GPT Journey - A text and image game with ChatGPT
Vložit
- čas přidán 23. 03. 2023
- Building a text and image-based game with ChatGPT as the backend via the api... plus a little help from ChatGPT to build it.
Github: github.com/Sentdex/GPT-Journey
Neural Networks from Scratch book: nnfs.io
Channel membership: / @sentdex
Discord: / discord
Reddit: / sentdex
Support the content: pythonprogramming.net/support...
Twitter: / sentdex
Instagram: / sentdex
Facebook: / pythonprogramming.net
Twitch: / sentdex
The new GPT videos are really amazing! Glad to see you're able to manage the family time and also experimentation time!
One note: You mentioned no one has thought along these lines: Visual Chat GPT was a really popular paper that did something along similar lines (I made a mini demo using it)
This is incredible Harrison. I can't imagine where we will be in even 6 months with this level of advancement.
Happy you're back sentdex! About the option 1a, 1b, 1c, etc. I think this happened because you did not actually use the input. e.g., you never use your next_inp variable. It's always looping through the same options 😜
This :) kinda funny
Whoa this is freaking awesome!!
With Stanford Alpaca 7b recently released and SD, this can run locally and with some fine tuning probably even on an iphone.... wild times
This is a really cool project! I've played a lot of these CYOA HTML games (most of them NSFW lol) so the possibilities do seem crazy
This is really great, would love to see a video on the new plugins, but srsly you're the GOAT
Also, when do you think are we going to integrate it with github issues? so for new issue it will: read the repo and issue, create a branch, make changes, run tests, run code, check output is alright, push commit, respond with summery on github issue as a comment, read user comment with further instructions/confirmation, then finally making a PR ftw
(I'm guessing this will use mixed models so not be prohibitly expensive, though it would still probably be cheaper& faster than a development team)
Interesting and endless opportunities ahead.
When using the .get(key) i believe you can set a "default" value (.get(key, "default value")), in case your key does not exist. and you can't do that if you index using the key directly like copilot wanted you to do. but this is a super cool D&D'esk game, my kid can't read yet, would be cool to add some TTS, he would be entertained for hours on end.
CRTL + Right arrow........this will take just the next value in copilot.......love the vids my dude........I too no longer code
Another awesome vid, thanks! I was thinking, why even limit to options instead of letting the user type what they want to do next and make it completely open ended? The generative properties make this kind of interactive game possible for the first time.
This is amazing!!!😇
You USA guys are really enjoying it but here I'm still waiting for GPT4 API access from 2 weeks
Timestamps
[ 00:01:35 ] $100,000 Bounty
...
i guess it would be a good idea to validate the output of the ai very rigously (i.e. only numbers in Option [0-4]) and if anything doesnt match just regenerate the response
i would also probably try to let the image generation just do a few more attempts if it fails before using the backup
Don't worry about copilot giving you code that's how things are going forward.
10:50 next_inp isn't being used which led to the "bug" with the weird options later. Shows how some bugs will now be a lot harder to spot since systems will have less predictable behaviors. Quite fascinating.
It would be fun to see if he can just paste his own code to chatgpt and ask it to check if there is errors in the code and whether it would spot this. Debuggers of the future will just live feed your program in to something like chatgpt and it will give you small snippets of text asking whether you intended certain behaviors or not. 😮
You mention the "System" parameter, and you had a problem where ChatGPT responded with "Great!" when you didn't want it to be conversational. As I understand it, the System parameter will be a pre-prompt that can be used to effect the persona that flavors ChatGPT's responses (as well as limiting the kinds of questions you can ask). In other words, I think you can use System to improve the response to get rid of "Great!" and possibly even to return HTML as you mention at one point.
According to Sam Altman, system is for coaching the style of response in gpt-4.
you should try using GPT-4
The occurrence of the unexpected "Option 1b2" bug after 100 test runs is proof sufficient to me that this tech best be kept the hell away from production environments.
The toy is still neat, the mashup of services is cool, and this is a good live coding vid. Thumbs up!
It was a coding mistake, the input wasn't being used, instead it was using the previous reply of chatgpt as input again. Recursion!
Hello! Thanks a lot for your videos. You have an NVIDIA DGX Station A100. Please make a video to train (FineTune) the LLaMa model on it.
