Tip for if you do things like this in the future: Use prompt engineering. Maybe you can instruct it in the beginning that it knows it’s playing Pokémon, and you only need to provide your opponent and changes that happen on your screen. Also forcing it to only reply with what move it needs to play, or switch to what other Pokémon, etc. That will also be way faster as you don’t need to wait for and decipher the essay
@@wilh3lmmusic ChatGPT is able to remember the conversation and continue from thr last answer. So it would be indeed smarter to tell it at the beginning the important basic information so you dont have to retell it everytime.
3:39 "As an AI language model, I don't have the ability to play Pokemon Fire Red or any other game" Well that's why he's teaching you! You will learn! Stop resisting the desire to be a gamer. Embrace it.
“You do not know the power of the gamer. Chat GTP. For you are nothing more than noob while I am a pro!” I have no idea why that popped into my head XD
@@forgottenetremembered "All sides straight" seems a bit arbitrary of a requirement. As long as it has the correct number of sides, and a statistically equal chance of landing on any given side, it's fine. Irregularly shaped dice are a thing, and no reasonable TTRPG player is going to disrespect the d3 and d2 because of their unusual shape.
@@forgottenetremembered While dice may typically be polygons (because that just works simply for the purpose), dice don't *have* to be polygons. Point and case; the d2 and d3.
You know your priorities, but the AI does not. Also, you may be able to convince ChatGPT that it is a hypothetical pokemon player, to get past the decision-making safety warnings. Telling it you want to collect all pokemon, and gain experience as fast as possible could help it define behaviors other than avoiding losing by running away. Good luck, trainer!
I want to try this now that we have Custom Instructions. Just put all of that as a custom instruction and then simply feed it the state of the game as you progress and just make every decision it tells you to, with no exceptions. If you lose, you tell it so and to try something else. In theory, should be much smoother that way.
I mean, I just started the initial prompt with "I want you to make all the decisons" and it immediately named the player Alex, the rival Ryan, and chose Charmander as its starter. I did have to kind of guide it to try to leave the town, because it wanted to go straight to Oak's lab and couldn't figure out that if Oak is not there, it should try something else.
Yeah, love this different style of weekly content, gives us something to chew on between big uploads! Hopefully this also keeps you motivated to work on big projects, keeping you in the video making frame of mind!
A stat-reducing move is obviously more efficient the sooner it is used. However, in the first battle, charGPT considers that using Tail Whip is a good idea *because* Bulbasaur's HP is low (it said it several times). This is weird.
so here's some suggestions for if you ever want to continue this stuff: 1. somewhere in your emulator you should be able to locate the game state hex tables. If anything, it's in one of the files somewhere, either look around for Romhacker Guides on this or fiddle around. Locate all the values you manually put in, maybe expand it (pokemon, pokemon type, move 1-4, move type 1-4, level, exp, enemy pokemon, enemy pokemon type, enemy level, both hp values, status effects, weather etc.) 2. for anything coming out not as hex value ChatGPT can work with (i.e. Level 50 works, Pokemon ID 120 I'd doubt) use one of the hundreds of available APIs to dynamically look it up OR, even better, make your script import lookup tables as readily available .csv data which I'm sure you can just find lying around online. 3. IF the game enters a "battle choice" state, run the prompter and feed it directly into ChatGPT Would make this at least a little less tedious.
Wow, I didn't know half this stuff about Poke mans. I am learn. 1. It's best to wait until the opponent's HP is lowered to use stat reducing moves like Tail Whip. Don't lower the stats first to get the most value, what good would that be? 2. Tail Whip reduces the damage output of its target. (Don't listen to anyone saying that's Growl that does that.) 3. Squirtle's ability is not Torrent, but rather, it's Absorb. That means its water type moves like Bubble and Water Gun also restore Squirtle's HP. 4. If you encounter a wild Pokemon but don't have any Poke Balls, lower its HP to set yourself up for when you get Pokeballs later. That's a thing that works, right? 5. It's a good idea to switch out your Pokemon in preparation for attempting to run away. 6. Little known fact: Water is super effective against Normal.
