Your channel is extremely helpful in the new AI race, I am surprised not many subs are here. LLMs are at least as large as bitcoins but not many people have caught on the craze, not at least in the AI producer space yet.
I appreciate it a lot! And to me it sure feels like a lot of subscribers (I'm super grateful) but yeah I'm certainly still a small channel in the grand scheme of things!
I'm a confident no-code operator of 6-7 years and #1 most helpful thing would be if you enabled transcripts of all your videos so I could pass the URLs to an agent to help me build a knowledge base around this information to draw from for different applications...
I appreciate the suggestion! Someone in the comments already said they scraped the transcript for my video, so I thought it was already enabled. I'll have to look into it!
@@ColeMedin I believe it was on another video of yours, can't remember anymore. But it said that a transcript was not available. Anyways, thank you! Would be incredibly cool if you also wanted to set up your own "Cole KB" that collates these learnings for us :D
Hey Cole, great video! During your video, I was thinking that I wish we had a stack for non-developers-I'm talking about all those like me who know about LLMs, prompt engineering, some coding, SQL, and low-code/no-code solutions, but whose main function is more like that of architects. I wonder if your solution at the end of the video will try to solve that!
That's a great thought! Yes, the tech stack I am working on is going to have a focus for not just technical people but also more "semi-technical" folks like yourself!
00:05 - AI advancements are accelerating rapidly, creating confusion about what to focus on. 02:05 - Focus on capabilities over tools for AI success in 2025. 06:03 - Prioritize capabilities over tools for effective AI utilization in 2025. 08:08 - Understanding key AI agent architectures for effective implementation. 12:01 - The future of AI involves combining reasoning and smaller language models for efficient workflows. 14:07 - Local large language models are rapidly closing the gap with closed LLMs. 17:53 - Choose between local hosting or cloud services for your AI tech stack. 19:27 - Test various AI models to find optimal performance for your needs. 22:57 - Key tools and skills for leveraging AI in 2025. 24:39 - Human-in-the-loop systems and massive context windows will reshape AI applications by 2025. 28:10 - Engage with the AI community and participate in competitions.
Started building my own open source agents last year for my business with sustainability in mind so I can leverage high AI capabilities, targeting lowest CO2 emissions with high impact on people. It's highly challenging as ML is not my main field but step by step I'll get there. Kudos for the roadmap Cole and happy new year 🎉
@aperson113 My system will be local, power efficient and powered mostly with renewable energy, I plan to use light docker images, optimized usage (serverless-like) and LLMQ that can acheive closest result compared to non quantizied models, this approach ensure that I can collect relevant data so I can improve the system, it's not the perfect solution but I think it's a good start. Then I'll do carbon offset based on what's left
Thanks Cole. Really solid advice and great channel. I think the multi-modality of Gemini 2.0 Flash is the current game changer due to all the possibilities now available with those ( Realtime vision and audio input ). Not to mention the superior conversational ability through audio input as well as the speed. Your message is spot on. "What is the application"? Then go from there. A Gemini 2.0 flash API agent using audio input would be a great video. Keep up the great work!
Absolutely love this content. Your Chanel in general is just great. Thanks for the effort in producing high quality content. I hope to join in the hackathon but just getting started in the game! Your content is helping and inspiring me a lot. Thanks!
Cole, great stuff as usual. I have all but given up on taking notes while watching your videos, there is SO MUCH content and value it is like standing in front of a fire hose to get a sip of water. I have started to track your videos using notebook LM (until I debug my RAG for youtube transcript agent). The problem is it appears to take several days for your transcripts to become available. I know you are doing a TON of work, but any chance you could run the auto transcription when you post your videos so it is available right away? And I know, give them an inch and they ask for a mile, but the worst you can say to me is no. 👍
Thank you very much! I'll have to look into running the auto transcription, I haven't seen that option before but I'm sure it's right in my face. I would love to do that for each video!
Very solid round up, this moves so fast its insane, I didn't know half of these tools you mentioned not even in mention, I have a whole different stack for a lot of this stuff, 2024 was the year of the eco system explosion. 2025 will test who will make it and who will be a passing fad.
@@ColeMedin As you comment this I am going through a bunch of your old content catch ing up on some of the tools I missed out on. Working on a fresh set up for local AI & Automation. Ive got to say man you have some of the best content on youtube in this niche I don't know how you stay so far ahead of the curve, manage a massive community, a youtube channel, and one of the hottest open source repos of the year.. absolutely crushing it brother.
Thanks for going through all my content and for the kind words, I appreciate it a ton! It's pure passion driving me with all of the things I've got going on haha. AI is my life :)
Thank you very much for making the video available in multiple audio languages! I understand that we need to focus on functions, not tools, but I found it interesting that in one part of this video you mentioned Supabase for databases (I imagine it is for RAG). In another video, where you put together a Docker package with tools to run AI locally (including Open Web UI), you included Quadrant. I would like to know if you currently recommend using Supabase or if they serve different purposes. I don't know how to use either of them, I would like to know your opinion so that I can start learning about one of them.
Good question! PGVector in Postgres (Supabase uses Postgres under the hood) and Qdrant both serve similar purposes for RAG. I included Qdrant in the local AI starter kit by n8n just because that is what the team put in. You'll get very similar performance with both, so I typically recommend PGVector with Supabase because that way you get RAG and your SQL database all in one platform. Just simplifies the tech stack a lot!
