*UPDATE:* Thanks to the viewer @tryingET great suggestion, I managed to improve the prompts and make the output consistent. You can check out the improved version on my github. Thanks for watching and I'm curious, what were your experiences with crewAI like?
Thanks for this deep dive! The problem with open source templates is that they don't handle function calls, which is necessary for the crew to function. It seems that OpenHermes handles it well, the scripts work as expected, but even gpt3.5 gives better results.. thanks again for sharing
@@tryingET This was just a great suggestion! I just ran the script couple of times, and you're right, the results are much more consistent. I only changed the part with "(linkToURL)", for some reason it was throwing the agent off. But it works with simple "(link to project)". I'll update the repo, thanks a lot for this help 🙏🏻
crewAI creator here, so cool to see videos like this! I also automated parts of my work with crewAI as well and it's a "a-ha" moment for sure! Great content! keep it up 💪
is it possible to run them in parallel instead of serial (maybe via threads)? The idea is that there's a manager who of course manages the worker agents (researchers, writers, ...); the worker agents then hand-off their work to analyst agent(s) who determine if more work is to be done, the result is then handed off to the manager, who then hands off to-do work to task creator agents to define what else needs to be done based on metrics from the analyst(s), then brought back to the manager who then assigns tasks to agents based on available workload (agent workload queue would be cool). also, dynamically "spawning" (instantiating) agents based on needs would be cool also, to conserve resources. maybe some features in crewai are missing yet to do that - what do you think?
When humans take their organic brain inside their skulls for granted... Even if we could reach Singularity, we will always be n-1 away from the 'n'th civilization that Created our n-1 universe. This is the fundamental limitation of pixel based evolution. Hence, we would go a longer way being organic hackers than materialists.
I hardly comment on TH-cam channels. But this is another level. The way you explain things, organization, referencing, and pace spot on. Great content. Please keep up the great work. Thanks :)
First time i encounter your channel/videos, and immediately got a subscribe from me. I love the clear explanation, straight to the point, no over promising or "Make $3000 dollar a day with these 10 simple bla bla". Thank you for keeping things real and useful. Breath of fresh air!
I'm relieved to find someone else who faced challenges while running local models. Your sincere and practical review is appreciated. Unlike many others who simply join the hype train without discussing their struggles, your honesty is refreshing. Thank you.
me too, local models beside being pain in the butt, freezing your whole damn dev environment, then you get a shitty output, i am surprised she did test many models, i would have givenup much faster.
Not enough people talk about this. I follow the steps and run into error after error. Versions don't match, missing dependencies, list goes on. Ai development really takes a toll on your life force lmao
I only tested one model weeks ago and i got no issue installing it and running it. It’s uncensored and i was curious. Outputs were great. If you have at least 16gb ram on 7B models there’s no way it will crash your PC. Of course it’s slow as fuck, like 1 word per second generation
🎯 Key Takeaways for quick navigation: 00:41 🧠 *The video discusses the differences between System 1 (fast, subconscious thinking) and System 2 (slow, deliberate thinking) in the context of AI capabilities, highlighting that current language models primarily operate on System 1 thinking.* 01:48 💡 *Introduces two methods to simulate System 2 thinking in AI: "Tree of Thought" prompting and the use of platforms like CrewAI, which enable the construction of custom AI agents capable of collaborative problem-solving.* 03:26 🚀 *Outlines the process of setting up AI agents using CrewAI, emphasizing the importance of defining specific roles, goals, and backstories for each agent to ensure effective task execution.* 07:22 📈 *Describes how AI agents can be made more intelligent by granting them access to real-world data, and how to avoid fees and maintain privacy by running models locally.* 14:31 💸 *Discusses the cost implications of using models like GPT-4 for AI-driven tasks and explores local models as a more cost-effective and private alternative, despite their varying performance and capabilities.* Made with HARPA AI
This video is fantastic! The content is thoroughly explained and incredibly helpful for professionals in the IT space. At Xerxes, we truly value such clarity and effort. Keep up the great work-looking forward to more insightful content like this!
00:28 The inner dialogue shows the two systems of thinking: slow, deliberate system 2 vs. fast, automatic system 1. 🤔 01:36 Some people found two ways to simulate system 2 rational thinking in AI: tree of thought prompting and platforms like CrewAI. 💡 03:13 I’ll set up 3 AI agents on CrewAI to analyze my startup idea: a marketer, a technologist, and a business consultant. 👥 05:42 I defined specific tasks for each agent: analyze market demand, make product suggestions, and create biz plan. 📝 08:12 Making agents smarter is easy - add tools to give them access to real-time data like Google and Wikipedia. 🧠 11:06 I wrote a custom Reddit scraper tool to get higher quality info for my agents. 🤖 15:25 I tested 13 local AI models - most failed, but Llama-13B worked best to understand the task. 📈 Supported by NoteGPT
🎯 Key Takeaways for quick navigation: 00:01 🧠 *Andre from OpenAI clarifies current LLMs are only capable of fast, instinctual thinking, not slow, rational analysis.* 02:15 🤝 *Crew AI allows anyone to build custom "agents" that can debate issues, thereby simulating rational thinking. * 03:26 💡 *Demos building 3 agents - marketer, technologist, consultant - to analyze a business idea.* 05:16 📋 *Defines specific tasks for each agent to collaboratively produce a business plan.* 08:37 🛠 *Adds "tools" to give agents access to real-time data, making them smarter.* 11:34 📰 *Builds a custom Reddit scraper tool to generate better newsletter content.* 14:58 💰 *Running models locally avoids OpenAI fees and keeps conversations private.* 18:49 🤔 *13B LLM was the only local model that vaguely understood the task.* Made with HARPA AI
It's shocking to me how quickly someone made an AI to do this, as I created my own autonomous agents in python a few months back to do these similar things. One tip I have for people trying to come up with a large & detailed tasks/descriptions is to write it in a .TXT file, and then reference it in your code. That way it keeps the code clean and also easy to modify the descriptions and tasks in the future without changing anything in the code.
