I use @pidotdev as my daily driver because it stays lean. Pi gives me an efficient, understandable core and lets me add capabilities only where my workflow needs them. These are my current favorite extensions: • Ponytail -- nudges the agent to reuse what already exists before writing more code. I use it to keep implementations focused and avoid unnecessary complexity. pi.dev/packages/@dietrichgeb… • pi-web-access -- gives Pi web search and access to pages, GitHub repositories, PDFs and videos. I connect it to my own SearXNG instance so Pi can research a topic and bring the relevant context into the session. pi.dev/packages/pi-web-acces… • pi-subagents -- lets me delegate focused work to different models. I use it to run research or reviews in parallel and return concise results to the parent session. pi.dev/packages/pi-subagents • pi-fff -- adds pre-indexed fuzzy file and content search with frecency ranking. It makes larger repositories noticeably faster to navigate. pi.dev/packages/@ff-labs/pi-… • pi-context-view -- estimates how the context window is being used, including the base prompt, tool definitions and extension injections. I use it to check the context cost of every capability I add. pi.dev/packages/pi-context-v… What extensions have earned a permanent place in your Pi setup? – at Kelsterbach, Deutschland

Liked the idea so much that I built a quick rust-based proof of concept for OSX. I love what AI enables me to do in minutes 🚀

Hard disagree. Local models make AI more accessible, support Sovereign AI and hybrid AI use cases. It helps push innovation.

I guess this is where we should thank the Pokémon Go community?

I see a lot of finger-pointing in the AI community, and it probably doesn’t help anyone. My mantra: A rising tide lifts all ships. @MiaAI_lab has credited me several times for tool-eval-bench even though she could have just pointed to a fork or her own project. Instead, she contributed to my project to help make it better for everyone. Let’s not forget that we’re all trying to make things a little easier and better for everyone.

If people think this is a joke - it’s not: ai.gov.uk/knowledge-hub/how-…

Restarting a large model in vLLM means waiting for every TP shard to load again, even when the checkpoint hasn’t changed. vLLM 0.30 adds a way to keep processed weights in per-node GPU daemons and map them into the new engine through CUDA IPC: –load-format ipc_cache It supports DP4 and multi-node TP. I’ll make sure to test this with my DGX Sparks! github.com/vllm-project/vllm…

Kudos to the llama.cop team and all contributors. A pretty big release with tons of optimizations across the board!

Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS just released and I knew instantly what I had to do. 🗺️ Given all the excitement for WoW Forever, I built a quick proof of concept using the Lite TTS model for an AI-powered questing addon for a more immersive time in Azeroth 🎙️ Generation is super fast and the audio quality is impressive 🎭 The model applies the persona and pacing very naturally ➡️ Learn more here: blog.google/innovation-and-a…

tool-eval-bench v2.7.0 just released: github.com/SeraphimSerapis/t… In addition to many fixes and improvements now also with support for --per-request-spec-decode-metrics which recently landed for vLLM: docs.vllm.ai/en/stable/featu… These metrics help you evaluate the effectiveness of your speculative decoding setup far more effectively. For those who love tinkering: tool-eval-bench picks up this flag when it is exposed and gives you more accurrate serving metrics for --spec-bench ⚡️

Swapped to iOS (iPhone 18 Pro Max) for the first time since the iPhone 4s to see what the other side of the fence looks like. I immediately miss the following features: - Granular sound and notification profiles (e.g., what should vibrate and how intense should vibration be?) - Schedules (aka focus modes) are present but less intuitive - Couldn't transfer my WhatsApp chat backup AFTER the initial setup. Well, I guess I don't need WhatsApp anymore 😅 What I like: - Stacked widgets are really well done - Flighty. So much nicer than other travel apps - USB C. Finally – at Großbeeren, Deutschland

This took way too long. Less fragmentation is always a good thing. – at Großbeeren, Deutschland

The M5 Ultra benchmarks look very promising. The question remains if prefill performance has caught up, too. Looking forward to real-world benchmarks as they start releasing next week! – at Großbeeren, Deutschland

Gemini 3.8 Live is truly impressive and lets you build low-latency, multilingual experiences such as this little app I built for my @pollenrobotics Reachy Mini.

I genuinely enjoy AI hardware posts. But whenever pristine RTX 6000s appear neatly lined up on a table rather than inside a server, I imagine the content workflow: - Stop inference and shut down the server - Remove every GPU - Take the photo - Rebuild the server - Debug the card missing from nvidia-smi - Check that no power connector is melting Respect the commitment. 🫡

This is pretty great work and helps to bring together two ecosystems in a way that will help drive local inference forward. The DGX Spark excels at prefill, Mac Studios bring excellent decode speed. Together, they'll be able to offer an excellent experience. I'll definitely follow the progress here! – at Großbeeren, Deutschland

Agent Skills have had an awkward deployment model: MCP delivers the tools, while the instructions teaching an agent how to use them live in a separately installed folder. The two can drift. SEP-2640 was merged into MCP on Sep 13. It defines a standard way for an MCP server to distribute tools and the skills for using them through the same connection. ⚙️ The useful mechanics: - skills/list exposes lightweight metadata and manifests - skills/get refreshes one skill’s manifest - SKILL.md and supporting files are retrieved only when needed - SHA-256 digests let hosts verify cached content 🔐 If any file or manifest changes, content-bound approval is revoked. A server can update its guidance without requiring a reinstall, but changed instructions do not silently inherit the previous approval. 🧭 For the skills manager I’m building, this suggests a practical boundary: - Local skill -> stable, standalone knowledge I own - MCP skill -> guidance coupled to live tools, authentication, tenant or server version This does not make every harness dynamically load skills overnight. Hosts still need to implement discovery, loading and consent. But it standardizes the missing wire between skill authors, MCP servers and agent clients. github.com/modelcontextproto…

So.. tempting. Must.. resist. All jokes aside: this and the Ocarina of Time release are serious flashbacks. – at Großbeeren, Deutschland

I love interacting with my agents via voice. Google's AI Edge Eloquent for OSX just got even more powerful: it supports a variety of languages allowing you to switch based on the necessary context. All powered locally by @googlegemma The app is available for free for OSX and iOS: developers.google.com/edge/e…

What's cooler than AI? AI in space. Now featuring @googlegemma 🚀 – at Großbeeren, Deutschland