Concise outputs beat verbosity — study behind the caveman prompting trend

I read this study on AI so you don't have to, so this interesting study came out last month, and it pretty much tells us that if we tell our large language models to be more concise, we will actually get better answers. Now this is really interesting because you have GitHub repo like caveman popping up right now, which got 5,000 stars in three days, and their whole thing is they're just telling cloud code to talk like a caveman stripping away all the filler words, and the idea is it reduces token count. But this study is telling us it doesn't just improve our token efficiency. It can improve actual outputs. We will get better and more accurate responses. So what they did is they tested 31 open-weight models against 1,500 problems, and what they saw in 8% of these problems, that these smaller models, sometimes significantly smaller models, think 2 billion parameters, outperform the big models, sometimes 100 times more parameters, so think a 2 billion parameter model versus a 200 billion parameter model, but if they're 30% by 30 percentage points. However, when they force these larger models to be more concise to talk like a caveman, that completely reversed. The models that were losing before these bigger models, now outperform the smaller models, simply by being more concise, they didn't change the thinking, the coding, nothing, just the text outputs. And they believe the reason for that was because these larger models tend to be more thorough, but to a fault, they'll pretty…