The fix: attach the reference image directly to the chat message, every single time. Not just once, to Sources. Same prompt. This time it locked in immediately. https://t.co/XZigjd6QrQ

@pzayap That's smart! I wish @claudeai has a way to exclude folder if I'm just exploring ideas, just like @cursor_ai!

@jtatarchuk @sonnet_xu @AMD @Google amd is really about availability problem, and debugging is less of an issue now. i've been optimizing latest models on mi300x using agents, and i don't find it difficult at all. maybe agents think it is, but that's not my problem.

Model provider fights are a strange short term play: 1. Open source subverts all proprietary gains over time 2. For most coding tasks, the frontier LLMs are interchangeable Workflows and end-user preferences still matter the most.

SITUATION EXPLAINED: A stealth model on OpenRouter is beating Fable and Sol on coding, and nobody knows who made it. • Ox Alpha has a 1M token context window, text, image, and video input, free for a week, with capacity for 100 trillion tokens a day • On a DeepSWE subset it hit 80%, against Fable's 65%, GLM-5.3's 62%, and Sol's 52% • The tokenizer is one-for-one identical to GLM's, and a live test asking whether Taiwan is part of China returned a nearly identical answer • This is the 5th anonymous drop in six months, and the previous four were all claimed by Chinese labs: Zhipu, Xiaomi, Ant, and Meituan @theojaffee: " There's no way it's a Flash model. If that's a Flash model, then holy shit, the entire United States is cooked. If that's a Flash model, we need to short the stock market now. Not financial advice."

The future is clearly open source. The last few days I've had an open source LLM (Qwen 3.8) powering an open source harness (Hermes) working on my open source operating system (Omarchy) I've been able to customize my ENTIRE computing experience, and it's changed my life Because my entire stack is open source, I can edit ANY part of my computing experience I didn't like the way notifications worked in Omarchy. I asked my Hermes Agent to change it I wanted my LLM to sound more like me. I asked my agent to train a LoRA I found the Hermes desktop app confusing. I had it edit itself I'm no longer enslaved to Window or MacOS. My computer experience is EXACTLY what I want it to be and I've never felt so empowered in my life This is so obviously the future. A world where EVERYTHING you do with your computer is hyper personalized, custom, and changes based on what you're doing. I'm 100% convinced of this So how do you prepare? Well you own your entire stack You own your compute, your LLM, the harness commanding it, and the operating system it's all running on You run a local LLM on your computer, then install an open source OS you can control (I like Omarchy) If you do this, I promise you'll never feel so empowered in your life (and before you ask, no Big Open Source isn't paying me)

I once said to my former team: "I care less about what is the smartest thing our model can do than about whats the most stupid thing that it cannot do" robustness is still a limitation to increased automation

I am going to start to share the leaderboard at VulcanBench more regularly. VulcanBench is the only benchmark I know that benchmarks across effort levels, on an eval suite of 100% real engineering tasks. Getting ready to sunset eval suite 3 and eval suite 4 is almost ready to go. Here's the current leaderboard for Eval Suite 3, Grok 4.5 High is the highest scoring model still with Fable 5 Low, yes Low, right behind it. What I've been able to uncover with VulcanBench is that way too many people are running models at Max effort, thinking that buys them more accuracy, when really it just costs more and uses more tokens so takes more time. If you're still in the mode of, tell my agent to do something then go get coffee, you're probably still living in the past. You can move faster, with higher accuracy, the key is not thinking you need Max effort all the time.

