Zero to MVP

In a recent video I showed GridCore, the scheduler I built so several local LLMs can share one GPU. I've now written up how it works under the hood, and what only broke once it ran on real hardware: • background jobs starved completely under constant chat • the scheduler evicted two models to free 200 MB • ~11 GB of "phantom" VRAM from a race between two nvidia-smi calls • random 502s from dead keep-alive connections Plus how it estimates a model's VRAM from the GGUF header before loading it, and why it never touches llama.cpp's slots. https://blokhin.us/notes/gridcore-gpu...

My GPU was 86% busy with a background job. I asked a 12B model a question anyway. First token: 200 ms. Time in queue: 0. The background job paused itself, waited for my answer, then picked up exactly where it left off. No model reload. No crash. No waiting. With Ollama, the same switch cost me 6 seconds. Every time. So I built GridCore, a scheduler that decides which local model gets the GPU right now. 4 models in 15.5 of 16 GB, and chat always comes first. Any app that speaks the OpenAI API works with it, with nothing to rewrite. Full demo: https://youtu.be/Mu3xzCVoHXc What do you run on one GPU at the same time?

I finally open-sourced my coding agent: Green Agent For the past few months I've been building an agent for myself: a terminal assistant that reads a project, searches the code, edits files, and runs the tests. Today the repository went public. It is heavily inspired by Pi agent and shares the same idea: a small, readable codebase instead of a huge framework. You can read the core in one evening and understand exactly what happens between you and the model. But it also has a few things of its own: - Full-screen mode: the project tree on the left, a syntax-highlighted file viewer, and the viewer reloads by itself when the agent edits the open file. You see what it does, not just what it says. - A browser: the `fetch` tool lets the agent open documentation or any web page; HTML is converted to clean text. - Smart approvals: inside a git repository the agent works without interruptions as long as git can undo what it does. Overwriting an untracked file, `rm -rf`, `sudo`, `git push`, `curl` always ask, even if you answered "always" before. - Any OpenAI-compatible API: OpenAI, Gemini, Ollama (cloud or local). If the model is busy, the agent switches to a fallback from your list on its own. - One dependency: `openai`. Textual for the full-screen mode is an optional extra. - Tools are just folders: `tools//` with a JSON definition and a Python file. Adding your own takes five minutes. - One-shot mode for scripts and CI: `python start.py -y -t "run the tests and fix what fails"`. Under the hood it is about 3,000 lines of Python and 500+ tests that run offline. MIT license. Repository: https://github.com/w512/green-agent Issues, ideas, and reports on which models worked best for you are very welcome. If there is interest, I'll make a video walking through the internals: the agent loop, the permission system, and how the tools are built.

II just published a new article on Patreon about GPT-6 Astra: https://patreon.com/ZerotoMVP The thing I find most interesting is not that it’s “smarter,” but that AI is getting better at working on its own for hours or even days. I wrote a short piece about why this shift matters, especially for coding, research, and the way we may start delegating real work to AI. If you’re curious, you can read it here: https://patreon.com/ZerotoMVP

Anthropic made Auto Mode the default in Claude Code — no more permission prompts, a classifier now checks every tool call instead. The numbers behind it: the classifier blocks ~89% of dangerous actions. Humans approving manually? 13.6%. We click "Allow" ~93% of the time without looking — that button was never control, just a ritual of control. But remember: a classifier is a prediction, not a mechanism. The last line of defense is still an isolated environment — not a model's best guess. More detail in the article: https://blokhin.us/notes/claude-code-... So — how do you run coding agents?

Auto mode / skip permissions — full speed

I review every permission prompt

Only in an isolated environment (VM/container)

I don't use coding agents yet

134 votes

A few thoughts on Kimi K3. Moonshot has released Kimi K3, and the release marks a notable step toward wider adoption of open-weight models — models whose parameters are available to developers. Interest turned out to be high: just 48 hours in, Moonshot had to temporarily suspend subscriptions due to a shortage of GPU capacity. Across numerous tests, Kimi K3 is already approaching the best closed models from OpenAI and Anthropic in coding and tool-use tasks. What matters most is that, for the first time, an open-weight model shows a comparable level on the kinds of tasks typical for AI agents. The infrastructure around such models is developing at the same time. vLLM added support for Kimi K3 on release day, including recommendations for running the enormous MoE model with a context of up to 1M tokens. DigitalOcean added K3 to its Inference Engine almost immediately, making it possible to use the model as a managed API without having to stand up complex infrastructure yourself. NVIDIA also used the moment to publicly back the open-weight approach, calling it an important part of US leadership in AI, and took part in founding the Open Secure AI Alliance, which is meant to address the security of open models. That is why, in my view, the significance of Kimi K3 lies not so much in its benchmark results as in the fact that the boundary between self-hosted and cloud models is starting to disappear. A developer can get the control of an open model while still using convenient managed infrastructure. The choice between local deployment and an API is gradually turning from a standalone infrastructure project into practically a matter of configuration.

Finally, the Qwen 3.8-27B model will be released next week! I think this will be a significant step forward for local models. But of course, we will only know for sure after the release and testing : ) https://x.com/Alibaba_Qwen/status/208...

🎉 Kivarion 0.8.1 is here — my first public release! Kivarion is a fast, private desktop password manager for KeePass `.kdbx` files. No cloud, no accounts, no tracking — your vault stays on your machine, and only you hold the keys. 🔐 What you get: 🔑 Full KeePass support — open, edit & create `.kdbx` vaults 👆 Touch ID unlock on macOS 💾 Safe auto-save with atomic writes + backups ⏱️ Auto-lock & auto-clearing clipboard 📎 Attachments with Quick Look Built with security by design: the app's UI has zero direct filesystem access, and macOS builds are signed & notarized. 📺 Here's a video on why I decided to build my own password manager — https://youtu.be/FOmuucZ4dfE Download it now and give it a try. Feedback and feature requests are very welcome — drop them in the comments! 👉 https://github.com/w512/Kivarion/rele... #Kivarion #PasswordManager #KeePass #OpenSource #macOS #Privacy

🚀 Texodus 4.2.0 is here — your files are safer, your workflow is faster 🔒⚡ Texodus is a free, open-source Markdown editor for macOS, Linux, and Windows (born out of one of my AI model tests). Highlights of this major release: 💾 Auto-save — flip it on and forget Cmd+S exists. External file changes are picked up too. ⚡ Quick Open — Cmd/Ctrl+P fuzzy-finds any note in your workspace. 🗂 Tabs that remember — your session is restored on launch; already-open files jump to their tab instead of making a twin. 🎨 Smarter preview — only the block you're editing redraws (no flicker, Mermaid diagrams keep their zoom), and code blocks get real syntax highlighting for 30+ languages. 🔒 Filesystem lockdown — the big one. Texodus can only touch files and folders you open; ~/.ssh, ~/.aws and Keychain are off-limits, period. 🛡️ Plus clickable local links in the preview, PDF export with embedded images, an honest word count, and more. 👉 Download: https://github.com/w512/Texodus/relea... Feedback, bug reports, and feature requests very welcome 🙌

ViewTube

Zero to MVP