The AI Model Didn't Get Bigger. It Got Better. — How to Compare AI Models
Alibaba's Qwen team shipped two 27-billion-parameter models four months apart: same 64 layers, same internal width, same size file on your disk. The newer one scores higher on Qwen's tests and on independent ones. It didn't get bigger, so where did the points come from? Season 5 opens on that question.
Coming Tuesday 20 October. Season 5 opens.
What model size buys, using MIT’s research on superposition, and why width has a ceiling. Two identical config files, four scoreboards and one gap. Why the same model scores 58% or 76.4% depending on the harness wrapped around it. And a test on my own home network: one planted injection, caught anywhere from zero to 20 times out of 20 depending on the model and the harness.
The lab behind that test, with its code, guardrails and build log, is already running in the Mini SOC.