My bubble keeps saying that in the AI era, the winner is whoever has "taste." I disagree. Successful things are usually built not by aesthetes, but by people who hit a wall and realized a tool was missing.

Found a local LLM setup for the M4 Pro that is fast and solid. Task: 10–20 docs, fix contradictions. Perf: 50–70 tps avg, 20–30 GB RAM, context 53K/262K. DeepSeek 4 Pro recheck — all perfect! Stack: mtplx (speed + UX) + Ornith-1.5-35B-A3B-MTPLX-4bit (qualified A3B) + pi (thin system context).

I wanted Opus quality for less. How do you cut tool calls? Speculative execution! After Opus's tool calls, a cheap model continues in the same session, right up until it needs to think. Thinking is Opus's job.

Another #reatom tip. We has the withAbort extension that allows you to stop a process easily, and there are a few features behind this 👇

Just finished Reatom adapter for @preactjs.com and it was so fun! You know why? The options hooks is super handy, I implemented transparent atoms reactivity with 5 lines of code 🤯 👇

I'm extremely happy to announce the beginning of a new era in state management! The new Reatom has way more than we're used to: - The simplest "signal"-like API - The most advanced concurrent effect management - The most extensible plugin system (after jQuery, lol) dev.to/artalar/reat...

Reatom: State Management That Grows With You

The Fragmentation Problem Modern frontend development has a familiar pattern: Start with...

dev.to