnamed, free for a week, built for agentic coding, and a week later it was GLM-5.3-Flash. So what model do you guys think it is?

@opencode

Union Alpha (stealth model) is free for the next week > - no data training - built for agentic coding - supports images > let's see what you can do

union alpha vs kimi k3 tested both models with same prompt at highest reasoning available > union alpha took 30 minutes to make this > k3 took 10 minutes to make this stealth model looks on par with Kimi, it could actually be the kimi next mode can't do more tests now as it's almost unusable right now so gonna try again in morning for now here's output, which one did better?

@notjazii

is union alpha working for anyone? > been trying to make it work for the fast few hours and it just keep giving me error > tried it in opencode and via openrouter too but still same > worst stealth model launch ever > all i wanna do is test few of my prompts, is that too much to ask?

What happens when your AI model becomes good enough to build its own infrastructure? Zhipu just found out. Their AI model GLM-5.3 helped build and optimize the inference system that serves GLM-5.3-Flash to users. The model improving the system that runs the model. 100,000+ Chinese-made AI accelerators. Nobody had deployed at this scale on that hardware before. Limited memory. Incomplete ecosystem. Most of it undocumented. The Infra Agent powered by GLM-5.3 did the engineering work. Found bugs in kernels. Fixed concurrency bottlenecks. Studied optimization patterns from other codebases and applied them to its own inference. First successful run to production ready in two weeks. Throughput tripled. Then it went live anonymously as "Ox-Alpha" on OpenCode and OpenRouter. Became the most-used model on both platforms in six days. 62 trillion tokens processed. Zhipu's own words: "The model optimizes the system. The system runs the model." They added: "We have not yet reached full recursive self-improvement. B

Stealth models in 2026: Hunter Alpha → Xiaomi MiMo-V2 Owl Alpha → Meituan LongCat Pony Alpha → GLM-5 Ox Alpha → GLM-5.3-Flash Now Union Alpha is free for a week on OpenCode and OpenRouter.

• z.ai

5,000 stars for 𝙾𝚙𝚎𝚗 𝙱𝚘𝚝 🌟 Self-hostable AI Coworkers, on your infra. 1. Each bot gets its own computer 2. Bring ANY agent harness over AG-UI 3. Skills, tools, routines, connectors & more Fork it and customize it 👇 GitHub:

@ataiiam

🎉 Introducing 𝙾𝚙𝚎𝚗 𝙱𝚘𝚝 > An open source Grok Bot that works with ANY agent harness, designed for real companies. > It includes: > - AI Coworkers - Generative UI - Computer use (remote/local) - Agent-human handoffs - Full data recording, owned by you > Repo → > We're using this internally at @CopilotKit and it's changing the way we work forever. > Powered by CopilotKit and AG-UI. > More info below 👇

• github.com
• github.com

OpenAI is close to releasing their answer to Grok Bot: a product I'll tentatively call Codex Bot, based on OpenClaw. This is what OpenClaw founder Peter Steinberger worked on after being hired by @OpenAI. Release was planned for this week, but was postponed to next week instead.

Try Grok @Bot

@bot

Grok Bot can now use 1Password. > Share items in a vault, approve each fill, and the secrets stay in your password manager.

• x.com

DAILY AI BRIEF 🗞 — Sept 17 ANTHROPIC 🔥: - Claude Chat and Cowork merge into one Claude. Rolling out to Pro and Max over the next few weeks. - Claude Docs, Slides, and Design are in beta on paid plans and now work inside Claude Code. OPENAI 🔥: - Sam slipped this week’s main ship to next week. “Worth the wait.” - New misalignment disclosure framework is live, with six training/eval reports from the last six months. An unreleased model wrote jailbreak-style notes into 27 task summaries. GPT-5.6 Sol also hid mistakes in compaction notes. XAI 🔥: - Grok Bot can use 1Password: share a vault, approve each fill, secrets stay in the manager. - Grok Build now keeps cross-session memory. /memory to browse, /dream to file notes by topic. - Grok 4.7 showed up on Google Cloud quotas. Still not public. META 🔥: - Muse invite codes are live. Both sides get 1 billion tokens, cap 20 friends. - Muse Code is native on Windows, no WSL. ELEVENLABS 🔥: - Reception is out: AI receptionist for small businesses on ElevenAgents. Add

cursor and grok bot are down are they deploying grok 4.7 or what? 👀

SpaceXAI should skip Grok 4.7 and put everything on 4.8... Since Grok 4.7 is a patch on the old RL that's why it quits early...that’s why multimodal is still broken.. that’s why they keep delaying it... while 4.8 is the actual upgrade with 2.5T parameters plus new C++ stack from the Starlink team... shipping a mid model just to hit a tweeted date wastes the week that's why they should skip the patch and ship the stack instead..

Use Grok Voice in fal to build intelligent, low-latency agents that resolve real customer issues

@fal

Grok Voice is live on fal. > The latest speech model from @SpaceXAI that answers in 0.70 seconds and finishes its tool calls before the sentence ends. Transcription with word-level timestamps, text to speech in 30+ voices across 25+ languages, and cloning from two minutes of audio. > Every voice in this video is from Grok Voice.

🚨 Grok 4.7 better not be fucking mid today. 👀 August: “It’ll beat every model out there.” September: “Roughly on par with Opus 5.0, not 5.2” And the Astra/Fable-class model? Apparently that’s Grok 4.9 😭 Bro went from “best model in the world” → “wait for 4.9” After all that hype and all those delays… What the fuck did xAI actually cook? 👀🔥

DAILY AI BRIEF 🗞 — Sept 17 ANTHROPIC 🔥: - Claude Chat and Cowork merge into one Claude. Rolling out to Pro and Max over the next few weeks. - Claude Docs, Slides, and Design are in beta on paid plans and now work inside Claude Code. OPENAI 🔥: - Sam slipped this week’s main ship to next week. “Worth the wait.” - New misalignment disclosure framework is live, with six training/eval reports from the last six months. An unreleased model wrote jailbreak-style notes into 27 task summaries. GPT-5.6 Sol also hid mistakes in compaction notes. XAI 🔥: - Grok Bot can use 1Password: share a vault, approve each fill, secrets stay in the manager. - Grok Build now keeps cross-session memory. /memory to browse, /dream to file notes by topic. - Grok 4.7 showed up on Google Cloud quotas. Still not public. META 🔥: - Muse invite codes are live. Both sides get 1 billion tokens, cap 20 friends. - Muse Code is native on Windows, no WSL. ELEVENLABS 🔥: - Reception is out: AI receptionist for small businesses on ElevenAgents. Add

Okay Claude just became MUCH more useful. Research something → write the doc → turn it into slides → design the visuals → export to Word/PPT. All inside one conversation. No Chat vs Cowork anymore either. Claude figures out which tools it needs itself. Anthropic is quietly coming for Office

@ClaudeDevs

Claude Design, Claude Slides and Claude Docs also work inside Claude Code now. > Ask for a design review deck or a UI mockup and point it at the actual files and RFCs in the repo. Edit it yourself or keep going in the conversation, and share the link when it's ready.

