Build Anything with Llama 3 Agents, Here’s How
ฝัง
- เผยแพร่เมื่อ 18 เม.ย. 2024
- If you're serious about AI, and want to learn how to build Agents, join my community: www.skool.com/new-society
Follow me on Twitter - x.com/DavidOndrej1
Please Subscribe.
In this video I will show you how to build AI Agents powered by Llama 3.
If you're serious about AI, and want to learn how to build Agents, join my community: www.skool.com/new-society
I cannot agree more with @milosjovanic803. I / We like your videos?
But you might want to think about your offer value @ 77/mth? (Ridiculous!)
I would like to join but i cannot pay using a Credit card. In Holland we use Paypal and Ideal mostly
So the first 60% of this video built up the expectation that we are gonna use offline llama3 py agents, but then at the very end you switch it to using the llama3 available through groq's api. Although you do get the agent working with llama3, its a bit misleading, and it would have been better to straight up say: I havent got Llama3 working offline*, but here is how I got it working through groq's API.
*Edit for clarity: llama3 working offline with crewai in the context of this tutorial.
I agree. Basically llama3 is not %100 open source as far as I know
@@ardagunay4699I think it has more to do with crewai not being correctly configured with the new llama model. It’s possible to use llama2 offline to do the same project, but if you repeat the same steps for llama3 there is a clear breakdown in the crewai step of the process
Thanks saved 12min of my life
you can have llama3 running local,, i used ollama and just got llama3 model, no problem
@@EccleezyAvicii It literally just came out. He probably hasn't done it yet.
I also noticed that as of 2024-04-27 the Llama3 (local LLM) does not work with CrewAI. However, you can replace Lllama3 with Erik Hartford's excellent model "dolphin-llama3" and you get the expected result. Dolphin-llama has the additional advantage of being uncensored.
Cheers! Keep up the good work!
Nice tip! thanks. I will try that.
Oh thank you !!
Love your enthusiasm.
more then anything i appreciate your showing when and where these processes dont work . the trouble-shooting is a critical part of process and the overhype of these systems is most deceitful when the user actually tries to integrate them and runs into all sorts of issues that were hidden by showmen . really excited for llama3 finetunes and more powerful agentic systems , thinking recursive self-debug and finetuning for the generation of the most understandable debugable code , with proofs and tests , could build a solid foundation .
Great Work!! Thanks for sharing with us.
Worked like a charm - amazing and Groq take a bow! Great videos as ever David..you the man!
Thanks 👍
Found you today with the Zuck news, and now I like watching you code. 👊
Nice work David! I appreciate your effort to get this code out to us so quickly.
the last one with groq is using open ais gpt4 and not llama correct? or do you still needs tokens from openai to get llama working? please explain
Nice video and tutorial! Thanks a lot! This gave me the head start I was looking for. Subbed and will def keep watching.
that's some top-tier level shit, keep it up
BRO - YOU ARE NEXT LEVEL - YOUR FUTURE IS BIG TIME
Bro I just want to say thank you for making this content. It's always super informative in an easy-to-understand format. Out of the 50 (exaggeration, but there's alot) I always find myself looking for your videos first. I've always had an interest in programming and hacking but didn't do much with machine learning. But know I'm a man obsessed. Mainly because of how critical it is for normal civilians to learn how create and train these things. I truly belive the future of humanity depends on it. If corporations kill us all building agi and it escapes especially Gemini we are so screwed because the odds are against us that it finds any value I'm something(humans)that's killing the thing it exists on or believing we will attempt to shut her down. Or government with runaway military ai because they couldn't wait until all the bugs were out when they deploy it.
If you are not in the new-society - check it out. Based on what you wrote - I think you would enjoy it.
Thanks. Very helpful. Waiting for 400b model
I'm praying that the 405B model is better than both Claude 3 Opus and GPT-4 Turbo
Because if it is, the world will no longer be the same.
Why use a Mac if you need a beefy pc? Just curious .. I know Apple has their ai chips besides cpu, gpu.. but a 3090 will smoke it out of the water.
@@Instant_Nerfnot really. Arm is just different
@@Instant_Nerf @Instant_Nerf With any big model, you're not going to be able to make much use of any consumer GPU like the 3090. He can run the 8B parameter model with it, but the most sensible route is cloud computing for stuff that's big, which he is doing with Groq. If you're going to run LLM's an absurd amount of time, sure. Get a rack of GPU's or get a high end server processor with large amounts of fast server memory. But for most people, this is not a good use of money.
