Hacker News Reader

Comment by arielweisberg | Hacker News Reader

I thought ChatGPT had pretty limited recall of past context of your conversation.

How can it remember the details needed across an entire campaign?

I've heard that GPT4's token limit is equivalent to a small novella. That's a lot of campaign details to take into consideration. Read the full transcript in OP to get a sense of how it does.

I don’t think the 32K context is widely available yet. I have limited GPT-4 access with 8K context after applying for it, and ChatGPT (even with Plus) has even less.

We get this prompt: "(In the interest of time, for the remainder of the game you decide the outcomes so that we don’t need to roll.)"

And at this point (which wasn't even through the end of the first combat encounter against some extremely basic enemies), we have no idea if it's actually playing D&D or if it's a freeform text adventure where narratively appropriate things happen at narratively appropriate times.

You could ask it to write a 5-10 bullet point summary at the end of each session, keep a separate file where you keep all the session summaries, and then include that summary at the beginning of your first prompt when beginning a new play session. But still, it would be difficult to get it to remember all the details of a lengthy campaign like a human would.

GPT-4 can have a context window of 8,192 or 32,768 tokens. GPT-3.5 has a context window of 4,096 tokens, GPT-3 has 2,049 tokens.

Maybe it doesn't really need to fully remember it in order to still do a pretty ok job? I think its ~50 pages of context?

To be fair, most players forget all important campaign details after an hour or two, also.

it has an extensive but constant part of the prompt?

Reply on news.ycombinator.com