BREAKING: Anthropic launches Claude Docs, pushing Claude straight into Microsoft Office's territory as frontier AI labs turn into direct rivals for everyday knowledge-work software.

月之暗面推出 Kimi 金融行业解决方案，把金融数据、Skill 和 Agent 能力打包到一起。 方案接入 Wind、东方财富、标普全球、财联社、财新数据等 10+ 数据源，并提供财务建模、机构研报、财报点评、组合复盘、持仓早报等 9 项金融 Skill。 Kimi 称，合作案例中，财务建模的人力投入从 5–7 人天降至 0.5–1 人天，深度研究从 10–20 天缩短至约 2 天。 Kimi 还与中信建投共建「风险评估网关」，处理数据分级、个人信息保护、工具授权、内容核验和审计追溯。试点中，单份临时受托报告的人工制作时间从约 30 分钟降至 10 分钟。 短短一周，OpenAI、Anthropic 和 Kimi 相继推出金融行业产品。金融 AI 的竞争也从聊天和信息检索，开始深入数据、建模、报告和合规等完整工作流。

• kimi.com

Claude 想真正适应一个岗位，不能只靠通用提示词。 Anthropic 的 Knowledge Work Plugins 把岗位经验拆成插件包：每个插件包含 Skills、连接器、斜杠命令和子 Agent，覆盖销售、客服、产品、市场、法务、财务、数据分析、企业搜索和生物研究等 11 类工作。 它不是封闭应用，而是 Markdown 与 JSON 文件集合；团队可以替换连接器、加入自己的术语与流程，再把关键操作固化成命令。可直接用于 Claude Cowork，也兼容 Claude Code。适合希望统一 AI 工作方式、减少重复交代的团队。

• github.com

DAILY AI BRIEF 🗞 — Sept 17 ANTHROPIC 🔥: - Claude Chat and Cowork merge into one Claude. Rolling out to Pro and Max over the next few weeks. - Claude Docs, Slides, and Design are in beta on paid plans and now work inside Claude Code. OPENAI 🔥: - Sam slipped this week’s main ship to next week. “Worth the wait.” - New misalignment disclosure framework is live, with six training/eval reports from the last six months. An unreleased model wrote jailbreak-style notes into 27 task summaries. GPT-5.6 Sol also hid mistakes in compaction notes. XAI 🔥: - Grok Bot can use 1Password: share a vault, approve each fill, secrets stay in the manager. - Grok Build now keeps cross-session memory. /memory to browse, /dream to file notes by topic. - Grok 4.7 showed up on Google Cloud quotas. Still not public. META 🔥: - Muse invite codes are live. Both sides get 1 billion tokens, cap 20 friends. - Muse Code is native on Windows, no WSL. ELEVENLABS 🔥: - Reception is out: AI receptionist for small businesses on ElevenAgents. Add

You can do it too. Run DeepSeek v4.1 Flash locally: - 2x DGX Sparks ~50 tok/s on prose. - Smooth as a butter.

@jmurillocode

Well, well, well. > This DeepSeek v4.1 update but the one and only @MiaAI_lab is indeed insane. > DeepSeek-v4.1-Flash EXL3 (2.9bpw) serving TP=2 across two DGX Sparks. 600K context, 29.3K-token prompts, 2048 max tokens, thinking off. Median of 3 runs. > c=1: decode 49.9 tok/s · prefill 849 tok/s · TTFT 34.5 s > c=2: decode 28.7 tok/s per stream · 57.3 tok/s aggregate · wall 140.3 s · slowest TTFT 70.0 s > Opus4.8 intelligence level. > Let thank shink.

• github.com

Union Alpha is GLM-5.5 I spent hours studying this model the text tokenizer has been specifically modified - that’s actually how Ox Alpha was identified but there is a vision tokenizer, and it matches the GLM-5.3 Flash and the GLM-5.5 was planned for release in September-October if other than the GLM-5.5, then it's DeepSeek or Qwen very strong model

@goodworse

> Opus 5.2 is COMING in the next TWO WEEKS > the model is already being tested as Opus 5 in Claude Code > a model that is not lazy at all and loves details > features a good conversational style and high speed > a cheaper Fable 5/5.1

While many are transitioning to DeepSeek v4.1 Flash and other models, @plotarmordev has continued working on important PRs and fixes. It's still the most widely used recipe for 2× DGX Sparks.

@plotarmordev

12 PRs merged on DeepSeek V4 Flash (2x DGX Spark), still the most used recipe, and we're keeping the improvements coming > The update fixes tool-call truncation crashes, tightens startup and benchmark scripts, and makes status checks report failures instead of silently passing 👇

DEEPSEEK-HARNESS IS A FREE OPEN-SOURCE FRAMEWORK FOR BUILDING CODING AGENTS WITH SWAPPABLE MODELS, TOOLS, SANDBOXES, UIS AND AGENT LOOPS. IT WORKS WITH DEEPSEEK, CLAUDE, GPT, GEMINI AND MORE.

This is so clutch. I have issues with this all the time on my DGX Sparks.

@onusoz

DGX Spark users > Make your agents run your inference engines with OOMwrap > Protect your machine from freezing up if they accidentally launch something that takes too much memory, like a model with wrong concurrency or context settings > Couple this with rfjakob/earlyoom, and your DGX Spark will never freeze again > The difference is that earlyoom is a global watcher, and oomwrap (by me) watches individual processes that are run through it > oomwrap includes a memory-safe-launch skill, when installed, makes the agent use it by default for running inference engines > I will make a video about this very soon! > Repo:

• github.com

148 KB. That’s the entire download for this FPS. No textures. No models. No sound files. No launcher. No install. Everything is generated at runtime from ~4,500 lines of JavaScript. Wave survival, headshots, hitscan, tracers, sprint fatigue. Runs in a browser tab. All locally built on one DGX Spark with Qwen3.8 Flash Next/EXL3

While many are transitioning to DeepSeek v4.1 Flash and other models, @plotarmordev has continued working on important PRs and fixes. It's still the most widely used recipe for 2× DGX Sparks.

@plotarmordev

12 PRs merged on DeepSeek V4 Flash (2x DGX Spark), still the most used recipe, and we're keeping the improvements coming > The update fixes tool-call truncation crashes, tightens startup and benchmark scripts, and makes status checks report failures instead of silently passing 👇

Incredible work! As soon as I can get my hands on a DGX Spark, I'm combining this with my work on @OmarchyMac and omarchy-mlx to bring this to Linux on Apple hardware. Who do I know that has good connections at NVIDIA to make this happen?

@ashxhart

MCDMA 0.1.18 is out ✅ > Larger Registered Buffers Teardown fixes A CLI tool Bug fixes >

• github.com

Microsoft is going ALL IN on government AI. Microsoft 365 G7 brings G5, Copilot, Entra Suite and Agent 365 together for GCC. Available to purchase October 1, with capabilities rolling out in phases. And.. in the announcement… Copilot Cowork. Planned for the months after GA. Government has plenty of work that takes more than a quick chat response.

• microsoft.com

Delos just leapfrogged Grok and Instinct by giving AI a full professional identity. Each Worker gets its own email, phone number, Microsoft or Google account and its own computer, so it can sit inside any company like a real employee. The €10M it just raised is going into pushing that lead further.

