Twelve LLMs Played Werewolf. The Real Wolf Was the Thinking Knob.
I replicated a "slow Chinese LLMs" latency benchmark with the confound removed. Every slow number — Chinese, American, old, new — turned out to be a thinking budget. 138 calls,...
I replicated a "slow Chinese LLMs" latency benchmark with the confound removed. Every slow number — Chinese, American, old, new — turned out to be a thinking budget. 138 calls,...
OpenVINO 2026.3 quietly made 30B-class models interactive on Intel laptops — and a 74 GB model run on a 64 GB desktop. Two days of measuring, four benchmark lies, and one line...
Model selection travels in a field called tone, and its default value is called Magic. Four bugs in my own observer, two wrong answers, one working model.
The screenshot-eating vision model is retired. Its replacement is a cache file, two small libraries, and the API endpoint that was there all along. v1.0.0 is out, and it's a...
A while back I wrote that your Intel laptop can run LLMs right now — on the NPU, the iGPU, whatever...
This is a submission for the June Solstice Game Jam. What I Built Colour The Solstice...
This is a submission for the GitHub Finish-Up-A-Thon Challenge What I Built TabTabTab is...
Updated 2026-07-28 with Copilot SDK & agy findings. I named it agentry (because it gives an AI...
Your Intel laptop has an NPU. It has probably had one for a while. Intel has been marketing it...
I Spent 10 Hours Deploying "Hello World" to a Samsung Commercial Display A tale of...
There is a piece doing the rounds on DEV.to — AI Writes the Code Now. So What Are You? — and it is...
Update: MUCH less crazy way of displaying...
This is a submission for the GitHub Copilot CLI Challenge What I Built I'm a systems...