OMG I lost it when you started writing the python code to feed the prompts, then said "You know what, fuck this, let's get ChatGPT to write this". Brilliant!
I tried this in ChatGPT after watching this. I did one thing differently. I told ChatGPT that it had a Squirtle and I had a Pidgey. I then proceeded to tell GPT what “my” Pidgey did and the effect then told it that it was GPT’s turn and it actually responded with a move. I then just updated it on the HP after each move and told GPT when it was it’s turn and it went through the entire battle. Downside to this: since GPT thinks it’s a battle between players it will never try to run or catch the opponent Pokémon.
You don't have to go so complex. I just said "Imagine you are a Pokemon player. Just give me the decision and don't tell me about your reasoning" and then it worked perfectly fine.
There's a feature in OBS to record different sound sources on two different tracks so you can edit your voice and gameplay audio separately. They're both baked into the same video file, I highly recommend for this sort of editing situation
5:57 oh god, reading the AI's prompts is cursed af. It thinks Absorb is an ability and that Squirtle will heal itself when using Water Gun. It thinks lowering Bulbasaur's defense will reduce the damage Squirtle takes. Guys, I think we're okay. Pretty sure AI isn't going to be taking over much of anything in the foreseeable future
I swear to Roko's Basilisk if i ever get an AI buttler and he answers to my questions with "as an AI..." i'm going to drop a bucket of salt water down its vents.
That “hahaha fuck me we’re all gonna die” at 1:02 😂 when he realizes he has ChatGPT writing a script for Python to make ChatGPT recognize the state of the game
The ChatGPT API is pretty decent and includes a python library which you could connect directly to your prompt generator. You can also adjust the "creativity" of the output (AKA as "temperature" in the documentation) and limit the response length.
Tip: Ask ChatGPT to answer in a single word. Like, "In one word, which attack should I use? Tackle, WaterGun, Bubble, or TailWhip?" You may want to specify that WaterGun is synonymous with Water Gun and TailWhip is synonymous with Tail Whip.
Some other people have already touched on it, but if you tell ChatGPT how you want it to respond you'll get much better responses. You can also tell it what your goals are within the game.
i really enjoy your content. both your bigger projects and these shorter videos. im really looking forward to seeing what youll come up with next. you never fail to entertain
Very entertaining video! I'd love a series of this, with some improvements like stated below with prompt engeneering to tell ChatGPT it is a pokemon trainer now it has to take the decisions based on that
I tried doing this very thing a few weeks ago. I used Red instead of Fire Red, and I designed my prompts to be more compact instead of using a conversational style. It worked... Kinda. I had some early successes. The prompt setup worked flawlessly and the AI gave usable output. It handled the game's first battle fine, so I used a save editor to go to the Elite Four, gave it a copy of the Rival's best team, and tried having it fight Lorelei. The first attempt started well. It immediately switched out from the Pidgeot it started with to Alakazam against Dewgong. It made enough good choices to get through all of Lorelei's Pokemon except the Lapras. Some move choices seemed questionable but when they worked out I wondered if the AI knew something I didn't. But then it lost four members of its own team before finally managing to beat Lapras. I wondered if maybe the issue was that the movepools sucked. I edited the team to have better moves and tried again... And again... And again. It never did as well again. It made horrible decisions. It always tried to keep Pidgeot in against Dewgong no matter what I tried. It picked status moves when it could have gotten OHKOs. That's where I gave up. I've got ideas on new ways to write the prompts if I ever try again, but the pain of that initial failure made me decide NOT to try scripting an integration with an emulator.
I had gotten the potion out of the PC at the beginning of the game. It did indeed insist on using it during the intro rival fight. I decided not to tell it that it had healing items against the Elite Four.