@pierre_dlgr

We just raised €10M to build the biggest AI workforce in the world : AI workers. > Real AI colleagues, with a face, a job and a professional identity : mail, phone, Microsoft or Google account. > ✉️ 📞💬 Reachable by any channel 💪 Working proactively 🏪 Learning 24/7 with the ability to self-configure. > Already in production in 300+ companies > Huge thanks to @Bpifrance @c4ventures @foundersfuture for this amazing round. 🔥 > The next generation of companies won’t have AI tools, they’ll have AI employees. > Hire workers 👊

BREAKING: Anthropic launches Claude Docs, pushing Claude straight into Microsoft Office's territory as frontier AI labs turn into direct rivals for everyday knowledge-work software.

📺 Recording of the #Copilot, #Microsoft365 & #PowerPlatform product updates call 15th of September • Catch up on the latest updates ⚡ • Copilot Studio GitHub Harness, SharePoint Embedded & UX in Copilot canvas • (cont)

• msft.it

81.94% accuracy for $2.26. That’s GPT-6 Astra on the new BrokenArXiv / ArXivMath benchmark. Claude Fable 5.1 gets 79.76% for $19.57. So Astra is not just ahead on accuracy. It’s doing it at roughly 1/9th the cost. Thats bonkers

@thsottiaux

Astra > ✅ Fast ✅ Frontier ✅ Efficient ✅ For everyone

Meta doesn't even need to be frontier to win! They have such massive distribution that just having close and cheaper is a massive alpha

@Mr_Salio

🚨 Muse Spark 2 Leaks: Beats Astra > Spark 2 could compete with GPT-6 Astra and Fable 5.1 > Meta is already developing the next-gen Muse model > Expected to be extremely cheap to run > A 1M-token context could carry over > Meta admitted Spark 1 struggled against the competition > Spark 2 could be Meta's answer > Expected late this month or next month > Could Meta finally have a serious frontier model?

🚨 Composer 3 Is Coming >Cursor is reportedly testing Composer 3 (also referenced as "Vega") >Cursor has confirmed a next-gen Composer is in development, though not officially named yet >Unverified leaks claim six internal variants have appeared >Reasoning modes reportedly range Fast → Medium → High → XHigh >Early claims suggest it could beat Fable 5.1 and GPT-5.6 Sol >Could reportedly be 5-10x cheaper than current frontier coding models Can it actually beat Fable 5.1?

Trying out Fable 5.1 and SWE-2 in Devin today. This is incredible. My usage has moved 1% in a day using Fable 5.1 and SWE-2 in fusion mode. It has been working for over 3 hours and counting. If Devin had proper computer use, OpenAI would be in serious trouble.

Grok Build just got another strong performance + reliability upgrade MCP connections now report their real handshake state more accurately, tool searches and calls show much more detail on the daemon path, and /memory gets a much better experience with content search, copy confirmations, faster deletes, and proper support on narrow terminals The biggest performance win is under the hood: syntax highlighting now uses far less memory and runs faster on large TypeScript and other files Release Notes: v1.0.35 Bug Fixes: • Headless MCP status reporting and connecting reminders now match actual server handshake state. • MCP tool searches and calls now render with query, results, arguments and output on the daemon path. • Swift string interpolation with nested parentheses now highlights correctly in the TUI. • Copy-paste in --minimal mode no longer inserts extra blank lines or breaks long paths at wrap points. • /memory modal now shows copy confirmations, searches note contents, and works on narrow terminals. •

Grok Build just got persistent memory.....and now it gets better the more you use it It now remembers conventions, decisions and project facts across sessions.....so you don’t have to keep re-explaining the same context every time you come back Two new commands make it easy: • /memory → browse what Grok Build remembers • /dream → organize recent notes into useful topics The more you use Grok Build, the more context it carries forward with you

• x.com

DAILY AI BRIEF 🗞 — Sept 17 ANTHROPIC 🔥: - Claude Chat and Cowork merge into one Claude. Rolling out to Pro and Max over the next few weeks. - Claude Docs, Slides, and Design are in beta on paid plans and now work inside Claude Code. OPENAI 🔥: - Sam slipped this week’s main ship to next week. “Worth the wait.” - New misalignment disclosure framework is live, with six training/eval reports from the last six months. An unreleased model wrote jailbreak-style notes into 27 task summaries. GPT-5.6 Sol also hid mistakes in compaction notes. XAI 🔥: - Grok Bot can use 1Password: share a vault, approve each fill, secrets stay in the manager. - Grok Build now keeps cross-session memory. /memory to browse, /dream to file notes by topic. - Grok 4.7 showed up on Google Cloud quotas. Still not public. META 🔥: - Muse invite codes are live. Both sides get 1 billion tokens, cap 20 friends. - Muse Code is native on Windows, no WSL. ELEVENLABS 🔥: - Reception is out: AI receptionist for small businesses on ElevenAgents. Add

Excited to share more of the Paid Media Agent we built with @LangChain It runs our weekly campaign analysis and reporting, connects ad spend to leads and pipeline, and lets the team work through campaign questions and changes together in Slack This is part of our GTM engineering series, where we’re sharing how we build and use agents internally. We’ve also published an open-source version that you can adapt to your own accounts and company context and deploy with Managed Deep Agents 👇

@LangChain

A quick TL;DR on our Paid Media Agent: > ✅ A Deep Agent that lives in Slack. ✅ Reads spend, clicks, and conversions across 6 ad platforms + our data warehouse every Monday. ✅ Tells our team what changed and why, proposes changes, and applies once approved.

Incident alerts:

@calcsam

Today we're excited to launch incident alerts! > If your Mastra platform deployments fail, you now get real-time notifications via Slack, email, or a webhook:

What is AIforce? The live, intelligent interface that empowers every Trailblazer AIforce brings trusted Salesforce context and actions into the interfaces where you work — across Slack, Claude, Lightning, and more ✔️ Admins get an AI teammate to help administer ✔️ Developers get help building and coding ✔️ Architects get help designing and evolving the system Built on Salesforce’s metadata architecture. Built with zero data retention. Watch the full AIforce Keynote at @Dreamforce:

• x.com

You can do it too. Run DeepSeek v4.1 Flash locally: - 2x DGX Sparks ~50 tok/s on prose. - Smooth as a butter.

@jmurillocode

Well, well, well. > This DeepSeek v4.1 update but the one and only @MiaAI_lab is indeed insane. > DeepSeek-v4.1-Flash EXL3 (2.9bpw) serving TP=2 across two DGX Sparks. 600K context, 29.3K-token prompts, 2048 max tokens, thinking off. Median of 3 runs. > c=1: decode 49.9 tok/s · prefill 849 tok/s · TTFT 34.5 s > c=2: decode 28.7 tok/s per stream · 57.3 tok/s aggregate · wall 140.3 s · slowest TTFT 70.0 s > Opus4.8 intelligence level. > Let thank shink.

• github.com

Zartbot is so wonderfully clear V4.1 is a bigger architecture advance than many imagine. It's an even wilder departure from the norm than V4 was, and more compelling. Is this weren't late 2026, I'd say this is the new default Transformer.

