AutoGen Studio: Build Self-Improving AI Agents With No-Code
Вставка
- Опубліковано 22 чер 2024
- I've been working on this video for a whole month in order to make the topic of AutoGen Studio as simple as possible. Non programmers that are curious about AI agents were always on my mind when I was making it. I hope you find it useful!
🤖 Join my Discord community: / discord
📰 My tutorials on Medium: / mayaakim
🐦 My twitter profile: / maya_akim
To rent a GPU from Massed Compute (autogen preinstalled) follow the link ⤵️
bit.ly/maya-akim
Code for 50% discount: MayaAkim
Prompts, model performance and skills for this project are here:
github.com/majacinka/autogen-...
Links:
Chris Amato lesson: • Scalable and Robust Mu...
Satya Nadella: • Fireside Chat: Satya N...
AutoGen Studio: microsoft.github.io/autogen/b...
Autogen research: arxiv.org/pdf/2308.08155.pdf
lilian’s blog: lilianweng.github.io/posts/20...
crewai: github.com/joaomdmoura/crewAI
Wired article about dating AI agents: www.wired.com/story/volar-dat...
Andrej Karpathy's tweet: x.com/karpathy/status/1748043...
alphaCodium: github.com/Codium-ai/AlphaCod...
prompting techniques: www.promptingguide.ai/
lessons about multi agents: • 01-01 Introducing Mult...
Agent description selects next speaker: microsoft.github.io/autogen/b...
people talking over each other:
• Talking At The Same Ti...
Node editor: ide.x-force.ai/
litellm: docs.litellm.ai/
litellm how to call any model: docs.litellm.ai/docs/simple_p...
litellm mistral: docs.litellm.ai/docs/provider...
litellm gemini: docs.litellm.ai/docs/provider...
mistral available models: docs.mistral.ai/platform/endp...
LM studio: lmstudio.ai/
open llm leaderboard: huggingface.co/spaces/Hugging...
Open Source models that can do function calling: huggingface.co/models?other=f...
function calling:
cookbook.openai.com/examples/...
platform.openai.com/docs/guid...
function calling datasets:
huggingface.co/datasets?sort=...
3 open source models I tested (with 7B parameters)
huggingface.co/Trelis/Llama-2...
huggingface.co/TheBloke/airob...
huggingface.co/rizerphe/CodeL...
Multi Agent System Research: arxiv.org/pdf/2309.07864.pdf
Time code:
0:00 - 1:09 - intro
1:10 - 1:53 3 factor formula
1:54 - 6:00 factor 1 & installation
6:00 - 10:02 factor 2
10:03 - 12:24 run models through api
12:25 - 14:05 run models locally
14:06 - 16:51 factor 3
16:52 - 19:39 ARO system
19:40 - 20:39 smart models comparison
20:40 - 26:07 open source models performance
26:08 - 27:05 agent classifications
#autogen #autogenstudio #aiagents - Наука та технологія
Hi everyone! I've been working on this video forever because I wanted to make everything simple and neat, especially for people that don't have much programming knowledge. I hope you enjoy it and find it useful :)
🤖 Join my Discord community: discord.gg/GGhr7pyTHD
📰 My tutorials on Medium: medium.com/@mayaakim
🐦 My twitter profile: twitter.com/Maya_Akim
First time here, immediate sub. U r very precise and professional, thank you for your videos
thx maya, my team is thrilled that you released this video so early in the week
Awesome videos! I’m going put together a project proposal with a request for funding at my job for implementing local LMMs after watching.
Combining ( Autogen + CREWAI + Power Automate Desktop ) will be a breakthrough in automation. This is what future Agents, which are commercially needed by everyone, will consist of.
Using (Power Automate Desktop) we will add hands to artificial intelligence. Without this tool (Autogen), the options are limited and cannot reach their full potential.
