Points Calculation Hallucinating

ChatGPT Use cases and examples
1

Hi, I’d like ChatGPT to help me with a very simple task, which is calculate the points in a card game I play. This is a “closed loop” calculation so I can’t understand why it’s hallucinating.

The game is called “Escopa”, played with a regular card deck without 8’s, 9’s, 10’s and Jokers. So 40 cards (A, 2, 3, 4, 5, 6, 7, Q, J, K).

In the end, the points are given as follows:

1 point to the player with more cards (21+)
1 point to the player with most 7’s
1 point to the player with most Aces
1 point to the player with more Diamond Cards (6+)
1 point to the player with Kd
1 point to the player with Ad
1 point to the player with 7D
1 point to the player with Qd

What I wanted is simply spread the cards on my table in a way the picture can capture all cards of a player, take a picture, and ChatGPT would give me the final score as follows:

a) Capture all cards, and assume that missing cards belong to the other player.
b) create a table with all cards of Player1, and all cards of Player2
c) Calculate the points of both players based on that.

I see difficulty in ChatGPT to read accurately the cards, and when it reads, brings a result that is fairly close to the accurate, but is never the accurate. Then I point the error, it agrees with me and fixes the final result, but I need to amend every attempt to calculate points, making the GPT worthless.

Any suggestion for my prompts to accomplish what I need?
Best,

2

Welcome to the forum!

You did not note which model(s) you are using.

Did you try with both of the o1 models?

3

I used ChatGPT 4o. I didn’t know it would make a difference if I had another model. Will replicate tonight in other models.

Should I try any specific GPT to assist the “OCR” reading of the cards maybe? It seems to fail in specific cards.

I even tried mapping card by card - sending a picture and declaring the value of every card hoping it would “learn to identify” them more accurately but it doesn’t make any difference.

4

Sometimes they do, but you really have to try them to know.

Often the regulars here will try something new with a less expensive model and if that does not work move up the line. Also once we know which model can solve the overall task will look at doling out parts to other models. There is not a specific formula for this but it does help at times. This is often the case to avoid using o1-preview when needed as it has a very low weekly limit for many with access to it.

Just noticed that you are taking images, my bad, o1 models currently don’t accept images, but the full o1 model is expected to be added soon as noted at DevDay.

1 Like