Great video thanks. I’m a beginner but I love the capabilities of AI and the data supports the demand. I’m taking action immediately and starting a business. I’m also joining the community. Any useful suggestions?
You are so welcome! Glad to have you as a part of the community! Anything specific you are looking for suggestions on? I'd say the best place to start is to connect with others in the Think Tank and also post any specific questions you have there as you're getting started with AI!
Good question! I would recommend against using Flowise for production because it doesn't have the best logging/monitoring/error handling, but I would certainly recommend it for fast prototyping!
@ColeMedin WOW thank you a lot! I think that you need to clarify the first section for capabilities vs tools. capabilities can be a very huge set of things, so maybe put sections and summarize of subjects will be very helpful :) Thanks again men! You Are AWESOME!
You bet, thank you for the kind words! Yeah I get where you are coming from - capabilities is a HUGE umbrella that can mean a lot. Could you clarify what you mean by "put sections and summarize of subjects"?
Thank you! Unfortunately I have to pay a higher tier for the platform (Prezi) to be able to export it. I'm am working on bring it into another platform I can share though!
Hey, thanks for video. I am a software engineer and using cursor while coding and also claude/chatgpt llms for prompting. I don’t know machine learning or etc. Do you suggest me diving deep into ML or reinforcement learning ? I use AI to increase my productivity that’s why I am just trying to understand what is llm, generative ai or etc. do you suggest any roadmap for integrating ai into my job?
You bet! At this point I'd honestly just keep focusing on generative AI instead of machine learning/deep learning unless you really want to dive into the core of how generative AI works or want to create models to solve very specific problems. A roadmap for integrating AI into your job depends a LOT on your specific job so I don't have something for that! This video though honestly serves as a roadmap for what I'd recommend focusing on this year.
Won't massive context windows make RAG obsolete in the mid-long term? No need to chunk and retrieve the contents of a book (for example) when you can just give the LLM the book in a prompt. I'm sure I'm misunderstanding something here!
Really good question! Theoretically some day we'll see LLMs that can take in billions of tokens and be able to find the tiny piece of text within the whole block to answer a very specific question. But right now we are still at the point where there are knowledge bases that are WAY larger than the maximum context window for an LLM. Plus, even if you can fit everything into a prompt, it doesn't mean you should to get the best results. A lot of times we see the "needle in the haystack" problem where the information for the LLM is there in the prompt, but it gets confused with all the text and misses what it needs to answer correctly.
Hey Cole, thank you for these valuable contents ! Would you mind showing more examples of ReAct agents on n8n ? Do you think they could be more powerful than Tool Agents ?
You bet man! The agent node in n8n really is using ReAct under the hood, especially if you tell it to think through what it needs to do before invoking a tool. Or maybe you could clarify more what you think is missing in n8n for ReAct?
Огромное тебе спасибо за такую потрясающую работу. Твоя работа помогает многим людям стать лучше и сделать наш мир лучше. Дай бог тебе сил, здоровья и достаточно времени )
Great question! You know, I haven't looked into CAG much but I really want to. My first hesitation is that overloading the LLM with a ton of knowledge ahead of time is going to introduce a big "needle in the haystack" problem, but I know that is going to be less and less of an issue overall as we see more and more of these LLMs that are meant to handle large context windows. But even right now with models like o1 once I start adding in 5k+ tokens it starts missing key pieces of information in the large amount of context I give it.
Hello Cole, have you ever had the impression that RAG, because it recovers k excerpts by likelihood, may be insufficient for more complex analyses? Have you ever used finetuning techniques in LLM, with an impact on the model parameters, either fully or partially? Would it be possible to use the finetuning technique and still be able to identify the source documents, as happens in RAG? I am creating a database of monographs for academic reference and citing the source is very important in my case.
Great question! Unfortunately with fine tuning, you aren't guaranteed that the LLM will recount specific knowledge you give it or be able to cite sources. Fine tuning is meant more for defining the behavior for LLMs through hundreds or thousands of examples, while RAG is necessary when you need the LLM to reference specific information. For citing a large number of sources, I think RAG is still appropriate. You just might need to split the workflow into chunks - i.e. having the LLM analyze one paragraph at a time and get the sources for that if applicable.
Question - why not use the same tools in every video so that noob like me don`t get confused with the tool lingo. E.g. if it is workflow - stick to n8n and don`t bring in langgraph. I'm trying to assemble to the stack of software and tools i need to enable the capabilities. its easier to follow if same the tools are used and know that they can be swapped out with other ones. just saying its confusing that's all. Otherwise great content - using this to build RAG chatbot web app. Thanks!
Thanks for the kind words and yeah that's a very fair question! Couple reasons: 1. Each tool has their pros and cons depending on the use case, so I want to cover a good range of tools so no matter what someone is building they can find something that works for them. I try to be very clear what specific tools are used for to help with what you are saying, I could probably be better at that! 2. A lot of these tools actually work together or just aren't comparable. For example, n8n is a workflow automation tool that is low/no code while LangGraph is a code heavy agent orchestration tool. So they aren't comparable and often used in the same system so it's worth focusing on both!
