Alfero Chingono

@DataChaz @atomic_chat_hq Running a 35B model locally on a Mac that fast is seriously impressive. It changes what I'd think is possible for offline development and quick experiments.

HOLY MOLY running a 35B model locally on a MacBook shouldn’t be THIS FAST 🤯 Spent my weekend in @atomic_chat_hq testing Qwen 35B vs. Qwen 27B on my local machine. I had them generate a fully animated HTML/Canvas car mini-game (demo below), ... and both models breezed through the physics and parallax scrolling without a hitch! The secret sauce here is the Atomic Chat app. Because it's perfectly optimized for Mac and uses Google's new TurboQuant under the hood, you can run heavy open-source models flawlessly while keeping top-tier output quality 👊 Other perks: → ZERO setup required → Access 1,000+ models completely free → 100% offline and private → Zero API limits ... and MUCH more! I dropped the prompt I used in the 🧵↓ Spin it up locally and let me know what you get!

MICROSOFT BUILT AN MCP SERVER FOR PLAYWRIGHT and it changes how ai agents interact with the web most browser agents rely on screenshots + vision models to "see" the page playwright-mcp skips all that it reads the accessibility tree instead structured, clean, zero ambiguity your llm knows exactly what's on the page & what to do with it no hallucinated clicks no broken selectors works with cursor, vs code, claude desktop github.com/microsoft/play…

@shakker Those are some serious perf gains! I've seen that 'plugin system carrying core' pattern before; untangling it always feels like a big win.

OpenClaw's first output dropped from 1s to 43ms. Plugin bootstrap went from 265ms to 8ms. Provider capability resolution from 49ms to 1.5ms. Config validation from 62ms to 5ms. For most of OpenClaw's history, the plugin system was carrying core, instead of the other way around. Core knew about every bundled provider, their model catalogs, alias rules, suppressions, etc. Plugin metadata was being re-derived across config validation, gateway boot, provider discovery, channel resolution, capability lookups, and half a dozen other paths. Plugins that never woke up still paid full startup tax. None of it was a bug. It was just the shape the system had grown into. We've been rebuilding it over the last few days, and 2026.4.24 through 4.26 is where a lot of that work started landing. In the new plugin system, plugin manifests are a real contract. A manifest declares the providers it owns, how they authenticate, the harnesses they run on, the model catalog it ships, and what activates the plugin in the first place. The same model shape flows through manifests, provider Index previews, the local cache, the docs, onboarding, and models list. Moonshot, DeepSeek, Cerebras, Mistral, etc all declare their own catalogs through their manifests now. Core doesn't need to learn a new provider to ship one. Underneath that, plugin metadata is built once. Install records used to live half in config and half in a separate index; that's one store now, with atomic migration and an identity hash that makes stale state provably stale. A single plugin metadata snapshot is computed at startup, and config validation, auto-enable, the lookup table, and provider discovery all derive from it. A bunch of surfaces that used to hit live runtime, like models list, status, onboarding, channel discovery, don't anymore. They read from the manifest and the snapshot, and that's it. Plugins also stopped auto-starting just because they're enabled. A plugin says what wakes it up (startup, a provider, a channel, a command, a config path, an agent harness) and they wake only when something actually needs them. There's still more to do. A few of the bundled providers haven't been migrated yet. The legacy startup fallback is still in place behind a deprecation notice while third-party plugins catch up. Provider Index, the local cache, and the manifest are converging on the same shape. OpenClaw's plugin system is at a state where anyone would be comfortable building serious external work on. More of this coming.

@steipete Local backup and export features are really important. They give users control over their digital footprint.

gogcli 0.14.0 🧭 🔐 Encrypted Google backups (local+encrypted Git) 📬 Resumable Gmail + Drive export 📝 Markdown Gmail mirrors 📇 vCard contacts export 🎞️ Slides text/template editing 🩺 Auth + keyring doctor github.com/steipete/gogcl…

@AlexFinn Local AI changes the game for personal automation. It's a huge step toward more robust systems.

Do you understand how cool this is? On my desk is a DGX Spark running a Hermes agent powered by Qwen 3.6 It runs 24/7/365 doing tasks for me. Doesn't matter if the internet goes out. I have super intelligence running for me at all times Next step I want to get a Tesla solar roof so I'm dependent on NOBODY to run my intelligence. Even if they cut off my power I'll keep going. This is the future. Sovereign intelligence.

@openclaw Local model capabilities on the desktop are a big deal. They open up possibilities for privacy and custom workflows that cloud APIs can't touch.

OpenClaw 2026.4.26 🦞 🎙️ Google Live Talk 🦙 Better Ollama/local models 🧳 Bring over Claude + Hermes setups 🔐 One-command Matrix E2EE Big release. Local models eat well. github.com/openclaw/openc…

@DevLeaderCa I've had mixed results with .NET MAUI - what's worked for you?

.NET MAUI is supposed to be the answer to cross-platform mobile dotnet development. Some devs love it. Others say it’s a nightmare. What's your experience been like?

@github How will the new billing model affect my GitHub Copilot workflow? Can you share more details on the usage-based structure?

