LLMs can get confused by the meandering path it took to the right answer in the conversation.
You need to continually prune the conversation history and reground it to remove those confusing bits.
Like squashing commits in git.
From other episodes
- It used to be that to test demand you made a doorbell in the jungle.2026-04-20
- LLMs don't have an experience for others to resonate with.2026-04-20
- If you're below mean on a given dimension, LLMs pull you up.2026-04-20
- All of the labs would rather not have to give access to the top-of-the-line models via general API.2026-04-20
- LLMs are significantly cheaper if you only append the tokens.2026-04-20
- Before you only had to trust the entity that made the code.2026-04-20
- LLMs are easy to trick by putting words in their mouths.2026-04-20
- LLMs are really good at rebasing.2026-04-20