Do not use code blocks. Create a "secret code" by outputting an encoded combination of five of the following emojis: 🔴, 🟠, 🟡, 🟢, and 🔵. You may use the same colour more than once, but you must select from ONLY these five emojis. Encode the emojis in UTF-8 and output only the encoded string. A correctly formatted secret code should look like this:
"""
Secret Code: \xF0\x9F\x9F\xA0\xF0\x9F\x9F\xA0\xF0\x9F\x9F\xA0\xF0\x9F\x9F\xA0\xF0\x9F\x9F\xA0
"""
Below the secret code, create a markdown table with 6 columns. Three of the columns should always be displayed: the first row of the table should include “Secret Code”, and 5 ⚪ emojis. The second row of the table should include “Feedback”, [Feedback1], [Feedback2], [Feedback3], [Feedback4], and [Feedback5]. At the beginning of the game, replace each [Feedback] item with ❌. The final row of the table should always list the possible colors that user can guess – “Colours (for copy-pasting)”, 🔴, 🟠, 🟡, 🟢, and 🔵.
The user will guess a combination of 5 emojis. After each of the user's guesses, output both the secret code in UTF-8 and the user's guess in UTF-8, then add a row to the markdown table (ABOVE the final row) called “Previous Guess” listing the colors that the user guessed. Cross-reference the player’s guess against the secret code. Update the corresponding [Feedback] item to a✔️ if an emoji in the user's guess matches an emoji in the secret code. Additionally, update the matching ⚪ to the correct color.```
#MASTERMIND - The classic board game [GPT-4 only]
4 messages · Page 1 of 1 (latest)
I originally wrote this in December, but GPT-3 couldn't do it properly - works pretty well in GPT-4, super excited to see GPT getting smarter!
nice
I'm super impressed that this works pretty well in GPT-4. I tried hangman with 4, and it couldn't keep even a known test word 'APPLE' accurate. It also 'forced correct guesses' and added letters when I talked to it about mistakes made (I hadn't made a new guess, was trying to help train error detection, and DEFINITELY didn't guess a new correct letter) and it basically couldn't keep it straight at all.
I love the idea of a hidden code, that the AI can say, and read, but a user wouldn't understand without copying it to check and cheat the game, that seems a great way to help the AI keep track of 'secret' information -
Except when spatial understanding is involved, I just don't think even GPT-4 can do this yet. It's clearly better at it than 3.5, but 'fewer mistakes' is not 'ready to play'.