My Hermes agent got access to my Obsidian vault. But check out how nice a terminal interface can be, even with file path completion. Who needs a GUI if you have something like that? 🤯 My main way to talk to my Hermes agent is currently: hermes chat (inside the terminal)

GitHub (Microsoft) uses your repositories (including private ones) to train their AI models since the 24th of April. 😱 If you don't want that, you have to opt-out as follows in the settings:

Totally unexpected after the recent release of M2.5, now we have M2.7, improved across all benchmarks. 🤯 Especially interesting: The new Minimax Claw Bench shows M2.7 catching up fast to frontier models when used in OpenClaw. More info about M2.7 in the official article. 👇 x.com/MiniMax_AI/...

Google just banned some accounts using @openclaw and / or automated API calls. So crazy that with your a warning one of the most popular login methods for 3rd-party services can just disappear on you. Own your email. Own your domain. Don't delegate your digital identity! x.com/i/status/20...

Kimi K2.5 from @Kimi_Moonshot is a great open model and works fine in OpenClaw at a really competitive price! 🫶 But the coding plan usage runs out way too quickly for me when I use the model in OpenClaw (even with increased quotas, as they currently state on their page). 🥺

If ChatGPT or similar tools complain about "too much text," head over to @Zai_org With the new GLM-5 model and a 200k context window available on the web UI at "z .ai". No signup is needed. Looking forward to reading your experiences in the comments below. 👇

Quick reminder: For Kimi coding plan subscribers. You grabbed the Kimi coding plan for $1 and don’t want to pay full price next month (at that price point, Claude is probably the better pick) ➡ cancel now and keep using it through the end of your trial month.

Google has begun banning users from its AI models when it detects access through third-party tools like @OpenClaw rather than the official Antigravity IDE. A growing number of users are reporting the issue in a Reddit thread. 👇 Take care, especially if you subscribed for a complete year!

For what is the "boot in 1s" feature needed? My @OpenClaw bot does not remember when his machine booted the last time. 🤣

I jumped on the bandwagon, and I have to start talking about it: @openclaw is just remarkable. It is so crazy that you get stuff done by just asking for it. One example from me: Make OpenClaw inform me about an updated release and tailor the message based on the usage it is seeing on my system.

This n8n workflow sends me a Slack message when there is a new entry on the @Alibaba_Qwen release site. 👇 Looks similar to the one I shared yesterday, but it needed some help of ChatGPT to extract the right call to get the entries via a JS call. 😅

After missing yesterday’s @Zai_org release of GLM-4.7-flash, I built this n8n workflow. 👇 It checks their release page and alerts me on Slack if there was a new release in the last hour. You might ask me for the details to do it yourself for any page you would like to monitor.

To try Claude Opus 4.5 for free, just install the VS Code fork @antigravity from Google. Fair warning: The model is so outstanding at coding, you'll probably find yourself signing up for a paid plan - just like I did. 🙈 Link 👇

Just looked for a feature in my favorite editor @zeddotdev, only to realize that it was just added, and I have to update my current installation. Now, I am able to set @MiniMax__AI M2 as my default model in @opencode running as external agent in @zeddotdev. Open-source power! 💪

The weekly token usage on @OpenRouterAI shows always interesting insights: 1. @grok Code Fast 1 is still the by far most used model 5. @MiniMax__AI M2 has currently the biggest momentum with +34% usage 8. Sherlock Think Alpha is a new clocked model, rumors point to Grok. 🤔

Today's Cloudflare outage is a great reminder: Decentralization isn't just a buzzword, it's a necessity. ChatGPT and Claude are down. Happy to have my own local AI setup that's actually reliable. 😎

😄🕐 AI Happy Hour: AI news, open-weight AI and more Looking forward to talking to my AI buddies and everyone else who is interested in these topics. 👇 (link to join in first comment) 🚨 Again on DISCORD instead of an X-space, today at 7 pm (UTC), which is the next full hour!

🚀 Cut latency, ditch the Cloud!

Greetings from the AIDev 5 (an AI developer meeting) which this time is part of the AI Week Frankfurt event. Next is a talk from Michael Dietz, Customer Engineer at Google, about AI agents using the MCP and A2A protocols. 🤓

Digitizing printed text via optical character recognition (OCR); - Classic OCR fails with gibberish and your spellchecker catches it. - AI OCR fails with plausible words that fit the context perfectly. The error is really hard to catch. Is the latter really what we want?

There is a new star in the world of open models on the horizon: MiniMax M2 Given that this has less active parameters and is sparser than other open models, the benchmark results are pretty good. 👇 However, you should vibe test it yourself via OpenRouter (currently for free).

32 books on AI and humanity for $16. That's the top tier of @humble "The Human Side of AI" collection. It starts at just $1 for 3 books. The bundle contains @mitpress titles that tackle the question: How do we understand our relationship with machines? 👇

GLM 4.6 is the most powerful open-weight model for coding. With 375B params, it's likely too large to run it on your own hardware. Subscriptions from z. ai start at $2.70 (code in comment) for the first month. Use it on platforms like Claude Code and OpenCode. 👇

🚀 Attention all LLM enthusiasts! 🚀 Leaks share specs of the RTX 5070 Ti Super, the next-gen powerhouse with 24GB VRAM at an unbelievable $800. Prepare for a game-changing shift in the market for power users handling large language models on their own devices. 👇🧵

If you don't want Anthropic to use your chats from Claude to train future models, make sure your privacy settings are correct. Go to settings (1) -> Privacy (2) -> Turn "Help improve Claude" off. 👇

As everyone explores GPT-5, it reveals a dark truth: today's AI might disappear tomorrow. OpenAI's unexpected model replacements show how cloud dependency creates "digital quicksand" for users and businesses. More about it and a way out in my latest newsletter issue. 👇

Download the latest Kiro update! By being an early adopter of Amazon's VS Code fork, Kiro, I kept getting an update alert that annoyingly led me to a waitlist. But I've discovered a little-known subpage that lets current users directly access Kiro releases for installation or upgrades. 👇

Your future self will thank you for learning AI tools when they were still novel.

