In this video I'll demonstrate the mobile game development process using ChatGPT.
Our goal was to check if ChatGPT 4 is capable of writing a simple game from scratch.
The project's codebase was written entirely by the AI, we just made a few minor adjustments to the final version.
AppStore:
apps.apple.com/ru/app/cobrax/...
Google Play:
play.google.com/store/apps/de...
You can support the channel on Patreon:
/ simulifehub
More support, more opportunities for new projects.
Thank you for the transparency on the process, your conclusion is spot on. There's an overwhelming amount of "Making X with ChatGPT" videos where they either make something extremely simple, cut out most of the prompts, or just get 10% of the code from the AI and do the rest themselves, giving a skewed view of what the AI is actually capable of. It's an amazing tool if you're experienced and need quick troubleshooting, since it's essentially just a very large encyclopedia, but it's incapable of creating by itself without a LOT of babysitting.
yes, it's amazing to do small specific questions, but don't use it to do stuff in your place, it is NOT sentient or intelligent.
While I'm sure you're right that many people make videos like this, I don't think ChatGPT is incapable of making a small to medium sized codebase/project at all. I think a big problem is that many people don't really understand or learn about neural nets and just jump in with ChatGPT and don't fully get how it works. Every time ChatGPT totally forgot something, or started making up functions/variables, it was because he accidentally didn't consider the token limit/context window/memory capabilities of ChatGPT. It's only 4k tokens, which isn't a lot if you're talking about even medium sized codebases. It's an annoying limitation, but you can certainly work around it and not have those specific issues. I think you just have to know how to work with it, and if you do, you can make it create a good working up to mediumish sized codebase/project. I think if you had two or more ChatGPT agents work together step by step to prompt eachother and continue on working, they could autonomously create a program. From what I've seen, it can do each step of the process, designing, coding, cleaning up, coming up with new features, you just need to organize a system where those flow together.
It would help if you used your ability to edit messages. If you have 2 questions to the last ChatGPT answer - you should make a message tree, instead of asking the second question after the answer to the first one. That way you keep the context clean. That said, this approach is still only good for playing around or getting small snippets of code, where you already know what you'll write, but don't want to do it manually. If you want the agent to properly take on writing code for the whole game - you need to take a more complex approach, something like what BabyAGI is doing.
*From my experience with chatgpt 3.5 and coding, it's best not to rely on it to build long lines of code... but if you must, then try using as much literal detail in your prompts as possible, and also if it's going to be extending code it has already written, tell it that it's not necessary to write out unchanged code. It seems to give me less troubling code that way.*
ur experience isn't bad become good experience I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want
From my experience, I can work with it but at early tries, the result is like you had in the video, when I understand it I can adjust the way to work with it, like starting with asking for a list of tasks to meet the goal, then ask one by one from the list, keep splitting the task if it is too large to chew, don't ask it to spit a whole bunches and keep reminding and correcting, finally, keep telling myself hallucination is real.
Interesting video - thanks for sharing your experience. I've not attempted anything as ambitious as getting it to write a whole app, but it has been useful for writing util functions. I've found that when it struggles I can usually get it to work by asking to write a unittest for the function first - i.e. getting it to fo Test Driven Development! Also, starting a new conversation session helps. Remember that these systems are firstly autoregressive language models designed to predict text continuations. When it was being trained on code, the code that had bugs was likely to have other bugs. So if it sees in the past conversation buggy code (that it produced), it is likely to produce bugs because that's what it's seen in the training data. In other words, it will reproduce the pattern that "buggy code tends to include more bugs".
Hmmmm? Don't you write the test first in Test DRIVEN development?
ur problem is decent code now I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want
@@messengercreator You have less braincells than you python script!
This would work a lot better with the Github Copilot version of GPT, but it also makes those mistakes where instead of creating an implementation it makes up a new class name and so on. It doesn't forget code in the project though.
ur complain is done now I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want
I was needing of a quick way to parse and convert some data, I wasn't willing to do it manually but unsure of an efficient way, so I asked and I got a simple python script(kinda incomplete). I remade it slightly and could save up work time that would've been wasted. Not as handholdy as an IDE in the end.
ur problem is answer now I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want
give us the source code for all of your projects please.
This is the trouble with working with an "AI" that doesn't actually have any true intelligence. As far as I know, it's just a pure, single-word language prediction model. Having to predict how a human would write coherent code one word at a time is difficult.
Yes, chatGPT is just finishing sentences and paragraphs but i don't saying "Predict how a human(which it doesn't attempt to) would write [...] is difficult" really explains the issue or how close we are to perfecting the solution. Landing on the moon is extremely difficult yet we successfully did it several times from several countries.
With how complex GPT4 is, I'd expect more than a one word prediction
ur problem is answer now I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want
Great now do autogpt !
Chatgpt has been getting worse as well. It never used to be this bad. I used it all the time but now its practically useless for conherent code of any length or complexity
ur complain is done now I'll introducing AI CHAT DEEPAI the AI CHAT DEEPAI is so powerful and much better than OPENAI since start 2015-2018 and u can show ur picture and video and anything u want