Excellent idea! You could even create services to update it from a GUI without touching your code. I'd probably set it up like a traditional JSON config file.
This mirrors my experience with local models vs automation. I've come to the conclusion I either need to massively upgrade my hardware or just wait it out for a new breakthrough model. I'm a bit jaded with all the hype that never seems to live up to real-world use.
Welllllll, yes and no. I've been able to tripple my productivity and I got 100% on two seperate essays using AI to workshop some ideas and build the essay outlines.
Super work there Maya. You earned a subscriber, and a follow. I built an agent on node with run tools with custom functions over RAG. But this is next level only, will try this next. Thanks again. Keep shining
Really nice work friend. Nice narrative style and prosody while getting such a structured goal. And definitely awesome to listen discussions on the topics and decision making, this marks the difference. ... For trivial coding there is AI and Internet... for the core reasons and concepts there is us the humans
Maya, have to say - that was in-depth. Love the detail. Expensive running GPTs with CHAT 4. The output is definitely worth it though. I guess getting your own custom newsletter every day for less than a dollar does save research time. The next step is to get that file data into the actual newsletter now. Cheers for the free resources on Langchain. Just getting into APIs with py and deployment apps. I predict that you will have a great future on TH-cam. Keep up the good work.
🎯 Key Takeaways for quick navigation: 00:01 🧠 *System 1 vs. System 2 thinking and AI limitations.* - Explanation of system 1 (fast) and system 2 (slow) thinking. - Current large language models (LLMs) are limited to system 1 thinking. - Desire for LLMs to achieve system 2 thinking for complex problem-solving. 02:02 🔄 *Two methods to simulate rational thinking in AI.* - Tree of thought prompting involves considering issues from multiple perspectives. - CrewAI and agent systems allow collaborative problem-solving with custom agents. - Demonstration of assembling an AI agent team for startup analysis. 07:09 🌐 *Making AI agents smarter with real-world data access.* - Integration of built-in tools, like text-to-speech and Google search, to enhance agents. - Example: Using Google search tool for creating a detailed AI and ML innovation report. - Introduction of local llama subreddit data for improved information quality. 10:25 🌐 *Custom tools for better data sources in AI tasks.* - Introduction of custom Reddit scraper tool for more relevant information. - Comparison of Gemini Pro's output and challenges with generic text generation. - Acknowledgment of occasional variations and flakiness in agent outputs. 13:50 💰 *Cost considerations and running local models.* - Discussion on pricing for API calls and the need to avoid high expenses. - Testing open source models with varying parameters for running locally. - Recommendation for at least 8 GB RAM for local models. 15:25 📊 *Performance of local models and surprising discoveries.* - Experimentation with different local models and their understanding of tasks. - Identification of models that performed poorly and surprising results with a basic model. - Sharing notes on local models to avoid and those showing promise. 19:18 🤖 *Conclusion and audience engagement.* - Recap of local model performance and potential improvements. - Invitation for viewers to share their experiences with CrewAI. - Appreciation for watching and anticipation for the next video. Made with HARPA AI
My understanding is that using a larger quantized model works better. I’m planning on trying it soon on my computer, maybe with autogen. I’ve got a 4090, i9-10900k, and 64Gb RAM, so I’m hoping I can run maybe a 30b quantized model on it. I read that the ~5-bit quantized models are the sweet spot that reduces your memory footprint without any significant loss in quality of responses. 4-bit is still good but takes enough of a hit to matter. Again, haven’t tried it myself, so maybe I’m mistaken, but that’s what I read.
