My favorites
- Bold take on why code review is the wrong place for knowledge-sharing and quality checks — shift pairing and design discussion left instead.
- An open-source CLI agent that autonomously reads papers, writes ML code, and ships models via the Hugging Face ecosystem — worth a spin if you’re prototyping ML pipelines.
- A reminder that psychological safety and trust beat any AI tool—if leadership treats AI as a headcount excuse, you’ll tank the culture that actually drives output.
- Homebrew’s creator tackles the credential problem agents inherit from your CLI tools — worth a look if you’re letting AI agents run commands with your access.
Agile, Leadership and Product
Architecture & Software Development Practices
- One dev approval plus an AI reviewer beat mandatory double human review — a pragmatic fix worth stealing before PR queues drown your team.
- Blender’s going all-in on making small studios capable of full-length films — worth a read if you ever wonder how far open source tooling can stretch.
- A CSS performance deep-dive turned meditation on frontend’s shrinking audience — worth reading for the style-recalculation debugging tips alone.
- Fun rabbit hole for anyone who forgets what a ‘monad’ or ‘currying’ means mid-conversation — an interactive glossary that turns FP jargon into a browsable graph.
AI, LLMs & Machine Learning
- Layer streaming trains an 8B model on a 4GB laptop GPU by paging frozen layers off VRAM — clever trick, though verify their bit-exactness claims before betting production on it.
- A hands-on walkthrough of building an on-device OCR receipt scanner with LiteRT.js and Gemma — no cloud API, your data never leaves the browser.
- A practical breakdown of engineering reliable agent loops from scratch — worth watching if you’re building anything beyond a simple prompt-response AI feature.
- Honest retrospective on building an internal AI agent that went from ClickHouse queries to company-wide use — great read on what actually drives adoption.
- A grounded look at why ‘move data to the cloud’ never actually solved the data-readiness problem — and why AI is about to repeat that same mistake.
- A free, open-source AI chat interface that auto-picks the best open model for your prompt — handy for testing without vendor lock-in.
- Colibri streams MoE experts straight from disk, letting a 2.8T-parameter model like Kimi-K3 run on consumer hardware — no fine-tuning, no massive VRAM required.
- A 27B-class reasoning model squeezed to ~6GB via ternary weights, retaining 98% of FP16 quality — worth a look if you’re serving LLMs on constrained hardware.
- A Danish-language Pong bot that reads one sentence about the ball’s position and reacts up to 10x/second — a fun demo of LLMs doing real-time control, not just chat.
- Thirteen multi-agent orchestration patterns, each rebuilt as a reactive reducer where agents only write context and a state machine owns control flow. Worth reading if agent token bills and non-reproducible runs hurt.
- Addy Osmani on letting agents loose in old codebases: mark green/yellow/red zones, limit blast radius, and lean on repeatable tests. Worth reading before pointing agents at your legacy code.
- A self-hosted, multi-user, multi-agent assistant with JWT isolation, tool approval, and pluggable workspace backends. Worth a look if you’re weighing how to run agents for a whole team without shipping data out.
- Agent memory that learns instead of just replaying chat history, with a retain/recall/reflect model and strong LongMemEval results. Worth a look if RAG-based memory is starting to feel thin in your agents.
DevOps, Observability & Security
- Great read on why user-level A/B tests break in two-sided marketplaces — budget cannibalization skews results, and splitting budgets per variant fixes it cleanly.
- Copy-on-write database branches give every PR, developer and AI agent an isolated, production-like database. The key point is that migrations stay the source of truth, not merges. Useful guardrails included.
- Microsoft’s new WSLC runs each container session in its own less-privileged process, with virtiofs volumes about twice as fast as plan9. Worth reading if you run Linux containers on Windows.
Developer Tools & Open Source
- An agent skill that turns a codebase description into self-contained, interactive architecture diagrams—useful for grounding PR reviews with real before/after topology diffs.
- A local-first agent workspace that logs every model message, tool call, and permission decision as an append-only record — worth a look if you care about auditable AI workflows.
- Swap port numbers for stable .localhost URLs in local dev — small DX win, especially handy for keeping agent and human workflows sane across multiple running apps.
- A local-first CRM for founders running a fundraising round—investor scoring with sources shown, pipeline tracking, and outreach, all in one SQLite file you own.
- A dbt package that unifies ad spend, clicks, and impressions across a dozen platforms into one reporting model—handy if you’re tired of stitching marketing dashboards by hand.
- A free, no-watermark alternative to Screen Studio for polished demo videos — auto-zoom, on-device captions, and full commercial use. Now archived, but a community fork carries it forward.
- Open-source guidelines that stop AI agents from writing outdated Go — teaches them max(), slices.Contains, cmp.Or, and other idioms your linter shouldn’t have to flag.
- An AI-powered data exploration tool that branches your questions into visual threads instead of burying them in chat history — worth a look if you’re tired of losing context mid-analysis.
- OpenAI’s spec for turning tickets into autonomous coding runs with proof-of-work attached — worth a read even if you just steal the harness-engineering pattern.
- Nice writeup on squeezing a 27B Qwen model onto a single 24GB card with vLLM patches and speculative decoding — solid read if you’re GPU-poor and tired of paying for API tokens.
- Turn a project into a polished launch video with one command — a fun way to close out a build and actually share it instead of letting it die in a repo.
- Google’s open-source runtime for running millions of agent sandboxes with 10x container density and sub-500ms resume — worth a look if you’re scaling agentic workloads on k8s.
- Google’s new agent orchestrator treats agents as first-class workloads — sandboxed, network-fenced, and resumable. Worth a look if you’re tired of duct-taping agent infra yourself.
- Alibaba’s battle-tested internal code reviewer, now open source — line-precise LLM review with a built-in ruleset for NPE, XSS, and SQL injection, 9x leaner on tokens than Claude Code.
- A native desktop app in ~6MB, no Electron bloat, no Rust build step. If you’re sick of shipping 150MB wrappers for a webview, this is worth a look.
- A Docker-compatible REST API layer over Apple’s native container framework — finally lets you point the regular Docker CLI at Apple’s containerization on macOS.
- One local CLI that combines ripgrep, BM25 and vector search, built for you and your coding agent. Fewer tool calls and tokens, and the index stays on your machine. Worth a try.
- A free library of UI sound effects built for modern web apps, shadcn-style. Handy when you want clicks and notifications that add polish without hunting through licensing pages or sourcing audio yourself.
- Devframe lets you define a devtool once and mount it as a Web Standard handler in Hono, Next.js, Vite and more, or ship it as a CLI or MCP server for coding agents.