Inception - A mind-bending science fiction thriller directed by Christopher Nolan, in which a thief is hired to enter people's dreams and steal their secrets.

The Social Network - A drama directed by David Fincher, based on the founding of Facebook and the legal battles that ensued between its co-founders.

Toy Story 3 - An animated movie from Pixar Animation Studios that follows Woody, Buzz, and the gang as they face their biggest adventure yet: being donated to a daycare center.

Black Swan - A psychological horror movie directed by Darren Aronofsky, in which a ballerina begins to lose her grip on reality as she prepares for a performance of "Swan Lake."

The King's Speech - A historical drama directed by Tom Hooper, based on the true story of King George VI's struggle to overcome his stuttering and deliver a speech to inspire his nation during World War II.

Here's a conversation like one I had recently:

Me: Take this list, cut off each item after the name of the movie, and delete the numbers.

Chat: [Gives a bulleted list of just the movie titles.]

Me: Great! Now can you take out the bullets?

Chat: [Generates a list, but with the movies and full descriptions.]

Me: No, I want the list with just the movie titles, no numbers, no bullets.

Chat: [Generates a list of something totally unrelated, like most nutritious fruits or some shit.]

It seems to be losing track of the referents of my requests -- when I say, for example, "this list" or "do it again but with..." after just a few iterations it will forget that I began by feeding it a list of something. I've had this happen, similarly, with an article I was asking it to rewrite. I asked it to rewrite some of my content in a different style, went back and forth a few times, until at one point it just spat out some clickbait about top 10 ways to earn money at home or something, which was totally separate from my article. This is even after I can give it instructions like, "Go back to the material I pasted in at the beginning..."

Does anyone have advice for getting around this? Suppose, for example, I start the first message in the convo with something like: "I'm going to paste an article here, and when I refer to the article, I'd like you to remember that this content is what I mean..."

Thanks!

(http://www.autoadmit.com/thread.php?thread_id=5675626&forum_id=2#48623944)

codergaard

•

2y ago

It has no memory. The website sends your current prompt and the last N tokens in your current conversation history to the (equivalent of) the API completion endpoint. If you use the API you control this context completely. The limit is 4096 token (including prompt) for 3.5 - it is much higher for 4 (also more expensive), but you will have to go on a waiting list for API access to gpt-4 (even if you access to ChatGPT-4 via plus).

A lot of users fail to see that these models have no memory. It feels like there is memory due to the context being passed with your prompt, but it's not actual memory on the server side. Your request might hit two completely different AI instances with successive prompts. It's the context that carries any information not already contained in the model. It cannot remember.

You can test this by taking an existing chat and copy-pasting a few prompts+replies into a new conversation - it will continue as if the reply was made in that new conversation. There is not state connected to you on the AI side. It is a transformer - it takes input and produces output. Nothing is stored anywhere.

We will "soon" get plugins that can function as memory, which will change this. It will allow storing conversation history server-side and having a specialized search retrieve the parts relevant to the current conversation - but even then it will be a different form of memory than many people expect as the AI's memory - a separate module which tries to guess which 'memories' are relevant and attach them to the prompt (out of sight). A few iterations down the road we may have the AI requesting memory retrieval and doing one or more "silent" iterations server-side to have a more realistic and effective memory, but we're not there yet.

This is not an entity which remembers. It's a fantastic tool and extremely promising technology. But right now it serves you best if you understand the stateless nature it has. If you carefully tailor the context, you can get much better results. If you ignore the limitations of the token limit, the conversation will quickly become weird.

(http://www.autoadmit.com/thread.php?thread_id=5675626&forum_id=2#48623946)