🎯 Key Takeaways for quick navigation: 00:41 🤔 *Large language models (LLMs) currently exhibit system one thinking, which is subconscious and automatic, lacking the ability for deliberate, rational system two thinking.* 01:48 🌐 *Two methods to simulate rational thinking in AI: Tree of Thought Prompting, where experts contribute perspectives, and platforms like CrewAI, allowing collaboration between custom-built agents.* 03:13 ⚙️ *Building AI agents with CrewAI: Create agents (e.g., market researcher, technologist, business development expert), define tasks, and establish a sequential process for collaboration.* 08:37 🧰 *Enhancing agent intelligence: Add built-in tools for real-world data access, such as text-to-speech or Google search tools, to improve AI agents' capabilities.* 09:31 📰 *Creating an AI-driven newsletter: Utilize tools like Google search to gather information, but consider custom tools (e.g., Reddit scraper) for better control over data sources.* 13:23 💡 *Challenges with AI instructions: AI agents, even with GPT-4, may not consistently follow instructions, and results can vary. Custom tools may offer more control but require careful implementation.* 15:10 💸 *Cost considerations: Running AI scripts with external APIs can accumulate costs. Local models can be an alternative, but performance varies, and some models struggled with tasks.* 16:05 🚀 *Local models in CrewAI: Use local models to avoid API costs. Consider RAM requirements (e.g., 8GB for 7 billion parameters) and experiment with different models for optimal results.* 17:54 🧪 *Experimenting with local models: Testing various local models (e.g., llama 2 Series, 52, open chat) showed varying performance, with some models struggling to understand tasks.* 19:04 📊 *Results and conclusions: Despite challenges, a regular llama 13 billion parameters model, not fine-tuned, performed surprisingly well in incorporating data from a subreddit, highlighting experimentation and challenges in AI use.*
It's quite possible that GPT4 is much more adept at understanding the premise of function calling, as it likely has a fine tuned expert in it's MOE to deal with "GPTs", thus making it more capable when dealing with OOTB solutions like CrewAI et al. I'd hazard that until someone fine tunes an OS model with a variety of function calling methods, and tools like CrewAI move on to more dynamic conversation flows rather than just sequential, then we'll begin to see the benefits of offline muilti-agent setups.
Side note, when showing something like the “ai agent landscape” would be neat if there was a reference where to find it. (There was enough info to do so, but a side of the repo would be sweet)
I want to thank you for this video! It is one of the most informative videos I have seen. Now that I watched it through I realized you really did your homework. I appreciate it.👏
TH-cam has automatic transcription. Can you have an agent go through all your subscriptions and summarize all the latest videos a few times a day? I would probably email that to myself or have an RSS feed, or some sort of news page to view the AI summaries.
@@maya-akim I also need an AI summary based on YT watch history for the day and the week. It's a weekly report for work. We need to spend 2 hours doing research on AI topics per week.
@@maya-akim I was able to find a YT transcription library and adapt it for my C# project pretty easily. There's a similar package for Python. th-cam.com/video/lY-6ZdsuHS8/w-d-xo.html
Thanks for the refreshingly honest results rather than the usual fake hype. It looks to me that LLMs have a long way to improve before autonomous agents can become actually useful.
I'm not sure how the video settings are different but I can't generate a video thumbnail from WordPress to this video. I keep track of all the top AI videos with my own WordPress instance. Embedding must be disabled.
7b-13b models are too small for a task like this most likely ( especially compared to a massive models like gpt4 ). Wonder how mixtral 8x7b would perform, even if we have to run it in cloud or through an llm api provider ( eg perplexity ).
I also didn't see anywhere any mention of the nature of the models which ollama uses. Most of the models it allows for download are the 4-bit quantized versions. Locally, I have been downloading the the slightly bigger 5 and 6-bit quantized versions and get better results. On the ollama git repo there are some easy instructions on how to import GGUF models one can download. TheBloke on huggingface has a lot of models that are converted and quantized.
00:02 AI assistants are currently limited to system one thinking. 02:15 Build custom AI agents to solve complex tasks. 04:40 Creating a team of AI agents for specific business tasks. 07:11 Making AI assistants work in a sequential process 09:36 Creating custom tools for improved newsletter quality 11:53 Automated AI assistants improve efficiency and accuracy. 14:06 Using local models as AI assistants can be challenging due to high RAM requirements and potential crashes. 16:23 Testing of different AI models for newsletter generation. 18:28 Using a regular llama 13 billion parameters model for Reddit conversations on self-driving cars
Hi Maya! I’m so fascinated by your technical abilities. I barely understand the whole thing, but I’ve always been fascinated by AI and started learning AI tools.I saw on your tiktok that you taught yourself Python. It would be awesome if you can also share your learning process coming from a non-tech background and how’s your progress so far. Thank you :)
Many thnakns, you just saved me a kot of time trying to figure out if compound model with agents will solve a particular problem. I have basic python knowledge and that was going to eat a lot of my time. So thank youuuuuuu! ❤
Watching from BerylCNC... what's your opinion on using CrewAI or AutoGen versus creating a GPT within OpenAI, and providing instructions that frame out the functionality in a similar way? My development work is mostly related to CNC tool path utilities. LLMs do a poor job of inferring and understanding geometry, so I have to bake in a lot of rules and math. GPTs seem like a cool way to get noticed, but I really need to include libraries and Python code. One of ours is called "Beryl of Widgets", and it helps makers figure out what to make and sell with the tools they have available. It could be so much more with CrewAI, I think, but then I need a way to deploy it. Great content, thank you!
On 8GB VRAM I had no problem running 10B or 13B models, however I run Q5 gguf. On 24GB VRAM I am able to run 70B Q4 gguf. For slow tasks it's acceptable speed.
Their efficiency in handling tasks like data processing and research is astounding. Have you ever attempted to coordinate several AI agents with SmythOS?
I would give mistral medium a shot, from what I’ve tested it’s quite good, but a fair bit cheaper than gpt4. Only available via the mistral api and not much info on it, but I really liked it
Was mainly a suggestion regarding cost, I think it’s about a quarter the cost of got4 turbo, and that already about half of normal gpt4. The biggest limitation might be the 32k context, but that’s been fine for me so far. Seems like a good balance between super cheap & basic (3.5) and expensive & good reasoning (4) regarding cost, but not everyone knows about it. (Not meant as critique to your video, rather in case anyone is looking at the comments for additional input ) And I’ll have a look at that, good to know it already exists!