### Roadmap to Dominate AI in 2025 This roadmap is designed to help you systematically develop the high-leverage skills and capabilities discussed in Cole Medin's video. It is divided into **phases** to guide your learning and implementation process. --- ### **Phase 1: Foundation (1-2 Months)** #### Goal: Build a strong understanding of AI fundamentals and adopt the "Capabilities Over Tools" mindset. 1. **Adopt the Right Mindset**: - Focus on mastering **capabilities** (e.g., problem-solving, architecture design) rather than specific tools. - Understand that tools evolve quickly, but core skills remain relevant. 2. **Learn AI Basics**: - Study the fundamentals of **machine learning** and **deep learning**. - Understand how **large language models (LLMs)** work (e.g., transformers, fine-tuning, embeddings). 3. **Explore AI Trends**: - Research the latest advancements in AI, such as reasoning LLMs, AI agents, and local LLMs. - Read articles like Anthropic's guide on **AI agent architecture**. 4. **Set Up a Basic AI Tech Stack**: - Choose a cloud provider (e.g., AWS, Digital Ocean) or experiment with local setups. - Install tools like Python, Docker, and basic LLMs (e.g., Llama or GPT-based APIs). --- ### **Phase 2: Master High-Leverage Skills (3-4 Months)** #### Goal: Develop expertise in the most impactful AI skills for 2025. 1. **AI Agents**: - Learn about **agent architecture** and best practices. - Experiment with platforms like **LangGraph**, **Pantic AI**, or **OpenAI Assistance**. - Build a simple AI agent to automate a task (e.g., email summarization or data retrieval). 2. **Reasoning LLMs**: - Study reasoning LLMs like **Gemini 2.0**, **DeepSeek V3**, and **OpenAI GPT-4 Turbo**. - Practice **prompt engineering** techniques: - Single-shot and multi-shot prompting. - Chain-of-thought (CoT) prompting. - Experiment with reasoning LLMs in agent workflows. 3. **Local LLMs**: - Learn how to set up and fine-tune local LLMs (e.g., Llama, Quen). - Understand hardware requirements for running LLMs locally. - Compare local vs. cloud-based LLMs for your use cases (e.g., privacy, cost, speed). 4. **Human-in-the-Loop Systems**: - Study how to integrate human oversight into AI workflows. - Use tools like **LangGraph** to implement approval mechanisms for AI agents. 5. **Massive Context Windows**: - Explore models with large context windows (e.g., Gemini). - Learn how to structure prompts to leverage extensive input data effectively. --- ### **Phase 3: Build and Optimize Your AI Tech Stack (2-3 Months)** #### Goal: Create a personalized AI tech stack tailored to your needs. 1. **Define Your Use Case**: - Identify a real-world problem you want to solve with AI (e.g., customer support automation, content generation). 2. **Select Tools for Your Tech Stack**: - **LLMs**: Choose between cloud-based (e.g., GPT-4, Gemini) and local models (e.g., Llama, Quen). - **Databases**: Use tools like **Supabase** (Postgres with PG Vector) for retrieval-augmented generation (RAG). - **Automation**: Experiment with platforms like **VoiceFlow** and **n8n**. - **Hosting**: Decide between cloud hosting (e.g., Digital Ocean) or local infrastructure. 3. **Prototype and Test**: - Build a proof-of-concept system using your selected tools. - Test different configurations to optimize performance, cost, and scalability. 4. **Follow Best Practices**: - Keep your stack **simple** (KISS: Keep It Simple, Stupid). - Reuse components across projects (DRY: Don’t Repeat Yourself). --- ### **Phase 4: Advanced Skills and Specialization (3-4 Months)** #### Goal: Deepen your expertise and prepare for cutting-edge AI applications. 1. **Advanced Prompt Engineering**: - Master advanced techniques for reasoning LLMs. - Learn to design prompts for domain-specific tasks (e.g., coding, creative writing). 2. **Fine-Tuning and Customization**: - Fine-tune LLMs on your own datasets for specific use cases. - Experiment with domain-specific LLMs (e.g., for healthcare, finance, or legal applications). 3. **AI Agent Ecosystems**: - Build complex agent workflows with multiple LLMs (e.g., reasoning LLMs + smaller, faster models). - Explore orchestration techniques like **parallelization** and **task routing**. 4. **Evaluation and Debugging**: - Use tools like **Phoenix** and **Ragos** to evaluate agent performance and reduce hallucinations. - Develop custom evaluation agents to test your systems. --- ### **Phase 5: Real-World Implementation and Community Engagement (Ongoing)** #### Goal: Apply your skills to real-world projects and grow with the AI community. 1. **Launch Projects**: - Build and deploy AI-powered solutions for personal or business use. - Examples: AI chatbots, automated workflows, or content generation tools. 2. **Participate in Hackathons**: - Join competitions like the **Live Agent Studio Hackathon** to showcase your skills and win prizes. 3. **Engage with the AI Community**: - Join communities like **Automator Think Tank** to network and collaborate. - Share your knowledge and learn from others’ experiences. 4. **Stay Updated**: - Follow advancements in AI tools, models, and frameworks. - Regularly update your tech stack and skills to stay ahead. --- ### **Phase 6: Future-Proofing (Ongoing)** #### Goal: Prepare for long-term success in the rapidly evolving AI landscape. 1. **Focus on Capabilities**: - Continuously refine high-leverage skills like problem-solving, architecture design, and prompt engineering. 2. **Experiment with Emerging Technologies**: - Stay ahead by exploring new tools, models, and frameworks as they are released. 3. **Build a Personal AI Enablement Stack**: - Create a reusable, open-source stack for your projects, including pre-built agents, coding assistants, and documentation. 4. **Collaborate and Innovate**: - Work with others to develop innovative AI solutions. - Contribute to open-source projects to give back to the community. --- ### Timeline Overview: | **Phase** | **Duration** | **Key Focus** | |----------------------------|--------------|-------------------------------------------------------------------------------| | **Phase 1: Foundation** | 1-2 Months | Learn AI basics, adopt the right mindset, and set up a basic tech stack. | | **Phase 2: High-Leverage Skills** | 3-4 Months | Master AI agents, reasoning LLMs, local LLMs, and human-in-the-loop systems. | | **Phase 3: AI Tech Stack** | 2-3 Months | Build and optimize your personalized AI tech stack. | | **Phase 4: Advanced Skills** | 3-4 Months | Specialize in advanced techniques like fine-tuning and complex agent design. | | **Phase 5: Real-World Use** | Ongoing | Apply skills to projects, join hackathons, and engage with the community. | | **Phase 6: Future-Proofing** | Ongoing | Stay updated, experiment with new tools, and focus on long-term capabilities.| --- By following this roadmap, you can systematically build the skills and expertise needed to dominate AI in 2025 and beyond.
