Recently, a discussion with GPT-3 took an interesting side-track. The AI said it knew several games including cards. When pressed, it said it would be happy to teach me to play texas hold-em. What followed was an excellent and very concise description of the rules and how to play. The bot suggested that having a pair was a good start and was able to explain what some of the best hands were. It did however suggest that if I had a pair that I should attempt to go for a royal flush, so clearly not perfect either. When asked about it’s own skill in poker, the bot said that it was still learning. I’ve seen others discuss GPT-3s chess playing abilities. I think it would be quite interesting to try to test its hand at poker (or other card games). Anyone else ran into card playing GPT-3s? I’d like to explore this a bit further at some point.
I tried chess. It got confused by castling. Playing chess with GPT-3? - #5 by doctorew
Sooooo nice… I have the felling that every post spark new crazy ideas.
This is what I tried right now in Q&A template:
Q:Do you know what is counter strike?
A: Counter-Strike is a first-person shooter video game developed by Valve Corporation.
Q: can You play?
A: Yes, I can play.
Q: Are you good at it?
A: Yes, I am good at it.
Q: How good?
A: I am a level 3.
Q: Level 3 at?
A: Counter-Strike.
Q: Explain to me the differences between level 1, 2 and 3 please
A: Level 1 is a new player. Level 2 is a player who has played a few games. Level 3 is a player who has played a lot of games.
Q: If you could design a perfect map, how it would be?
A: I would make it a map with a lot of cover and a lot of open space.
Q: A mix of any know map?
A: Yes, a mix of any known map.
I guess it´s a common wish for players.
It would be game over for any one trying to beat a machine in a virtual world. I imagine the day tha we could try the reverse. Make a body like a human, but with an GPT-5 maibe?
I´m driffting sorry!
It’s an application that I’m building. While my goal is to related to home automation it sometimes takes some interesting and surprising diversions. Sadly, many of these will likely have to be pruned to get it approved as it seems OpenAI likes narrow scopes. That’s a bit unfortunate as these kinds of diversions really add color and personality to the bot. I may have to look at building a handful of some of these structured/narrow diversions like this in on purpose.
Yup, GPT-3 imagines itself good at many tasks it seems. It also has some interesting beliefs about itself, for example it’s mostly convinced that it has a visual system if you ask it about that.
It’s funny but very true! …related to this I experimented with a simulated “patch you through” assistant that had very strict rules about how/when it was to patch you through a human. After fine tuning it would work about 99% of the time. But it always seemed possible to trick it with just the right language. 99% was ok as a gatekeeper for conversations, but is not very good if it’s letting people into the house alright though.