I this is a very important observation. What it means becomes more clear when you consider the absolute failure of all models to code in esoteric languages:

"The rate of improvement is very small, close to 0 points per month. AI models have not meaningfully improved at EsoBench since launch."

Anyone who really knows how to code knows that anyone who can really code can pick up a new language. The implications or LLMs not having that capability should be obvious, IMO.