Thanks for this great overview, it is really overwhelming and unfortunately most of the tools require some time investment before they can be used. Are there tools out there that make it really easy and fast for someone to get at least a hello world example?
You are so welcome! I totally get the barrier to entry for a lot of AI tools can be pretty high. My recommendation to get something up fast that is still impressive to some extend is to build an agent with n8n!
in the right track, could have used better examples than c++ one also after going through anthropic arch, it's missing the point to go back and use langgraph
Good question! The best use of AI is to build things to save you time, which obviously indirectly can make you a LOT of money. Things like an AI agent to organize your tasks, customer support agents, agents to manage your inbox, etc. And then anything you build for yourself and prove is useful for yourself is probably something other people would want to pay for too! So it makes sense to then turn it into a service/SaaS.
Arguable statement of Capabilities not tools. I see myriads of feature packed products which are full of bugs and poor performance. Mainly result of it is a culture of making things fast (and thanks to AI even faster) mixed with one’s poor knowledge of a programming language (tool), platform (iOS, android, windows , etc.), how memory, rendering, etc. working. AI won’t tell you to focus on such things unless you ask and you won’t ask since you don’t have knowledge to ask correct question since you’ve been focusing on capabilities.
I get where you are coming from! And maybe I should have clarified this more in the video, but when I say capabilities I'm not just talking about "features" at a surface level user level. I also mean capabilities in a very technical sense. Like "is my system capable of reporting errors and monitoring API usage so I can actually resolve bugs and performance issues". So I'm certainly not arguing to focus just on features and avoid spending the time really making the product bullet proof!
So why are you the only one who is pushing n8n my thought is you own the company cause no one else even mentions it, you know you would get more views if you built agents without it all the time.
I see a TON of other content creators using n8n myself! I don't own it at all, I just appreciate it a ton because it works great for me, it's free to use, and it's open source!
Including the acronym itself? I've only heard it before under the context of software development! If you mean just the word itself though then yeah it's been around for ages haha
Register now for the oTTomator AI Agent Hackathon with a $5,000 prize pool! My gift to you going into 2025 :)
studio.ottomator.ai/hackathon/register
Ai in #MASSIVESQUIDGAMEEVENT! DE #Morgan! de #666! #Cumvorarătabuletineleelectronice! in 13 2025
solid video. I appreciate you eliminating all the noise out there with a new tool coming out every day
Bro, your are a gift for AI enthusiasts. :
Your channel is extremely helpful in the new AI race, I am surprised not many subs are here. LLMs are at least as large as bitcoins but not many people have caught on the craze, not at least in the AI producer space yet.
I appreciate it a lot! And to me it sure feels like a lot of subscribers (I'm super grateful) but yeah I'm certainly still a small channel in the grand scheme of things!
Man you are hands down the best content creator i’ve come across in this space . Thank you for everything
I appreciate it, that means a lot! You are so welcome :D
Love your work Cole definitely onboard! Kia Ora from Aotearoa, looking forward to changing the game in 2025. Happy New Year Mate....
Thanks man! Happy new year!!
Bro the information you provided is useful. Honestly the knowledge you dispensed is high level
I appreciate it a lot!
Best videos so far I’ve seen about AI. Can’t wait to see your upcoming videos. Happy new year 🎆
Thank you very much! Happy new year!
I'm a confident no-code operator of 6-7 years and #1 most helpful thing would be if you enabled transcripts of all your videos so I could pass the URLs to an agent to help me build a knowledge base around this information to draw from for different applications...
I appreciate the suggestion! Someone in the comments already said they scraped the transcript for my video, so I thought it was already enabled. I'll have to look into it!
@@ColeMedin I believe it was on another video of yours, can't remember anymore. But it said that a transcript was not available. Anyways, thank you! Would be incredibly cool if you also wanted to set up your own "Cole KB" that collates these learnings for us :D
Yeah I am actually looking to make this!
Automated process, of course. And show us how, please. I just subscribed to your channel. Thank you @@ColeMedin
Happy New Year! Thank you for all the good content!
You are so welcome! Happy new year!
Hey Cole, great video! During your video, I was thinking that I wish we had a stack for non-developers-I'm talking about all those like me who know about LLMs, prompt engineering, some coding, SQL, and low-code/no-code solutions, but whose main function is more like that of architects. I wonder if your solution at the end of the video will try to solve that!
