Mark Wickersham observes that large language models (LLMs), despite large context windows, exhibit a human-like memory quirk: they preferentially retain the beginning and end of presented material (primacy and recency) and tend to forget or ignore content in the middle. He notes this is surprising because one might expect LLMs not to suffer the same middle-item forgetting as humans.
One of the things that's interesting. AI, these large language models are trained on human data. They're trained by humans, and they have some strange quirks as humans. So, for example, if I asked you to read a book and then afterwards, a few years later, ask you to tell me about what you read in the book, what happens is the law of primacy and recency. We tend to remember best the things we learned first and the things we learned last, whereas the stuff in the middle we tend to forget. Now, you'd have thought that large language models AI wouldn't suffer from that. Interestingly, they do. So you can have a huge context window. You can feed ChatGPT or Gemini a huge amount of information. It still tends to forget and ignore the middle stuff. Okay, which is really interesting.
Effective Pricing Training: Mastering Software Settings for Optimised Pricing
Can Pricing Be Done in Writing? How to Handle Value Conversations Without Calls
