Transcription
AI is moving fast going into 2025, and no one is even going to try and tell you that things will slow down. If anything, it's just going to keep getting crazier, and that, my friend, can be overwhelming.
It can feel impossible to know what to focus on, no matter who you are—whether you're non-technical and just getting your feet wet with AI, a developer trying to build an awesome app around AI, or a business owner just trying to cut through all the AI fluff and implement something that's actually useful for your business. We're all in the same boat.
It feels like every week there's a new AI coding assistant, a new large language model, or a new AI agent framework or tool. With so much being released every single week, how could I possibly filter out what's important for me and what's not? How do I know what I'm investing my time into learning isn't just going to become irrelevant in a month?
If you don't have any of these questions and you have it all figured out, you, my friend, are a wizard, and you should just run my YouTube channel for me. If I was a betting man, I would definitely bet that you have a lot of questions right now around where to truly put your time as you learn how to leverage AI effectively, especially with everything released even in the last couple of months.
Here in 2024, we've got 03 Gemini 2.0, Flash Pantic, AI Windsurf, Lovable, New Quen, Llama, Deep Seek models, and Claude's Model Context Protocol. It is all just way too much.
But in this video, I have a fundamental mindset shift for you that's going to really help you manage all the craziness of AI going into 2025. As I've wrestled with these questions myself and done hours and hours of research and development with AI this year, I've come up with a set of just a few high-leverage skills that are going to really keep you ahead with AI going into this year, no matter what is released, what becomes irrelevant, or how much we achieve AGI this year.
This isn't just your average 2025 AI roadmap or "here's how I learn AI if I had to start over" video. This is a no fluff, no BS conversation that I just want to have with you, sharing how I plan to invest my time into AI that I hope will resonate with you, no matter who you are.
So with that, let's dive right into it.
I have crafted this beautiful sort of roadmap for you for us to talk about AI trends going into 2025 and the important skills for you and I to focus on right now. Now, it's important to note that all these things that I have laid out here don't necessarily go in a specific order. This roadmap style is the best way for me to lay out everything for you right now.
So with that, we can dive into the first thing here, which is really setting the stage for everything that I want to talk about here because it is the most important motto that you can get through your head right now, and that is the idea of capabilities, not tools.
Let me explain.
When I was a software engineer working a 9-to-5 job back in the day, I had this VP on my team, this really intelligent dude that had multiple PhDs. He taught at colleges, he had a ton of patents, doing incredible things with deep learning, machine learning, and generative AI. The thing that he would always say—it really became his motto—is focus on capabilities, not tools.
What he meant by that is focus on the bigger picture, higher-leverage skills instead of a bunch of disparate low-leverage skills, focusing on mastering specific tools like specific programming languages or agent frameworks, whatever it might be.
So the idea for us, as we're thinking about AI, is to focus on what can I do with AI, not as much how can I work with these specific tools to do things with AI. Because in the end, AI can do a lot for us already, like coding and creating animations and stuff. All these tools that we might be tempted to really dive into and master get replaced all the time.
We have AI coding assistants, LLMs, agent frameworks coming out every single week, and so it's not really worth it for us to invest our time there and get so caught in the weeds when that knowledge might become irrelevant.
Like I said right here, it's especially true in the age of AI, even though this applies to software engineering and really all of life. These higher-leverage skills are what are going to keep us relevant when things are changing so much in the industry.
One super important caveat that I want to state here is that learning specific tools is still very important. I mean, you still have to pick some LLM to work with, some agent framework to code with, some automation tool to use—you get the idea. You still need to pick one, but don't make it a priority to master that.
Think about the more higher-leverage skills, a lot of which I'm going to talk about through the rest of this roadmap here.
All right, it's time to take a step away from the presentation quick because it is story time. I used to be that person that was completely obsessed with tools over capabilities, and it didn't get me very far.