That's a great thought! Yes, the tech stack I am working on is going to have a focus for not just technical people but also more "semi-technical" folks like yourself!
00:05 - AI advancements are accelerating rapidly, creating confusion about what to focus on.
02:05 - Focus on capabilities over tools for AI success in 2025.
06:03 - Prioritize capabilities over tools for effective AI utilization in 2025.
08:08 - Understanding key AI agent architectures for effective implementation.
12:01 - The future of AI involves combining reasoning and smaller language models for efficient workflows.
14:07 - Local large language models are rapidly closing the gap with closed LLMs.
17:53 - Choose between local hosting or cloud services for your AI tech stack.
19:27 - Test various AI models to find optimal performance for your needs.
22:57 - Key tools and skills for leveraging AI in 2025.
24:39 - Human-in-the-loop systems and massive context windows will reshape AI applications by 2025.
28:10 - Engage with the AI community and participate in competitions.
Started building my own open source agents last year for my business with sustainability in mind so I can leverage high AI capabilities, targeting lowest CO2 emissions with high impact on people. It's highly challenging as ML is not my main field but step by step I'll get there. Kudos for the roadmap Cole and happy new year 🎉
Nice David!! Happy new year!
@@ColeMedin Thanks 😀
What is your strategy to keep co2 emissions low?
@aperson113 My system will be local, power efficient and powered mostly with renewable energy, I plan to use light docker images, optimized usage (serverless-like) and LLMQ that can acheive closest result compared to non quantizied models, this approach ensure that I can collect relevant data so I can improve the system, it's not the perfect solution but I think it's a good start. Then I'll do carbon offset based on what's left
Thanks Cole. Really solid advice and great channel. I think the multi-modality of Gemini 2.0 Flash is the current game changer due to all the possibilities now available with those ( Realtime vision and audio input ). Not to mention the superior conversational ability through audio input as well as the speed. Your message is spot on. "What is the application"? Then go from there. A Gemini 2.0 flash API agent using audio input would be a great video. Keep up the great work!
Thanks Jeff, that means a lot! And yes I'm with you that Gemini 2.0 Flash and multi modal models like it are a game changer!
this is gold! you're my go-to creator in the new AI era
Thank you! I appreciate it a lot!
Thanks!
Thank you so much for the support Eddie!!
Wow, this was so informative!
I'm glad you think so - thank you!
Absolutely love this content. Your Chanel in general is just great. Thanks for the effort in producing high quality content. I hope to join in the hackathon but just getting started in the game! Your content is helping and inspiring me a lot. Thanks!
Thank you so much - that means a lot to me! :D
You bet!
Incredibly helpful video. Thank you so much my friend! 😌😌🙏
You're very welcome - glad you enjoyed it!
Yo I just wanna say this is straight fire bro, thanks for work you put into these vids🙏🙏🙏
I appreciate it Paul, glad you are enjoying the content man!
Cole, great stuff as usual. I have all but given up on taking notes while watching your videos, there is SO MUCH content and value it is like standing in front of a fire hose to get a sip of water. I have started to track your videos using notebook LM (until I debug my RAG for youtube transcript agent). The problem is it appears to take several days for your transcripts to become available. I know you are doing a TON of work, but any chance you could run the auto transcription when you post your videos so it is available right away? And I know, give them an inch and they ask for a mile, but the worst you can say to me is no. 👍
Thank you very much! I'll have to look into running the auto transcription, I haven't seen that option before but I'm sure it's right in my face. I would love to do that for each video!
Nice mate! Appreciate the effort put into this.
I appreciate it, thanks!
Best video of 2025 without a doubt and thank you so much!! 🔝🔝👏👏
Thank you so much, you're so welcome!
Very solid round up, this moves so fast its insane, I didn't know half of these tools you mentioned not even in mention, I have a whole different stack for a lot of this stuff, 2024 was the year of the eco system explosion. 2025 will test who will make it and who will be a passing fad.
Thanks man! And yeah you're 100% spot on!
@@ColeMedin As you comment this I am going through a bunch of your old content catch ing up on some of the tools I missed out on. Working on a fresh set up for local AI & Automation. Ive got to say man you have some of the best content on youtube in this niche I don't know how you stay so far ahead of the curve, manage a massive community, a youtube channel, and one of the hottest open source repos of the year.. absolutely crushing it brother.
Thanks for going through all my content and for the kind words, I appreciate it a ton! It's pure passion driving me with all of the things I've got going on haha. AI is my life :)
Stay informed is the most important topic for me.
I would love to thank you for this great info to start with ai for 2025
You're welcome!
Great topic, thanks 👍
Thanks! You bet!
Nice Work Mate. But can you provide this roadmap image 2:16 in the description or comment section if possible ?
Thank you! Unfortunately I have to pay a higher tier for Prezi to share this presentation or export it...
Thank you very much for making the video available in multiple audio languages! I understand that we need to focus on functions, not tools, but I found it interesting that in one part of this video you mentioned Supabase for databases (I imagine it is for RAG). In another video, where you put together a Docker package with tools to run AI locally (including Open Web UI), you included Quadrant. I would like to know if you currently recommend using Supabase or if they serve different purposes. I don't know how to use either of them, I would like to know your opinion so that I can start learning about one of them.