Good luck finding a computer that isn't over 15k to do that.
Thanks for the video!
One question: any clue why do you set OPENAI_API_KEY with groq api key? I found it a bit confusing. Especially when using openai's API_KEY for authentication. Is OPENAI_API_KEY a placeholder in crewai for groq api key? I know a bit nonsense, so what do I missing? Thanks!
Is it possible to use agents to build an AI powered question answering system for pdf documented for academic research purposes? I'd like to build my own so I can avoid the costs of those currently available commercially.
I have made state of the art automation scripts for me work and i also added some stealth web scraping methods, how can i train lama model to use my coding methods??
Hahaha thanks so much brother. I want to stay on the edge with help of a friend such as you.
Did you get llama 3 running with CrewAI without using Groq though? Or did I miss that in the video?
Yea, if you still have to pay for some damn service... I'm having issues getting autogen working with llama3:8b-instruct-fp16 and the teachability module(runs at 42+ t/s though!) It almost never decides to flag things as important/worthy of remembering!... But just started messing with that today. If you have a solution to using agents with only localLLM, no api keys, please let us know!
TL;DR - it fails to understand it's asked to basically form a question about what it's supposed to store in the db, so that it could be found that way, and the analyzer just keeps asking this same question every time. Probably need a better analyzer?..hm
--------------------------------------------------------------------------------
teachable_agent (to analyzer):
Imagine that the user forgot this information in the TEXT. How would they ask you for this information? Include no other text in your response.
--------------------------------------------------------------------------------
analyzer (to teachable_agent):
What is the context or background information mentioned in the provided text that I should be aware of? Can you remind me what important details are missing from the passage and need to be recalled?
--------------------------------------------------------------------------------
For me, it works well with LM-Studio.
@@PrinzMegahertz I'm interested in your setup. I'm using it with LMStudio and getting the same result ... the Executor kicks off, and the GPU ramps up, but nothing happens.
@@richardchinnis I tried with crewai and llama3:8b locally on my computer, 15 minutes later, it still is stuck at > Entering new CrewAgentExecutor chain...
@@PrinzMegahertz i have the same output, like in the video agents going nuts repeating, any idea?
is groq api anytime adopt command r plus?
Thanks david for thw consistency and simplicity.. we still learning about agents, thanks and touch on foundation agents by nvidia dr Jim Fan
Thanks for sharing
So this started as a good tutorial, ran into some issues and kinda just ended. I did manage to get the groq to work in the end. I do have Llama running on a Docker container. Now i would like to combine both of those . Thanks for the tutorial
If I bought your course, do you teach how to make sophisticated ai chatbot?
It's possible to create an agent that talk like a novel character? It have to talk in Italian, no English. I wanna know if your curse you explain that
Nice video do you have a video where you create and use tools ?
I'm a 3d artist and not a programmer by any stretch of the imagination. Is there a change we could have CrewAI with a nice user interface and an installer?
Happy you created course , hope you continue .
What the difference from CREWAI?
@DavidOdrej Can you get any LLM model to understand what a magical square is, how to create it and create an working example.... ? i bet not
would really be cool how you could get agents to understand it and create a working sample.. its weird as its knows what it is, but can calculate simple.. please give it a try or anyone else ai interested
Awesome tutorial David. I think GPT4 as a benchmark is quite old now. I wonder what agents with GPT5 will look like.
Please can you list me the capabilities the PC needs to run this well
Nice video 🎉
so did this end up using llama or openai at the end...?
good simple examples showing Groq capabilities
awesome thanks, I wasn't sure if I understood correctly, does downloading the local model work with crewai or only through the API?
I suggest you trying with LMStudio, cause with Ollama and crewai, it seems to be problematic.
@@mayorc sounds great thank you!
Pc recommendations for 70b model?
quantize it at 4bit to 1bit so that you are running 13b param model with not much accuracy life then buy a 4090 or maybe even a a100 in your pc build from nvidia, honestly don't better to not buy that much hardware so many 1000s of $
Does anyone know if you can run this on in iPad locally and upload documents in order to answer queries? For example if you made an app for allergies on a food menu would you be able to upload ingredients of the food menu into the LLM and have it RAG answers similar to "I have a gluten allergy, can i have the Ceasar salad?"