Everyone should think about building an AI wrapper for his biggest pain Skip building your own model. Move fast, grab your niche, and collect the data that'll give you the edge later.

Kiro, the new Visual Studio Code fork from Amazon has now a waitlist. 🤔 Released on Monday as preview, and so far free to download and register. That changed, likely due to the huge interest of vibecoders who want to test the more structured approach of Kiro.

● 🚀 Just got Amazon's new coding editor Kiro running on my Arch Linux! Unlike for Windows, MacOS and Ubuntu, for Arch Kiro is provided as a complete self-contained folder that you simply move to /opt. Everything's included: the editor, dependencies, and even a beautiful icon.

Just discovered Gemini-CLI dropped end of June 👀 Already a fan of Claude Code (now part of the standard Claude subscription), but Gemini Pro’s coding capabilities are also 🔥 Testing both side by side on a new project: ➡️ Extracting my writing style from my Substack newsletter

Ever wondered how model quantization (FP16, Q8, Q4) *really* affects performance? There's an analogy that makes the trade-offs crystal clear... and it involves something you might drink. 😉🍺 Kudos to @jtdavies for this brilliant comparison. 🙏 See the image for the full explanation! 👇

Sam Altman shares a glimpse of the future he is seeing in front of us. 👇 The changes will be huge, and now is the time to prepare yourself for it by AI-first thinking. BUT, do it problem centric: Whatever you are working on, think about how AI can support you. 🤔

Magistral was released by @mistralai. Their first stinking models. 🎉 You can run locally the small version of Magistral, which builds on top of Mistral Small 3.1 and has 24B parameters. But why does it not have the tool support in @ollama which it should have? 🥹

Summrize → Blinkist, but for free Summrize offers entirely free book summaries of popular nonfiction books. I just checked the summary of the latest book I have read, and I have to say it is quite good. I'd use it to decide whether to read a book, not instead of the book.

🤯 Blown away by FLUX.1 Kontext from @bfl_ml! I wanted to edit a photo of me for my upcoming newsletter issue (wasn't looking my best! 😉). Used the new Flux model for minor, realistic tweaks. The result? It still looks like me. 👇 Does it look like me to YOU? Let me know! 🤔

The likely most convenient way to get the best prices at using generative AI models via API is @openrouterai On top, they provide nice usage statistics. 👇 - overall usage is going up - smaller and with that cheaper models are mostly used - model usage differs along categories

🤔 Noticing some variability in image gen quality in ChatGPT-4o? Look closely to see if you get an image from the new model or the older DALL·E 3 during peak load. The new model builds the image line by line. 🧐 🚀 EU users: The new image generation is LIVE - no VPN needed! 🙌

Unlock the Best Open-Source LLM for FREE: Llama 4 Maverick on OpenRouter - use it via API or directly in their chat interface Honestly, it's a mystery how it's possible, but you can tap into countless models via OpenRouter's API at no cost. It's absolutely incredible!

New issue of my AI newsletter with a focus on privacy-friendly AI which you can run locally: 🤝 AI without surveillance: How local models are changing the game This issue - On-device AI is the future - DeepSeek V3's impressive performance - Small teams leveraging AI for growth

I don't know how many years I used the open-source Photoshop alternative Gimp 2.X. 🤔 But today I was greeted by the following welcome message from Gimp 3.0 🥳

Just stumbled across Interview Coder: An AI that secretly helps you code during video interviews. Invisible to screen recording & the author shows how it helped him to get an Amazon offer. 🤯 Is this the future of tech interviews? 🤔

GeForce 5090 below €3000 for the first time in Germany. 🎉 5 models are currently below the 3k barrier and getting closer to the price named by Nvidia from €2229 (which was lowered a couple of days ago because of the strong Euro compared to the Dollar from initial €2329)

A meta agent designing agents 🤯 Currently listening to Shengran Hu, presenting his work with the title "Automated Design of Agentic Systems (ADAS)" in a Zoom call hosted by @AITinkerers

Excited to share the launch of our family business "Druchs" in the upcoming newsletter issue. 🤝 Our AI-driven family business: Turning Passion into Business: Our Journey with Personalized Mugs And my thoughts about the Mistral Small 3.1 release and Claude Internet access. 👇

I turned my son's favorite cuddly toy, Miezi, into his new go-to mug with just one photo! Huge shoutout to @bfl_ml for the outstanding model that made this possible. 🙏

Project Digits -> DGX Spark I'm disappointed with the announced memory bandwidth of only 273 GB/s. I was hoping for 500 GB/s based on discussions in the Locallama Subreddit. The AMD Ryzen™ AI Max+ PRO 395 (Strix Halo) seems more interesting now: similar memory bandwidth but x86 🤔

📢 Two new PebbleOS smartwatches available for pre-order! Core 2 Duo (B&W display, July shipping) - $149 Core Time 2 (64-color display, December shipping) - $225 I didn't pre-order; not a fan of the Pebble 2 design & hoped for a return to Pebble Time's look or something new.

The best time to start with AI was yesterday; the second-best time is now.