Good question! PGVector in Postgres (Supabase uses Postgres under the hood) and Qdrant both serve similar purposes for RAG. I included Qdrant in the local AI starter kit by n8n just because that is what the team put in. You'll get very similar performance with both, so I typically recommend PGVector with Supabase because that way you get RAG and your SQL database all in one platform. Just simplifies the tech stack a lot!
Wealth of knowledge 🎉
Hard not to be scared for the future!
Great new year introduction thanks
You are a legend! Thx for your contribution 🎉
Thanks Peter, you bet! :D
What's the tool used for these presentations?
I use Prezi for this!
Great video as always!
Thank you Paul! :D
you are amazing. thank you thank you
You bet man! I appreciate it!
Great video thanks. I’m a beginner but I love the capabilities of AI and the data supports the demand. I’m taking action immediately and starting a business. I’m also joining the community. Any useful suggestions?
You are so welcome! Glad to have you as a part of the community! Anything specific you are looking for suggestions on? I'd say the best place to start is to connect with others in the Think Tank and also post any specific questions you have there as you're getting started with AI!
What if i use flowise for production ?
Good question! I would recommend against using Flowise for production because it doesn't have the best logging/monitoring/error handling, but I would certainly recommend it for fast prototyping!
Ok ottomater pages read aloud. Thank you.
@ColeMedin WOW thank you a lot!
I think that you need to clarify the first section for capabilities vs tools.
capabilities can be a very huge set of things, so maybe put sections and summarize of subjects will be very helpful :)
Thanks again men! You Are AWESOME!
You bet, thank you for the kind words!
Yeah I get where you are coming from - capabilities is a HUGE umbrella that can mean a lot. Could you clarify what you mean by "put sections and summarize of subjects"?
Excelent video. I have a question, can you share de roadmap in image, link, pdf or simething like that?
Thank you! Unfortunately I have to pay a higher tier for the platform (Prezi) to be able to export it. I'm am working on bring it into another platform I can share though!
Very useful information. Thanks!
You're welcome! Glad you found it useful.
A must see video
Hey, thanks for video. I am a software engineer and using cursor while coding and also claude/chatgpt llms for prompting. I don’t know machine learning or etc. Do you suggest me diving deep into ML or reinforcement learning ? I use AI to increase my productivity that’s why I am just trying to understand what is llm, generative ai or etc. do you suggest any roadmap for integrating ai into my job?
You bet! At this point I'd honestly just keep focusing on generative AI instead of machine learning/deep learning unless you really want to dive into the core of how generative AI works or want to create models to solve very specific problems.
A roadmap for integrating AI into your job depends a LOT on your specific job so I don't have something for that! This video though honestly serves as a roadmap for what I'd recommend focusing on this year.
Won't massive context windows make RAG obsolete in the mid-long term? No need to chunk and retrieve the contents of a book (for example) when you can just give the LLM the book in a prompt. I'm sure I'm misunderstanding something here!
Really good question! Theoretically some day we'll see LLMs that can take in billions of tokens and be able to find the tiny piece of text within the whole block to answer a very specific question.
But right now we are still at the point where there are knowledge bases that are WAY larger than the maximum context window for an LLM. Plus, even if you can fit everything into a prompt, it doesn't mean you should to get the best results. A lot of times we see the "needle in the haystack" problem where the information for the LLM is there in the prompt, but it gets confused with all the text and misses what it needs to answer correctly.
Kudos Cole! future must be opensource, depending on corporate AI is a moronic decision
Hey Cole, thank you for these valuable contents ! Would you mind showing more examples of ReAct agents on n8n ? Do you think they could be more powerful than Tool Agents ?
You bet man! The agent node in n8n really is using ReAct under the hood, especially if you tell it to think through what it needs to do before invoking a tool. Or maybe you could clarify more what you think is missing in n8n for ReAct?
Огромное тебе спасибо за такую потрясающую работу. Твоя работа помогает многим людям стать лучше и сделать наш мир лучше. Дай бог тебе сил, здоровья и достаточно времени )
Cole any thoughts on CAG over RAG? It seems like the future as context windows get bigger but also it seems like a weird/underdeveloped technology.
Great question! You know, I haven't looked into CAG much but I really want to. My first hesitation is that overloading the LLM with a ton of knowledge ahead of time is going to introduce a big "needle in the haystack" problem, but I know that is going to be less and less of an issue overall as we see more and more of these LLMs that are meant to handle large context windows. But even right now with models like o1 once I start adding in 5k+ tokens it starts missing key pieces of information in the large amount of context I give it.
How you make this beautiful roadmap animation
Hello Cole, have you ever had the impression that RAG, because it recovers k excerpts by likelihood, may be insufficient for more complex analyses? Have you ever used finetuning techniques in LLM, with an impact on the model parameters, either fully or partially? Would it be possible to use the finetuning technique and still be able to identify the source documents, as happens in RAG? I am creating a database of monographs for academic reference and citing the source is very important in my case.
Great question! Unfortunately with fine tuning, you aren't guaranteed that the LLM will recount specific knowledge you give it or be able to cite sources. Fine tuning is meant more for defining the behavior for LLMs through hundreds or thousands of examples, while RAG is necessary when you need the LLM to reference specific information.
For citing a large number of sources, I think RAG is still appropriate. You just might need to split the workflow into chunks - i.e. having the LLM analyze one paragraph at a time and get the sources for that if applicable.
U of M ❤ I didn't go there, I live down the street though!