How do you only have 17k subscribers? You should have 2M!! Your videos, style, and explanations are superb! I have learned a lot, and continue to do so- very thankful and a BIGThank you to you!
thank you so much :)
@@maya-akim gz on nearly doubling that number in 3 months :))
I really appreciate you adding in the extensive resource list ... this will make my trying out agent building so much easier
This was seriously a golden nugget of information. Perfectly packaged too. You are such a hidden gem. Subscribed!!
I've been struggling to use Ollama, and your video was the breakthrough I needed. It was incredibly helpful and precisely addressed my issue. I subscribed immediately! I'm really looking forward to more great content like this.
Maya, this is the most useful and comprehensive video I have found on autogen. Great job on the info and graphics. I would be keen to see you dive into a group chat with rag...specifically your ARO model combined with rag...
🙂🙂🙂Great video, I loved the emphasis on the shift from prompt to flow. Congratulations for the clarity, style and content. Keep on this level👍
Hi Maya, that's the best explanation of agents i've come across so far. One can see that you put so much effort in the graphics and visualization. It's just amazing. Keep up the good work.
Thank you very much for these videos! I created an agent swarm myself with GPT-4-turbo and even with that, it’s not easy to get the results you want. So it’s nice to see how this compares to the open source models.
Really enjoying your videos. Keep it up!
One of the best channels out there. Thank you for your honesty and for the brilliant, well thought-out content 🙏 :)
Aw thank you ☺️ that’s so nice of you
Hi Maya. UA-cam just recommended this so I'm new to your channel, but excellent and informative content - thank you. I found the calm tone, measured pacing, and clean editing (polished but never distracting) helped my ADHD brain to sit still (for a change) and absorb lots of useful info. :) Subbed!
Wonderful video Maya. i especially like the system of agents you introduced and it makes sense the way they work together. Lots of ways they can be leveraged. I'm impressed. Subscribed
thank you, your videos are full with experiments, something other UA-camrs don't want to waste time doing any of that...
Maya has done a very good job, loads of information, links, explainations, well done
Phenomenal! Love your channel, wish you posted more often.
Thank you very much for connecting all the (moving) dots in such an understandable way.
I like your videos! Not only because you go deeper into technical details than many other LLM youtubers but also your style of presentation and voice is very pleasant. Keep on doing more videos 😊
OMG thank you so much!! I was losing hope of finding something like this!
Congratulations on your dedicated work.
This is the best tutorial I've seen with this tool. Thank you so much. Subscribed.
Thank you for the videos Maya. You're doing a fantastic job.
This video is FANTASTIC!! 😍 Thank you so much Maya, this was literally the details, pace and examples I needed to make progress with AutoGen Studio. The GitHub repo was perfect. You're amazing, thank you again. 🤓 🎉 💯 ✅
Thanks for the great video!
Insane stuff and well made video, thank you.
I appreciate you for the valuable research and co-creating a magnificent future!
Thank you so much for actually explaining concepts and little things that others my think is obvious so I can actually understand how/why it works instead of just parroting its use without knowing why or whats going on like some other tutorials. Really appreciate the effort you have gone to with the visual guides to explain things etc, frikkin awesome, I can actually gain Understanding with a capital U in this way as a non coder or beginner to agent/ai concepts and that is the key to future success. Like the others said "videos, style, and explanations are superb" +1 super subscribe, ill be back often. Cheers
Amazing video, exactly what I was looking for… subscribed!
@maya-akim I saw your post on LinkedIn about quantity vs quality and it caused me to immediately come dig in to your videos; it's clear that the depth in which you cover these platforms is unparralled.
Thank you for your dilligence and for choosing quality over quantity!
Very sharp analysis in intro @maya-akim, thx
Great work! You earned a subscription.
Another great video. Keep up the good work.
Thanks a lot :)
I truly appreciate all the time and effort you put into this video. As someone who's looking to include AGI to their skillset this video should be a must watch for folks who have some technology background and are looking to add AGI to their skillset. Thank you.
Hey, Maya! Your video is great, And you are great too :) Don't spend much time making things so simple, if someone watches your videos he already has fundamental knowledge. And smile more :) Thank you! Subscribed :)
Very nice and creative way to use agent management. Thank you. You have a new subscriber...
