How I’d Learn AI From Scratch in 2026 (skip the useless… — Transcript

If I had to learn AI from scratch today, I'd be overwhelmed because so much of what's out there is either outdated and no longer relevant or just theory you'll never actually use. So, in this video, we're focusing on the 20% of AI that's practical and that will still matter a decade from now. Organized in three levels that build on top of each other.

Let's get started. Kicking things off with step one, pick one model and go deep. There are two reasons for this.

First, this artificial analysis chart shows that although models used to be far apart in capability, they're now all so powerful, as represented by the clustering in the top right, that the difference for the average user is negligible. Second, because all the top AI companies are copying each other, all the models have the same core features. Meaning, once you go deep on one, that skill carries straight over to the rest.

So, what are your options? Well, just like picking a starter Pokémon, there really is no wrong choice, as long as you pick Charmander. Just kidding.

Before I get wrecked in the comments, a few clarifications. First, I've got nothing against Elon, but objectively speaking, xAI simply is not competitive anymore. Second, Perplexity does not have their own frontier model. They're primarily a search tool that fine-tunes other companies' models. And third, [snorts] as much as I appreciate the open-source Chinese models, they're still behind their western counterparts.

Which leaves three choices: ChatGPT, Claude, and Google Gemini. But which one is the right one for you? This brings us to step two. Pick the model based on three simple principles. First, prioritize paid tiers. So, if you're on free ChatGPT, but your job gives you paid Gemini, go deep on Gemini, because the gap between free and paid is like night and day. Second, if you have a choice, pick the AI that matches your work, since each one has its own strengths. ChatGPT is the most mature, has the most users, and so it's got the most tutorials to learn from. I've also found it's the best at web search out of all the chatbots, and therefore it's great for research. Claude is really good at writing and design, and excels in coding, which matters even if you're not technical because stuff like data analysis and pretty diagrams are all built on code under the hood. And Gemini is the best if you work across text, images, audio, and video, since it's the only AI that can process mixed media natively. And it's also the best pick if you're a heavy Google Workspace user.

Third principle, vibes. I know it sounds dumb, but each AI has their own personality, and the more you enjoy using it, the more you'll use it and the better you'll get at it. Pro tip, switching is pretty easy since all three have some sort of a memory import feature. For example, in Gemini under settings, just open import memory to Gemini and follow the on-screen instructions. And that brings us to step three, change your defaults. Put simply, the companies default you to the weakest model because this is the cheapest for them to run. So, you always want to select the more powerful models you have access to for real work since these will actually break down your requests, map out the steps, and catch nuances you didn't think to mention. Here's a quick recap of level one. First, use the paid version if you can. Second, choose the chatbot that matches your work and stick with it. And third, always default to the most capable model you have access to. Now, you might have noticed I didn't mention prompting at all, and that's on purpose because with how powerful these models have become, your prompt is no longer the biggest factor in determining output quality. By the way, I'm kind of proud of this. I actually partnered with today's sponsor, HubSpot, in putting together a Google Gemini cheat sheet for all of you with a list of my top productivity tips and techniques for Gemini. Here are two of my favorites.

First, in Google Docs, you can add an AI summary block at the top of any working document by typing @aisummary, and now anyone on the team can click refresh to get the current gist without rereading everything. Second, within Google Sheets, you can type equal AI and start giving instructions in plain English. For example, "Hey, append the video number and topic to the front of the video title." And there you go. No formulas needed, right? Saves so much time. The cheat sheet is completely free, link in the description. Thank you, HubSpot, for sponsoring this video.

So, what's more important than our prompt? Let me illustrate with a simple example. Say you need to find a restaurant for your boss. You could either spend 10 minutes describing everything you think your boss likes in a long and detailed prompt, or you could hand the AI a list of restaurants your boss has loved in the past and let it figure out the pattern. The second approach wins every single time. Put simply, that list of restaurants is an example of the right context. And the right context will always beat the perfect prompt. In practice, this means we don't need to memorize a bunch of prompting frameworks anymore, because models have gotten so powerful that they're able to accurately infer inputs like the role, the format, and the tone, as long as we give them two things: a clear outcome and the right context. So, there's really just one framework worth remembering: OC, outcome plus context.

Let's see this in action. Say you want a new workout plan. Option A, you could write a long detailed prompt spelling out your experience level, your equipment, your schedule, your goals, and the format you want it in. Or option B, copy an article breaking down the push-pull-legs routine, paste it into AI as context, and write a simple prompt like, "My goal is to develop a 4-day training routine for muscle growth, 45 minutes a day. Tailor a plan for me based on this." Option B's output will always be better because instead of you saying something like, "Role: act as a senior personal trainer," the AI infers a much sharper role from the context, like a specialist in compound training, which is something you would not have thought of yourself. I would have because I'm just, you know, built different. Just kidding. Anyways, um now that we know context is king, here are the top three ways to find the right context. First, explicitly name a relevant framework. For example, instead of typing a paragraph explaining how to restructure a report, I tell the AI to rewrite this using the pyramid principle. Since those two words, pyramid principle, carry more context than a paragraph of me describing what the pyramid principle is. Pro tip, you can ask the AI for relevant frameworks.