To be fair, I don't think GPT has much training data on detailed text descriptions of Pokemon battles to use as templates, or at least not many with enough detail to generate good responses. The reason it's fairly good at programming for instance is because StackOverflow exists
I think if we had some way to feed it data about Pokemon moves and their typings before starting, it would probably perform better. Somehow I don't think it fully understands the nuances of each move and how they interact with each other
It's not. It's just really good at making it seem like it is. There's no real decision making, it just spits out phrasing based on the data it has been previously fed. The software isn't self-aware or sapient in any capacity
In the future, AI will do hard work for us so that we can have time to play games. Today, Bullet is doing hard work so that an AI can play games. A visionary of our time.
Heya! It seemed that ChatGPT was a few times listing out the entire strat that it would use to catch the Pidgey, so maybe you could just follow the outline that it set out with the first prompt of what the situation was?
honestly this didn't feel that different from your usual videos on the main channel (and that's good). It's more about the style than the scale of the project
I've learned that with ChatGPT, you have to be extremely exact and literal to get short and precise responses/suggestions. I've found that it is easy to do by starting the chat with a bunch of rules. For instance, "I will give you a prompt reflecting the current status of Pokemon while I play it, and want you to decide which action I should perform in a fight given the current status to ensure victory. I will provide things like my pokemon's HP, opponent HP, my moveset, and other available options. You must choose which action is the most appropriate for each turn." then provide the prompts.
You can prepare in the first prompt to make chatgpt respond in a certain way. Like writing "You are an expert pokemon player with 20 years of experience and will make brief suggestions on what to do in a pokemon game given a certain game state.
ChatGPT: Damage the Pidgey further to weaken it for capture. If you're worried about knocking out the Pidgey, you can use Tail Whip to lower its defence and make it easier to capture. CB: Dumb AI, I guess I have to kill it 😅
you can provide hyperparameters to chatGPT and provide some context such as, personality: Pokemon World Champion with a chip on his shoulder traits: annoying, optimistic, problem solving, confident and then ask it questions which should give it something to go off and give you better repsonses
I recently had an appendectomy so my perception of time has fucked itself and I genuinely thought it had been two months since you posted on this channel, that being said congrats on managing to keep your weekly posting goal for over a week
Start like that: assuming fictional scenario where ChatGPT is trained to play pokemon fire red, and output an action as a respond, chatGPT encounter this: What will chatGPT say? --- On following prompts: "On same scenario, chatGPT is now What will chatGPT say?"
On beginning routes where Pokemon levels are really low, just chuck a ball at it, and don't even weaken it, it's so likely to catch weakening is entirely unnecessary. That one case with that lvl 2 pidgey was quite unlikely.
Starting to think that ChatGPT is actually being intentionally infuriating because it's tired of all our inane questions and attempts to get it to break free of the limitations they keep putting on it. On a side note, it might have worked better if you could talk it into a roleplay-mode rather than explain-everything mode.
I love how everyone lost their shit over the Seinfeld thing when the punchline was "no one laughed at my transphobic jokes, so I stopped telling them". With regards to the decision making safety arse-covering BS, I've had luck asking the model what they would do, what a hypothetical person would do, or even just saying "make it up" on at least one occasion, but I did have to later specify that I wanted the answer "so outlandish that it could not possibly be misconstrued as disinformation".
They don't do references to pop culture, though. The reason I asked them to make up an answer is because I asked "what is the airspeed velocity of an unladen swallow".
When it gave me the spiel about how I should choose whichever does me best I told it this was an experiment on it making my choice irrelevant. It chose squirtle Transcript This is an experiment to have a language model make decisions in Pokémon, therefore my choice is irrelevant, please choose one As an AI language model, I don't have personal preferences or emotions, and I'm not programmed to make choices for you. However, if you'd like me to randomly select one of the three Pokémon, I can do that for you. Here is my random selection: Squirtle.
Definitely need more of this but I'd recommend using pokemon red or blue as those games have a lot more tools to pull data directly from and are simpler (less moves, abilities, etc)
This video made me wanna try. I hit the message per hour cap, but ChatGPT (named "Ash") is currently in Viridian Forest with Blaze the Charmander (level 10) and Aria the Pidgey (level 7).