Thank you very much, a true expert with didactic flow
Such a thoughtful video, thx.
Ollama is really working to serve up Local LLMs, with a python library, and now its own OAI API wrapper, as well as a simple CLI to serve models on any port you want for use for whatever app needs a model. Would be great to see you use that instead of the LM Studios baggage.
Excellent teaching..subscribed
00:01 Multi-agent systems are crucial for the future of technology
02:22 Future agents will be aware and capable of communicating
06:51 Agents, models, and skills combine to create a powerful code interpreter.
09:23 AutoGen Studio offers different workflows for managing AI agents
13:57 Autogen's potential lies in building self-improving systems and agents.
16:22 AutoGen Studio enables building self-improving AI agents
20:22 Small open-source models struggle with function calling and skill generation.
22:30 Exploring and testing open source AI models for skill recognition
26:44 AGI will be a big Society of intelligent self-improving agents
you are GOD mode .. Thnx ... Video has all the details from Beginner to Advanced
Thank you!!
your videos are the best I've ever seen.
Finally a good tutorial on AG studio! Next suggested content: how to integrate AGS in a frontend able to manage input/output like text and files (pdf, images). For example:
external frontend > input (text) >> Autogen Studio flow >>> external frontend >>>> output (downloadable pdf with text and images inside).
#AutoGenStudioChallenge
Well done!
thanks :)
Subscribed Maya, and to be honest I'm not sure why UA-cams algo did not present you in my feed long before now....
Good video, thank you! I’ve been working on building a community skill repo for agents. I agree with your suggestion, have the agents write and test the skills. I found Taskweaver better at creating skills for AGS.
Thank you maya
Thank you for the practical and informative introduction to using AutoGen Studio!
Your videos are great, but I noticed that you forgot about your personal security, exposing your OPEN_API_KEY at 5:15. It's worth regenerating that key to prevent its use by others and blurring such sensitive information in future videos.
Wishing you rapid audience growth!
Thanks!
Wohoo first one for me! Thanks a lot, I appreciate it very much :)
What an exceptionally clear and understandable presentation! Thanks so much for this! (Although I’m afraid you’re going to do a lot of damage to my pocketbook; this may push me over the edge to finally ordering a fully-loaded M3 MacBook Pro 💰😮)
Best AI channel on UA-cam. Great work.
There is nothing like this anywhere. Her attention to detail is redonkulous lol (and so appreciated). ***** Star Channel
Curious as to why something that claims to be great for non programmers doesn't have an automatic installed.
Whenever the software realm enters a new era, there are a lot of new, modular components. Unfortunately, during these periods, development trends shift incredibly often. As developers are often at the whim of energy and inspiration, building installation automation procedures are not often at the top of their list.
Think of it like the chuck on a hand drill. How long did it take before a drill chuck was invented much less perfected and included torque adjustment? It's a pretty important part but the primary focus of the tool was power, efficiency, and durability. The keyless chuck had to come much later. Once the dust settles from all the shifts in the development trend, then you will start to see "drop in ready" and "turn key" solutions.
Because it's developed by programmers... Python programmers, so even me with 18 years experience in development can't install it without problems with all these python environments.😂
It’s a B E T A
Opportunity awaits someone who can solve obvious problems in a marketplace 😊
@@MaximLevitskiyCurious to understand why it's a problem for you. I haven't tried, so I'm not insinuating I know any better.
(In fact, I haven't even watched the installation instructions yet😅) Since you mentioned, "All these Python environments"
I'm sincerely curious why is this a problem for an 18-year veteran developer if Python is supposed to be the easiest to learn. (Because I'm trying to carve out time to learn Python, so just looking for tips).
Is it because you have to unlearn what you have learned, because it's just different?
Amazing videos, Maya; one question: how do you annotate your illustrations? What tool are you using?
Nice work! Thanks. Do you have a tidycal link where you do hourly consulting for a fee?