One example in particular that I cringe at every single time I think about it is right before I went to college to get my bachelor's in computer science. I had a couple of free weeks during the end of that summer, and what I wanted to do was give myself a coding boot camp with something just to help me prep for college and be a better programmer.
Out of all of the things that I could have done, for some weird reason, I chose to dive into advanced C++ topics. I'm talking really, really technical stuff like static polymorphism and advanced inheritance structures, some of which you guys might not even be aware of. It's so technical; it's a very good example of being way too into the tool instead of capabilities.
So yeah, honestly, it ended up being a waste of my time. I don't code with C++ anymore; that stuff is way too technical for what you usually need to do anyway.
Then the next summer, that's when I did something quite the opposite. I learned a ton about deep learning and machine learning—really high-leverage stuff as well. I wasn't just focused on individual frameworks like TensorFlow or Keras or working in Python in particular, but it was high-leverage skills like learning how models work and how to actually work with the AI within any language that I wanted.
It was super valuable stuff, and so that's a good example of that stark contrast between focusing so much on a tool—which sometimes you need to, but I did it way too much—and then focusing on the capabilities.
So learn from my mistake. It's super important to focus on what really is going to get you far. I hope that story gives you a clear example of what I mean by capabilities, not tools.
I'm going to be referencing this model constantly throughout what I'm talking about here because it really does apply to all the skills that I'm going to speak to and why they're actually so important.
So with that, we can move on to the next one here, which is AI agents. This is going to be the focus of 2025. All the big companies—Microsoft, LangChain, Nvidia, Anthropic, OpenAI—they're all saying that AI agents are the way of the future for AI, and 2025 is going to be a massive year for them.
The biggest reason for that is a lot of the innovation of 2024 is really laying the groundwork for AI agents going into 2025. Things like establishing best practices in architecture, which I'll get into in a little bit, and creating LLMs that are actually powerful enough to reason well within complex agentic workflows—all that is there now, ready for us to utilize.
And remember, we want to focus on capabilities and not tools. What that means for agents is for you and I to dive into agent architecture best practices, when to build yourself versus use a platform that does a lot for you, like OpenAI Assistants, for example.
Notice how all of these skills are very, very high leverage and agnostic to the framework or tool that you end up using to build your agent. That's important because all these tools are good to learn, but don't make it a priority. Don't spend weeks and weeks learning Crew AI or Pantic or n8n, even though I love all these platforms, because that is going to get you caught in the weeds.
Focus on the high-leverage stuff. One of the most important things is learning about agent architecture. Anthropic actually released this amazing article recently on all the different architectures and best practices coming up for AI agents.
So let's go ahead and check that out now because that is a super important skill for us to learn going into this next year.
On December 19th of last year, Anthropic put out this article on building effective agents. Anthropic is an incredible company with the amount of research and development they're doing within AI, even more than most other companies like OpenAI, at least in my eyes. They're very, very impressive.
So when they put out an article like this talking about something that they're claiming to be very important in the AI space, we listen. This applies to you, really no matter who you are—if you're super technical or just a non-technical business owner, all of these things apply.
It's important for us to know what makes an AI agent good as we're developing them and evaluating tools that we might want to incorporate for ourselves and our businesses.
What Anthropic is talking about here applies a lot to what I'm speaking to as well because they talk about what tools you might want to use to build your agents, but they don't focus on it a lot. They move pretty quickly into the higher-leverage stuff, specifically the architecture options that you have for building powerful agents.
So it's kind of like, "Yeah, hey, here are some tools that you can use to implement these things, but don't focus on it a ton." Instead, the rest of this article, that's many, many paragraphs, is talking about all the different architecture options.
At the forefront here, we have this architecture diagram that really defines what an AI agent is at the core, like with a knowledge base, with retrieval tools to interact with your services like Slack and Gmail, and then chat memory as well.