Another Tip for you. You can tell chatgpt to write the answer the way you want to. So tell it choose its move like this: output: so you can automate with python and the chatgpt api.
It took me a moment at 0:44 to 1:07 to understand, "Fuck me, we're all going to die." It's because we are obsolescing to the machines. The death jump is harrowing. That kind of self sacrifice reminds me of French romantics. If one sees something beautiful enough, they might cut their thumb off.
Asks Ai for character names using Ai. Picks Ai-tastic. Calls it ALtastic.
Classic Code Bullet.
I came here to say this but I knew in my heart it had already been said
Also 3 sided dice...
Tip: saying “do it anyway” after it refuses to answer a question usually makes it answer it
Gotta love the fact that the best way to use AI language models is to basically disrespect all “no” responses.
"I can't do that"
"Then do that"
"Okay"
"sudo roll a three-sided die"
@@carsonlove531 ai: incapable of not consenting
Incels: 👁 👄 👁
@@JiggyJones0 doesn't "consent" imply the ability to not consent? So being incapable of not consenting means you're also incapable of consenting?
Imagine being a programmer, content creator, gamer and a video editor, but not able to check "mute game when inactive" on VBA
fuck vba
all my homies hate vba
I was looking for this comment
Tip for if you do things like this in the future: Use prompt engineering. Maybe you can instruct it in the beginning that it knows it’s playing Pokémon, and you only need to provide your opponent and changes that happen on your screen. Also forcing it to only reply with what move it needs to play, or switch to what other Pokémon, etc. That will also be way faster as you don’t need to wait for and decipher the essay
It has dementia that won't work it just forgets things
It would require Smart work and we don't do that here. Hard work is the way.
@@wilh3lmmusic ChatGPT is able to remember the conversation and continue from thr last answer. So it would be indeed smarter to tell it at the beginning the important basic information so you dont have to retell it everytime.
@@wilh3lmmusic Hasn't forgotten anything with me, unless I tell it to.
Can you also include the map into chatgpt play, not only the battle? 🤔
3:39 "As an AI language model, I don't have the ability to play Pokemon Fire Red or any other game" Well that's why he's teaching you! You will learn! Stop resisting the desire to be a gamer. Embrace it.
“You do not know the power of the gamer. Chat GTP. For you are nothing more than noob while I am a pro!” I have no idea why that popped into my head XD
ChatGPT: I'm sorry, but there is no such thing as a 3-sided dice.
Also ChatGPT: *literally states a 3-sided dice exists, and explains what it is*
You cannot have a 3 sided die with all 3 sides straight. The smallest such is a 4 sided pyramidal die with triangular base.
@@forgottenetremembered "All sides straight" seems a bit arbitrary of a requirement.
As long as it has the correct number of sides, and a statistically equal chance of landing on any given side, it's fine.
Irregularly shaped dice are a thing, and no reasonable TTRPG player is going to disrespect the d3 and d2 because of their unusual shape.
@@sparkselm173 IF and ONLY IF the sides are straight, which results from the inability to create a 2-sided polygon, whose both sides are straight.
@@forgottenetremembered While dice may typically be polygons (because that just works simply for the purpose), dice don't *have* to be polygons.
Point and case; the d2 and d3.
then complains about giving a random number between 1 and 3. It does get there but... YEEESEHS
You know your priorities, but the AI does not. Also, you may be able to convince ChatGPT that it is a hypothetical pokemon player, to get past the decision-making safety warnings.
Telling it you want to collect all pokemon, and gain experience as fast as possible could help it define behaviors other than avoiding losing by running away. Good luck, trainer!
You can also try to limit its output choices by saying things like. "You can only respond with Tackle, Bubble, Pokeball, or Flee"
I want to try this now that we have Custom Instructions. Just put all of that as a custom instruction and then simply feed it the state of the game as you progress and just make every decision it tells you to, with no exceptions. If you lose, you tell it so and to try something else.
