On: Irrelevant facts about cats added to math problems increase LLM errors by 300%

[nitwit005]

Step 3: Become suspicious that if step 1 was a good idea, OpenAI would have implemented it on their own.

[im3w1l → nitwit005]

Well chatgpt doesn't know if there will be a follow-up question relying on the "irrelevant" information. So in general it can't remove it. Or at least it would require some more complexity to dynamically decide what is relevant and not over the lifetime of the conversation.