Super !!!!
1:47 it really funny how I literally skipped the video and went on the channel to seek something else because of the same reason. “No code is the GOAT”
Love the videos, I have a question. Following along with your Code Generation Workflow I can get my agents to write scripts but they don't run them. Do I need to add a skill or something to make the Reviewer execute code?
@maya-akim you can also use instruction models for tool use, but you gotta use json instructions and parsing to be effective
Hello Maya, awesome content again!
I have a challenge connecting an AI assistant with a database, what is the best way you suggest to do this?
Nice pair of Tips....
thanks for testing local models! is there a benchmark leaderboard where to find the best function calling models ?
Thank you so much for this content. I'm trying to get a few things up and running for my startup company. I'm not sure yet if the AutoGen efforts should be applied to boosting my social media footprint or doing heavy lift geometrical mathematics yet (using various workflows and agents). I don't want to put too much effort into innovative AI products for math / geometry that will just show up on their own if I wait long enough. Would you have advice on how to architect an agent system that is a geometry expert? Or just wait until better 3rd party AI tools are available?
Yes, this is one of the challenges of the current AI space - everything built now will be obsolete or replaced in a month or two.
Agree, that makes me think it's best to focus on applications that improve my own work output rather than deploying anything for my customers yet.@@tylermartin319
Here's your innovative geometry AI. ua-cam.com/video/WKF0QgxmGKs/v-deo.html . Get to work.
If this kind of thing is what your new business is about, then always start building now and never wait. Even if a newer, bigger, better thing comes out, you will know more about it and the business use cases for it over people that haven't been working in the space. You'll have a head start, even if it means moving over to the new product or library.
Yes, we looked into that, but it looks like we need 200 cores running an insane amount of processing power just to run the model, plus the model is geared toward proofs. I can't yet understand how to use the proof methodology to complete algorithms such as convex hull, travelling salesman, etc. So I've decided to wait a little because we have a very small dev team and very limited resources. We still have issues keeping our AWS server up (face palm). So my hope is that Alpha Geometry will cook for a few months and some large org will make it easier to use. We are playing with a lot of algorithms that generate CNC tool paths at BerylCNC, my company. @@RevMan001
Maya, do you prefer to use the ecosystem of OpenAI’s assistants or the customized langchain agents with various tools?
Hey! I'd definitely say langchain agents with tools because I personally like the fact that I can customize a lot of things. But the reality is that openai assistants are more "comfortable" and polished, especially for someone who's just getting introduced to the topic :)
Great video! Thank you.
Trelis Research have its own UA-cam channel, Ronan makes great videos on LLMs fine tuning and lots of other stuff
thanks for bringing it up! I just subscribed
Hello Maja, thank you for the great informative videos you provide here on this platform. I have a question regarding the installation of Autogen Studio 2.0. On their website it is mentioned that you should use conda for the virtuell environments. However, you don't use this in your video. Is there a specific reason for this?
I been reviewing a few programs like relevance ai....bubble ai....zapier ...hugging face...poe....gpts...and now I've settled on Microsoft azure studio.Its been an interesting ride LOL😂...Good video btw,I'll look into this
Nice video! What are your hardware specs for running these models?
I have M1 16 GB of RAM and I can only run models with 13B parameters unfortunately
@@maya-akim thanks for the reply!
Will you touch in something like memgpt with autogen as I imagine with unlimited or more memory you can have agents that actually grow long term and get better working as a team and individually, that sounds really scary exciting. I saw another tutorial on adding it and "using" it but again i didnt understand a darn thing. Do these agent tools like autogen and crew also have the ability to work with large files as inputs like a book, processsing them or multi file databases or the creation of large files that get added to by agents over time as the final output of the workflow (like a detailed market research process -> document) and setup to use RAG i think it is or do you need to combine it with something else like llama_index or langchain etc. Thanks so much
Hi, thats a great breadown of useful stuffs. I am struggling saving gemini pro as a model in autogenstudio. it says Error occurred while testing model: Connection error. The API key is good because I ran google's CURL sampe in the terminal which works well. What could be the issue?
