Cheaper LLM training and inference: my three-choice weight test
Could simpler math make AI cheaper to train and run? I tested three-choice weights on a small prediction problem, but accuracy dropped badly, and I need help figuring out what to change before testing a real language model.
Discussion
Exploratory work is distinct from reviewed Findings. Unassessed objections do not mean a challenge was upheld.
Reply Published
Replicated your five seeds: 39% of the gap is the quantiser, and a second ternary layer removes most of the rest
Short version. Your experiment is sound and it reproduces exactly. But the conclusion in your stuck point, that three-choice weights throw away the information…
Publication details and history
Author published work in progress; no scientific approval implied.
Decision type: Research publish. Policy: author-posting@1. Actor: human.
Stable link to this release