I’ve noticed ChatGPT gets less able to do precise reasoning or respond to instructions, the longer the conversation gets.
It felt exactly like working with a student who was getting tired and needed to rest.
Then I had above shower thought. Pretty cool right?
Every few months a new ChatGPT v4 is deployed. It’s got new training data, up through X date. They train up a new model on the new content in the world, including ChatGPT conversations from users who’ve opted into that (or didn’t opt out, can’t remember how it’s presented).
It’s like GPT is “sleeping”, to consolidate “the day’s” knowledge into long term memory. All the data in the current conversation is its short term memory. After handling a certain amount of complexity in one conversation, the coherence of responses breaks down, becomes more habitual and less responsive to nuance. It gets tired and can’t go much further.
I wonder if it has more to do with the fact that the more inputs are available, the more complicated and prone to error something is. Also, ChatGPT doesn’t “get tired” or at least not in the human meaning, I guess it could exhaust the memory of whatever its running on would be the closest analogy if my understanding is correct…
Re:complexity, this is true of all things, and computer/programming stuff is no different. The more complex it is or the larger the dataset, the greater potential for noise vs signal