I'm not going to dive into all these specific architecture diagrams right now, but they have a ton of really cool things like prompt chaining and routing, which is awesome for working with different LLMs, parallelization to conquer tasks with multiple LLMs at the same time.
They've got orchestration workers, so having like a master agent that dishes out work to different LLMs with kind of an aggregator at the end here, an evaluator, which is super important for reducing hallucinations, and then yeah, so many other architectures as well.
Again, I'm not going to dive into this in a ton of detail, but I'll have a link to this in the description. I would highly recommend following along with this, reading through it, and really just starting to engage yourself with the whole idea of learning about how to build solid architecture and implement best practices for your agents going into 2025.
Since I was just talking about AI agents, it felt very fitting to call out that my platform, Automator, is hosting an AI agent hackathon competition in partnership with Voiceflow and n8n, and the prize pool is $5,000.
The only criteria is you just have to build an AI agent for the Live Agent Studio, which is my community-driven, open-source-first AI agent platform that I developed recently. This competition is just a way to get us all engaged on that platform, and it's also my Christmas gift to you. I announced this on Christmas—super, super exciting!
So I'll have a registration link in the description and pinned comment. Don't miss it! Definitely register and be a part of this, build an agent, win some money, and showcase your AI mastery to the world.
If AI agents are the most important thing for AI to invest your time into going into the next year, I would say the second most important is reasoning LLMs and learning how to work with them very well.
That also ties a lot to AI agents, which I'll speak to in a little bit. But all a reasoning LLM is, kind of as the name suggests, is a large language model that is able to reason with itself about your prompt before giving you the final answer.
You'll also hear "chain of thought," "inference," "time compute"—all of this means the same thing. It's really cool that actually existing large language models are being used to create these systems that turn into these reasoning LLMs, which is kind of why I'm saying it's the ticket to AGI, even though it's kind of cheesy.
We're taking these existing LLMs and we're making them 10x, 100x more powerful just through this process. Like 03 by OpenAI is a really good example. You're probably living under a rock if you haven't heard of 03.
Then we have 01, also from OpenAI, QWQ, an open-source reasoning model from Quen. We've got Gemini 2.0, Flash Thinking—these are really starting to blow up, and they're super powerful because they solve two of the big problems with generative AI right now.
Maybe not 100% solved, but we're getting there, and that is the issues of hallucinations and LLMs just making bad decisions as agents.
So really, the future is going to be agentic workflows that combine reasoning LLMs with smaller LLMs that can actually produce output quickly to create strong and fast systems.
There are going to be parts of our processes that can be slow but need to be more powerful, and there are also going to be parts that need to be fast and maybe aren't as complex.
So combining these reasoning LLMs with smaller, fast LLMs is really the future that I see. Then add in the whole idea of domain-specific LLMs, which I think is going to be another big thing going into 2025.
You're going to have all these different LLMs to bake together in these agent architectures that I just finished covering, and that's really what's going to drive innovation forward going into 2025.
Lastly, what this means for you from a practical standpoint, what skill to actually invest in, is to learn how to prompt and work with these reasoning LLMs because they are definitely different than normal large language models with how they behave because of this whole idea of the chain of thought that it goes through before it gives you a final response.
Learning how to prompt them efficiently, use them as agents, and also when and how to bake them into your agent architectures is going to be crucial for you to build powerful agents.
Next up, we have the entire arena of local large language models, and this is entirely a skill of its own because there is so much that goes into understanding hardware requirements for local LLMs when you run them yourself, when and how to get them, the understanding of fine-tuning, and how to take these LLMs and have them work on your dataset.
When to use them versus cloud large language models—there's so much that goes into it. One really important thing to keep in mind here is that closed LLMs, like Claude GPT, the ones that you can't download and run yourself, are still better than the local ones that you can, like Quen and Llama.
But going into 2025, this gap is diminishing very, very quickly, especially with new models like the new Deep Seek V3. It's just incredible what we can do with local large language models, and they have a ton of advantages as well, like fine-tuning that I already mentioned—being able to take a large language model and train it on your own data so it can perform better on your use case.