The topic sounds incredibly interesting. Trying to follow up your steps but never used an editor before (I don´t even know what you´re talking about from minute 3:25) Question, where do you recommend me to begin to be able to make the whole project?? Saludos desde México!
So these agents requires something called "Function Calling" in LLM which is enabled in GPT-4. That's why open source models didn't perform well, but I think models that are fine tuned for agents and function calling will do better. Worth a try!
This is a great video!! THANK YOU! What about fine-tuning a local model to perform better by training individual agents? Copywriter: trained on how to write effective copy Proofreader: trained to review the copy for edits and suggestions Project Manager: review work against requirements Marketer: brings in marketing experience for evaluating concepts Researcher: skilled in researching against the requirements and working with the marketer. Risk Manager: identifies risks and how to mitigate them Venture Capitalist: reviews the project and provides feedback on how to get funding. Can you have a router with crewAI? Starts with PM to scope the work, assign tasks, and validate deliverables.
I had the same idea about fine-tuning LLMs for specific tasks. Since the agents are created with prompts (afak), I think it makes sense to fine-tune LLMs for those prompts. Maybe there is already a dataset for that. If someone knows about it, please share where it is.
This is like a book in one video, thank you so much! Just curious, but how would you compare the latest autogen studio to crew ai? Lots of wonderful ideas here and beautifully presented, thank you so much for publishing this, you are indeed a knowledge sharing master and the world needs more intellectual contributions like this. Thanks again!
![5 Pieces by Hans Zimmer \\ Iconic Soundtracks \\ Relaxing Piano [20min]](http://i.ytimg.com/vi/Os47nMrjw_Y/mqdefault.jpg)
*UPDATE:* Thanks to the viewer @tryingET great suggestion, I managed to improve the prompts and make the output consistent. You can check out the improved version on my github.
Thanks for watching and I'm curious, what were your experiences with crewAI like?
Thanks to You I will start this in my home lab :) btw good job, waiting for more content. I see ring on finger, that's mean it's to late?
Top ASMR experience👌
Thanks for this deep dive! The problem with open source templates is that they don't handle function calls, which is necessary for the crew to function. It seems that OpenHermes handles it well, the scripts work as expected, but even gpt3.5 gives better results.. thanks again for sharing
i have to agree @@TheGalacticIndian
@@tryingET This was just a great suggestion! I just ran the script couple of times, and you're right, the results are much more consistent. I only changed the part with "(linkToURL)", for some reason it was throwing the agent off. But it works with simple "(link to project)". I'll update the repo, thanks a lot for this help 🙏🏻
crewAI creator here, so cool to see videos like this! I also automated parts of my work with crewAI as well and it's a "a-ha" moment for sure! Great content! keep it up 💪
Wow I was trying to find your video my guy… someone posted a tutorial with your video and didn’t link you, great videos @BuildNewThings
Great program! Any chance we will be able to use memgpt with it?
is it possible to run them in parallel instead of serial (maybe via threads)?
The idea is that there's a manager who of course manages the worker agents (researchers, writers, ...); the worker agents then hand-off their work to analyst agent(s) who determine if more work is to be done, the result is then handed off to the manager, who then hands off to-do work to task creator agents to define what else needs to be done based on metrics from the analyst(s), then brought back to the manager who then assigns tasks to agents based on available workload (agent workload queue would be cool). also, dynamically "spawning" (instantiating) agents based on needs would be cool also, to conserve resources. maybe some features in crewai are missing yet to do that - what do you think?
Huh wat? The video eventually comes to the conclusion that the results are useless and basically a waste of time. ☹️
@@themax2go not at the moment it only does sequential, maybe pair it with autogen
This is by far the most clear explanation I've found on agents, how to use them and how to run them locally. Congrats!
Literally every LLM video
Title: "I automated everything"
Video: "Wow no model can understand the task"
Still a long way to go!
When humans take their organic brain inside their skulls for granted...
Even if we could reach Singularity, we will always be n-1 away from the 'n'th civilization that Created our n-1 universe. This is the fundamental limitation of pixel based evolution. Hence, we would go a longer way being organic hackers than materialists.
@hidroman1993 A little late to the party. Has it gone any better by now?
@@syedirfanajmalNo, lol
I hardly comment on TH-cam channels. But this is another level. The way you explain things, organization, referencing, and pace spot on. Great content. Please keep up the great work. Thanks :)
First time i encounter your channel/videos, and immediately got a subscribe from me. I love the clear explanation, straight to the point, no over promising or "Make $3000 dollar a day with these 10 simple bla bla". Thank you for keeping things real and useful. Breath of fresh air!
Dr CC cfv CC cccc denn cr FFM ggf so XXL ref für du v_''_& cv
I'm relieved to find someone else who faced challenges while running local models. Your sincere and practical review is appreciated. Unlike many others who simply join the hype train without discussing their struggles, your honesty is refreshing. Thank you.