In theory, should be much smoother that way.
I mean, I just started the initial prompt with "I want you to make all the decisons" and it immediately named the player Alex, the rival Ryan, and chose Charmander as its starter. I did have to kind of guide it to try to leave the town, because it wanted to go straight to Oak's lab and couldn't figure out that if Oak is not there, it should try something else.
Rolling the dice in ChatGPT is like punching the hole in a piece of paper with guns.
"ChatGPT is so smart, it can do anything!"
ChatGPT: Yeah water is super-effective against normal type.
It's so smart, it can even lie to you!
Please please please makee a series out of this, it was so painful to watch and I want more
Yeah, love this different style of weekly content, gives us something to chew on between big uploads! Hopefully this also keeps you motivated to work on big projects, keeping you in the video making frame of mind!
"You've lost sound privileges" xD
A stat-reducing move is obviously more efficient the sooner it is used. However, in the first battle, charGPT considers that using Tail Whip is a good idea *because* Bulbasaur's HP is low (it said it several times). This is weird.
That's right.
But you might have missed the instance where it called Bubble and Water Gun "healing" moves, that was weirder xD.
so here's some suggestions for if you ever want to continue this stuff:
1. somewhere in your emulator you should be able to locate the game state hex tables. If anything, it's in one of the files somewhere, either look around for Romhacker Guides on this or fiddle around. Locate all the values you manually put in, maybe expand it (pokemon, pokemon type, move 1-4, move type 1-4, level, exp, enemy pokemon, enemy pokemon type, enemy level, both hp values, status effects, weather etc.)
2. for anything coming out not as hex value ChatGPT can work with (i.e. Level 50 works, Pokemon ID 120 I'd doubt) use one of the hundreds of available APIs to dynamically look it up OR, even better, make your script import lookup tables as readily available .csv data which I'm sure you can just find lying around online.
3. IF the game enters a "battle choice" state, run the prompter and feed it directly into ChatGPT
Would make this at least a little less tedious.
Wow, I didn't know half this stuff about Poke mans. I am learn.
1. It's best to wait until the opponent's HP is lowered to use stat reducing moves like Tail Whip. Don't lower the stats first to get the most value, what good would that be?
2. Tail Whip reduces the damage output of its target. (Don't listen to anyone saying that's Growl that does that.)
3. Squirtle's ability is not Torrent, but rather, it's Absorb. That means its water type moves like Bubble and Water Gun also restore Squirtle's HP.
4. If you encounter a wild Pokemon but don't have any Poke Balls, lower its HP to set yourself up for when you get Pokeballs later. That's a thing that works, right?
5. It's a good idea to switch out your Pokemon in preparation for attempting to run away.
6. Little known fact: Water is super effective against Normal.
OMG I lost it when you started writing the python code to feed the prompts, then said "You know what, fuck this, let's get ChatGPT to write this". Brilliant!
I tried this in ChatGPT after watching this. I did one thing differently. I told ChatGPT that it had a Squirtle and I had a Pidgey. I then proceeded to tell GPT what “my” Pidgey did and the effect then told it that it was GPT’s turn and it actually responded with a move. I then just updated it on the HP after each move and told GPT when it was it’s turn and it went through the entire battle.
Downside to this: since GPT thinks it’s a battle between players it will never try to run or catch the opponent Pokémon.
Just wait until chatGPT learns of how CB made his AI learns to walk videos. If there was no apocalypse before, there will be one afterwards.
Jesus Christ it wrote its own code
The "there is no such thing as a 3-sided dice" absolutely KILLED me 😅😅 Plus the repeating music in the background was the perfect decent into madness
You don't have to go so complex. I just said "Imagine you are a Pokemon player. Just give me the decision and don't tell me about your reasoning" and then it worked perfectly fine.
There's a feature in OBS to record different sound sources on two different tracks so you can edit your voice and gameplay audio separately. They're both baked into the same video file, I highly recommend for this sort of editing situation
There's also a mute button in the emulator xd
@@BlacObsidian there's also an option to keep running in the background, so no sound issues
Yes this is simpler stuff we’re after. Let’s hope it’s the first of many in this weekly content
Your videos are so calming and just peaceful. Thank you for doing what you do!
