Chicken Genes for AI
Animal-breeding math pointed at AI fine-tuning, with five predictions written down in advance. Most of them failed.
This started as a joke. The Costco rotisserie chicken has cost $4.99 since 2009, so it might be the most optimized object on Earth, and maybe superintelligence should be modeled on it. Then I looked at what actually made that chicken: seventy years of breeding for growth, and a pile of side effects nobody selected for, like birds outgrowing their own hearts. Breeders built real math for that problem. The G-matrix measures how traits are wired together, and the breeder’s equation uses it to predict the side effects of selection before you breed a generation.
Fine-tuning a language model has the same problem. Train a model on one narrow kind of bad advice and it can go bad across the board, which is called emergent misalignment. So I measured a G-matrix for Qwen2.5-0.5B on my laptop’s 3060: 21 behaviors, hundreds of tiny weight perturbations as “siblings”, and real LoRA fine-tunes as ground truth, including matched pairs like bad medical advice against good medical advice.
The first round looked promising: once you subtract what the harmless twin also does, the harmful part of fine-tuning seemed to lie along one of the model’s easiest directions. That came out of exploring, so I wrote the analysis and five predictions into a preregistration, trained 16 new adapters on a rented A100, and had a separate model grade the open answers blind.
The predictions mostly failed. Every difference between two fine-tunes lies on easy directions, so misalignment isn’t special there. The breeder’s forecast couldn’t tell which dataset would cause misalignment, and neither could a persona-vector baseline. The forced-choice test sheet only partly agreed with graded answers, partly because my recipe made small models collapse into talking about their training topic. And a “price lock” that projects locked traits out of each update made the locked traits move more, not less.
The write-up keeps all of it: the catches, the failed predictions, related work, and a lab notes window with every number.