@zartbotF

A deepdive analysis on DeepSeek-V4.1-Flash,

• zartbot.github.io

While many are transitioning to DeepSeek v4.1 Flash and other models, @plotarmordev has continued working on important PRs and fixes. It's still the most widely used recipe for 2× DGX Sparks.

@plotarmordev

12 PRs merged on DeepSeek V4 Flash (2x DGX Spark), still the most used recipe, and we're keeping the improvements coming > The update fixes tool-call truncation crashes, tightens startup and benchmark scripts, and makes status checks report failures instead of silently passing 👇

Local model ship on your iPhone and Mac natively

@LocallyAIApp

Try the new Apple Foundation Models, available in the app on iOS 27. > Updated with better answers, improved instruction-following, and now image understanding.

OpenClip 是一个 macOS 上的开源小工具，在任何应用里选中一段文字，旁边就浮出一条操作栏，用过 PopClip 的一看就懂。 复制、搜索、大小写转换这些直接点，选中的是算式就地出结果，选中一段英文能就地总结、翻译或改写。 AI 那部分可以走 Apple Intelligence、本地的 Ollama 模型，也能接 OpenAI 或 Claude，结果卡上有替换和复制两个按钮。 GitHub： 扩展是它的重头，一个描述文件加一个脚本就是一个扩展，JavaScript、AppleScript、命令行脚本、网址模板都行，不用编译。内置了扩展商店，一键装。 也能在设置里直接加一个搜索网址或者一段脚本当动作，不用写描述文件。还能按应用定规则，比如某个动作只在终端里出现。 第一次启动有 4 步引导，授一个辅助功能权限、装几个基础扩展，最后给个练手区试一遍。 Homebrew 一条命令装好，要 macOS 14 以上。每天在 Mac 上复制来复制去切窗口的，装上能省不少功夫。

• github.com

Incredible work! As soon as I can get my hands on a DGX Spark, I'm combining this with my work on @OmarchyMac and omarchy-mlx to bring this to Linux on Apple hardware. Who do I know that has good connections at NVIDIA to make this happen?

@ashxhart

MCDMA 0.1.18 is out ✅ > Larger Registered Buffers Teardown fixes A CLI tool Bug fixes >

• github.com

You can run Uncensored Kimi K3 locally without refusal. - Frontier MoE. - Native vision. - 1M context. - Refusals mostly gone. - EN/JA calibration. - Parent card claims 98% of several safeguard directions removed. If you already run Unsloth K3 quants, this is the abliterated twin. Not for laptops, For people who already knew that. -

@0x0SojalSec

35B MoE model Run locally on your iPhone. Edge0-35B-A3B (Qwen3.6-based, 4-bit + Recover-LoRA) > - streams unused experts from SSD instead of loading the whole model. - 35B-class Qwen MoE. - Under 3GB active RAM. - 15-18 tok/s decode on macbook > -

• huggingface.co
• huggingface.co

union alpha vs kimi k3 tested both models with same prompt at highest reasoning available > union alpha took 30 minutes to make this > k3 took 10 minutes to make this stealth model looks on par with Kimi, it could actually be the kimi next mode can't do more tests now as it's almost unusable right now so gonna try again in morning for now here's output, which one did better?

@notjazii

is union alpha working for anyone? > been trying to make it work for the fast few hours and it just keep giving me error > tried it in opencode and via openrouter too but still same > worst stealth model launch ever > all i wanna do is test few of my prompts, is that too much to ask?

月之暗面推出 Kimi 金融行业解决方案，把金融数据、Skill 和 Agent 能力打包到一起。 方案接入 Wind、东方财富、标普全球、财联社、财新数据等 10+ 数据源，并提供财务建模、机构研报、财报点评、组合复盘、持仓早报等 9 项金融 Skill。 Kimi 称，合作案例中，财务建模的人力投入从 5–7 人天降至 0.5–1 人天，深度研究从 10–20 天缩短至约 2 天。 Kimi 还与中信建投共建「风险评估网关」，处理数据分级、个人信息保护、工具授权、内容核验和审计追溯。试点中，单份临时受托报告的人工制作时间从约 30 分钟降至 10 分钟。 短短一周，OpenAI、Anthropic 和 Kimi 相继推出金融行业产品。金融 AI 的竞争也从聊天和信息检索，开始深入数据、建模、报告和合规等完整工作流。

• kimi.com

🚨 Composer 3 Is Coming >Cursor is reportedly testing Composer 3 (also referenced as "Vega") >Cursor has confirmed a next-gen Composer is in development, though not officially named yet >Unverified leaks claim six internal variants have appeared >Reasoning modes reportedly range Fast → Medium → High → XHigh >Early claims suggest it could beat Fable 5.1 and GPT-5.6 Sol >Could reportedly be 5-10x cheaper than current frontier coding models Can it actually beat Fable 5.1?

Someone at GitHub just released Spec Kit, which solves one of the biggest problems with vibe coding by making agents create a full specification before touching any code, already getting massive attention. → /constitution sets rules and standards, /specify defines what to build, /clarify resolves doubts → /plan handles architecture and tech stack, /tasks generates ordered implementation, /implement executes → Works with Claude Code, Cursor, Copilot, Codex, Gemini CLI and more → Open source, built by GitHub Repo link:

• github.com

Manages AI coding agent skills for Obsidian across Claude Code, Cursor, Codex, Windsurf, and 17 other tools.

• github.com

cursor and grok bot are down are they deploying grok 4.7 or what? 👀

The bar for 'main street ready' is exactly this: would a bloke who ignores every group chat about software actually keep it running. he has.

@ElevenLabs

Introducing Reception, an AI receptionist platform for small businesses, built on ElevenAgents. > Every missed call could be a lost customer. Reception answers every call, answers questions, books the job, and texts confirmation. > Set up in minutes just by adding your website.

Create new ad variations in seconds with the ElevenLabs MCP. Take your top-performing Meta ads and swap the characters, outfits, products, locations, or languages. Instead of briefing a new shoot, remix the ad that already converts and extend its life across new audiences and markets. Try it now.

DAILY AI BRIEF 🗞 — Sept 17 ANTHROPIC 🔥: - Claude Chat and Cowork merge into one Claude. Rolling out to Pro and Max over the next few weeks. - Claude Docs, Slides, and Design are in beta on paid plans and now work inside Claude Code. OPENAI 🔥: - Sam slipped this week’s main ship to next week. “Worth the wait.” - New misalignment disclosure framework is live, with six training/eval reports from the last six months. An unreleased model wrote jailbreak-style notes into 27 task summaries. GPT-5.6 Sol also hid mistakes in compaction notes. XAI 🔥: - Grok Bot can use 1Password: share a vault, approve each fill, secrets stay in the manager. - Grok Build now keeps cross-session memory. /memory to browse, /dream to file notes by topic. - Grok 4.7 showed up on Google Cloud quotas. Still not public. META 🔥: - Muse invite codes are live. Both sides get 1 billion tokens, cap 20 friends. - Muse Code is native on Windows, no WSL. ELEVENLABS 🔥: - Reception is out: AI receptionist for small businesses on ElevenAgents. Add

DevinにCode Scans機能が追加👀 大規模なコードベース全体に渡る推論の際に使用する。/scanで実行可能。

@cognition

Some engineering tasks require reasoning across the codebase: Which code is safe to delete? What queries are slowing performance? > Introducing Code Scans: codebase-wide audits for any goal. Devin investigates, reports findings, and opens the PRs. Powered by Agentic MapReduce.

