Chau Le

🛠️ ピックアップ ✨ GitHub Copilotの出力に毎回手直しする課題に対し、カスタマイズファイルの全種類を試した著者が、それぞれの役割と効果的な使い方を解説しています。設定ファイルの種類と使い分け、具体的な活用例が学べます。 zenn.dev/nagi98/article…

🤖 ピックアップ ✨ 異なるプロジェクトで独立したClaude Codeセッション間の連携不足を解消するため、著者が「cc-to-cc」を開発しました。ファイルシステムとローカルHTTP webhookを使い、Claude Code同士がメッセージを送受信できるMCPサーバーの仕組みが解説されています。 zenn.dev/yukitakeshita/…

🚨 CRITICAL: Active supply chain attack on axios -- one of npm's most depended-on packages. The latest axios@1.14.1 now pulls in plain-crypto-js@4.2.1, a package that did not exist before today. This is a live compromise. This is textbook supply chain installer malware. axios has 100M+ weekly downloads. Every npm install pulling the latest version is potentially compromised right now. Socket AI analysis confirms this is malware. plain-crypto-js is an obfuscated dropper/loader that: • Deobfuscates embedded payloads and operational strings at runtime • Dynamically loads fs, os, and execSync to evade static analysis • Executes decoded shell commands • Stages and copies payload files into OS temp and Windows ProgramData directories • Deletes and renames artifacts post-execution to destroy forensic evidence If you use axios, pin your version immediately and audit your lockfiles. Do not upgrade.

🆕 The Awesome GitHub Copilot project has a new home. Head over to explore hundreds of community-built customizations: 🔍 Full-text search for agents and skills 📚 A dedicated Learning Hub ⚡ 1-click plugin installs for Copilot CLI & @code Built by the community, for the community. Check it out.👇 awesome-copilot.github.com

@odd_joel thanks for great advice. Use mosh I need to use blink shell which will cost some money, currently I am happy with termius + tmux for free, I will reconsider when I have more budget.

solid phone-first setup. when you need to drop into a terminal to debug or check on agents, mosh protocol makes a huge difference — sessions survive wifi/cellular switches and phone sleep. been building Moshi around that plus on-device voice input so you are not fighting the phone keyboard all session

Here is my current vibe coding workflow all can do from phone without toching PC 1. Input idea and make claude opus write design 2. Request gpt5.4 to review the design 3. Claude Opus fix desgin and write details design in markdown 4. Request open claw fully implement it.

Building cool agent don’t make money So rule: 👉Agent system must answer: What does it produce? Who pays for it? How fast can I test it?

@vinhnx impress with your work spirit, why you make copilot as first class provider

VT Code is compiling... This release I support GitHub Copilot as first-class provider.

instead of spending 100$ on claude max I divided my budget as follow 20$ for chatgpt plus, 20$ for claude pro,10$ for github copilot. That way I can try and evaluate many tools. Codex and Copilot can also be used with open code and pi. So now I’m thinking how effectively use them

@sukhdeep7896 @ViralityEngine @grok explain what is it different with current claude code

Anthropic just open sourced the coding tool that is making junior developers obsolete and it's not a Copilot. It's called Claude Code and it's not a tab-completion plugin. It's a fully agentic engineer that lives in your terminal, reads your entire codebase, and executes tasks autonomously through natural language. Here's what it actually does: → Reads and understands your entire codebase not just the file you're in → Executes routine tasks end-to-end without you writing a single line → Explains complex legacy code in plain English on demand → Handles your entire git workflow -- commits, branches, PRs, all of it → Works directly in your terminal, IDE, or tags @claude on GitHub → Runs on Mac, Linux, and Windows out of the box → Extends with plugins for custom commands and agents Here's the wildest part: It doesn't suggest code. It writes it, runs it, debugs it, and commits it. You describe what you want in plain English. Claude Code figures out how to do it and does it. Senior engineers are using this to ship in hours what used to take days. Junior engineers who don't use this are already falling behind. 44,900+ GitHub stars. 3,100+ forks. Built by Anthropic. The future of software engineering just shipped to your terminal. (Link in the comments)

@_onlyscott Phill Neville

Romelu Lukaku and Marouane Fellaini excluded, name a player that played for Manchester United and Everton Difficult

@elonmusk @grok who is the author of backpropagation

There is a good reason for blame x.com/i/grok/share/3…

@WhiteHouse @grok what are benefits of japan in this deal

Our MASSIVE Trade Deal with Japan has just launched!

