jawerty

This looks like it's exactly what i need for a project im working on. Right now Im actually starting a thread and raising an exception to stop it and restart which is not very safe.

Most people don’t need to write their own OS or compiler. Knowing the details of how things work is apart of getting to gaining expertise

Hey this is a pretty cool idea. I tried exploring this on a live stream a few months ago where I made an endless RPG with an llm i believe I was using llama 2 at the time. Check it out if you have any ideas to improve let me know --> https://www.youtube.com/watch?v=xPbM5Rs1_8Y

Try 2captcha https://2captcha.com/2captcha-api

Super fun! Great work.

Totally agree. I recently did a livestream/tutorial building a choose your adventure with llama 2 and React and had genAI for characters. A lot of fun! Hoping to add loot and fights in my next stream about it https://www.youtube.com/watch?v=xPbM5Rs1_8Y&t=16s

Lifting, animal fats, 10k steps a day, cold showers and no porn

I love the pitch however my brain would hurt after a day of this. Maybe it's simply because no one is metric time woke.

Yea I wanted to automate some basic story outlining with a valuable template and then see how far it can get writing dialogue.

Working on giving it the basic story as input. Also want to use a multi modal expression to inform the story generation soon.

Its pretty fun to play with not really trying to change the world with this one.

true the most useful people I know you'll never hear in NPR they're typically highly skilled workers who just do their job everyday.

Has anyone utilized multi agents in production curious the use cases

Great to hear!

I find LLMs to be pretty incredible for knowledge but any tool for learning is never going to replace a solid apprenticeship. Leadership will always come from people and early students need good leaders

Incredible did one about beef tallow and honestly might be cool to use this to learn concepts depending on how complex can break down topics

Hi thanks for the feedback! It’s opinionated since it specifies those principles in the prompt. However I should probably make the prompt more extensible. I do have a prompt class with the prompt in there for anyone to edit ofc.

Absolutely, I'll add llama.cpp support soon.

Define "original" technically generative AI is always doing something original it's just the novelty and "creativity" that's in to question. Those are likely in the human realm. Even the prompts we cater are themselves what hold the spark of creativity.

This is what excites me the most. It’s such a simple interface (prompting) with unlimited capabilities once you look at it as a logic and pattern engine rather than magic

A lot of it is based on what you’re aiming to generalize. Typically for data that is fairly “conceptual” I find a DPR type solution to be better for most people with a similarity search. But if you want to generalize on style/data extraction/basic pattern matching fine tuning is incredible. I talk through some of these concepts on a fine tuning llama 2 live stream I did a couple months ago https://www.youtube.com/live/TYgtG2Th6fI?si=7VVv0Qooe2o1wuzD

Man I love puppeteer

Killian, you're killin it!

Thanks many more videos to come!

Not sure if this is relevant for a maker Youtube channel but I recently (past 6 weeks) started a live coding channel where I code projects from scratch (various topics from Machine Learning, Web scraping, web dev to general software engineering). I try to walk through my thought process while keeping it funny. So far getting some good feedback

The channel if you're interested: https://www.youtube.com/channel/UCVTtYLn03CmwX-PUfZ3RF0A

Last night, after 8 hours, I was able to get this working OK. It looks at GitHub issues, finds related "APIs" to the GitHub issue and finds which files need to be fixed in a cloned repo, then updates them with WizardCoder.

So far WizardCoder actually has solid reasoning compared to using the Llama 2 React Coder I made last week https://github.com/jawerty/10x-React-Engineer

I highly suggest if you want to use 10x-React-Engineer you move the prompts to WizardCoder and load that LLM. Or I may just do it after I rest from this one.

Overall this is my 3rd live streamed LLM Agent in the past week (check my latest videos) and seems like getting good reception. What should I build next?

Ah I got goofed. Clickbait

How will they enforce this...Prohibition without any threat of enforcement is just a piece of paper.

Haha thanks for the feedback good to see you’re liking it. Many more to come. Lmk if you have any ideas for projects to stream

Last night, I live streamed myself coding this Llama 2 Agent on a Single GPU (Colab). After 6 hours it actually has some good results.

How it works is it takes in your intuition (e.g. "I think x would be cool") and develops a business idea (with a name and branding colors) and a business plan. After the business plan is developed, it criticizes this plan recursively until the "Investor" prompt is satisfied with the plan. After all this it will generate the final MVP idea and pass it to a the React Engineer Agent I live coded 2 days ago (https://github.com/jawerty/10x-react-engineer). Finally once the code is generated you have an option to pivot.

As for the AI's memory, I have a small vector db for history of good intuition+idea examples to use as context (builds as you generate) and a simple backlog of business idea criticisms.

Overall, this experiment showed me that you definitely don't need OpenAI keys to have an effective LLM Agent for you. This is only using the Llama 13b parameter chat model and doing a wonderful job. I spend maybe 5 dollars of Colab compute building this for 6 hours which is way more than it would be to mess around with this.

One caveat is the final react code base often needs some modifications to the package.json/webpack setup. However, It's generating the components/folder surprisingly well starting from the entrypoint file. I addressed most of these concerns in this video but still going to fine tune a model that's way better in my next video or just move to a fine-tuned starcoder for this portion.

Let me know what you think and play around with the Colab if you can!

The live stream: https://youtube.com/live/ZCyxDRSw0wE The 10X-React-Engineer: https://github.com/jawerty/10x-react-engineer

If you want to see me live code more let me know any ideas you'd like to see built. I've been streaming projects like these most days this past month and don't feel like stopping haha.

I have a part of the “dev loop” here that reads each file and “fixes it” based on user instruction. Look at the modification prompt I have. It certainly needs work however I think it’s the start of what could help you write a script to fix your code. I suggest using starcoder instruction fine tuned model instead however