Bingo! Love seeing other people from MN :)
Question - why not use the same tools in every video so that noob like me don`t get confused with the tool lingo. E.g. if it is workflow - stick to n8n and don`t bring in langgraph. I'm trying to assemble to the stack of software and tools i need to enable the capabilities. its easier to follow if same the tools are used and know that they can be swapped out with other ones. just saying its confusing that's all. Otherwise great content - using this to build RAG chatbot web app. Thanks!
Thanks for the kind words and yeah that's a very fair question! Couple reasons:
1. Each tool has their pros and cons depending on the use case, so I want to cover a good range of tools so no matter what someone is building they can find something that works for them. I try to be very clear what specific tools are used for to help with what you are saying, I could probably be better at that!
2. A lot of these tools actually work together or just aren't comparable. For example, n8n is a workflow automation tool that is low/no code while LangGraph is a code heavy agent orchestration tool. So they aren't comparable and often used in the same system so it's worth focusing on both!
### Roadmap to Dominate AI in 2025
This roadmap is designed to help you systematically develop the high-leverage skills and capabilities discussed in Cole Medin's video. It is divided into **phases** to guide your learning and implementation process.
---
### **Phase 1: Foundation (1-2 Months)**
#### Goal: Build a strong understanding of AI fundamentals and adopt the "Capabilities Over Tools" mindset.
1. **Adopt the Right Mindset**:
- Focus on mastering **capabilities** (e.g., problem-solving, architecture design) rather than specific tools.
- Understand that tools evolve quickly, but core skills remain relevant.
2. **Learn AI Basics**:
- Study the fundamentals of **machine learning** and **deep learning**.
- Understand how **large language models (LLMs)** work (e.g., transformers, fine-tuning, embeddings).
3. **Explore AI Trends**:
- Research the latest advancements in AI, such as reasoning LLMs, AI agents, and local LLMs.
- Read articles like Anthropic's guide on **AI agent architecture**.
4. **Set Up a Basic AI Tech Stack**:
- Choose a cloud provider (e.g., AWS, Digital Ocean) or experiment with local setups.
- Install tools like Python, Docker, and basic LLMs (e.g., Llama or GPT-based APIs).
---
### **Phase 2: Master High-Leverage Skills (3-4 Months)**
#### Goal: Develop expertise in the most impactful AI skills for 2025.
1. **AI Agents**:
- Learn about **agent architecture** and best practices.
- Experiment with platforms like **LangGraph**, **Pantic AI**, or **OpenAI Assistance**.
- Build a simple AI agent to automate a task (e.g., email summarization or data retrieval).
2. **Reasoning LLMs**:
- Study reasoning LLMs like **Gemini 2.0**, **DeepSeek V3**, and **OpenAI GPT-4 Turbo**.
- Practice **prompt engineering** techniques:
- Single-shot and multi-shot prompting.
- Chain-of-thought (CoT) prompting.
- Experiment with reasoning LLMs in agent workflows.
3. **Local LLMs**:
- Learn how to set up and fine-tune local LLMs (e.g., Llama, Quen).
- Understand hardware requirements for running LLMs locally.
- Compare local vs. cloud-based LLMs for your use cases (e.g., privacy, cost, speed).
4. **Human-in-the-Loop Systems**:
- Study how to integrate human oversight into AI workflows.
- Use tools like **LangGraph** to implement approval mechanisms for AI agents.
5. **Massive Context Windows**:
- Explore models with large context windows (e.g., Gemini).
- Learn how to structure prompts to leverage extensive input data effectively.
---
### **Phase 3: Build and Optimize Your AI Tech Stack (2-3 Months)**
#### Goal: Create a personalized AI tech stack tailored to your needs.
1. **Define Your Use Case**:
- Identify a real-world problem you want to solve with AI (e.g., customer support automation, content generation).
2. **Select Tools for Your Tech Stack**:
- **LLMs**: Choose between cloud-based (e.g., GPT-4, Gemini) and local models (e.g., Llama, Quen).
- **Databases**: Use tools like **Supabase** (Postgres with PG Vector) for retrieval-augmented generation (RAG).
- **Automation**: Experiment with platforms like **VoiceFlow** and **n8n**.
- **Hosting**: Decide between cloud hosting (e.g., Digital Ocean) or local infrastructure.
3. **Prototype and Test**:
- Build a proof-of-concept system using your selected tools.
- Test different configurations to optimize performance, cost, and scalability.
4. **Follow Best Practices**:
- Keep your stack **simple** (KISS: Keep It Simple, Stupid).
- Reuse components across projects (DRY: Don’t Repeat Yourself).
---
### **Phase 4: Advanced Skills and Specialization (3-4 Months)**
#### Goal: Deepen your expertise and prepare for cutting-edge AI applications.
1. **Advanced Prompt Engineering**:
- Master advanced techniques for reasoning LLMs.
- Learn to design prompts for domain-specific tasks (e.g., coding, creative writing).
2. **Fine-Tuning and Customization**:
- Fine-tune LLMs on your own datasets for specific use cases.
- Experiment with domain-specific LLMs (e.g., for healthcare, finance, or legal applications).
3. **AI Agent Ecosystems**:
- Build complex agent workflows with multiple LLMs (e.g., reasoning LLMs + smaller, faster models).
- Explore orchestration techniques like **parallelization** and **task routing**.
4. **Evaluation and Debugging**:
- Use tools like **Phoenix** and **Ragos** to evaluate agent performance and reduce hallucinations.