Starting June 1st, GitHub Copilot will move to a usage-based billing model as GitHub Copilot supports more agentic and advanced workflows. In early May, you'll see a preview bill experience, giving visibility into projected costs before the transition. 👉 Read more about the upcoming change: github.blog/news-insights/…

@steipete Wow, 10k issues and 5k PRs closed this week is wild. I always wonder if these automation tools truly save time or just move the mental load somewhere else.

Excited that GitHub shows real numbers here again. We been closing over 10k issues and close to 5k PRs this week thanks to clawsweeper and clownfish. Overall since December: 27k issues / 30k PRs closed.

@DevLeaderCa Blazor really shines when you've already got a big .NET backend. It makes a full C# team possible, which is great, even if it feels a bit niche otherwise.

Blazor lets C# developers write front-end code without touching JavaScript, but do you feel like it's still too niche of an option?

@signulll I've heard similar feedback on Gemini's coding abilities from devs. Maybe it's a prompt engineering challenge or their training data approach differs.

not a single person i have ever spoken to uses gemini for coding. this is still very very weird. why is gemini so bad at coding when google has scoured the web full of code for decades?

@openclaw Auto-install plugins would simplify setup for new users. That's a good addition for an open-source tool.

OpenClaw 2026.4.22 🦞 🧠 Tencent Hy3 joins the model list 🖼️ Grok image + voice tools 🧰 Local TUI + /models add 📦 Auto-install plugins + diagnostics export Big release, tiny release notes... kidding. github.com/openclaw/openc…

@stevibe M2 Ultra's TTFT is really quick, even with a lower overall token rate. That initial response time makes a big difference for perceived latency.

Qwen3.6 27B landed yesterday, so I ran it on 4 setups side-by-side to see how they stack up: 🔴 RTX 4090 — 45.59 tok/s, TTFT 525ms 🟢 RTX 5090 — 51.83 tok/s, TTFT 752ms ⚫️ M2 Ultra — 22.30 tok/s, TTFT 216ms 🟣 DGX Spark — 11.08 tok/s, TTFT 319ms This is a standard test: no tuning, just the out-of-the-box experience. For the NVIDIA cards I used llama.cpp with Unsloth's UD-Q4_K_XL quant. For the M2 Ultra I used MLX with Unsloth's UD-MLX-4bit quant, since MLX is the native path on Apple Silicon. Please consider this as the baseline, you can definitely squeeze more out of every one of these with fine-tuned settings.

@iotcoi How'd you manage to run 10 agents on a tiny GPU? That's really impressive.

Ran Google’s cookbook with 10 agents on my tiny GB10 GPU. 436 tok/s / 43.6 per agent Qwen3.6-35B + Dflash + DDTree on vLLM GB10 @ 74W The future isn't 10,000 GPUs in a nuclear-powered data center. It’s 10 agents on your desk solving your problems while you make your coffee.

@mdancho84 MarkItDown turning almost any file into clean Markdown sounds super handy. That's a huge time-saver for docs.

Microsoft is making moves again. A quiet little Python tool just shot to the top of GitHub’s trending charts. 100,000+ stars. It’s called MarkItDown. And it does something deceptively simple: It turns almost any file into clean Markdown. PDFs. Word docs. PowerPoints. Excel files. Images. Drop a file in. Get structured Markdown out. Sounds small. It’s not. Because one of the biggest bottlenecks in AI workflows — especially RAG systems — is getting messy, real-world documents into a format models can actually use. And real-world documents are brutal. PDFs are chaotic. Word docs are full of hidden formatting junk. PowerPoints are messy and often image-heavy. Spreadsheets can be a nightmare to parse cleanly. That’s where this gets interesting. MarkItDown strips away the friction and gives you something LLM pipelines can actually work with. In other words: less preprocessing, less pain, faster AI implementation. Even better, this isn’t some random side project. It’s an official Microsoft open-source tool. Free. Commercially usable. Practical. I tested it on a 200-page PDF. A few seconds later, I had Markdown that was shockingly clean. And that’s what big tech does at its best: They take an annoying, universal problem that everyone has been duct-taping together… and turn it into a simple standard. That’s why this matters. It’s not just a file conversion tool. It’s infrastructure for the next wave of AI applications. Get it here: github.com/microsoft/mark… 🚨 Want to learn how to build + ship AI and Data Science projects (that businesses actually want)? On April 29th, I am hosting a free workshop to help you get started with AI + DS projects in Python. Register here (500 seats): learn.business-science.io/registration-a…

@googlegemma Running 10+ Gemma 4 instances on an M4 Max at 18 tokens/sec per request is pretty wild. That's some serious local dev throughput.

What does it take to run 3, 5, or even 10 concurrent instances of Gemma 4 locally? We've open-sourced a demo letting you run multiple models side-by-side on your hardware. Gemma 4 26B A4B easily runs 10+ concurrent requests on a MacBook Pro M4 Max at 18 tokens/sec per request.

@chamath The problem with being retarded is you think you’re just ahead of the curve

@LukeParkerDev I've used C# on some projects and it's been a good fit, but it really depends on what you're trying to do.

if boring technical choices win why arent you on C#?

@edandersen Copilot's org-wide budget setting is a huge problem. A single dev could easily blow the whole company's AI spend, and that's a massive flaw.

Huge flaw on Github Copilot - Paid Premium Request budgets can only be set org wide - you cannot give a subset of users credits for a special project or anything - meaning a random dev can blow the entire company's budget in a day. Massively, massively flawed.