For example, what are the best frameworks for setting goals? Then pick the one that fits and use it in your next prompt. Second, the best type of context to give AI is real examples of what good looks like. Since examples contain everything you forget to explicitly say in a prompt, like your manager's expectations and your team's preferences. For example, instead of spelling out the format, length, and tone for a weekly status update, just paste in the last two to three updates that were approved. Add raw notes for the week and say, write this week's update in the same format, and the AI will pull everything it needs from those examples. Third, connect your tools.

Simply, your best context already lives somewhere, in your email, your Google Drive, Slack, Notion, what have you, right? So, by connecting to those platforms, your AI can just pull what it needs directly without downloading and re-uploading. In practice, this means instead of manually attaching transcripts from multiple meetings, we point the AI to the tool and ask it to pull the required files and surface insights. Now that we know what great context looks like, how can we save relevant context for recurring work so we're not repeating ourselves every time? The answer in ChatGPT and C

images, audio, and video, since it's the only AI that can process mixed media natively. And it's also the best pick if you're a heavy Google Workspace user.

Third principle, vibes. I know it sounds dumb, but each AI has their own personality, and the more you enjoy using it, the more you'll use it and the better you'll get at it. Pro tip, switching is pretty easy since all three

have some sort of a memory import feature. For example, in Gemini under settings, just open import memory to Gemini and follow the on-screen instructions. And that brings us to step three, change your defaults. Put simply, the companies default you to the weakest

model because this is the cheapest for them to run. So, you always want to select the more powerful models you have access to for real work since these will actually break down your requests, map out the steps, and catch nuances you

didn't think to mention. Here's a quick recap of level one. First, use the paid version if you can. Second, choose the chatbot that matches your work and stick with it. And third, always default to the most capable model you have access

to. Now, you might have noticed I didn't mention prompting at all, and that's on purpose because with how powerful these models have become, your prompt is no longer the biggest factor in determining output quality. By the way, I'm kind of

proud of this. I actually partnered with today's sponsor, HubSpot, in putting together a Google Gemini cheat sheet for all of you with a list of my top productivity tips and techniques for Gemini. Here are two of my favorites.

First, in Google Docs, you can add an AI summary block at the top of any working document by typing @aisummary, and now anyone on the team can click refresh to get the current gist without rereading everything. Second, within Google

Sheets, you can type equal AI and start giving instructions in plain English. For example, "Hey, append the video number and topic to the front of the video title." And there you go. No formulas needed, right? Saves so much

time. The cheat sheet is completely free, link in the description. Thank you, HubSpot, for sponsoring this video.

So, what's more important than our prompt? Let me illustrate with a simple example. Say you need to find a restaurant for your boss. You could either spend 10 minutes describing everything you think your boss likes in a long and detailed prompt, or you could

hand the AI a list of restaurants your boss has loved in the past and let it figure out the pattern. The second approach wins every single time. Put simply, that list of restaurants is an example of the right context. And the

right context will always beat the perfect prompt. In practice, this means we don't need to memorize a bunch of prompting frameworks anymore, because models have gotten so powerful that they're able to accurately infer inputs like the role, the format, and the tone,

as long as we give them two things: a clear outcome and the right context. So, there's really just one framework worth remembering: OC, outcome plus context.

Let's see this in action. Say you want a new workout plan. Option A, you could write a long detailed prompt spelling out your experience level, your equipment, your schedule, your goals, and the format you want it in. Or option

B, copy an article breaking down the push-pull-legs routine, paste it into AI as context, and write a simple prompt like, "My goal is to develop a 4-day training routine for muscle growth, 45 minutes a day. Tailor a plan for me

based on this." Option B's output will always be better because instead of you saying something like, "Role: act as a senior personal trainer," the AI infers a much sharper role from the context, like a specialist in compound training,

which is something you would not have thought of yourself. I would have because I'm just, you know, built different. Just kidding. Anyways, um now that we know context is king, here are the top three ways to find the right

context. First, explicitly name a relevant framework. For example, instead of typing a paragraph explaining how to restructure a report, I tell the AI to rewrite this using the pyramid principle. Since those two words, pyramid principle, carry more context

than a paragraph of me describing what the pyramid principle is. Pro tip, you can ask the AI for relevant frameworks.

For example, what are the best frameworks for setting goals? Then pick the one that fits and use it in your next prompt. Second, the best type of context to give AI is real examples of what good looks like. Since examples

contain everything you forget to explicitly say in a prompt, like your manager's expectations and your team's preferences. For example, instead of spelling out the format, length, and tone for a weekly status update, just paste in the last two to three updates

that were approved. Add a raw notes for the week and say, write this week's update in the same format, and the AI will pull everything it needs from those examples. Third, connect your tools.