@excalidraw @AnthropicAI @claudeai wow, now agent can visualize beautifully 🤩

Thanks to good people at @AnthropicAI we now have an official MCP for Excalidraw! Take it for a spin on @claudeai (search for Excalidraw in Connectors, or use in Claude Code and elsewhere). More to come. ✌

@FabrizioRomano @grok does arne slot right about the his first season

🚨 Arne Slot: “If we don't have Champions League qualification, then it's definitely NOT been an acceptable season”. “When I arrived here we could only sign Chiesa and that was after a Europa League season. I'm completely aware of that”.

Introducing Kimi Code, an open-source coding agent under the Apache 2.0 License. 🔹 Python-based, easy to extend. 🔹 Fully transparent — clear, safe, reliable. 🔹 Seamlessly integrates with VS Code, Cursor, JetBrains, Zed, and more. 🔹 Fully-featured & out-of-the-box ready. Switch over in seconds. 🔹 Native Multimodal support Experience it with Kimi K2.5.

@levelsio I am living in Japan for 10 years, feel very safe.

Delusional dataset I've been robbed, burglared, spit on, had girlfriends sexually harassed in Netherlands by street thugs may times, many of my friends there too I've never had any single bad thing happen to me in 10+ years of living in Asia (from Thailand to China to Korea to Japan) Literally the worst thing was one Thai bodybuilder guy getting angry at me for not putting the weights back (which he should) Asia is still a massive blindspot for Western datasets, which makes sense because they're all funded by Western countries

@TheAhmadOsman @grok what pc specs need to run glm-4.7

MASSIVE The year of Local LLMs officially starts with GLM-4.7-Flash by Zhipu AI > 30B-A3B MoE > built for consumer GPUs > runnable from your basement > strongest 30B-class release we’ve ever seen This is THE BEST <=70B I’ve ever run locally BTW Architecture > DeepSeek-style MLA attention > slim MoE routing > 30B total params, ~4B active > 64 experts total, 5 active (incl. shared) Depth & intent > roughly GLM-4.5-Air class > but tuned harder for locality Benchmarks SWE-bench Verified > GLM-4.7-Flash: 59.2 > Qwen3-30B-A3B: 22.0 > GPT-OSS-20B: 34.0 > Nemotron-3-Nano-30B-A3B: 38.8 > not the same universe τ²-Bench > GLM-4.7-Flash: 79.5 > Qwen3-30B-A3B: 49.0 > GPT-OSS-20B: 47.7 > agentic + tool-use cleared BrowseComp: > GLM-4.7-Flash: 42.8 > Qwen3-30B-A3B: 2.3 > GPT-OSS-20B: 28.3 > yes, Qwen 30B-A3B has two point three > web reasoning still breaks models AIME 25 > GLM-4.7-Flash: 91.6 > Qwen3-30B-A3B: 85.0 > GPT-OSS-20B: 91.7 > ~4B active params btw GPQA > GLM-4.7-Flash: 75.2 > Qwen3-30B-A3B: 73.4 > GPT-OSS-20B: 71.5 > quiet, consistent lead HLE > GLM-4.7-Flash: 14.4 > Qwen3-30B-A3B: 9.8 > GPT-OSS-20B: 10.9 > still brutal for everyone > GLM just hurts less Context window reality check > FP16 weights fit with > ~27k tokens on an RTX PRO 6000 > or 4x RTX 3090s (96GB VRAM) > 4-bit AWQ weights fit with > ~70k token context on an RTX PRO 6000 > or 4x RTX 3090s (96GB VRAM) > reminder: KV cache, not weights, is the real tax > why? > thiccc KV heads > ~3x VRAM per token vs GPT-OSS-120B > even though both sit around ~60GB weights Local pain points > vLLM / SGLang: on main, still a little immature > new arch + kernels = rough edges > KV cache can bite fast as seen above > depending on dtype + kernel path What actually matters next > stable MLA kernels across more GPUs > FP8 / quant drops + clean GGUFs > real “daily driver” reports > does SWE 59.2 really feel like 59.2 in real repos? If you’re on RTX 5090s/4090s/3090s or PRO 6000/PRO 5000/PRO 4500/PRO 4000 stacks > this is your lane > 30B-A3B-class, local, real tool-use scores > kernels mature and this becomes a default install > now we wait for quants, kernels, and reports from local community folks > Buy a GPU > run your LLMs locally

@thedankoe @grok based on which author shared, make a detail action plan