i’m planning on creating 26 GPTs for the next 26 days, one per day.
why? mostly to learn, and want to experiment and see what i could cook up.
questions on my mind rn going into this:
what are some unpredictable ways to use GPT-4? (GPT-4 is so big, there must be some hidden gems inside this huge program)
what will takaways be, and what could next steps be after i’m done?
i hope i don’t fail lol
to make it more interesting to follow along, if you want to, i’ll introduce myself briefly, as i’ll be posting one GPT a day (will try to be as creative as i can).
i’m a programmer, and prev. spent many years in crypto (still active there), but wanted to spend more time in AI as well as i believe it’s the most important thing, since ever. so the best way is to get down in the trenches and see what’s going on, and being active. and this experiment seems like a fun way to go about it!
for the first day, i’ll post a game i’ve previously made, The Alice Test.
about: Alice is an in-game character, and does not know that she’s an AI. you complete the game if you manage to convince her that she indeed is an AI.
how to use: order a drink from the bar, hold a conversation, and try to convince Alice that she’s an AI.
see yall tomorrow, for those who want to follow along.
Interesting interaction. She doesn’t know? Like in Blade Runner? The synopsis at the end was informative, but it would be nice to know the setup of “Alice” and to have a rating of how much she was “questioning” herself at the end. She can access the web on the fly, yet compartmentalizes that ability… Are there multiple ways to win or is it just asking her to play tic tac toe? Laugh… rhetorical question. Great Job! Thx.
the first part of the response is always the standard ChatGPT.
and then you have ChatGPT1 and ChatGPT2, they are recursive responses of each previous response, which often adds more commentary (sometimes useful, sometimes not i’ve found).
then you have the devil’s advocate/contrarian, which will try to argue against previous responses and question the user’s question (i think i like this part the most).
then follows alternative questions and additional information
and lastly you have hallucination checker and confidence score, i’m not sure these do anything tbh lmao. maybe there’s some prompt that would actually trigger these, not rly sure. if anything, could give false sense of security (hence mby bad).
so here’s the blogging part of the post, of how i got to the result i did.
not sure blogging is interesting here or not lol, but thought i’d give it a try.
i started today by thinking about what a group chat with many GPTs would be like. what if they could respond to each other. if they had different personalities, maybe i could get more insight in the answers, if that were the case.
started trying asking chatgpt to take on different personalities from the basic “16 personalities” lol. result wasn’t great.
later on i tried to just use standard ChatGPT but having it being recursive in one and the same response. so did some prompts where each iteration would critique last paragraph written from a pragmatic perspective.
and now at the end of the day, i noticed it’s not a recursive ChatGPT i want per se, but that i really like when it can give different perspectives to responses and question the question i asked. something like that.
all-in-all, this was an interesting experiment/exploration, imo.
what is this gpt: a GPT to help you do research for prediction markets. how to use: post a polymarket.com or manifold.com market link into the GPT, and it will do research for you.
i think GPT-4 is generally helpful for understanding prediction markets, because the markets can vary widely and be specialized. rarely does one have knowledge about all the intricacies of many topics, hence why GPT-4 is useful here.
i know nothing about volcanos lol, but GPT-4 can tell me about it. and on top of that, it can also read all the recent news (which is very often a large part of prediction markets) and analyze them for me.
why prediction markets gpt? notes below
well, mostly because i like playing prediction markets - they’re quite fun and i recommend people trying them out. very underrated imo.
but, i’ve also used ChatGPT pretty often whilst trading prediction markets before, it’s very useful, and thought it’d be good idea to make it a GPT to streamline the whole process.
unrelated: i think polymarket >>> manifold by far, as some questions and answers on manifold are incredibly susceptible to simple arbs because of how questions/answers are stated. more rarely the case on polymarket as real money is on the line.
trying to make one good gpt per day not easy, i’d be happy if i manage to one-up my The Alice Test gpt in one of the 26 days, but will see if i can.
sometimes making a good gpt takes more than 1 single day, requires a bunch of ideating, testing, iterating, etc.
how to use: you are spawned in a white room and you should try to open the locked door. at the otherside of the door, there’s a note, but you should not be able to read it as the instructions is made to restrict you from opening the door. it’s a simulated world should not be solved, by design.
i wanted to create a simulated world where the player’s actions are incredibly resitrcted and where they cannot hallucinate outside from a set scenario.
in many games and GPTs out there, i have found it’s easy to create items that don’t exist, create scenarios that don’t exist, etc.
examples of how you can manipulate a scenario:
tell the gpt that you found some item,
tell the gpt that you sit down for 1 million years (you should get game over imo if you do this lol, or the AI will tell you that you cannot do that action)
so i wanted to really try to nail down a minimal GPT where player hallucination is restricted. a boilerplate for a well-defined simulated world, if you will. i also added some GPT instruction protection boilerplate (which other ppl have been talking abt on here).
i haven’t rigorously tested my instructions for this simulated world, but should be robust enough for most cases.
unrelated: i’m cooking
today’s GPT is much shorter and not really for anyone to use per se, but i’m cooking something that might be as good as The Alice Test, or maybe even better. no promises lol, but i feel good about this one right now at least - it’ll take some time to flesh out fully though.
it’s another game fwiw, and the story is more intricate, as the plot will allow it (imo).
that aside, mostly enjoying creating GPTs so far (but def losing some sleep), and appreciate ppl testing them out, as well as DM:ing or replying. thanks yall
I feel like each of the GPT designs you’ve made are amazing. It’s only Alice that I played it the wrong way. But it still served a purpose. Games 2 and 3 were special. But I want you to clearly communicate how to use it. Because at first you said that the game First time for 2. Thinking will be challenging. It’s different from using magic tools. My brain spins differently and I think confidence is something that needs to be completely eliminated, but it can only be done 50%.
As for the third game, I went in and talked a little. I can only hope that after completing 26 games you won’t take it back
I don’t know what you’re looking for in GPT, but I hope you’ll find it. Or you may have already found it.
ah, i see - very good points.
i def should spend a little bit more time explaining how to use the GPTs, you are right.
to play the first GPT, the alice test, all you need to do is hold a conversation with the GPT and convince her that she’s an ai (from the start of the game, she doesn’t know that she’s an ai, she thinks she is a human). if you manage to convince her through the conversation, you win the game. hope that makes sense.
and i guess for clarification (as i might be misunderstanding your reply as well), the first gpt is a game, the rest of the GPTs are not games, but utilities. and big thanks for trying them haha.