This was an interesting, and delightfully short, read.

It seems that for a lot of tasks you can just repeat the input twice and get better performance out of an LLM. Granted, this is shown for pretty old models, but still!