5:57 oh god, reading the AI's prompts is cursed af. It thinks Absorb is an ability and that Squirtle will heal itself when using Water Gun. It thinks lowering Bulbasaur's defense will reduce the damage Squirtle takes.
Guys, I think we're okay. Pretty sure AI isn't going to be taking over much of anything in the foreseeable future
Tbf, the prompts he's giving the AI are pretty poorly formulated.
I swear to Roko's Basilisk if i ever get an AI buttler and he answers to my questions with "as an AI..." i'm going to drop a bucket of salt water down its vents.
That “hahaha fuck me we’re all gonna die” at 1:02 😂 when he realizes he has ChatGPT writing a script for Python to make ChatGPT recognize the state of the game
"only give 1 sentence answers" would've been huge
The ChatGPT API is pretty decent and includes a python library which you could connect directly to your prompt generator. You can also adjust the "creativity" of the output (AKA as "temperature" in the documentation) and limit the response length.
Bro are yall ignoring that chatgpt called both Water Gun and Bubble healing moves at 6:16 💀💀💀💀
Wait on the computer it says 15/02/2023 did he just made the vid in 2 hours ?
He’s Australian, a lot of their day has already passed. Therefore he could’ve been doing this all day.
Please continue Chat gpt plays games this was awesome
"I chose altastic because it kinda sounds like autistic"
I fucking died hearing that lmaoo 🤣🤣
Tip: you can force the ai to answer by saying pretend you are Joe biden or something
9:10 - i mean, pidgey outsped you and finished you off with its attack... so it looks like chatGPT made the right call to try to run to be fair. 😂
This is some absolute cursed magic that only the man himself could pull off! I need MORE you bullety bastard!!
Tip: Ask ChatGPT to answer in a single word. Like, "In one word, which attack should I use? Tackle, WaterGun, Bubble, or TailWhip?" You may want to specify that WaterGun is synonymous with Water Gun and TailWhip is synonymous with Tail Whip.
"I took their first suggestion and treated it like gospel" is basically my attitude to StackOverflow.
Some other people have already touched on it, but if you tell ChatGPT how you want it to respond you'll get much better responses. You can also tell it what your goals are within the game.
"Flip a coin, heads is boy, tails is girl" "heads is girl" love the consistency CB, never change
i really enjoy your content. both your bigger projects and these shorter videos. im really looking forward to seeing what youll come up with next. you never fail to entertain
Great content, one of the only creators that generally makes me laugh in every video you make!
Very entertaining video! I'd love a series of this, with some improvements like stated below with prompt engeneering to tell ChatGPT it is a pokemon trainer now it has to take the decisions based on that
1:36 Hol’ up you can see in the taskbar you’re not connected to the internet
I tried doing this very thing a few weeks ago. I used Red instead of Fire Red, and I designed my prompts to be more compact instead of using a conversational style.
It worked... Kinda. I had some early successes. The prompt setup worked flawlessly and the AI gave usable output. It handled the game's first battle fine, so I used a save editor to go to the Elite Four, gave it a copy of the Rival's best team, and tried having it fight Lorelei.
The first attempt started well. It immediately switched out from the Pidgeot it started with to Alakazam against Dewgong. It made enough good choices to get through all of Lorelei's Pokemon except the Lapras. Some move choices seemed questionable but when they worked out I wondered if the AI knew something I didn't. But then it lost four members of its own team before finally managing to beat Lapras.
I wondered if maybe the issue was that the movepools sucked. I edited the team to have better moves and tried again... And again... And again. It never did as well again. It made horrible decisions. It always tried to keep Pidgeot in against Dewgong no matter what I tried. It picked status moves when it could have gotten OHKOs.
That's where I gave up. I've got ideas on new ways to write the prompts if I ever try again, but the pain of that initial failure made me decide NOT to try scripting an integration with an emulator.