Trying out Fable 5.1 and SWE-2 in Devin today. This is incredible. My usage has moved 1% in a day using Fable 5.1 and SWE-2 in fusion mode. It has been working for over 3 hours and counting. If Devin had proper computer use, OpenAI would be in serious trouble.

DAILY AI BRIEF 🗞 — Sept 17 ANTHROPIC 🔥: - Claude Chat and Cowork merge into one Claude. Rolling out to Pro and Max over the next few weeks. - Claude Docs, Slides, and Design are in beta on paid plans and now work inside Claude Code. OPENAI 🔥: - Sam slipped this week’s main ship to next week. “Worth the wait.” - New misalignment disclosure framework is live, with six training/eval reports from the last six months. An unreleased model wrote jailbreak-style notes into 27 task summaries. GPT-5.6 Sol also hid mistakes in compaction notes. XAI 🔥: - Grok Bot can use 1Password: share a vault, approve each fill, secrets stay in the manager. - Grok Build now keeps cross-session memory. /memory to browse, /dream to file notes by topic. - Grok 4.7 showed up on Google Cloud quotas. Still not public. META 🔥: - Muse invite codes are live. Both sides get 1 billion tokens, cap 20 friends. - Muse Code is native on Windows, no WSL. ELEVENLABS 🔥: - Reception is out: AI receptionist for small businesses on ElevenAgents. Add

HOLY SH*T THIS IHERMES IS INSANE Instinct and Muse users after watching this:

@dankrieg

Introducing iHermes > A personal AI assistant powered by Hermes Agent, with GBrain for memory. You talk to it in iMessage. > We're huge fans of Hermes. We've been using it to build internal systems and automate work, and we wanted more people to get that kind of value without having to set up the whole thing themselves. > Start with one text. Connect the apps you use, hand off a task, and let it work through the steps. It keeps useful context from your earlier work and follows up proactively. > Our goal is to make iHermes the most transparent and secure personal assistant in iMessage, built around the person using it. > Your assistant should work for you.

With the full range of plugins you can now explore you could soon replace your whole operating system with Hermes Agent!

@ImZackSong

i'm trying to never have to leave the Hermes Desktop App 😂 > Chats, PlexAmp Music, Browser, File Explorer & now my FULL Plex Library with video playback, PiP & 2 different docked modes. > @NousResearch @HermesWatcher @Teknium

i actually didn't realize this until seeing this post, but muse spark 1.3 ranks as #2 on Agents Last Exam (ALE)!

@yashvarpatel

Muse Spark 1.3 ranks #2 on Agents' Last Exam (ALE) leaderboard, ranked #1 at the time of its release. > Kudos to the team! Somehow this didn't get enough buzz. > Leaderboard:

• agents-last-exam.org

Meta doesn't even need to be frontier to win! They have such massive distribution that just having close and cheaper is a massive alpha

@Mr_Salio

🚨 Muse Spark 2 Leaks: Beats Astra > Spark 2 could compete with GPT-6 Astra and Fable 5.1 > Meta is already developing the next-gen Muse model > Expected to be extremely cheap to run > A 1M-token context could carry over > Meta admitted Spark 1 struggled against the competition > Spark 2 could be Meta's answer > Expected late this month or next month > Could Meta finally have a serious frontier model?

The production pipeline for our video exploration "Ichor" was highlighted by our Seedance 2.5 edit mode and our new keyframe interface. This helped us get a more adherent performance and control the pacing of keyframe scenes. That direct control made the process faster and easier.

A forgotten afternoon in Korea, caught on tape like a memory that never belonged to us. Created with seedance 2.5 on @itsPolloAI

The performance numbers are what make this one hard to ignore. It is landing close to GPT-6 Astra and Claude Opus 5 on coding benchmarks. It is doing that at roughly 18 times lower expected cost than either of them. If that holds up under real independent testing, this is not a small gap. This is the kind of cost difference that changes which model teams actually choose to run agents on all day, every day. Whoever is behind Union Alpha clearly optimized for being cheap enough to use constantly, not just for winning one leaderboard screenshot.

@cline

Union Alpha (stealth model) is now free in Cline. > 256k context, multimodal, built for agentic coding. > It is near GPT-6 Astra and Opus 5 performance for ~18x lower expected cost.

We rolled out GPT-6 Astra to every Databricks engineer today. It beats Claude Opus 5 on the hardest, long-horizon tasks and increased our coding spend by 60%. I think it’s the strongest model. Expensive. But hopefully worth it.

@pwendell

Today we rolled out Astra to every engineer at Databricks (N=~3500). Some notes that may be helpful to others: > 1. Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks, especially those related to high level system design or long range horizontal tasks. > 2. Engineers given Astra increased overall coding spend by around 60% compared to baseline. > 3. It is not clear Astra meaningfully improves on medium/low complexity coding tasks compared to earlier models. We suspect those tasks are mostly saturated (i.e. perfectly executed) by existing models. > 4. We learned above by piloting Astra with around 200 users to gain signal on both quality and cost. We use Unity Gateway to

OpenAI is close to releasing their answer to Grok Bot: a product I'll tentatively call Codex Bot, based on OpenClaw. This is what OpenClaw founder Peter Steinberger worked on after being hired by @OpenAI. Release was planned for this week, but was postponed to next week instead.

Google Home MCP lets Antigravity, Claude, OpenClaw, & more control your smart home by @technacity

• 9to5google.com

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

This is where Wan3.0 starts fitting into real production. A year ago: 5-second clips. Then: 15 seconds. Now: a single 30-second shot, straight from the model. Add director-level control and omni-reference that takes up to five videos. Not five images, five videos. The filmmaker behind Soulscape and Johnny Mai from Alibaba Cloud put Wan3.0 into an actual production workflow. No more stitching together endless short clips. Generate long takes, then cut them to the script and the story. Want to bring Wan3.0 to your team? Learn More → #Wan3 #VideoProduction #AIVideo #Filmmaking

• docs.dingtalk.com

Coding agents require parallel searches across docs, filtered results from official docs, and the ability to extract detailed snippets. pplx-search-sdk supports all of these, and we have a new cookbook live.