The easiest way would be to deploy all this stuff on a server or a home PC, expose an end point, then write an ipad app to upload docs and chat with doc via your app.
Now we need to find a way to pack everything so we can sell those personal assistants and install those assistants in any website.
Ur a legend!
What are system requirements for running this model(8b) locally?
I have ran it with a nvidia 3070 with like just a few tokens a second at float 16 i believe not sure but using recent branch text generation webui I was able to do it. I was also able to run the 8b model in colab for free using some code I got/came up with on huggingface face model card discussions you can see it. I would say 3090 would run it faster though
How did you download the 40gb model? Do you have that much disk space?
Yes... I think almost all modern computers have more than 40GB of free storage.
This isn't your RAM, this is how much space you need on your hard drive.
@@DavidOndrej But if i'm using Groq, I don't need to download it right? I can just run it using the API?
Can i make it live for others to use it?
Looks promissing!
I see David berman replicated this today David
I would like to know your thoughts on this: With a limitation that restricts the potential of Llama materials to enhance other major language models.
Researchers and developers often want to compare or fine-tune different models to improve their performance or tailor them to specific task?
However, due to the restrictions in the licensing terms, they cannot freely utilize the Llama materials to do so unless they specifically use Llama 3.
I want something to experiment with agents and get the handle of it and experiment on how much better it can help me at work, without spending any money or having particular premium keys. I've watched a lot of videos but i still don't understand what agent builders allow free to use agents, even if it's on a daily token limit.
Which are the bare Minimum specs for my hardware to be able to run this?
If you are using groq, none of the processing is done locally. So basically any hardware would do.
How to self host llama 3 on cloud?
why u need openai key if u using free llama?
He stored the Groq API key as an OpenAI API key variable with os.environ["OPENAI_API_KEY"], so when llm = ChatOpenAI(model = "some model") is called, it will automatically switch out "some model" against the variable defined in os.environ["OPENAI_MODEL_NAME"], which he set it to be "llama3-70-b-8192". Finally, he had to specify the url from which the model is accessed, so he set os.environ["OPENAI_API_BASE"] to some Groq related url.
@@wenhanzhou5826 thx for clarification mate
wouldve been nice to see you successfully set up and use llama3 the first time without the use of an openapi key etc
When I use llama 3 8B on ollama or LM Studio, it is much dumber than on OpenRouter. Even after resetting all parameters to factory and loading the llama 3 preset. Even with the full non-quantized 8-bit version on LM studio.
What kind of computer can run llama3:70b locally?
My computer
@@DavidOndrej Share specs please
What did you achieve ?
A video pointing to his chorus
GREAT VID! FINALLY THE INSTRUCT!!
...may I just ask a sidebar question Re. Your VS Code editor window behavior? PLEASE?!!
How have you set your VS Code preferences so that the longer length strings you've written for the classifier and responder classes (specifically, the strings stored in 'goal' and 'backstory') when they reach the edge of the editor window they wrap to the next line down, WITH THE NEXT WORD CONTINUING FROM THE CORRECT INDENTATION POSITION (Directly beneath the declaration like:
To demonstrate/explain here:
|| = indicated the edge of the editors window
Your editor looks like this:
responder = Agent(
(\t) goal = "qwer||
tyabcdefghijklmnop",
)
My editor looks like this:
responder = Agent(
(\t) goal = "qwer||
tyabcdefghijklmnop", #
Love your work, but 77 dollars is a bit delusional. Not sure what kind of value you would have to pay for that. With such a price you'll only get idiots. I pay 4.99 for Control Alt AI subscription and I get incredible complex Comy UI Stable Diffusion workflows etc. You are asking almost twice the money that I pay for my GPT Plus and Claude 3 combined. That's a lot. You have good ideas etc but the content you are sharing is not that complex or innovative to be worth that amount of money, per month? Haha. Anyway. Wish you all the best though.
Agreed
I was excited about signing up until I saw the price, $77 a month!
Assuming the count on the site is accurate, at around 450, multiply that by 77 and you've got some serious coin being made, which is fine, but with that sort of money one would expect the video's to be far more polished
ctrl+D on all OS"s exits anything in the terminal
10:00 thanks for being real
What Did you even show I missed it😂 . You just kept saying it was broken
strong ever
Is LLAMA 3 uncensored?