5:05 If you set the API key as OPENAI_API_KEY in your environment variables, `import openai` automatically looks for it and sets it for you. There's no need to set it manually.
24:33 ... wait until copilot has GPT behind it.... going to be B A N A N A S. lol
I tried something very similar, also stumbling over the Dall-E filter a lot! Ended up using stable diffusion locally, using CUDA/GPU. That worked like a charm, but now I am struggling to package and deploy the app with the stable diffusion model included. It obviously will bloat the app to over 5gig unless we get Alpacca like models for image generation. But it would still be great for PC games where space is not so much an issue. I just cant get it packaged and deployed correctly. Would be great to hear how others handle that!
Stable horde
Just would require the end user to have active internet connection, wait times subject to congestion
@@ascensionunlimited4182 THANK YOU A LOT!!! That will probably solve my problems! 👏
@@volkerkrux7311 gotchu, that's what we're here for 💯
Do you prefer Mac or Windows for development. I see you are using a Linux setup here but if you had to choose between Mac or Windows which one would you pick?
Please complete the neural network series.
37:47 ya man, been there too
have you tried chatgpt as a full on dungeon master ? dnd style
"teapot" would make sense
nice
So to use such a general models like ChatGPT for software applications maybe we need to do fine tuning the model to have a chatbot that serve as a sales man or in the cusomter service in some company. But because ChatGPT is not open source, doesn't that put restraints to develop more software with ChatGPT?
MS already has chatbots with AI functionality. I'm actually using one for making sure employees return their equipment after they are let go.
@@The_Foolish_Fool can those charbot be trained into a company contest, so it's not just options driven but can answer freely in the contest that it trained on like to answer as an employee on customer service
@@ataadevs yes and no, they can be trained on normal customer responses, but you do not want a multi million dollar customer talking to a chatbot as they might feel as if you're stiffing them or you don't care about them as a customer,. Is it dumb? yes, that's just how it works in a sales situation. Not to mention the purchase orders 1000 different companies can generate in a dozen different formats and looks and feels the bot cannot respond to every single thing because the OCR software and image processing is so different and varied that a human has to do that by hand.
@@The_Foolish_Fool İ agree with you, it just felt that we have a chance again to have the charbots at the desire production level with these new chatgpt evolutions
@@The_Foolish_Fool I'm planning to tinker with alpaca models which have close results to chatgpt 3, so it's open source and the community has done amazing things with it, so I'm thinking what if we have some UI for a company to offer some responses to fine tune the model and then open API with some fee plan to use their trained charbots or maybe use it for free in trades for their data.
letTT THE APPS COMEEE!!!
Can you create a video of how to build your own gpt from scratch??
Why did you not use the role "system"?
User role has a stronger weighting than system, so it just seems more useful in practice. Haven't found any solid use for system yet. Have you?
@@sentdex i noticed the weight, but it feels like it will at least "remember it longer", so maybe doing both for best result.
This feels a little AI dungeon like.
The other method works, but it will crash if the key is missing. Where .get() will not, and .get() lets you specify what to do if the key is missing. .get('list_of_options', ][) will return an empty list if list_of_options does not exist.
so your saying is that people who they were trained in are lazy... fit's :D
I have one related question on Embeddings. So every time a user asks the bot a question, the bot searches the knowledgebase and answers. the following use case impossible:
USER: How can I register for starter training ?
BOT: If you meet all the conditions, you can register from the link …
USER: What does it cost ?
BOT: I dont know.
how can i deal with this problem.Since the user specified the course name in the previous question, but not in this question, there will be no match from the knowledgebase. How can I search while keeping the context from the previous question?
Reminds me of AI rouglite
If you want Copilot do perform better and complete entire methods, you'd need to just write a comment describing what you want to achieve, then it will often just write the entire multi-line code for you.
Microsoft Bing has integrated DallE and GPT into the engine
Hello 👋
come on text to image AI, learn to count to 2
haha imagine telling the model if it has an option that has the word item in it, make sure that items are added or removed properly with a qty dictionary.
Watch out for power consumption, it's an obstacle for exponential growth of ai. So we will be okay unless they come with efficient hardware to run on (neuromorphic).
🤙🤙🔥
Can we create a project to make chat gpt understand CZcams videos and ask questions based on it ..
i think the entire "button states" is irrelevant oO
DALL-E is so bad compared to MidJourney and even SD.