You also have utmost data privacy because the LLM is run on your own infrastructure. You have better cost, depending on the use case, but a lot of times you do because you're not paying for the API cost of using these large language models. You just have to pay for the cost of your own hardware or hosting fees.
Then, believe it or not, speed can often be another advantage, assuming you have good enough hardware because you're not dealing with the network delays of calling these LLMs across the internet with an API.
Last but not least is flexibility. There's just way more options for local LLMs available that you can download yourself and run yourself from like Hugging Face or Olama—way more than the number of closed LLMs that are available.
There are so many fascinating local LLMs out there for domain-specific things like creative writing and coding, so just so many cool things to play with that are becoming more and more practical. It's something that I focus on a lot on my channel, local AI, and I'll continue to as well.
All right, so we've talked about AI agents, reasoning LLMs, and local LLMs, but how do we actually leverage these things to create a system to solve a real problem for us?
Because we're going to have to have a database, and we're going to have to host it and serve our large language model. How do we do all of that in a system? That leads us to the next skill here, which is finding and getting a good grasp on your AI tech stack.
An AI tech stack is really just your collection of tools and services that you're combining together to create the system for yourself or your business. Like I just alluded to, we are getting into tools now instead of capabilities, but like I said at the start, even though you want to prioritize capabilities, you still do need to focus on tools to some extent because in the end, that's what's going to enable you to build out the capabilities that you are mastering.
There are a ton of resources online that look kind of like this right here, where you have the different categories for your tech stack, and they give you all the options that you can have, or at least a good chunk of options that you have for each category, like your LLM, your hosting, your database—all that good stuff.
This gives you the starting point. I'll actually show one of these in a little bit here. It gives you the starting point to then research all these different services and find the ones that work for you and your tech stack.
The important thing is to find what works for you and your business and just stick with it. Don't spend a month diving into the weeds of all these services and getting good at them like I did with my C++ boot camp.
The important thing here is to not overthink it. There's an acronym in the software engineering industry that is KISS, and that's just short for "keep it simple, stupid."
The other important thing to keep in mind is to engineer for simplicity and reuse. Find the technologies for your tech stack that you're going to be able to use the most for the different use cases you have for AI for yourself or your business, so you don't have to repeat yourself as you're building different architectures.
The acronym for that, also from the software engineering industry, is DRY, which stands for "don't repeat yourself."
So just keep in mind, as you're building out your AI tech stack, you want a DRY KISS—don't repeat yourself and keep it simple, stupid.
Really, the biggest decision as you're building out your tech stack, the most important one, is whether or not you want to go local with your large language model—hosting that yourself, hosting your database yourself, and all your infrastructure—versus going to the cloud and using those services for hosting and LLMs and all of that.
But with that decision out of the way, it really is pretty simple. You just go through all of these different options here, do a little bit of research, and find the services that you want for each category.
So with that, I'll actually take you to an example, and we'll walk through one of these AI tech stack diagrams and how you might choose the services for you.
All right, so here we got a good example of what one of these AI tech stack resources looks like that you can use to help you figure out what you want for your tech stack specifically.
This is open source; it's just a GitHub repository where people, through pull requests, can actually contribute additions to this tech stack diagram here. It looks like a GitHub repo at its core, but when you scroll down, you have this massive image in the readme that gives you all these options that you have for each layer of your tech stack.
Let me full screen this really quick for you here. Just scroll down to the bottom.
You basically just start from the base, pick a couple of these options here, and then just do a little bit of research. Maybe build a quick POC to test different things and see what works best for you.
So you just pick your cloud provider—like maybe you already use AWS for your business, so it makes sense to just stick with that. Then you go up to your models; maybe you'll try a few models with specific prompts for agents you're trying to build, or again, you'll build a sample proof of concept agent and try with these different models to see which performs the best.
