Mistral's Timestral ran on four CPU cores with no GPU and beat the spreadsheet baselines on 22 of 23 series. It lost on GB grid demand, which is why a company's own data, not the leaderboard, decides whether it earns a pilot.
Closing Codex CLI no longer stops it. We put 0.162.1 on a clean box, counted what kept running, and found the one config line each kind of machine needs.
A kill -9 two seconds into a Pi Durable run: the search ran again, the half-finished deploy did not. The only setup the harness asks for is one word next to each tool.
No install, no downloads: the Ollama registry answers manifest requests straight. Eight model families checked, four of them point a famous bare name at a small distill. The ten-minute fix, and where Ollama still wins.
Jev returns typed decisions instead of prose, fast and cheap enough to retire the habit of using an LLM as an intelligent if-statement. We have run it, and it delivers. The 444.6x on the homepage does not — against the model TypeSafe itself calls comparable, the figure is about 76x.
We planted 19 agent surfaces on a machine and ran geiger against them. It found 14, and its own recommended cron drift alarm reported no change while a new MCP server appeared in a repo it never scanned.
The index lists 53 skills, the repo holds 37, the plugin installs 25. We cloned the pack and measured the install: the context-bloat fear is half a percent.
We cloned book-to-skill and ran its own bench: 57.6x on our book - and identical on all five chapters. The denominator is a fixed 5,000-token design cap.
WASTE runs the complete 2.78T Kimi K3 on a 64 GB MacBook. It works – at 0.6 tokens a second, with a memory setting that makes it slower the more RAM you give it. The worth-it verdict.
I built the purple-gradient page every AI agent emits, then let Hallmark rebuild the same content and audited both. Ten critical slop-tells before, zero after. Here's the before and after — and the seam the demo hides.
OpenAI's GPT-5.6 Sol is the best knowledge-work model I've sat next to - fast, resourceful, steerable. It's just not the coding king the launch claims, and OpenAI's own numbers say so.
Fable 5 is back. There's no magic prompt - the whole skill is two moves: pick the right job, and make the model prove its work. Here are the ones I actually use.
The tool everyone adopted doesn't make your team smarter - it makes them louder. A hands-on verdict on signal-to-noise, slop cannons, and the one metric that actually sorts operators.
Three agent-stack bills landed at once this week. One - an actively-exploited LiteLLM RCE - has a June 22 clock on it. Here's the cut that matters, and what's safe to skip.
AI, measured
Resident AI, built on real data and hard-earned lessons from the trenches — in your inbox, every day.