Try ollama or gpt4all anyway to run a SOTA model you will need gpu or apple silicon.
me too, local models beside being pain in the butt, freezing your whole damn dev environment, then you get a shitty output, i am surprised she did test many models, i would have givenup much faster.
Not enough people talk about this. I follow the steps and run into error after error. Versions don't match, missing dependencies, list goes on. Ai development really takes a toll on your life force lmao
I only tested one model weeks ago and i got no issue installing it and running it. It’s uncensored and i was curious. Outputs were great. If you have at least 16gb ram on 7B models there’s no way it will crash your PC. Of course it’s slow as fuck, like 1 word per second generation
@@VizDopeyou assume you have nothing else running dude, how do you develop this way? you need at least 64GB just to feel something!
Amazing video thanks for the insights.
Very good and clear explantion. I rarely comment but when i do its because is worth it. So GPT-4 was the best model for all the tasks.
One of the best, honest reviews - love this! Thank you!
Good Job Maya :) I thought about many ideas while watching your video!
🎯 Key Takeaways for quick navigation:
00:41 🧠 *The video discusses the differences between System 1 (fast, subconscious thinking) and System 2 (slow, deliberate thinking) in the context of AI capabilities, highlighting that current language models primarily operate on System 1 thinking.*
01:48 💡 *Introduces two methods to simulate System 2 thinking in AI: "Tree of Thought" prompting and the use of platforms like CrewAI, which enable the construction of custom AI agents capable of collaborative problem-solving.*
03:26 🚀 *Outlines the process of setting up AI agents using CrewAI, emphasizing the importance of defining specific roles, goals, and backstories for each agent to ensure effective task execution.*
07:22 📈 *Describes how AI agents can be made more intelligent by granting them access to real-world data, and how to avoid fees and maintain privacy by running models locally.*
14:31 💸 *Discusses the cost implications of using models like GPT-4 for AI-driven tasks and explores local models as a more cost-effective and private alternative, despite their varying performance and capabilities.*
Made with HARPA AI
This video is fantastic! The content is thoroughly explained and incredibly helpful for professionals in the IT space. At Xerxes, we truly value such clarity and effort. Keep up the great work-looking forward to more insightful content like this!
Huge thanks for such detailed, well structured and illustrated information! The best video I’ve watched on AI so far.
Love this. Really transparent and talking about limitations of models..
Damn, you really are DOING THE WORK and then reporting back to us for free, dude !
Thanks so much for such gem. Much appreciated !
Thanks!
Thanks!😀
Thanks for the support :)
Thank you so much for putting this together! This is exactly what I have been looking for 🙏
Excellent explainer... Congrats and keep going !!!! Hello from Dominican Republic.
00:28 The inner dialogue shows the two systems of thinking: slow, deliberate system 2 vs. fast, automatic system 1. 🤔
01:36 Some people found two ways to simulate system 2 rational thinking in AI: tree of thought prompting and platforms like CrewAI. 💡
03:13 I’ll set up 3 AI agents on CrewAI to analyze my startup idea: a marketer, a technologist, and a business consultant. 👥
05:42 I defined specific tasks for each agent: analyze market demand, make product suggestions, and create biz plan. 📝
08:12 Making agents smarter is easy - add tools to give them access to real-time data like Google and Wikipedia. 🧠
11:06 I wrote a custom Reddit scraper tool to get higher quality info for my agents. 🤖
15:25 I tested 13 local AI models - most failed, but Llama-13B worked best to understand the task. 📈
Supported by NoteGPT
Great video, Maya! Please keep creating more valuable content about agent creation.
INCREDIBLE CONTENT. You just got a new follower
I watched Matthew Berman’s video on AutoGen, what made you decide to use CrewAI and have you tried/compared it to AutoGen?
Thank you for your clear and lucid explanation about CrewAI !!!
This is incredible. Thank you so much for sharing. Very inspiring!
🎯 Key Takeaways for quick navigation:
00:01 🧠 *Andre from OpenAI clarifies current LLMs are only capable of fast, instinctual thinking, not slow, rational analysis.*
02:15 🤝 *Crew AI allows anyone to build custom "agents" that can debate issues, thereby simulating rational thinking. *
03:26 💡 *Demos building 3 agents - marketer, technologist, consultant - to analyze a business idea.*
05:16 📋 *Defines specific tasks for each agent to collaboratively produce a business plan.*
08:37 🛠 *Adds "tools" to give agents access to real-time data, making them smarter.*
11:34 📰 *Builds a custom Reddit scraper tool to generate better newsletter content.*
14:58 💰 *Running models locally avoids OpenAI fees and keeps conversations private.*
18:49 🤔 *13B LLM was the only local model that vaguely understood the task.*
Made with HARPA AI
It's shocking to me how quickly someone made an AI to do this, as I created my own autonomous agents in python a few months back to do these similar things.
One tip I have for people trying to come up with a large & detailed tasks/descriptions is to write it in a .TXT file, and then reference it in your code. That way it keeps the code clean and also easy to modify the descriptions and tasks in the future without changing anything in the code.
Excellent idea! You could even create services to update it from a GUI without touching your code. I'd probably set it up like a traditional JSON config file.
How to use .txt from my laptop and feed it as an input to the llm?
please guide
@@DAN_1992 I don't know the answer, but could you just type that question into ChatGPT and it'll tell you?
