// news
News
Links to articles worth reading. No aggregator noise — just things that are actually useful.
Running Google Gemma 4 Locally With LM Studio’s New Headless CLI & Claude Code
How to run Gemma 4 26B locally on macOS using LM Studio 0.4.0’s headless llmster daemon and lms CLI. Gets 51 tok/s on an M4 Pro with 48 GB, plus a shell alias to use it as a backend for Claude Code.
read on georgeliu.com →Dropping Cloudflare for bunny.net
A step-by-step migration from Cloudflare to bunny.net — motivated by centralization risks, outages, and dependency on a single US company. Covers CDN pull zones, SSL, DDoS protection, and Origin Shield config.
read on jola.dev →Harness Engineering: Leveraging Codex in an Agent-First World
OpenAI shipped an internal product with 0 lines of manually-written code — ~1M lines generated by Codex agents in 1/10th the time. Lessons on shifting humans from writing code to designing environments and feedback loops.
read on openai.com →Signing Data Structures the Right Way
Canonical encoding and domain separation for cryptographic operations. Real-world vulns in Bitcoin, Ethereum DEXs, TLS, JWTs, and AWS via type confusion. Introduces Snowpack — a serialization system with embedded domain separators.
read on foks.pub →OpenClaw: The Complete 2026 Deep Dive
The definitive OpenClaw guide: installation across VPS/Mac/Windows, real cost breakdowns, hardware recs, security warnings, user stories, and comparisons to AutoGPT and CrewAI. Covers personalities, heartbeats, and skills config.
read on virtualuncle.com →The Weight of Remembering
How the KV cache gives every AI conversation a physical weight in silicon. Traces the evolution from GPT-2’s multi-head attention to Gemma 3’s sliding window, and what happens when memory runs out.
read on future-shock.ai →The Revenge of the Data Scientist
Despite LLMs reducing the need for model training, data science fundamentals — experimentation, debugging, metrics design — are more essential than ever. Covers five common eval pitfalls and how to fix them.
read on hamel.dev →How We Built a Virtual Filesystem for Our Assistant
Mintlify replaced expensive sandboxes with ChromaFs — a virtual filesystem over Chroma that lets their docs AI assistant explore documentation using UNIX commands. ~100ms session creation, zero marginal compute.
read on mintlify.com →10 GitHub Repositories to Master Vibe Coding
Covers 10 repos for shifting from one-off AI queries to orchestrated workflows — context templates, prompt style guides, MCP tools, and kanban boards built around AI agents. The main takeaway: vibe coding isn't just autocomplete, it's a structured practice with its own toolchain.
read on kdnuggets →