Hacker News Reader

Comment by nopassrecover | Hacker News Reader

This is pretty impressive.

I’ve been similarly impressed by ChatGPT’s ability to both absorb and playback strategic advice on complex games based on a snapshot mid-game scenario (eg Twilight Struggle) and more impressively to strategise in completely novel games that it wouldn’t be able to lookup rules or advice for.

For example I described an invented hidden identity party game and ChatGPT without further guidance inferred the dynamics and proposed great starter strategies depending on whether you were the “hider” or the “catchers” including deceptive approaches to fit in, lure out other players, cast doubt, play it cool etc.

It’s been a little trickier as they’ve nerfed its abilities as public uptake has increased, but the glimpses of underlying inference and intelligence is astounding.

View on HN

>> It’s been a little trickier as they’ve nerfed its abilities as public uptake has increased

this is what I've observed as well. There were some things that ChatGPT was doing spectacularly well in the initial days. Now, for the similar queries it is just pointing to the source for more info. Or telling me that it cannot access the web and / or just asking me to do more research by using other platforms.

I wish there was a stable diffusion equivalent for ChatGPT alternative.

>I wish there was a stable diffusion equivalent for ChatGPT alternative.

I agree, we'll only really start to get an idea of how far we can push this technology when there's a version of it that's available to use as the user sees fit. I wonder if there's any projects like that in the works already?

Same here.

However, just like chatgpt struggles with arithmetic, it also can’t figure out simple logic puzzles. I gave it some puzzles like the ones here: https://www.woojr.com/printable-logic-puzzles-for-kids/ and it has no idea how to solve for it.

It kind of makes sense if the model is purely generative, as if it has not seen the exact puzzle before, it can’t just “regurgitate” the answer. For example, I gave it the classic four gallons of water in a five gallon container puzzle from die hard. It solved it perfectly, probably because that exact puzzle was in its training set, but could not generalize it to other bucket sizes.

Yes it's really impressive. I've noticed that it stays more consistent if I ask it to automatically print the player's inventory on every command. I'm working on improving the prompt.

Reply on news.ycombinator.com