- Develop custom evaluation agents to test your systems.
---
### **Phase 5: Real-World Implementation and Community Engagement (Ongoing)**
#### Goal: Apply your skills to real-world projects and grow with the AI community.
1. **Launch Projects**:
- Build and deploy AI-powered solutions for personal or business use.
- Examples: AI chatbots, automated workflows, or content generation tools.
2. **Participate in Hackathons**:
- Join competitions like the **Live Agent Studio Hackathon** to showcase your skills and win prizes.
3. **Engage with the AI Community**:
- Join communities like **Automator Think Tank** to network and collaborate.
- Share your knowledge and learn from others’ experiences.
4. **Stay Updated**:
- Follow advancements in AI tools, models, and frameworks.
- Regularly update your tech stack and skills to stay ahead.
---
### **Phase 6: Future-Proofing (Ongoing)**
#### Goal: Prepare for long-term success in the rapidly evolving AI landscape.
1. **Focus on Capabilities**:
- Continuously refine high-leverage skills like problem-solving, architecture design, and prompt engineering.
2. **Experiment with Emerging Technologies**:
- Stay ahead by exploring new tools, models, and frameworks as they are released.
3. **Build a Personal AI Enablement Stack**:
- Create a reusable, open-source stack for your projects, including pre-built agents, coding assistants, and documentation.
4. **Collaborate and Innovate**:
- Work with others to develop innovative AI solutions.
- Contribute to open-source projects to give back to the community.
---
### Timeline Overview:
| **Phase** | **Duration** | **Key Focus** |
|----------------------------|--------------|-------------------------------------------------------------------------------|
| **Phase 1: Foundation** | 1-2 Months | Learn AI basics, adopt the right mindset, and set up a basic tech stack. |
| **Phase 2: High-Leverage Skills** | 3-4 Months | Master AI agents, reasoning LLMs, local LLMs, and human-in-the-loop systems. |
| **Phase 3: AI Tech Stack** | 2-3 Months | Build and optimize your personalized AI tech stack. |
| **Phase 4: Advanced Skills** | 3-4 Months | Specialize in advanced techniques like fine-tuning and complex agent design. |
| **Phase 5: Real-World Use** | Ongoing | Apply skills to projects, join hackathons, and engage with the community. |
| **Phase 6: Future-Proofing** | Ongoing | Stay updated, experiment with new tools, and focus on long-term capabilities.|
---
By following this roadmap, you can systematically build the skills and expertise needed to dominate AI in 2025 and beyond.
Thanks for this great overview, it is really overwhelming and unfortunately most of the tools require some time investment before they can be used. Are there tools out there that make it really easy and fast for someone to get at least a hello world example?
You are so welcome! I totally get the barrier to entry for a lot of AI tools can be pretty high. My recommendation to get something up fast that is still impressive to some extend is to build an agent with n8n!
Amazing high value content, thanks so much!!
I find myself in this situation too. Would you recommend a solid tutorial to get started with n8n?
I cover a good overview of setting up a RAG AI agent in n8n in this video!
ua-cam.com/video/PEI_ePNNfJQ/v-deo.html
in the right track, could have used better examples than c++ one
also after going through anthropic arch, it's missing the point to go back and use langgraph
I appreciate the feedback! What specifically do you think isn't the best about the C++ example?
top top top
11:59 Actually, you’ve probably been living under a rock, isolated in a bubble, if you have heard of o3! Me included! 😂😢
What are good ways you've seen people make money leveraging AI?
Good question! The best use of AI is to build things to save you time, which obviously indirectly can make you a LOT of money. Things like an AI agent to organize your tasks, customer support agents, agents to manage your inbox, etc. And then anything you build for yourself and prove is useful for yourself is probably something other people would want to pay for too! So it makes sense to then turn it into a service/SaaS.
How to build this
Why you dont haaave on tour stock Oracle or IBM ?
Arguable statement of Capabilities not tools. I see myriads of feature packed products which are full of bugs and poor performance. Mainly result of it is a culture of making things fast (and thanks to AI even faster) mixed with one’s poor knowledge of a programming language (tool), platform (iOS, android, windows , etc.), how memory, rendering, etc. working. AI won’t tell you to focus on such things unless you ask and you won’t ask since you don’t have knowledge to ask correct question since you’ve been focusing on capabilities.
I get where you are coming from! And maybe I should have clarified this more in the video, but when I say capabilities I'm not just talking about "features" at a surface level user level. I also mean capabilities in a very technical sense. Like "is my system capable of reporting errors and monitoring API usage so I can actually resolve bugs and performance issues". So I'm certainly not arguing to focus just on features and avoid spending the time really making the product bullet proof!
So why are you the only one who is pushing n8n my thought is you own the company cause no one else even mentions it, you know you would get more views if you built agents without it all the time.
?
a lot of people use N8N...
see the UA-cam video and community talking about it.
I see a TON of other content creators using n8n myself! I don't own it at all, I just appreciate it a ton because it works great for me, it's free to use, and it's open source!
KISS existed long before software was a word. Or maybe it was a word used to describe pillows.
Including the acronym itself? I've only heard it before under the context of software development! If you mean just the word itself though then yeah it's been around for ages haha
ua-cam.com/video/wAzBl6xllzE/v-deo.htmlsi=HFHKVayP1x_SCX-P&t=292 University of Minnesota Twin Cities ?
Bingo!!