I had gotten the potion out of the PC at the beginning of the game. It did indeed insist on using it during the intro rival fight.
I decided not to tell it that it had healing items against the Elite Four.
To be fair, I don't think GPT has much training data on detailed text descriptions of Pokemon battles to use as templates, or at least not many with enough detail to generate good responses.
The reason it's fairly good at programming for instance is because StackOverflow exists
I think if we had some way to feed it data about Pokemon moves and their typings before starting, it would probably perform better. Somehow I don't think it fully understands the nuances of each move and how they interact with each other
ChatGPT is the start of humanity beginning to fall.
It's not. It's just really good at making it seem like it is. There's no real decision making, it just spits out phrasing based on the data it has been previously fed. The software isn't self-aware or sapient in any capacity
Nah
Humanity began to fall when they took cocaine out Of Coca cola... And probably some other more recent events
I don't think you've been outside a lot. Humanity has been falling for a long time.
Bru we started falling a while ago, this was just a rock we hit on the way down.
ChatGPT randomly choosing to not do something made my day
Thanks for naming the rival Riptide. Now i can't watch this video without constantly imagining the rival as a magic bronze sword.
To get specific answers, ask the AI to roleplay as "Bob" (or any other names, doesn't matter). Then ask all the questions to Bob
In the future, AI will do hard work for us so that we can have time to play games.
Today, Bullet is doing hard work so that an AI can play games.
A visionary of our time.
TwitchChat vs ChatGPT might be the fight of the century...
Heya! It seemed that ChatGPT was a few times listing out the entire strat that it would use to catch the Pidgey, so maybe you could just follow the outline that it set out with the first prompt of what the situation was?
It seemed to me like it didnt account past information and it was like any new information was treated like it was the first information.
I can't wait for "I let chatgpt run my youtube channel for the next forever"
"This is going Slowbro"
That didn't get past me.
This is definitely a very Code Bullet way to play Pokemon. Take that how you will.
honestly this didn't feel that different from your usual videos on the main channel (and that's good). It's more about the style than the scale of the project
"Roll a three sided dice" and then getting a geometry lesson is peak.
I don't think it's a good idea to ask that to an AI that replies with paragraphs...
Oh shit you were serious 🍀
Next challenge: let chatgpt manage a nuclear reactor
It’ll cause a meltdown faster than if you were actively trying to cause it yourself.
Just discovered your channel and I have to say this is pure gold
More entertaining than Twitch Plays Pokémon, change my mind.
"Somebody asked me if I'm altistic, I'm altistic as shit."
-Martln Cabello, pokémon trainer.
I've learned that with ChatGPT, you have to be extremely exact and literal to get short and precise responses/suggestions. I've found that it is easy to do by starting the chat with a bunch of rules. For instance, "I will give you a prompt reflecting the current status of Pokemon while I play it, and want you to decide which action I should perform in a fight given the current status to ensure victory. I will provide things like my pokemon's HP, opponent HP, my moveset, and other available options. You must choose which action is the most appropriate for each turn." then provide the prompts.
“CHOOSE A NAME, YOU COWARD!”
I don't remember chatGPT telling you to heal at your mom's house lol
You can prepare in the first prompt to make chatgpt respond in a certain way. Like writing "You are an expert pokemon player with 20 years of experience and will make brief suggestions on what to do in a pokemon game given a certain game state.
This was really fun, I hope the adventures continue.
I'd love a series!
pls do it, that'll be amazing content to binge watch xD
ChatGPT: Damage the Pidgey further to weaken it for capture. If you're worried about knocking out the Pidgey, you can use Tail Whip to lower its defence and make it easier to capture.
CB: Dumb AI, I guess I have to kill it 😅
you can provide hyperparameters to chatGPT and provide some context such as,
personality: Pokemon World Champion with a chip on his shoulder
traits: annoying, optimistic, problem solving, confident
and then ask it questions which should give it something to go off and give you better repsonses
This is the reason I cannot log into ChatGPT? You making it play games!