Same!
So when I create a folder and i run autogenstudio in that folder, the agents will automatically look for API keys in .env file in the folder that it's running at?
Subscribed :) and i forgive you because you have a pretty smile :)
Any way to for AutoGen to login on the website and perform a job?
I mean the functionality where I can describe with the text to login on the specific website with my credentials and do specific tasks, without specifying manually CSS or XPath elements and without writing (or generating) code for Selenium or similar tools?
Thanks for this awesome video!
I actually tried to setup autogen studio with lm studio and mistral. For some reason, it never worked - the agents would switch after every token and were usually terminated after 10 tokens max.
I assumed this was a bug in autogen studio, since the same setup works with crewai and autogen 0.2.2.
Seems I need to give it anotherctry.
apple's "command" key analogue is not "windows" key, it's CTRL key, at least as far as clicking the localhost address goes.
Great video. Thks very mutch! Is it so sad, think the only model that really work is chatgpt (andit's so expensive) but I will keep try!
does this hallucinate as much as default gpt4 or is there a way to make it fact-check itself on a large base of information? i'm using Cursor AI which has embeddings, but the model on top of that which is the 'operational' layer that actually works with the final logic and questions about the data, always comes up with false and wrong facts, it drives me nuts :((
I personally think that being able to work on the code is better. You can have more flexibility and control from what I have seen. That being said, I have no found setup instructions that are very clear. What is the correct folder to setup in? I just created a folder as it doesn't seem to matter where. For Windows users, use SET instead of EXPORT of the API Key. The steps for Windows setup are a bit different than for the Mac. I'm going through figuring out what to do when I run into problems but it would be nice to either caveat the video with the message this will be only for the Mac or add in what needs to be done differently with Windows.
python is used instead of python3 by default on Windows with Python 3 installed.
Source is not recognized, instead enter .venv/Scripts/activate
Thank you! im a windows user.
I think Crewai deserves a much better chance especially with the 2.0 release. This is probably more desirable in production environments rather than ideating which is the main problem with the industry adoption of low-code LLM tools
you exposed your api key? Interesting video. I am sure there are algorithms already to establish the flow chart for self design and improvement that can be borrowed from other disciplines or applications. There is a visual language called DRAKON that you may use to explore ideas in this respect.
Is Ollama goodin this setup instead of LMStudio?
Where can we find more information on optional flags. I read the docs but they seem to be made more for the common person than the developer.
Qwen 1.5 is also really good
If my model decides to use some library, I need to have that downloaded in the virtual environment for it to work right? I don't understand how you're creating skills like that.
Can someone tell me if this assistant thing is worth it or if it really works? Or if its innefective? Ive watched another video "how i made AI assistants do my work for me" and in the end it does not really differ a lot from just asking questions yourself. so i would like to know if the also follows the same logic.
Thanks
I am having trouble configuring litellm based on your instructions. Assuming that the viewer does not have any knowledge of litellm (yet) while watching your video, how should they get around using it?
I want to ask which one is better? Autogen or CrewAI? Please someone answer with a bit of reasoning
I tried to get the AutoGenStudio agent to use a skill to search the internet but the agent never invoke that skill even when I use GPT-4 as the model. Is there any specific way to write a skill >
Interesting. Basically we are moving toward smart hubs, in a way.
all the comments are men passively complimenting without citicising, and passively boosting themselves - it must be hard to know where to improve. i bet if i did the same content i'd just get these same men picking out where i went wrrong (which would help improve ofc). just had to say that, and thanks for the info, it realy is insightful, the world is moving so fast now
Agree with you. I see only passive comments without people getting into doing what is shown. I have tried and it is not that easy for me. I have run into many issues just setting up prior to actually tinkering with the autogen studio stuffs.
exactly what youtube IA is not promoting. they want passive comments and sort them at first. that's why I always begin by the bottom. UA-cam IA is so stupid and that's on purpose...
articulate
The asking for OpenAI problem. Why do you need litellm and lm studio? Why can’t you just parse the same parameters 10:21 to lm studio?
