For developers
AI news for developers
Models, dev tools, open source and coding agents.
Picked from this week's feed.
48 stories this week · top 9
Lessons for developers
Fragments: November 20
Messy code costs more for AI agents to review and fix, so keeping code healthy is a practical way to control agent spend, not just a style preference.
Why your fine-tuning job takes longer than the FLOPs say
A CatBoost model using only pre-dispatch job settings predicted fine-tuning time far better than a FLOP estimate (R² 0.98 versus 0.31), which helps shared GPU clusters schedule jobs with less idle capacity.
Running the full server-side Datasette Python app in the browser with service workers and Pyodide
Simon Willison rebuilt Datasette Lite on service workers, so a full server-side Python app runs in the browser and can be shared as a single URL demo or bug reproduction.
Inference engineering for software engineers
Teams cutting AI spend can swap frontier APIs for open models on an inference provider first, then move to dedicated GPUs only after hitting rate limits, uneven performance or overpaying.
agent-autoloop
Agent Autoloop runs command line agents like Claude Code or Codex in the background on event triggers and pings you for approval only when unsure; for solo builders and small teams.
magpie
A local gateway and menu bar app that lets coding agents like Claude Code and Codex use any model or provider, for people juggling several agents, keys or accounts.
Try it in your browser
Today's AI news and lessons, one card per swipe.
Free, no account needed.