People are working on it ;)
dolphin 2.9 (llama 3 8b) already appeared ;-)
Anyone who is not ultra-rich will never pay 77 USD just to be in your community.
It is simply insane. I suggest you take a different approach. Because it wont work.
I appreciate the value you offer with your community, but I want to be honest about my perspective. The current membership fee of 77 USD is simply too high for many, including myself.
I understand that there are costs associated with maintaining the community and providing value to the members, but I wonder if there is room for a more accessible membership fee.
A fee that is feasible for more people and enables them to participate and expand their knowledge.
Yes. Reduce the price to 5 dollars
We always have a choice - we can either stick our nose into other people's business and give unsolicited criticism, or we can start with ourselves, like earning more and not making ourselves look like a victim.
Almost burned down my GPU :D but thanks for tutorial
ofc solved with Groq :D enjoy everyone this great tutorial
Dude you got a ShoutOut By the Man himself Alex Hormozi. Go watch his recent Thumbnail
Can You make some better example of this agents. Something that is really helpfull, you always say thay for questions of time you do something basic, But it will be really fascinating if you spend more time doing something that have some realistic value. Thanks
Ahhhhgh I was focused on making happen until saw 40gb, blast
More videos pls
i'm uploading daily brother
lost me at the beginning. did everything you said only to immeadiately get error when i paste ollama run llama3 into terminal. you flew that part too fast. you missed something.
You are not doing this in enough detail you go back and forth. Which makes it hard to follow for allot of people
im so pissed im tried joining at 38$ but it wouldn't work now its 77$ and i still cant join, please make a patron
Soon it will be $97. The price grows together with the value. When it was $37 we only had 1 or 2 courses and not much else.
Now we have 10+ trainings, two weekly calls, 400+ active members, and so much more.
thanks cuz,, gg
dont forget to delete that groq api key xx
Hey brother, thank you very much for your channel. I’m a single father, I’ve been following this AI stuff closely. I’m also in school, and have so many coals in the fire it’s not even funny lol.
Thank you for your posts, because when I get a decent computer I’ll be able to quickly jump on board. I grew up very poor, my son will have a better life. I need to be on top of this. My next pay check I’ll be joining your community. Any advice on how to get my hands on a decent computer? To run this stuff? What should it have? I don’t want to miss the opportunity to provide for my son
I was going to join and then I saw it was $77!?!? Thats mad.
Soon it will be $97 ;)
Groq like websites and other alternatives
so you didnt get it working locally gg
TLDR: Don't bother with this video if you need to run locally
He gets 9:38 in, can't get it working with a local Ollama model so just gives up and switches to a remote model.
Really annoying if you're coding along with the video then realise it's useless for your purposes.
I hope his premium content is better than this otherwise a bunch of people are getting taken for a ride.
i got it working with ollama run llama3
llama3 umí hrozně špatně česky. To je hrozně velká nevýhoda
But I'm not nerd
What can I do. My friend?
Become nerd or give up
Thumbnails?🤣
I just received an interesting response from llama 3, claiming something quite unique
I bought the course and it's underwhelming. full of fluff. I asked for a refund and he is ignoring my messages. He's a new age scammer
sponsor video?
not sponsored
your accent is so sexy
Stop faking Llama3 as better than GPT4. Llama3 is at position 5 below GPT4 (proprietary) at number 1, Mistral (proprietary), Gemini Pro (proprietary) and Claude 3 (proprietary.
Llama 3 is governed by the Licence Llama 3 community, which means it is not Apache 2.0 opensource for anyone use for anything. You are still governed by Facebook! How useless.
We literally have a GPT-4 level model that we can run locally, on our computers. That is the exact opposite of "useless".
Actually, you have a point. I feel the Open source argument is overhyped sometimes.
@@DavidOndrej Any chance you can run this locally on an ipad and upload documents for it to read to answer queries?
you can use it for free unless you have nmore than 700 Million monthly users.
There are so many benefits to this I can't even describe. For one, the dystopian future of AI God gets further away the stronger these open source LLMs are. On top of that, people wanting to start businesses or even just wanting some privacy will not need to make huge sacrifices in order to access these LLMs. There is so much redemption in this I'm almost willing to forgive Meta for their past.
Waste of time
Seriously Bad Content!
That’s way too complicated.
We did not even get the punch line of that Long Joke :( just kidding check your email David