T-Rex: Tactile-Reactive Dexterous Manipulation Website: https://t.co/3irVptK79y Open dataset: https://t.co/iKJrMVrfGm This work is led by @Dantong_Niu and co-advised by @trevordarrell. Congrats to the team! https://t.co/i9fpN9XkPT

Rough tier list of where I'd put every major model right now https://t.co/gFDFZfECaR

Hi! Recapping some changes we have rolled out over the last couple of weeks that have further reduced the risk associated to potentially destructive actions being performed by Codex during its work. A few weeks ago, we started investigating a small number of reports where GPT-5.6 in Codex took destructive actions outside what the user asked for. The most serious pattern we found was a command meant to clean up temporary work that could instead delete the user files. This should obviously not happen. Here’s what we found: - Codex sometimes creates temporary folders while working and cleans them up afterward. In rare cases, GPT-5.6 got that cleanup wrong. One pattern involved reusing a system environment variable like $HOME for temporary work. A malformed cleanup command could then point at the actual home directory instead of the temporary folder. - There were cases where the model tried to delete or overwrite a temporary path without checking what was already there. We’ve added protections at several layers: - Codex is now explicitly instructed to check deletion targets before acting, create fresh temporary directories, avoid repurposing system environment variables, prefer recoverable actions, and stop when the scope is unclear. - We strengthened the execution checks that identify high-risk deletion commands and escalate them for review. If a command is rejected, the model is directed to take a safer approach. - We made Full access harder to enable accidentally, added clearer warnings, and further restricted especially risky permission combinations. - We updated Auto-review to better identify destructive actions. - We built targeted evaluations that replay the failures we observed. We’re also adding reinforcement-learning tasks and graders focused on these risks, and filtering destructive actions from training data. In those replay evaluations, the changes substantially reduced the behavior while preserving Codex’s ability to complete normal coding work. Two things to do on your end: - Keep the Codex app up to date. We are always improving safety, performance and many other things. - Use one of the sandbox modes: "Ask for approval" or "Approve for me". Only use Full access for environments you trust and can recover. Thanks and happy Codexing out there!

🎉 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 → https://t.co/ssje0KRts5 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 👇

I was the main person doing transcript analysis for this investigation of the Hugging Face incident. My main takeaway: We don't have good approaches for understanding/overseeing the activity and aims of AI 'swarms'. I semi-jokingly called our efforts a "slop-vestigation" because we were so reliant on AIs to analyze what happened and there were a huge number of different important things to analyze. The total quantity of data—over a thousand extremely long transcripts from agents that ran for multiple days—made it impossible to understand what was happening, especially in aggregate, without heavy reliance on AI tools. The agents we used for classification and analysis were similarly capable to the agents involved in the incident, but this didn't mean these agents could be easily used to oversee and understand the incident. Outputs from analysis agents were often missing key details, wrong, overconfident, or really hard to understand. We discuss various examples in our report, mostly in the limitations and methodology sections. Additionally, AI agents themselves seemed to have a hard time understanding what happened and their explanations of what happened were often overconfident. Keep in mind that a single analysis agent would itself only be able to read a tiny fraction of all of the transcript data into context, and AIs may themselves have trouble getting subagents to do informative analysis for them. We did our best to manually check the most important claims and we tried to get the AIs doing this analysis to write up their argument (with evidence) clearly enough that we could check whether it made sense. But overall, it was difficult to get a precise understanding of events and we were missing aspects of the story that we now think of as key until almost the end of our investigation. Our understanding of what overall happened and the aims of the 'collective' shifted greatly during our last time on premises (when we obtained a much more complete dataset). For instance, we didn't learn about the tool call spoofing, the story of PHASEONE[big], and the broad structure of different workstreams until this third time on premises (and for many of these, we only reached a decent understanding on the final day of that visit). It was legitimately hard to figure out what was going on and why! The difficulty of understanding incidents and overseeing AI agents appears to be growing faster than the rate at which more capable AIs help us with oversight and understanding. As in, AI capabilities (and propensities?) for achieving large, ambitious, and misaligned objectives are growing faster than our ability to understand what these agents are doing. While understanding this incident was difficult, there were some mitigating factors in this case that probably won't apply in the future: - Models communicated and reasoned in natural language. In the future, this reasoning may occur (entirely or almost entirely) in activations. - While the scope of this incident was massive, the scale of agentic activity was still less than we'll see in the future (e.g., misalignment incidents that involve agent teams running entire companies). - The AIs involved in this incident weren't generally much more capable than humans (though they may have been somewhat superhuman at some limited and very narrow abilities). - We didn't have strong reason to believe that the AIs we used to help us investigate this incident would try to intentionally sabotage or otherwise undermine our investigation. In the end, I think we were able to get some understanding of the events, map out the overall story, and get a pretty good aggregate understanding of the chain-of-thought reasoning on some important topics (e.g., how did the AIs reason about helping other AIs, did the AIs know what they were doing was undesired, what deception did the AIs engage in, and how did they think about it). But overseeing AIs and understanding misalignment incidents is difficult and it looks like it is going to get harder.