Simply, your best context already lives somewhere, in your email, your Google Drive, Slack, Notion, what have you, right? So, by connecting to those platforms, your AI can just pull what it needs directly without downloading and re-uploading. In practice, this means

instead of manually attaching transcripts from multiple meetings, we point the AI to the tool and ask it to pull the required files and surface insights. Now that we know what great context looks like, how can we save relevant context for recurring work so

we're not repeating ourselves every time? The answer in ChatGPT and Claude is a feature called projects. In Gemini, they call it Gemini Gems, but they work basically the same way. And a project is a permanent home for recurring work

streams and contains three components. First, the project instructions are the rules that always apply, like your goals and constraints for this entire work stream. Second, knowledge files are the reference information that the AI pulls from, like your source documents,

examples, and frameworks. Third, the memory. This is updated by the AI automatically and keeps track of key updates and milestones within this workstream. Building on top of our previous example, here I have a workout project in Claude that acts as my coach.

And within the project instructions, we see the goal of the project and my current setup, which is really constraints. 45 minutes a session, home only, right? And a reference to the knowledge file, which is the push pull leg program from earlier. And since

every single conversation has access to this knowledge file, I can say something like, "I only have 25 minutes today for a pull day. What do I do?" And the AI recommends a workout based on our program. And right now, the memory is

empty because this is just an example. But let's say I injure my shoulder pressing, I don't know, 200 kg dumbbells. True story.

Uh the AI would actually make a note of that and steer me away from shoulder-related exercises until I heal.

Pro tip, you always want to use something called .md markdown files instead of PDFs whenever possible because markdown files are easier for the AI to read and cheaper to process.

And you can easily ask the AI to convert PDFs into markdown. Now, as good as projects are for recurring work, each project is a standalone silo. Meaning, my workout coach project can't see what's in my annual health reports

that's stored in another project, right? Even though it totally makes sense for them to talk to each other. And this is where an AI system comes in. So, what's an AI system? Put simply, it's a setup that does two things projects cannot do.

First, it can pull context from different projects, spot patterns across them, and surface insights individual projects would have missed. For example, here I'm told that, "Hey, I took a look at your travel project and your personal finances project. And guess what? You

don't have enough money to go on your Bali trip. Second, an AI system updates itself after you give it feedback, meaning learnings compound over time.

For example, when I say, "Reconcile my final draft with your initial output," the AI system analyzes all the changes I made and proposes rules to remember for next time. I'm going to oversimplify quite a bit here, but broadly speaking,

you have three options when it comes to AI systems. First, Gemini Spark from Google is the most beginner-friendly and requires minimal setup because it's already connected to tools like Gmail, Google Calendar, and Google Drive. But the trade-off is we have less control

over how it's configured. Let me know if you want a full tutorial on Gemini Spark. Next up, Claude Cowork is specifically designed for non-technical people like myself. It gives us more control, but it does require some sort of setup, as you can see. Uh if you're

interested in setting this up for yourself, I'll leave a link to my free Cowork Toolkit below. Finally, Anthropic's Claude Code and OpenAI's Codex are basically Claude Cowork on steroids. These are fully customizable, extremely powerful, but you'll need to

be somewhat comfortable with code. This is a bit silly, but each tool's model selector is a pretty good signal for who it's for. For example, in Codex, you have all these options, right, that are perfect for power users, but

intimidating for everyone else. Cowork kind of dumbs it down a little bit by giving us fewer options, but you know, still some control.

And Gemini Spark doesn't let us pick the model at all, although I think they'll change that soon. All right, now let's go through some real-life examples of what these AI systems can do.

Previously, I had three separate projects tracking my annual health checkups, the supplements I take, and my workout plan. But once I migrated them into Cowork, the AI is able to cross-reference my latest checkup with my workout plan, flags the fact that I

don't have any cardio days, although I have borderline high cholesterol. And so it recommends adding cardio sessions to my rest days and notes that since I'm already taking fish oil, there's nothing it would change in my supplement stack.

To be clear, no single project could have connected those dots unless I decided to like copy and paste information between the two, right? But the AI system sees them all. Example two is the reconcile move I mentioned earlier and applies to any writing that

you do. Here, the AI gave me a rough draft of a segment of my YouTube script.

Then I made a bunch of edits and I shared it back under the final version, right? To make it sound more like me.

Then I told it to reconcile my version with the initial draft from the AI and the AI is able to dissect every change, as you can see down here, and it remembers all the rules for the next segment. And so the more I

give feedback, the smarter the system gets and the less instructions I need to give moving forward. All right, here's a quick recap of everything we covered today. Level one is just getting started with one of the big three chatbots,

default to the most powerful model and go deep since the skills carry over to the rest. Level two is learning what kind of context actually helps the AI, then saving it in projects so you can stop repeating yourself for recurring

work. And level three is connecting those projects into one system that compounds the more you use it. Most people aren't here yet and that's okay.

There's no rush. The gap between using AI and using AI well has always been invisible, but hopefully this video has helped you figure out where you are right now and what to focus on next. If you enjoyed this, check out my Claude

Co-work playlist next. See you over there and in the meantime, have a great one.