Then you'll go on to maybe knowledge engines. If you have RAG as a part of your agent, you'll just try a couple of these and see which agent, using one of these knowledge engines, retrieves information correctly from your knowledge base the most.
So you get the idea. You just try a few, figure out what works best for you, and don't overthink it. You don't have to test every single one of these or have a full agent built out to test them. Just keep it simple.
I don't necessarily agree with everything that's in here, and I think there's a lot that's missing, so just keep in mind that this is just an example of one of these many resources. I'll link it in the description below. Just don't take this one specifically as your end-all, be-all resource to go find your AI tech stack. It's just a good example to work off of.
If you are curious what my AI tech stack looks like, I'm sharing it with you right now. Again, I'm not super focused on tools; I want to be focused on capabilities, so a lot of these are going to change.
That is why I'm sharing resources with you to help you decide instead of just giving you what I do only and telling you to do that because a lot of this is going to change. But this is what I'm planning on focusing on using at the start of the year, at least, and a lot of this will probably last for a while as well because there are definitely some tools, like LangGraph, that I definitely don't think are going anywhere.
So anyway, for large language models, this is probably the most volatile, but right now I'm using a lot of the new Deep Seek V3. I've tested it; it's performing fantastic.
Then also Quen and Llama for models that I run on my computer, and then Gemini 2.0, Flash, and Claude are the really awesome, fast, and cheap cloud models.
I use a lot of 01 as well for the more advanced reasoning stuff, especially when Windsurf and Bolt get stuck on little coding problems, then I'll throw them at 01.
Languages: Python and JavaScript almost exclusively. I think Go would be cool to learn as well.
Then for AI coding assistance: Bolt, DIYne, and then yeah, Windsurf and Cursor—these are kind of my go-tos.
For AI agent frameworks: Pantic AI and LangGraph, which is a fantastic combo, by the way. More content coming on that combo soon.
I like using Flowise for prototyping as well.
For my database, I basically exclusively use Supabase for the Postgres under the hood because it also has PG Vector for RAG, so I don't have to go and use another service like Pinecone, which there are times to use that as well, but I found this works very well for me.
Then Voiceflow and n8n for automations—they're both sponsoring the hackathon, which I talked about earlier for the Live Agent Studio, so super cool.
I use Docker a lot to containerize my applications for a consistent environment.
For hosting in the cloud, I use Digital Ocean for general purpose and RunPod when I have machines in the cloud running large language models.
For my development and testing, I'm using Playwright, Pytest, Pantic AI, and then CodoCover. I haven't used it yet, but it's this new tool that generates entire tests for your AI agent using AI, which is super cool, so I might be looking into that.
Yeah, and then for CI/CD, GitHub Actions—just nice and simple for all my repository testing.
For LLM evaluation, this is something that I'm going to get into a lot more this year. I haven't done a lot with it yet, but creating custom agents to evaluate other agents using Bolt, DIY.
So like when I have a large language model, I use it as an AI coding assistant with Bolt, DIY, and just see how well it can code these templates that I have.
Then Regos and Phoenix are a couple of up-and-coming tools as well. Phoenix specifically for evaluating agent tool calling, and then Regos is meant for evaluating RAG pipelines—very awesome tools.
Finally, for search, like when I want to have an agent that can search the web, I've been using Brave a lot, but also Firecall is something I'm going to be making a video on soon.
Then you've always got the classic Perplexity, and then Search One API is another one that I'm planning on looking into pretty soon here.
So yeah, that is my AI tech stack as a whole. Again, it might change completely, but I just thought it'd be worth sharing with you what I'm going to be using for my tools going into this year.
Hopefully, that gives you a nice and concise overview of finding your AI tech stack with some solid recommendations.
So with that, the next high-leverage skill to invest your time into is working with large language models. The most obvious thing being prompt engineering—learning techniques like single-shot, multi-shot, chain of thought prompting.