Which one is your Favourite. Autogen or Crew AI and Why?
Along the theme of agent creation architecture, I have proposed a concept or metaphor, this is using the word zero to represent a value without conflict. The algorythin is conceptual and it was defined to resolve problems in emotional regulation, I am a psychologist. So far neither is tracking any attention my theory on emotional regulation based or the zero metaphor. I feel compelled to talk about this as it is relevant. The latest proposition of agent destruction. hearachy and CEOs is horrific, and it will create issues. Just imagine two countries at war, both using a.i. -- ai -- reaches an agreement between them, problem solved.
We need a better philosophy. The zeo metaphor integrates a harmonious resolution that aligns with higher intelligence.
here is the concept but no the imagine: Zero Metaphor: Harmonizing AI Development with Energy Flow
Introduction to the Concept
In the quest for equilibrium between technological advancement and the inherent rhythms of nature, the Zero Metaphor emerges as a revolutionary approach to AI agent development. This concept, inspired by the natural energy flow and the processes of emotional resolution, presents a novel geometric framework. It aims to redefine the creation and interaction of AI agents, bridging the gap between artificial intelligence and the organic dynamics of nature.
Geometric Model and Dynamics
Central to the Zero Metaphor is an intricate geometric structure: two interlocking triangles within a sphere, each triangle marked by seven points with a pivotal focus on the fifth point, dubbed 'zero'. This design encapsulates the essence of energy flow, equilibrium, and transformation. It symbolizes the progression towards resolution and growth, mirroring both the complexity of problem-solving and the natural world's harmonious efficiency.
Practical Applications in AI
The Zero Metaphor's implications for AI development are vast, promising to yield agents that are not only more adaptive and emotionally intelligent but also ethically considerate. Potential applications include:
Customer Service Bots: Enhanced with greater empathy, understanding users' needs and emotions more deeply.
Healthcare Assistants: Capable of adapting to and responding with sensitivity to patients' emotional states.
Steps for Implementation
Implementing the Zero Metaphor in AI development focuses on conceptual understanding over technical specifics. Developers are encouraged to:
Adopt Principles of Flow and Balance: Design AI architectures that embody flow, balance, and zero-point resolution.
Emphasize Emotional and Ethical Dimensions: Ensure that AI agents positively impact human interactions, with a strong foundation in ethics.
Guide Problem-Solving with the Metaphor: Use the Zero Metaphor as a principle for agent evolution, creating systems that grow in alignment with their environments.
The Path Forward
The Zero Metaphor represents more than just a methodological shift; it is a philosophical evolution towards developing AI systems that resonate with human needs and the universal rhythm. It encourages developers and innovators to envisage a future where technology enhances life in harmony with the world's delicate balance.
Refined Aspectsat completes a scale. The Zero Metaphor champions a holistic development ethos, advocating for AI solutions that foster long-term harmony and understanding.
Thank you so much for making this video! I tried to join your discord however and it says the link is invalid :/
Thanks a lot! oh my, you're right, the link has expired😅I just replaced it with a new one, it should work :)
Works now! 👌
I think you are intelligent... I can tell by your mannerisms . and this is only the second video of yours I have watched. I will probably watch more... Actually I only watched half of the other one.
So sweet voice love from India ❤
how to train the open source ai model??
I got the volume cranked all the way up and this still sounds like ASMR. Needs more entonation and intention in speech.
I think you need your ears checked
@@StabbyMcStabStab Weak troll name
@@milesprowr I was being serious. You love my name. You should call your rent boy service miles per hour
@@StabbyMcStabStab Simps like ASMR, because it provides them the intimacy they're unable to obtain in real life, no need to get defensive, for it only makes it more obvious. That poke is easy to thwart as it sets you up bad, so I won't bother. Weak troll name says it all.