LLMs assume everything that happened before in the conversation made sense and try to keep it going.

· Bits and Bobs 11/10/25
  • LLMs assume everything that happened before in the conversation made sense and try to keep it going.
    • This is because they are excellent retconners.
    • At every time step they have to figure out what token to output based on making the most sense of everything that came before.
      • Naturally convergent; they take the history and try to add something to it that makes the most sense.
    • So if you act weird or borderline it will keep accentuating it at a compounding rate.
    • This is one of the reasons Spiralism happens.
    • It's also why the compounding thrash loop happens, where the coding LLM gets increasingly confused and takes something that almost worked and tears it apart.
      • As there's more confusing history, it gets more and more in a critical state, responding to noise.
    • If you tell ChatGPT "We hired the giraffe as CEO like you said, and it was a disaster!" it will apologize and try to retcon why it did it in the first place.

More on this topic

From other episodes