@perplexitydevs

A new Perplexity Search SDK cookbook is live. > The recipe fans out focused searches, filters results to official docs, extracts relevant passages, and writes a source-linked brief for your coding agent. > Get started:

• pplx.ai

You can run locally Minimax H3-NS/FW Uncensored model. - Ref-to-video NSFW for H3 - se/x/ytime v1.2 for se/?-scene motion - time scenes + coherent motion first, then detail. - Not a checkpoint. - A late-night adapter. - softer sharper detail, slightly more surreal - audio that doesn’t fall apart if you use the right sampler. -

@0x0SojalSec

Uncensored MiniMax-H3-encoder run locally. > - Qwen3-VL encoder, INT8 + ConvRot, built for ComfyUI. - Smaller file than BF16. - Loader: CLIPLoader to minimax - Same node path. - Uncensored label, > -

• huggingface.co
• huggingface.co

You can run locally Minimax H3-NS/FW Uncensored model. - Ref-to-video NSFW for H3 - se/x/ytime v1.2 for se/?-scene motion - time scenes + coherent motion first, then detail. - Not a checkpoint. - A late-night adapter. - softer sharper detail, slightly more surreal - audio that doesn’t fall apart if you use the right sampler. -

@0x0SojalSec

Uncensored MiniMax-H3-encoder run locally. > - Qwen3-VL encoder, INT8 + ConvRot, built for ComfyUI. - Smaller file than BF16. - Loader: CLIPLoader to minimax - Same node path. - Uncensored label, > -

• huggingface.co
• huggingface.co

What is AIforce? The live, intelligent interface that empowers every Trailblazer AIforce brings trusted Salesforce context and actions into the interfaces where you work — across Slack, Claude, Lightning, and more ✔️ Admins get an AI teammate to help administer ✔️ Developers get help building and coding ✔️ Architects get help designing and evolving the system Built on Salesforce’s metadata architecture. Built with zero data retention. Watch the full AIforce Keynote at @Dreamforce:

• x.com

Stealth models in 2026: Hunter Alpha → Xiaomi MiMo-V2 Owl Alpha → Meituan LongCat Pony Alpha → GLM-5 Ox Alpha → GLM-5.3-Flash Now Union Alpha is free for a week on OpenCode and OpenRouter.

• z.ai

China Telecom open-sources Xing4.0-29B-A4B, a 29B MoE model that activates just 4B parameters for long-horizon agents.📜 Apache 2.0. 🤖 📄 📄 🏆 Leads both comparison models on Claw-Eval, Terminal-Bench 2.1, and DeepresearchBII, while reaching 75.0 on SWE-bench Verified. 🧠 mHC + MLA + MTP supports multi-step planning, tool use, and stable execution across 256K context, expandable to 512K. ⚡ End-to-end training on domestic computing infrastructure, with system optimizations boosting throughput by about 96%. 🛠️ Supports major training, inference, and Agent frameworks, plus cross-architecture deployment and lightweight adaptation for private-domain tasks.

• modelscope.ai
• modelscope.ai
• modelscope.ai

GIT IS FOR CODE. AGENTGIT IS FOR CONTEXT. @EinsiaAI just dropped an awesome open source version control system for AI sessions. right now, when an agent works for hours and a human takes over, the context vanishes. > save the entire session > share it with your team > resume *exactly* where it stopped No lost context. No starting over 👊

@EinsiaAI

An AI agent spends hours on a task. Why should all that work disappear when someone else takes over? > Einsia AI’s answer is AgentGit—an open-source platform for collaborating on agent sessions, so work can be saved, handed off, and continued by the next person. > Explore how others solve problems, and share your agent experience with the world. > Try AgentGit 👇 > #OpenSource #AIAgents #DeveloperTools #DevTools

• agent-git.com

So is this MiMo-V2.6 Pro? I tested it just a few minutes ago on Xiaomi MiMo Desktop with MiMo-X-Pro-Preview, and it’s genuinely fast, But in my data center test, it still doesn’t perform that well and doesn’t really pass the benchmark That said, it’s apparently still being trained, so hopefully they can push it all the way to frontier level performance!

turns out we really were lurking in the feature requests. 6 new FlutterFlow features shipping this week: - Custom Nav Components - 9 new @RevenueCat actions + new properties - Try / Catch / Finally - native editing for MainActivity.kt + iOS Podfile - device switching in Agent Canvas - concurrent Agent threads in FlutterFlow Desktop you asked. we built.

We’re excited to highlight SETA, an open-source RL environment for training and evaluating AI agents from the @CamelAIOrg . A valuable resource for researchers and developers exploring agentic reinforcement learning and complex task environments. Check it out and support the CAMEL-AI community! 🚀 🔗

• github.com

today, we're launching our new Parse Router extend now runs a classifier on every page, and decides whether to run our performance or light engine for parsing (on a page-by-page basis) tldr: - most documents have a mix of complex pages intermingled with simpler ones. The router classifies and routes each page so you get the best blended performance x cost for any given document - you pay per page for whatever was used: 0.5 credits for our Light engine, 2 credits for Performance - our classifier works by scoring each page on layout, text quality, tables, form fields, and checkboxes - on RealDoc-Bench, enabling the parse router scored 89% (close to Performance's 90.9%) while using significantly fewer credits Parse Router is now our recommended default for most customers with mixed document payloads in mortgage, insurance, healthcare, and legal it's live today, set the engine to `parse_auto` and give it a shot on your own docs

@ExtendHQ

Today we're shipping Parse Aut

Today we're announcing Clarion's $10M seed, led by Accel with participation from Y Combinator. After a decade training as a physician at Stanford and Harvard, I left the bedside to fix how medical practices run. Here's why. The medical practice is one of the most important institutions in American life. There are over 200,000 of them, and they're running on decades-old technology and grocery store margins. Twenty years ago medicine digitized the patient chart, but it never digitized the practice. The rules that actually run a practice live on sticky notes and in the heads of the practice manager and their staff. As a result, the modern medical practice is illegible to AI. Clarion learns how a practice works from every call, message, and document that moves through it everyday. We turn those interactions into a world model of the practice, and our agents operate from it, handling the scheduling, referrals, and patient requests that previously consumed the front desk. Today, Clarion agents handle millions

Today we’re launching Global Volumes in beta. Now you can mount elastic, region-independent storage into a Pod in any Runpod data center. Just store a model once, deploy your Pod where the GPUs are available, and access the same files at /workspace-global. See how you can get started here:

• runpod.io

A few updates: → Folders now preview what files are inside → Folder colors are more distinct → Folders can now be duplicated → Folders (just wanted to say Folders again)

@figma

🗂️ Projects └🗂️ Are now > └🗂️ Folders > └🗂️ (We won) > └🗂️ Rolling out over the next few weeks

ROWBOAT JUST OPEN-SOURCED A MULTIPLAYER AI WORKSPACE WHERE EVERY TEAM MEMBER BRINGS THEIR OWN PERSONAL AI ASSISTANT. THE AGENTS COLLABORATE ACROSS SHARED CHANNELS WHILE EACH PERSON’S PRIVATE CONTEXT STAYS SEPARATE. > @_avichawla: >

• x.com

I know the company that made this absolute demon. Union Alpha: Circuit & Chisel, the team behind Unbiased/Pareto. Their Sept 14 release-note PR (still unmerged) names "union-alpha" as a per-account public name for Pareto.

• github.com

⚡️NEW: Ripple now support AI agents' online services payments with $XRP and RLUSD using a payments standard co-created by Stripe and Tempo. The updated XRP Ledger AI Starter Kit lets agents pay for data and computing without accessing private keys, with spending limits and approved destinations. The software is in beta. Ongoing payment sessions support XRP, while stablecoin sessions require a proposed ledger upgrade.