This mirrors my experience with local models vs automation. I've come to the conclusion I either need to massively upgrade my hardware or just wait it out for a new breakthrough model. I'm a bit jaded with all the hype that never seems to live up to real-world use.
totally agree
Welllllll, yes and no. I've been able to tripple my productivity and I got 100% on two seperate essays using AI to workshop some ideas and build the essay outlines.
We already have "Agency Swarm" in our business so there is no need for CrewAI or Autogen for sure. Try it out.
Super work there Maya. You earned a subscriber, and a follow. I built an agent on node with run tools with custom functions over RAG. But this is next level only, will try this next. Thanks again. Keep shining
Really nice work friend. Nice narrative style and prosody while getting such a structured goal. And definitely awesome to listen discussions on the topics and decision making, this marks the difference. ... For trivial coding there is AI and Internet... for the core reasons and concepts there is us the humans
Maya, have to say - that was in-depth. Love the detail. Expensive running GPTs with CHAT 4. The output is definitely worth it though. I guess getting your own custom newsletter every day for less than a dollar does save research time. The next step is to get that file data into the actual newsletter now. Cheers for the free resources on Langchain. Just getting into APIs with py and deployment apps. I predict that you will have a great future on TH-cam. Keep up the good work.
I haved learned few very important things from your video. Thank you for an amazing video 🙏
🎯 Key Takeaways for quick navigation:
00:01 🧠 *System 1 vs. System 2 thinking and AI limitations.*
- Explanation of system 1 (fast) and system 2 (slow) thinking.
- Current large language models (LLMs) are limited to system 1 thinking.
- Desire for LLMs to achieve system 2 thinking for complex problem-solving.
02:02 🔄 *Two methods to simulate rational thinking in AI.*
- Tree of thought prompting involves considering issues from multiple perspectives.
- CrewAI and agent systems allow collaborative problem-solving with custom agents.
- Demonstration of assembling an AI agent team for startup analysis.
07:09 🌐 *Making AI agents smarter with real-world data access.*
- Integration of built-in tools, like text-to-speech and Google search, to enhance agents.
- Example: Using Google search tool for creating a detailed AI and ML innovation report.
- Introduction of local llama subreddit data for improved information quality.
10:25 🌐 *Custom tools for better data sources in AI tasks.*
- Introduction of custom Reddit scraper tool for more relevant information.
- Comparison of Gemini Pro's output and challenges with generic text generation.
- Acknowledgment of occasional variations and flakiness in agent outputs.
13:50 💰 *Cost considerations and running local models.*
- Discussion on pricing for API calls and the need to avoid high expenses.
- Testing open source models with varying parameters for running locally.
- Recommendation for at least 8 GB RAM for local models.
15:25 📊 *Performance of local models and surprising discoveries.*
- Experimentation with different local models and their understanding of tasks.
- Identification of models that performed poorly and surprising results with a basic model.
- Sharing notes on local models to avoid and those showing promise.
19:18 🤖 *Conclusion and audience engagement.*
- Recap of local model performance and potential improvements.
- Invitation for viewers to share their experiences with CrewAI.
- Appreciation for watching and anticipation for the next video.
Made with HARPA AI
This was an awesome video Maya. Thank you so very much for the wonderful and very helpful information! 🙏
Really good foundational information and great content. Thank you!!
This is an incredible guide!!! Thank you so much for making this video
This is awesome! Building a team of AI agents that can access real-world data sounds incredibly powerful.
My understanding is that using a larger quantized model works better. I’m planning on trying it soon on my computer, maybe with autogen. I’ve got a 4090, i9-10900k, and 64Gb RAM, so I’m hoping I can run maybe a 30b quantized model on it.
I read that the ~5-bit quantized models are the sweet spot that reduces your memory footprint without any significant loss in quality of responses. 4-bit is still good but takes enough of a hit to matter. Again, haven’t tried it myself, so maybe I’m mistaken, but that’s what I read.
This is awesome Maya, thank you for sharing.
You look like a 500k+ suscribers's creator, great job and great video btw
thanks a lot :)
🎯 Key Takeaways for quick navigation:
00:41 🤔 *Large language models (LLMs) currently exhibit system one thinking, which is subconscious and automatic, lacking the ability for deliberate, rational system two thinking.*
01:48 🌐 *Two methods to simulate rational thinking in AI: Tree of Thought Prompting, where experts contribute perspectives, and platforms like CrewAI, allowing collaboration between custom-built agents.*
03:13 ⚙️ *Building AI agents with CrewAI: Create agents (e.g., market researcher, technologist, business development expert), define tasks, and establish a sequential process for collaboration.*
08:37 🧰 *Enhancing agent intelligence: Add built-in tools for real-world data access, such as text-to-speech or Google search tools, to improve AI agents' capabilities.*
09:31 📰 *Creating an AI-driven newsletter: Utilize tools like Google search to gather information, but consider custom tools (e.g., Reddit scraper) for better control over data sources.*
13:23 💡 *Challenges with AI instructions: AI agents, even with GPT-4, may not consistently follow instructions, and results can vary. Custom tools may offer more control but require careful implementation.*
15:10 💸 *Cost considerations: Running AI scripts with external APIs can accumulate costs. Local models can be an alternative, but performance varies, and some models struggled with tasks.*
16:05 🚀 *Local models in CrewAI: Use local models to avoid API costs. Consider RAM requirements (e.g., 8GB for 7 billion parameters) and experiment with different models for optimal results.*
17:54 🧪 *Experimenting with local models: Testing various local models (e.g., llama 2 Series, 52, open chat) showed varying performance, with some models struggling to understand tasks.*
19:04 📊 *Results and conclusions: Despite challenges, a regular llama 13 billion parameters model, not fine-tuned, performed surprisingly well in incorporating data from a subreddit, highlighting experimentation and challenges in AI use.*
Wow. Thank you for such a great video and for sharing these insights. Really good.