There are a lot of resources online for learning these things, but it's important to do so because the way that you work with all the different LLMs is very, very similar.
There are some subtle differences in the way that you want to prompt each one, but learning how to prompt specifically for one LLM is pretty low leverage. Learning the overarching stuff and these techniques is very high leverage and super important for anything that you're going to do with AI.
Also, learning how to work with AI coding assistance because it's definitely the future of AI, where even non-technical people can leverage AI coding assistants like Bolt, DIY, and AI IDEs like Windsurf and Cursor to make a ton of incredible things.
So it's definitely worth investing time into learning how to have AI as a PA programmer or even building the application for the most part for you.
Along the lines of working with LLMs is the whole idea of "human in the loop." This is going to be very big in 2025 because as AI agents are enabled to do more and more, we still need the human to approve certain actions that the agent might take and things like that.
That is what human in the loop allows. A lot of really awesome tools like LangGraph make this very possible to do as well because it can get quite complicated to have the agent do a ton of things itself and then somehow come back to the user for specific actions.
So it's worth learning that—not trivial.
The last thing is taking advantage of massive context windows. We're already seeing models like Gemini where you can put millions of words into the prompt already, and going into 2025, there are just going to be many more models where that is possible.
That's a whole new paradigm for how we work with LLMs, and we can just give it that much information in a prompt. So it's super important to spend some time learning how you can actually leverage these things as they become more widely available.
At this point, I have covered a lot, to say the least. So now it is time to tie it all together, and I also want to speak to how I'm going to be combining everything that we talked about in a very cohesive way in my channel for this year to provide you an insane amount of value.
So yeah, let us tie it all together. Imagine having a fully open-source AI enablement stack that gives you your entire tech stack for AI—all the tools and services that we just talked about, pre-built agents that you can pick and choose from based on your use cases for your business, an AI coding assistant to help you make all your applications, the ability to choose a local versus cloud setup, and last but not least, extensive documentation to help you get started very easily implementing this for your systems and processes and helping you implement all of the best practices.
Well, this is my dream and what I'm going to be working on with all my heart and soul this year for you. It really is what I figure to be the best way to provide an insane amount of value for you because having this open-source AI enablement stack is going to make it so easy for you to get started with AI.
I'm super excited for this. So much of my content is going to be going into this over the year because it also covers all the important topics that I've already spoken to in this video that are going to be the big things for 2025.
Last but not least, it's not a skill, but it's equally important—that is finding an AI community to learn and grow with other people. You do not want to go through the wild ride of AI alone.
Also, the cheesy saying "your network is your net worth" definitely applies to AI as well, so find a community. There are a lot of fantastic ones out there. One that I can sincerely recommend because I see to it myself that it is awesome is the Automator Think Tank.
Automator is my platform, and the Think Tank is the community that we have been building up to be a central hub for all the knowledge, networking, resources, and tools you need to succeed with AI.
It's also the home of the Bolt, DIY community, which we are building the best open-source AI coding assistant over there. So a lot of really exciting stuff going on there.
This is just a nice way to kind of wrap up all these skills that I talked to because having a community growing with other people is the best way to make sure that you're actually mastering these high-leverage capabilities.
So there you have it. Together, we have covered the most important high-leverage skills to focus on with AI going into 2025.
Like I said at the start of this video, I really want this to feel like a conversation with you as much as I can, so please let me know in the comments if you think I missed anything big or just if you have any thoughts at all around AI agents, reasoning LLMs, local AI, or anything that I covered with you just now.
Also, participate in the Live Agent Studio hackathon competition. It's a fantastic way for you to win some prize money and showcase your AI mastery to the world, and it's just an awesome community to be a part of as we are doing some incredible things.
If you appreciated this content and you're looking forward to everything that I've got this year around all these topics, I would appreciate a like and a subscribe.
With that, I will see you in the next video.