[bensyverson]

Exactly. The model is exquisitely sensitive to language. The idea that you would encourage it to think like a caveman to save a few tokens is hilarious but extremely counter-productive if you care about the quality of its reasoning.

[andai]

Does this imply that if you train it on Gwern style output, the quality will improve?

[gwern]

Unfortunately, that is an oversimplification for a highly RLed/chatbot trained LLM like Claude-4.7-opus. It may have started life as a base model (where prompting it with correctly spelled prompts, or text from 'gwern', would - and did with davinci GPT-3! - improve quality), but that was eons ago. The chatbots are largely invariant to that kind of prompt trickery, and just try to do their best every time. This is why those meme tricks about tips or bribery or my-grandmother-will-die stop working.