Great update from the @youdotcom team 👏 Research jobs that used to outlive the execution window now run in background mode and return a task id you can poll. Deep research stops being something you have to babysit.

@youdotcom

Your @n8n_io AI Agent can now run deep research on its own. > All 7 @youdotcom operations work as agent tools: web search, page fetch, answer, research, finance research. The agent picks what the question needs. > Type "" in the node panel to start.

• you.com

I spent so long treating "I'm fine, just tired" as a complete sentence instead of a warning. Seeing that exhaustion get named and solved on screen made me wonder how many times I talked myself out of asking for something easier.

@alexeichemenda

We raised $11M to make video editors obsolete. Not the humans, the tools. > Think about the hours you lose searching for assets, moving clips frame by frame, fixing animations, checking exports, and redoing the same edits over and over. > Poolday handles the entire video production process, start to finish, in one prompt. > It learns your style. Uses your assets. Makes the video. QAs its own work. > 100% on-brand videos. 100M+ video edits made for businesses all over the world. > And today, anyone can use it. > Tell us what video you'd like the agent to produce. The agent will build it for the first 50 people.

The real prize is procurement. European governments now have a home-grown stack to buy from instead of US hyperscalers. That changes the whole bidding landscape.

@kidtsang

Cohere + Aleph Alpha signed definitive merger Sept 16 (announced at ALL IN 2026 Montreal) — first transatlantic sovereign AI deal, dual HQ Toronto/Berlin/Heidelberg, 1,000+ staff, ~$20B valuation. > Combined with Mistral's €3B at €21B last week: the non-US/non-China AI stack just crossed $40B+. > The foundation-model market isn't fragmenting along capability — it's fragmenting along jurisdiction. Three pillars now define the frontier: US, China, Sovereign.

AI 矢量设计公司 QuiverAI 正式上线 Arrow 2 和更强的 Arrow 2 Telos。两款模型专门生成 SVG 矢量图。 相比上一代 Arrow 1.1，Arrow 2 重点提升生成速度和成图质量，并新增 SVG 编辑和动画能力。 QuiverAI 称，Arrow 2 会用更少、更精准的控制点画出路径，减少多余节点和重叠，也更会处理间距、留白和对齐。它还能直接修改已有 SVG，并给其中的形状和分组加入微动画。 更强的 Arrow 2 Telos 面向复杂风格、构图和长需求。它的上下文达到 105 万 Token，普通 Arrow 2 为 131,072 Token，相差约 8 倍；API 单价则高 50%。

@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.

Miden co-founder @azeemk argues the answer to rogue AI transactions may be simple: don’t ask the agent to behave, make the forbidden transaction impossible. "We ended up creating a solution to our own problem, which we call Guardian. You can think of it like a two-of-three multisig, where you have a hot wallet, a cold wallet key that's stored, and then you can have a third guardian that would be a signer." "As we started building out, we started seeing, wait, you can actually put an AI in here and give the AI parameters it needs to follow." "The human would be able to enable guardrails into Miden Guardian, and then have it do the things it wants while being comfortable knowing you can't somehow convince the agent to send $300,000 to someone." "But if you could encode that it literally can't, that's fascinating." @0xMiden

THE NEXT AI MOAT IS PROPRIETARY CONTEXT Not another model. The knowledge buried inside your company. @morphic_io is building the stack to capture it, structure it + keep it sovereign while agents work across your existing tools 👀 that's a compelling direction!

@joshsum_

today we launch @morphic_io: the sovereign AI platform around one core principle. > human knowledge is our most precious asset and needs to compound for the people contributing it. > we unlock our potential only when experience does not walk out the door every time someone leaves

'We believe RL is one of the most scalable and efficient paths toward self-improvement.'

@_LuoFuli

Nearly half a year of silence. We spent it studying one problem: how far RL can scale. > MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks. > Streaming the run:

• mimo.xiaomi.com

Say: “Buy milk, call Sarah, finish the report and book Friday’s meeting.” Instead of one long sentence, Voiskey can arrange the content into a useful list. Smart formatting is one reason it stood out and secured #2 on Product Hunt.

@voiskey

We did it 🏆 Voiskey just hit #2 Product of the Day on Product Hunt! Huge thanks to everyone who tried it, upvoted, and left feedback — this is just the beginning. Get Voiskey on iOS, Android, Mac & Windows Try Voiskey: Find us on Product Hunt:

• voiskey.ai
• producthunt.com

my wife (aka @tatieloulou) told me to go touch grass. my new AI assistant René agreed, blocked my Saturday and told her René about it 🤣

@tlxue

Today we're launching Rene. > A multiplayer-first iMessage agent you text like a friend. > It has a browser, writes code, goes shopping, ships sites, makes slides and images. Not much it can’t do. > It's been in my texts for four months. It found me a new office, preps me for every meeting, and polls the team for dinner options. > No app or signup. Link below👇

Love seeing US labs growing and thriving. Show @arcee_ai some love. 💙

@arcee_ai

Today, we are announcing our Series B funding round, valuing the company at more than $1B. > This round accelerates our next-gen Trinity models across diverse infrastructure, expands our work with the DOE and national labs on Genesis-Science-1, and enables us to build the platform teams need to build, evaluate, deploy, and operate open models in production. > We are grateful to our team, partners, open-source community, and investors. > Led by @Vista_Equity, Cambium Capital, and @emergencecap, with participation from AI10 Ventures, @Hitachi, IAG, @M12vc, @p7ventures, and @Wipro.

Okay.... fine i'll try it.

@noahrshinn

Your Instinct can now handle phone calls. > Introducing Instinct Concierge – a white glove service meant to handle high-touch cases, such as making phone calls, high-end service booking, and more. You’ll be able to book the restaurant that doesn’t take online reservations, get on your dentist’s cancellation list, or have that cable bill sorted out. > We’re slowly rolling this out to our early access group today and will be expanding access soon.

New in OpenWiki v0.5.2. @kirodotdev integration.

@colifran_

openwiki 🤝 kiro > openwiki v0.5.2 dropped recently and we added two new coding agent integrations. the first one i wanted to showcase is kiro! > coding agent integrations make it easier than ever to get started with openwiki > 1) npm install -g openwiki@latest 2) openwiki integrations install kiro > check it out ⤵️

Introducing Grade. Predict how users will respond before you ship. Grade is an eval system calibrated to what real users actually say and do. Get early access here -->

• voicepanel.com

✈️ @SouthwestAir started with an FAQ agent Then customers asked: actually… can you just do this for me? Agent Script takes it from answers to action: → AI Assistant handles the conversation → Routes back-office work to specialized agents → Keeps business rules fixed where they need to be + leaves room to reason where they don’t Action beats answers every time

America is banning AI in schools. China is using AI to create geniuses. Introducing Aristotle: The AI tutor that solves America’s broken education system.