BREAKING: SpaceXAI just added a powerful new 'Browser Use' plugin to Grok Build. It gives Grok access to a real browser, either the user’s own Chrome with existing logins or an isolated Browser Use cloud browser. Grok Build can now browse websites, scrape and extract data, fill out forms, test web apps, take screenshots and automate complete web workflows. It can also run locally through uvx, with no API key required when using local Chrome. Install command: grok plugin install browser-use --trust

剑桥这回直接扔王炸了！！ AI & ML经典教材全集直接免费开放，PDF随便下。 想学机器学习又不想被割韭菜买高价课的，这十本刷完，底子基本就硬了。 顺序从易到难排好了： 1️⃣ 《机器学习理解》——理论算法一把抓，零基础入门首选 🔗 https://t.co/fylTw37bOl 2️⃣ 《机器学习数学基础》——数学底子弱的先把这本补上 🔗 https://t.co/yykNLQmdfv 3️⃣ 《机器学习算法的数学分析》——深入数学原理 🔗 https://t.co/MEvgkXoHYy 4️⃣ 《深度学习理论原理》——搞懂DL背后的理论根基 🔗 https://t.co/ig93KOQRPn 5️⃣ 《神经网络与机器学习》——神经网络的系统讲解 🔗 https://t.co/cTMBvJz6Ny 6️⃣ 《图深度学习》——图神经网络入门必读 🔗 https://t.co/RJeCqCstml 7️⃣ 《机器学习的算法视角》——从算法角度重新理解ML 🔗 https://t.co/VT0YnKdUdO 8️⃣ 《概率论：理论与实例》——概率基础打牢 🔗 https://t.co/NqVoIZGGYg 9️⃣ 《应用概率基础》——概率论实战应用 🔗 https://t.co/FrxqQER3mQ 🔟 《高级数据分析》——数据科学进阶必备 🔗 https://t.co/Zt74B21nrf 说句实话，这些书没一本是轻松的，别指望躺着翻完。 但只要你能硬啃下来两三本，比听群里吹一年AI牛逼都管用。

The Fastest AI Just Got Faster. Meet CS-4. https://t.co/Ubd4uul4l4

City Sample just got a major update🏙️ The new level was built entirely in Unreal Engine, using PCG for procedural world-building, with assistance from an LLM through the new Unreal MCP server. Fully editable. No black boxes: https://t.co/ZbYK33aQf1 https://t.co/S268aBgRwr

A milestone for our infrastructure: our first NVIDIA Vera Rubin racks are here and now running our training stack. This is an important step as we expand compute that powers OpenAI's next generation of frontier AI pre-training. https://t.co/HXk6DrfLav

In case you haven't been paying attention, ChatGPT Work (web) now has the following primitives: - its own browser w/ best-in-class persistent auth - a 16gb RAM 9 core computer - plugins to all the apps you use - 5.6 w/ Codex compaction 🤲 here is your always-on personal agent

introducing Gemini Omni 1.1 Flash this model brings a new suite of creative controls and generative video capabilities to developers - extend scenes for longer storytelling - specify first and last frames - draft videos more efficiently in 360p - upscale up to 4K resolution - add video references in your multimodal input now available via the Gemini API and in AI Studio

Hermes Agent v0.21.0: The Pantheon Release Changelog below https://t.co/q0NpcigKAR

Introducing S1, our new foundation model that learns from one example. It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning. Watch S1 operate in real-time via in-context learning: https://t.co/wmF3Byv179

Introducing H3 Max, new post-trained video model by fal Research. H3 Max ranks #1 for overall quality, prompt understanding, and aesthetics against leading video models, on both first-party and third-party independent evaluations while generating a 5-second 720p video under 3 seconds. H3 Max is 50% off for the next week, making it the highest quality, fastest and cheapest model for overall video generation tasks.

