I don't like being directly critical, people learning in public can be good and instructive. But I regret the time I've put into both this article and the last one and perhaps someone else can be saved the same time.

This is someone with limited knowledge of chess, statistics and LLMs doing a series of public articles as they learn a little tiny bit about chess, statistics and LLMs. And it garners upvotes and attention off the coat-tails of AI excitement. Which is fair enough, it's the (semi-)public internet, but it sort of masquerades as being half-serious "research", and it kind of held things together for the first article, but this one really is thrown together to keep the buzz going of the last one.

The TL;DR :: one of the AIs being just-above-terrible, compared to all the others being completely terrible, a fact already of dubious interest, is down to - we don't know. Maybe a difference in training sets. Tons of speculation. A few graphs.