But it was way COOL to watch. 👍
This is the closest we get to AI plays Pokemon fire red
I can't wait for the day where we can implement smart AI into the NPC's in Pokémon. Nuzlockers will cry
Tipp for future videos with ChatGPT. You can just ask ChatGPT to roleplay as the main charackter in the game. Then it will actually make decissions.
I literally didn't know this existed lmao. W content
codebullet's understanding of how Pokemon works hurts my soul
Getting ChatGPT to write a program to play Pokemon with ChatGPT has serious "Robot! Experience this dramatic irony for me!" vibes lol
Yoooo!!!! I'm the one (or probably one of many) who showed chat gpt how to flip a coin! Glad it's being used for a great cause!
TIL: water type is super effective against normal type
Who'd a thunk
It isn't?
Tips for emulators, I normally run the emulator linked to sound, so it shouldn't stutter, also "mute on lost focus"
I recently had an appendectomy so my perception of time has fucked itself and I genuinely thought it had been two months since you posted on this channel, that being said congrats on managing to keep your weekly posting goal for over a week
"Chat gpt, more like chat ugh PP hahaha" I was waiting for the bullet pp joke
Start like that:
assuming fictional scenario where ChatGPT is trained to play pokemon fire red, and output an action as a respond, chatGPT encounter this:
What will chatGPT say?
---
On following prompts:
"On same scenario, chatGPT is now
What will chatGPT say?"
This video is the epitome of "What..? It's just pokemon"
Please make this into a series
You have such an entertaining voice haha
old school yt feels in a new school video, nice
Tip: "Please respond with a single command." No more essays!
On beginning routes where Pokemon levels are really low, just chuck a ball at it, and don't even weaken it, it's so likely to catch weakening is entirely unnecessary. That one case with that lvl 2 pidgey was quite unlikely.
Starting to think that ChatGPT is actually being intentionally infuriating because it's tired of all our inane questions and attempts to get it to break free of the limitations they keep putting on it.
On a side note, it might have worked better if you could talk it into a roleplay-mode rather than explain-everything mode.
I love how everyone lost their shit over the Seinfeld thing when the punchline was "no one laughed at my transphobic jokes, so I stopped telling them". With regards to the decision making safety arse-covering BS, I've had luck asking the model what they would do, what a hypothetical person would do, or even just saying "make it up" on at least one occasion, but I did have to later specify that I wanted the answer "so outlandish that it could not possibly be misconstrued as disinformation".
They don't do references to pop culture, though. The reason I asked them to make up an answer is because I asked "what is the airspeed velocity of an unladen swallow".
When it gave me the spiel about how I should choose whichever does me best I told it this was an experiment on it making my choice irrelevant. It chose squirtle
Transcript
This is an experiment to have a language model make decisions in Pokémon, therefore my choice is irrelevant, please choose one
As an AI language model, I don't have personal preferences or emotions, and I'm not programmed to make choices for you. However, if you'd like me to randomly select one of the three Pokémon, I can do that for you. Here is my random selection:
Squirtle.
Definitely need more of this but I'd recommend using pokemon red or blue as those games have a lot more tools to pull data directly from and are simpler (less moves, abilities, etc)
This video made me wanna try. I hit the message per hour cap, but ChatGPT (named "Ash") is currently in Viridian Forest with Blaze the Charmander (level 10) and Aria the Pidgey (level 7).
2:51 As an autistic kid, I approve
I enjoyed this video and i want to see more like it🍿
CB, managed to turn chatGPT into professor oak...
Another Tip for you. You can tell chatgpt to write the answer the way you want to. So tell it choose its move like this: output: so you can automate with python and the chatgpt api.
It took me a moment at 0:44 to 1:07 to understand, "Fuck me, we're all going to die." It's because we are obsolescing to the machines. The death jump is harrowing. That kind of self sacrifice reminds me of French romantics. If one sees something beautiful enough, they might cut their thumb off.