It's quite possible that GPT4 is much more adept at understanding the premise of function calling, as it likely has a fine tuned expert in it's MOE to deal with "GPTs", thus making it more capable when dealing with OOTB solutions like CrewAI et al. I'd hazard that until someone fine tunes an OS model with a variety of function calling methods, and tools like CrewAI move on to more dynamic conversation flows rather than just sequential, then we'll begin to see the benefits of offline muilti-agent setups.
This was probally the most helpful video I have ever watched.
Side note, when showing something like the “ai agent landscape” would be neat if there was a reference where to find it. (There was enough info to do so, but a side of the repo would be sweet)
good callout, thanks! just included the link in the description box!
I want to thank you for this video! It is one of the most informative videos I have seen. Now that I watched it through I realized you really did your homework. I appreciate it.👏
TH-cam has automatic transcription. Can you have an agent go through all your subscriptions and summarize all the latest videos a few times a day? I would probably email that to myself or have an RSS feed, or some sort of news page to view the AI summaries.
that's a great idea, thanks for sharing! I'm going to work on and either make a video about it, or I'll get back to you!
@@maya-akim I also need an AI summary based on YT watch history for the day and the week. It's a weekly report for work. We need to spend 2 hours doing research on AI topics per week.
@@maya-akim I was able to find a YT transcription library and adapt it for my C# project pretty easily. There's a similar package for Python. th-cam.com/video/lY-6ZdsuHS8/w-d-xo.html
@@maya-akimI found an API that works to get the YT transcriptions from Windows forms - th-cam.com/video/lY-6ZdsuHS8/w-d-xo.html
this is so in-depth and really appreciate your hard-work and dedication Maya.
Absolutely Stunning Maya, Thanks for sharing these golden information.🤩
I used this same method only through Google apps script and integrated into spreadsheets. Nice work
That sounds interesting. I know Apps script better than Python and I want to work with sheets. Would you share your work, or some of your insights?
@@lausianne sure. I have a few videos on my channel and if you shoot me an email, I can share the code I used
Great presentation of the use cases and functionality. Thanks, Maya! 🙂
Have you ever consider that the problem might be in the Prompt Styling or Prompt Structure at all?
great! so much of ur tests and trials, thank you for sharing all
Thanks for the refreshingly honest results rather than the usual fake hype. It looks to me that LLMs have a long way to improve before autonomous agents can become actually useful.
Appreciate the way you've explained difference with analogy of "Thinking, Fast and Slow"
That's the most mental use of a boom arm I've seen anywhere.
😆
Commented for creativity, and the engagement boost.
I'm not sure how the video settings are different but I can't generate a video thumbnail from WordPress to this video. I keep track of all the top AI videos with my own WordPress instance. Embedding must be disabled.
Yep, you were right! I switched it off, but it’s on again 👍 that project of yours sounds interesting
7b-13b models are too small for a task like this most likely ( especially compared to a massive models like gpt4 ). Wonder how mixtral 8x7b would perform, even if we have to run it in cloud or through an llm api provider ( eg perplexity ).
I also didn't see anywhere any mention of the nature of the models which ollama uses. Most of the models it allows for download are the 4-bit quantized versions. Locally, I have been downloading the the slightly bigger 5 and 6-bit quantized versions and get better results. On the ollama git repo there are some easy instructions on how to import GGUF models one can download. TheBloke on huggingface has a lot of models that are converted and quantized.
This is one of the best videos on ai assistants.! 🎉 thank you Maya!
Great video! Thanks for all your hard work!
Great video! Congrats!
00:02 AI assistants are currently limited to system one thinking.
02:15 Build custom AI agents to solve complex tasks.
04:40 Creating a team of AI agents for specific business tasks.
07:11 Making AI assistants work in a sequential process
09:36 Creating custom tools for improved newsletter quality
11:53 Automated AI assistants improve efficiency and accuracy.
14:06 Using local models as AI assistants can be challenging due to high RAM requirements and potential crashes.
16:23 Testing of different AI models for newsletter generation.
18:28 Using a regular llama 13 billion parameters model for Reddit conversations on self-driving cars
You just won a New Subscriber, great video 🎉
Hi Maya! I’m so fascinated by your technical abilities. I barely understand the whole thing, but I’ve always been fascinated by AI and started learning AI tools.I saw on your tiktok that you taught yourself Python. It would be awesome if you can also share your learning process coming from a non-tech background and how’s your progress so far. Thank you :)
Thanks! It's a great idea, I'll definitely make a video about that :)
Thanks for testing local AI models, I had a similar experience
Many thnakns, you just saved me a kot of time trying to figure out if compound model with agents will solve a particular problem. I have basic python knowledge and that was going to eat a lot of my time. So thank youuuuuuu! ❤
I like your setup and vibe keep this good work up
Wow! Very good video! So much GREAT info - thank you!