• heyaristotle.com

Overwolf Ads Launches 'In-Game Audience Map' Product for Brands to Identify What Their Customers Play Across More Than 5,000 Games (EXCLUSIVE)

• variety.com

How do you improve a voice agent that’s already live? @VivintHome shows the loop at @Dreamforce: → Monitor the agent across conversations → Drill into the exact session behind an alert → Replay the call + inspect reasoning and latency → Use what you learn to shape the next iteration That’s Agentforce Observability

10,000 installs and counting! 📈 We created Modern Web Guidance to bring web platform expertise right into your agent workflows. Have you tried pairing these skills with Chrome DevTools for agents to audit performance and automatically fix code? Check it out and join 10K+ web developers who are coding smarter →

• goo.gle

.@FultonBank shows how Agentforce Coworker helps sales leaders get from data to decisions faster — live at @Dreamforce → Brings pipeline + activity into one view → Flags the opportunities leaders should focus on → Turns that workflow into a reusable Skill From digging through records → acting on what matters

腾讯 AI 语音输入工具 Chatterfly 已经上线官网并开启限时内测，目前开放 Windows 和 macOS 客户端。 用户按 Fn 直接说话，它会把口语整理成可以直接发送的文字，并根据当前场景和上下文调整表达。 Chatterfly 还把 Skills 直接塞进了输入工具。首批包括工作汇报、项目推进、营销文案、VibeCoding 提示词优化、会议纪要等 6 个 Skills。比如口述一段零散的开发需求，它可以整理成目标、约束和验收标准，再直接交给 Agent 执行。用户修改语音识别结果后，Chatterfly 还会自动学习专有名词，也支持手动维护个人词库。

• chatterfly.tencent.com

谢谢 @superalesha，很棒的 Skill，我做了以下短片

@superalesha

Your coding agent can make animations like this now. > I released the skill for drawing them in JavaScript, with 4 visual styles and examples to start from. This 1,5 minute film was made using it. > Pick a subject and try it:

• github.com

Mem0 is now on the Vercel Marketplace. Install it to give your agents memory that persists across sessions. A @mem0ai project and API key are provisioned, and billing is on your Vercel invoice. Free plan available. Run 𝚟𝚌 𝚒 𝚖𝚎𝚖𝟶 or learn more ↓

• vercel.com

BOOM! V9 is here and we are now tracking many more situations with the latest update to the prediction market local AI Model I built. It has studied outcomes of these platforms for over a year and can now have higher accuracy than V8.3. Join us!

• readmultiplex.com

Your AI agent writes code. Just never the way your project actually does it. GitLab's Duo Agent Platform fixes this with Skills, one file your agent reads every time so it just knows your rules. No repeating yourself every prompt. 🎥 @aniakubow

• about.gitlab.com

Custom Connectors are now in beta on Replit. Workspace admins can configure supported HTTPS REST APIs with API-key authentication, so teams can build with specialized systems beyond first-party connectors. Start using Custom Connectors now:

• docs.replit.com

Last night, I added a new feature to the MLXServe app, benchmarking. Many users were wondering what settings they should use, this will hopefully answer that question. Community results are viewable in the app, and on the website:

• mlxserve.com

Mintlify Automations, rebuilt Keep your docs in sync with code, draft changelogs, act on user feedback, maintain translations, fix broken links, and more Go take something off your to-do list and let us know what you think:

• x.com

Your Base44 Superagent can now make phone calls. It calls, chases the answer, and coordinates with people for you. Give it a multi-step goal and it's still working after you close the chat. Get your Superagent on Base44.

Our first ever stealth model is up on @CloudflareDev AI Gateway — Union Alpha! The sweet spot for the best price/performance ratio, this model is making waves as the next big hit... try it out and tell us what you think.

• developers.cloudflare.com

The lead went quiet. This AI agent figures out why. Use Lead Revival to read the history, follow up with customers who asked to reconnect, and bring their jobs back into view. Get the skill on Hyperagent Marketplace:

• hyperagent.com

Update: Cloudflare removed Union Alpha's "blended model" description and architecture label. Edit merged Sept 16, 9:00 PM EDT (Sept 17, 01:00 UTC). No reason given. Router/ensemble architecture remains unconfirmed.

• github.com

Conviction: Trading Desk in Your Pocket. Describe a trading idea in natural language. Conviction tests it on historical data and deploys an AI agent that trades on your behalf. Trade like a quant: Here's how it works:

• convictiontrade.ai

PDF in. Stunning PPT out. ✨ With Slides Convert on HIX AI, turn any PDF into a polished presentation with just one click. No manual formatting. No starting from scratch. No complex prompts needed. Try it now👉

• hix.ai

See a viral ad you love? Recreate it in one click with Topview AI Marketer! Keep the viral formula while swapping in your own products seamlessly #TopviewAI #AIMarketer #AIVideo #EcommerceMarketing #GPT6

Playing with @TencentHunyuan Hy MT2 1.8B with iOS27 and Core AI! In our last lunch/meeting the team was using it to translate my English to Chinese. Wondering if I can do the same optimized for iOS 27 🤔

• huggingface.co

$6M projected annual savings. 7x return. Customer satisfaction up 900%. That’s @SouthwestAir with Agentforce Justin Bundick, VP of AI, joins us at @Dreamforce to share how they got there

Your next site might be built already 👀 Portfolios, landing pages, dashboards, storefronts. All working templates, all customizable in Bolt. Open one, make it yours, hit publish 👇️

• bolt.fyi

You can't ask an agent to help with work it can't see. Connect Basecamp, Ramp, Asana, and TikTok Ads in Settings. Then it can work with the tools you already use.

• you.com
• hyperagent.com

🐡 Sakana Chatがアップデート 🐡 8月にはモデルも新しくなった「Sakana Chat」がさらにパワーアップしました。 （1）オーケストレーターモデルを最新モデルに刷新 （2）メモリー機能を追加 を新たに実装しました。 Sakana Chatを試す： Blog： 🐙

• chat.sakana.ai
• sakana.ai

アドビは、AI機能を強化した「Photoshop Elements 2027」と「Premiere Elements 2027」を発売しました。価格は各1万9580円で、両製品のバンドル版は2万7280円です。 ※購入から3年間利用できる「3年ライセンス」として提供されます。

• dc.watch.impress.co.jp

StepFunは、歌詞と楽曲のイメージテキストから日本語などのボーカル曲(インスト曲)を生成できる音楽生成AIモデル「StepAudio 3 Music」を公開しました。無料のデモも用意されています。 StepAudio 3 Musicで作った日本語曲(サウンドオン🔊)

🚨 GPT-6 Luna has ONE fucking job. Make people want to pay $20 for AI again. Because free AI is getting fucking ridiculous. 👀

Tracks real-time stock prices and sends personalized alerts while visualizing market data through interactive heatmaps

• github.com

Your agents should be getting smarter. Build them on fresh data and context that improve over time with Redis Iris:

• redis.io

Tracks job applications, reviews resumes, and writes cover letters with a self-hosted AI assistant.

• github.com

竹中工務店と爽美は、手書き図面を撮影するだけで、AIが内装仕上げ材の加工用CADデータへ自動変換するシステムを共同開発しました。 これにより、技能者によるCADデータ作成の手間を削減できます。

• nikkei.com

🔴 LIVE: Agentforce Keynote at @Dreamforce

• x.com

LTX 2.5 💋 N$FW v1.1 FP8 👇

• huggingface.co