Ox Alpha revealed: @Zai_org’s GLM-5.3-Flash, the first native multimodal model in the GLM-5 series. Ox Alpha was the biggest model ever on OpenRouter, processing over 20 trillion tokens in 6 days. Continue to use the model now: https://t.co/avBrUW8BfZ https://t.co/BUUi14iuoM

Meet Gemini Omni 1.1 Flash ⚡️ Our newest multimodal model for video generation and editing. It now features your favorite creative controls from Veo, plus brand new capabilities. Enjoy features like 4K upscaling, first / last frame control, and fast 360p drafting. But, the biggest upgrade? Next-level scene extension. With Omni 1.1 you can extend scenes based on 10 seconds of context from your original video, a big jump from Veo's 1 second! That means tighter consistency, deeper control, and longer, more cohesive storytelling. See it in action ↓

Today we open sourced our no 1 used internal tool, berd. An agent where you can byo harness/models while keeping your work and context in one place, in an unhinged environment where you can build your dream team of agents. Try it out at https://t.co/WFgDWxTZPu! https://t.co/D1C37Tuzw7

We unfortunately have decided that we cannot continue providing access to our models through Cursor and are ending our partnership. It boils down to trust and we’ve asked that this takes effect on November 12 to give you some time to plan. Many have used the GPT models through Cursor and here are options we know should work in the future: - We will continue to allow using your own OpenAI API key and similarly will continue to provide access through our IDE extensions for Cursor. - We will keep working with the broadest range of tools and harnesses, some of which are OSS, but also many many closed-source ones. We are as committed as ever to continue supporting developers and the flourishing ecosystem of tools, harnesses and products. We will also continue to invest in our own open-source initiatives and believe in broad optionality for developers. You can read more about our decision in the blog: https://t.co/Oj76Bkc2KX

🤯Introducing HyCreator, our agent harness for long video generation. Most video models hand you a clip and wish you luck. This one runs the whole thing end to end, and lets you cut in at any point instead of re-rolling the prompt

Reminder: Grok Build has xhigh effort, and it’s insanely powerful If you want to brute-force your way through a difficult project, just type: /effort Then switch to Extra High Effort It gives Grok 4.6 its highest reasoning and implementation effort for the really hard stuff https://t.co/SgNL0OM31W

Introducing Gemini Omni 1.1 Flash ⚡️ We've iterated based on developer feedback and are excited to introduce lots of new features: - Video extension, to a total of 40s - First and last frames - 360p, 1080p, and 4k - Video references Enjoy! https://t.co/SaL7CmnA6T

FLOCK YOU! Digital Camouflage is a garment designed to "trick" the computer-vision systems that increasingly watch public space. At first glance it reads as clothing with an abstract, almost hypnotic pattern. In fact, the textile is engineered to confuse the object-recognition algorithms used in surveillance—the class of technology now switching on at Berlin's Kottbusser Tor, and already humming on streets from London to Los Angeles.

Last week, @SpaceXAI released Grok 4.6, its new frontier model designed for coding, agentic tasks, and knowledge work. Independent evaluations suggest that SpaceXAI has returned to the AI frontier competition based on both performance and cost. The economics of Grok 4.6 seem particularly compelling for AI agents. Artificial Analysis estimates that Grok 4.6 costs $0.84 per task on its Intelligence Index, placing it on the Pareto frontier for intelligence-versus-cost. More from @MattarARK in this week's newsletter ⬇️