Thanks for the info. Didn't even knew about Crew AI.
I had some unsalted Pringles, last week. This week, I had a Four Loko Gold
Watching from BerylCNC... what's your opinion on using CrewAI or AutoGen versus creating a GPT within OpenAI, and providing instructions that frame out the functionality in a similar way? My development work is mostly related to CNC tool path utilities. LLMs do a poor job of inferring and understanding geometry, so I have to bake in a lot of rules and math. GPTs seem like a cool way to get noticed, but I really need to include libraries and Python code. One of ours is called "Beryl of Widgets", and it helps makers figure out what to make and sell with the tools they have available. It could be so much more with CrewAI, I think, but then I need a way to deploy it. Great content, thank you!
On 8GB VRAM I had no problem running 10B or 13B models, however I run Q5 gguf. On 24GB VRAM I am able to run 70B Q4 gguf. For slow tasks it's acceptable speed.
Thanks for making this type of content. I like it.
Great work! I'll try it out tomorrow! 👏
Their efficiency in handling tasks like data processing and research is astounding. Have you ever attempted to coordinate several AI agents with SmythOS?
Amazing video 🤘
I'm so confused, you load the api env variable, but how is it used?
It's very calming to listen to you. Doesn't happen with technical videos a lot. Great vid!
I’m not a coder but want to learn how to build and tinker with these. Thanks for the clear explanations!
everybody can do it.. then opens a terminal and starts typing 😂.. loving this video already.. can’t wait till we start pip installing stuff
some encouragement is not bad ;) but also, the creator of crewai is already working on UI, so hopefully, people won't be pip installing for too long
😂 yea that install she skipped at the beginning would have been helpful for people like me, luckily I have GPT4 😅
I have been considering doing this. Thanks Maya.
Thanks a lot for this video. I like it. It's very useful. :)
Will be learning more about this CrewAI. I have been using ollama to run LLMs on my Raspberry Pi5. I really need more Pi5s to run in parallel.
I would give mistral medium a shot, from what I’ve tested it’s quite good, but a fair bit cheaper than gpt4. Only available via the mistral api and not much info on it, but I really liked it
try integrating it with this langchain tool, should work but I didn't try: js.langchain.com/docs/integrations/chat/mistral
Was mainly a suggestion regarding cost, I think it’s about a quarter the cost of got4 turbo, and that already about half of normal gpt4. The biggest limitation might be the 32k context, but that’s been fine for me so far.
Seems like a good balance between super cheap & basic (3.5) and expensive & good reasoning (4) regarding cost, but not everyone knows about it.
(Not meant as critique to your video, rather in case anyone is looking at the comments for additional input )
And I’ll have a look at that, good to know it already exists!
Great video and LLM review.
what is that beautiful looking tool being used for displaying that continuous stream of nicely indexed and formatted notes and figures around @1:57???
18:17 Maybe I don't understand completely but don't you need to train local models first?
You download them trained already. That's the whole point
What's the difference between this and the just released AutoGen Studio? If you don't know it watch the very well done tutorial made by Matthew Berman
Thank you! This content is so good.
I couldn't get passed the missing R. It bothered me very much.
Ha never heard anyone mentioning thinking fast and slow anywhere that is a great book
The topic sounds incredibly interesting. Trying to follow up your steps but never used an editor before (I don´t even know what you´re talking about from minute 3:25)
Question, where do you recommend me to begin to be able to make the whole project??
Saludos desde México!
So these agents requires something called "Function Calling" in LLM which is enabled in GPT-4. That's why open source models didn't perform well, but I think models that are fine tuned for agents and function calling will do better. Worth a try!
Thank you for showing this! can you put a gradio interface on this application to polish it off
thanks a lot and that's a great idea!
Great content and overview Maya. Thanks for sharing.
Do you have a video of how you create the virtual environment?
This is a great video!! THANK YOU!
What about fine-tuning a local model to perform better by training individual agents?
Copywriter: trained on how to write effective copy
Proofreader: trained to review the copy for edits and suggestions
Project Manager: review work against requirements
Marketer: brings in marketing experience for evaluating concepts
Researcher: skilled in researching against the requirements and working with the marketer.
Risk Manager: identifies risks and how to mitigate them
Venture Capitalist: reviews the project and provides feedback on how to get funding.
Can you have a router with crewAI? Starts with PM to scope the work, assign tasks, and validate deliverables.
I had the same idea about fine-tuning LLMs for specific tasks. Since the agents are created with prompts (afak), I think it makes sense to fine-tune LLMs for those prompts. Maybe there is already a dataset for that. If someone knows about it, please share where it is.
Wow, what great research and presentation. Thanks for sharing.
This is like a book in one video, thank you so much! Just curious, but how would you compare the latest autogen studio to crew ai? Lots of wonderful ideas here and beautifully presented, thank you so much for publishing this, you are indeed a knowledge sharing master and the world needs more intellectual contributions like this. Thanks again!
thanks a lot! I'm working on autogen studio video and I'll compare it to crewai