Super cool video on mixing Clause 3 Opus API, ChatGPT 3.5 Turbo, and Python to create a self-correcting AI coding buddy. The walkthrough is insightful and shows some neat tricks.
I wanted to ask you if it's possible to create an entire system with the description of a project and the creation of various files for that project just by using a context window of one million tokens. And after creating this complete system with various functionalities, would it be possible to do something like what you did in this video retroactively, improving the system by fixing bugs and adding new features until it becomes exactly as it was written in the initial project? This would be revolutionary and wonderful, and I'd like to know if you could be the first on RU-vid to do it
Dude, you have access to different llms. You ask them how to do it. Tell it to break it down into steps. Then break down those steps down because it will likely give you steps that to high level. Anything you don't understand, you ask it to explain it to you at 7th grade level. It's more important for you to learn how to use the tools. Tell it what you want to do and ask it if it's possible. Then tell it to give you an example of what it explained. Also you going to have to go into more detail than just a description of a project. You need to go down to user story level for the system to plan. Then even deeper to write the code.
Insane! Can't believe I'm alive to see AI that can self correct code. Kris is the guy to learn from. I've been a member for a while now and definitely feel that it's worth it!
5 месяцев назад
You rock, Kris! Thanks for your content and experiments.
Thank you for putting the time into figuring this out. It gives me a good idea on how to approach something like this! Keep on giving the great content!
Great video, I know the snake game has been one of the tests we all try at some stage, but I have never seen it work first time and then iteratively improve. Awesome!
This was going to be my next project but using local LLM's. Anthropic Opus is really good though, so I may have to use it for code writing and self correction. Nice work!
Wow again Kris! You mentioned in passing about the video or possibly screen shots overread. I think what you did with Elden Ring combined with this style of agents could really open some amazing use cases. Guided education for new software or games of course. But, possibly a broader automation/RPA method for almost any current software available. Examples: Quickbooks to do your taxes, video editing automation, spreadsheet visualizations. Thx!
Thanks Kris! You mentioned vision model to deal with screenshots. And the results were not so good. I recommend to try some OCR API. I tested one of them and it worked great even with photos of monitors.
Kris, thank you for sharing the video. I think the verge website is too easy to crawl (verge is also used in the open Interpreter video). Can you try a more difficult website crawl next time? Thank you
I am curious to know which improvements this code would suggest whilst looking at itself (to have it analyze the code from which it is made) Perhaps it would build a friendly GUI and such .. who knows really. Very useful code solving real issues elegantly! - The compliment! 🖖
You can make this more efficient by getting the llm to output the code in json. Then deserialise it into a variable. That way we dont need to use a llm to clean the code. { "Code":"" }
cool :) yeah i think a lot of ppl or working on latecy, i guess if you gonna have a consumer facing product it needs to be low latency, ppl dont have any patience and are used to instant responses
I suppose you are right. As an individual, it is harder to reach a wide audience at the same time but you still pull it off! I think you would be happy to know that I have solved the bugs myself and now adding some spice to the sts assistant :) One thing I wanted to ask was - how do i get my assistant to stop listening as soon as I stop talking ? RIght now I press 'ctrl+c' to interrupt and stop recording.
This is really cool , although , it is pretty much AUTOGEn but basic, isnt it? Do you see any advantage using this scope for self correcting AI instead of using something like autogen or crew ai?
