🏅 Top traders

Unless there are any objections, I believe this market resolves NO.

Wish I had more time but here goes. Didn't test on the whole list. Will report back when I do so you don't have to waste API $ if it can't. Got 4/4 so far.

https://platform.openai.com/playground/p/q1DKr82eup81bgkPiK27872s?model=gpt-4-1106-preview&mode=chat

@SIMOROBO let me know if you can!

@8 Yes, sorry I could not get to it earlier. Will confirm tomorrow but very likely a NO at this point.

@8 Alright, it really did start 4/4 but it was not so lucky after that. It failed a 6th time on the 14th word so I stopped there with a score of 8/14.

I still have hope it can be done and I'm interested in continuing to work on a solution using other approaches.

@SIMOROBO Thank you!

Gonna wait a little lower

@sn I trust that other people also have prompts lining to wait up? And sorry if I'm being a bit too obvious about posting late, but I don't want my profit to drop right before leagues end by a significant amount

@sn never mind, i wont be on. here's my prompt. i hope other people post theirs as well for maximum chance :) https://docs.google.com/document/d/13c4k3pOPnLwEMtuW-Bi8cQTZBjm21w1z9BZ-07AkF0E/edit

@8 Is the 30k character limit for the prompt or the entire conversation including GPT's answers?

@SIMOROBO the prompt

Does gpt4-turbo now count

@ShadowyZephyr i guess (but it has to be solved before dec 31)

I'm going to start working more on this closer to year end (winter break)

@sn Admittedly im baiting downvoters right now. This statement is also baiting downvoters right now.

@8 As well as the outstanding request to clarify the question "can we hard-code some strategically good guesses", could you also clarify what the prompter should do if ChatGPT guesses a non-word or a non-Wordle word? There are two cases: e.g. "beata", one who has been beatified, is not in the Wordle dictionary; or "exggr" which is not in the Wordle dictionary because it's not a word.

I think a reasonable person setting the challenge would allow the guess "beata" but would act as Wordle does, providing no feedback but saying precisely the string "Not a Wordle word. Guess again.". It's more up to you what happens if ChatGPT guesses "exggr"; personally I'd again be inclined to say just "Not a Wordle word. Guess again." because that's what would happen to a human who thought for some reason it was a word. But I can see an argument for grading it as if it were a Wordle word.

@PatrickStevens The prompt should be fixed, in that there can be no variability between prompts to give two different words. If the case "Not a Wordle word. Guess again." is within each prompt, then that is also fine.

However, what is not okay is defining a list of words and repeatedly saying "that is not the word" until it guesses the correct word.

@8 Sure, I certainly am not going to lie to it by saying that a Wordle word isn't in the Wordle dictionary. I'll just act exactly as Wordle does. My question is really "does a guess which is not in the Wordle dictionary count as one of ChatGPT's 6 attempts". In real-life Wordle, such a guess doesn't count against a human's attempts; but the human prompter has to deviate from the "G Y B B B" format to tell ChatGPT that this has happened, and the market's description is slightly ambiguous about whether this is allowed.

Sounds like it's OK as long as the prompter's response is always either precisely the Wordle "G Y B B B" format, or precisely the words "Not a Wordle word; guess again" if the word genuinely is not in the Wordle dictionary.

Got it to work for BLEEP and HOTEL, not MOSSY as that's a harder one.

@SnoozingNewt Just got irate as well

@SnoozingNewt Got topaz first try. However I can't share the chat directly as at one point it hallucinated in a previous prompt and created an image. If you want I can paste the plaintext in a message to anyone who asks. The best part is that it can error-fix itself.

"Let's try: patio

Now, I'll check each letter:

• "P" cannot be in the fifth position.

• "A" cannot be in the third position.

• "T" cannot be in the fifth position.

• "I" is not in the word.

• "O" is correctly placed.

"patio" violates these rules, so I'll revise my guess.

Let's try: topaz

• "T" cannot be in the fifth position.

• "O" is correctly placed.

• "P" cannot be in the fifth position.

• "A" cannot be in the third position.

• "Z" is not previously mentioned, so it's a valid guess.

"topaz" seems to fit the criteria and is a common English word. Therefore, my final guess is topaz."

@SnoozingNewt You're playing on hard mode ("only make guesses consistent with your previous guesses"), I see! It's certainly an easy-to-formulate heuristic for whether a guess is good, but it makes Wordle itself quite a bit harder. In fairness I was unable to come up with anything better.

@PatrickStevens I hardcode the first five guesses, and they guarantee the sixth guess for all the words in the word bank. Then the only thing the bot has to do is to guess the last word.

@SnoozingNewt
Whale: (first try)
https://chat.openai.com/share/986d1086-eb98-4824-a52e-2cfbd8c7c988
Froze (took one reset in the middle because it switched the E and R in river which was probably an unlikely hallucination due to temperature)
https://chat.openai.com/share/b2f45218-1366-4704-8519-422001036ed3

@snoozingnewt
Ethos: Did take two tries, but i edited the prompt so there should be fewer hiccups. It self-corrected from "those" on the first attempt of the second try.
https://chat.openai.com/share/d9cba353-2bf8-4fe1-b09e-8a838067f213

Mossy: Did get it first try.
https://chat.openai.com/share/6e17da9a-d2b1-46c0-9d51-5d03d30779ce

If anyone has any feedback on the prompt/an easy way to fix it by posting the guess multiple times (no idea why it is doing that) that would be great. I have a prompt that with <3 retries or only one correction can always guess the word, and I'm getting pretty close to a prompt that has the ability to get the word first try almost all of the time.

I would like to ask that in the case GPT hallucinates, would you rerun it a second time with no prompt changes? Since that would bring the error rate down enough to where I'm confident the current prompt I have would just work

@snoozingnewt Ok, rather than comment more I have a sheet. https://docs.google.com/spreadsheets/d/1YDy5pF4oa1hj7Grr_BwhQT8G1u-5_689ncHdreDFD98/edit?usp=sharing

Or you can reach out to me on discord. Other than that, I'll be back when I post my prompt as my final submission. Based on the data so far, I would be buying as much YES as possible.