A one-week curriculum that starts at the 1950s perceptron and ends at selling AI work, with the full reading list attached.
Posted
today
Duration
Format
Essay
educational
Views
16.1K
597 likes
57 · 43
Big Idea
The argument in one line.
Most people learn AI tools without learning AI, and one focused week on history, model mechanics, the hardware stack, the labs and the economics makes every later tool and opportunity decision obvious.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You've used ChatGPT or Claude for a few months and can feel that you're pattern-matching prompts without understanding what the model is doing underneath.
You're an operator or consultant who has to judge whether a new model release is a real shift or marketing noise, and you currently can't tell.
You want a sequenced reading and watching list instead of another scattered feed of AI takes, and you have five to seven evenings to spend on it.
You're trying to sell AI or automation services and keep pitching ideas that sound impressive but don't map onto anything a business already pays for.
You keep hearing terms like pre-training, context window, chain of thought and recursive self-improvement and nod along without a working definition.
SKIP IF…
You already know what a Transformer is, why sampling temperature exists, and who ASML is. This is foundations, not frontier research.
You want hands-on build instructions. This is a curriculum and a reading list, not a tutorial you follow along with in an editor.
You're looking for specific tool rankings or benchmark numbers. The video deliberately avoids them so it doesn't date.
TL;DR
The full version, fast.
Learning AI tools without learning AI leaves you unable to judge what matters, so the fix is a sequenced week of foundations. The plan runs seven stages: the history from Rosenblatt's perceptron through the AI winter, AlexNet and the 2017 Transformer; the eight ideas that explain model behavior, including tokens, next-token sampling, pre-training, human feedback, chain of thought, context windows, hallucination and the fact that models are grown rather than built; the physical stack of chips, fabs, data centers, energy, labs, APIs and harnesses; the differing philosophies of OpenAI, Anthropic, Google DeepMind, Meta and xAI; and the race toward recursive self-improvement. It closes with the practical layer: pick one assistant, one harness and one automation tool, delegate rather than prompt by handing over context, goal, what great looks like and how to verify, and remember that AI mostly made old profitable processes cheaper rather than making new things possible.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Seven years of operating AI businesses sets up the promise: a week of structured study that beats years of scattered tool use.
01:20 – 05:42
02 · AI history, 1950s to now
Perceptron to AI winter to expert systems to AlexNet to the 2017 Transformer, with the argument that complicated terms become simple once you know what problem they solved.
05:42 – 12:49
03 · How LLMs actually work
The eight ideas that explain model behavior: tokens, next-token sampling, pre-training, human feedback and sycophancy, chain of thought, context windows, hallucination, and grown-not-built.
12:49 – 18:46
04 · The AI stack
The layers underneath a chat box: chips, fabs and the machines that build fabs, data centers, energy and permitting, model labs, APIs, harnesses, and finally implementation.
18:46 – 23:49
05 · The major AI labs
How OpenAI, Anthropic, Google DeepMind, Meta and xAI differ in philosophy, lineage and bet, and why those differences show up in their products.
23:49 – 26:42
06 · Recursive self-improvement
What RSI means in plain terms, why compute and energy become the only bottlenecks, and the optimist, cautious and skeptic readings to balance against each other.
26:42 – 28:51
07 · The tools to learn
Pick one chat assistant, one harness and one automation tool, on the grounds that the leading options are close enough that commitment beats comparison.
28:51 – 31:09
08 · How to prompt
Delegate instead of prompting. Hand over context, the goal, what great looks like, and how the work will be verified.
31:09 – 34:55
09 · Turning AI into money
AI made old processes more profitable rather than making new things possible, so the test for any idea is whether the underlying process already made money.
34:55 – 36:15
10 · AI opportunities are coming
Opportunities are getting larger and more frequent while the window to act shrinks, which is the case for doing the foundational week now.
Atomic Insights
Lines worth screenshotting.
A token is a fragment of a word, not a word, which is why models behave strangely on spelling, counting and unusual names.
Models do not pick the single most likely next token. They sample from a distribution, and that deliberate jitter is what keeps output from being perfectly predictable.
The 1950s neural network architecture did not fail on theory. It failed on compute, and AlexNet in 2012 proved the hardware had finally caught up.
Expert systems were just complicated if-then trees. What we now call a normal computer program was called artificial intelligence for about thirty years.
Model sycophancy is an artifact of training: human raters consistently preferred the answer that flattered them, so that behavior got reinforced.
Hallucination follows directly from learning language statistics instead of facts, because the text models learned from was never required to be true.
Researchers do not build these models, they grow them, which is why no one can point at the code that performs reasoning.
NVIDIA designs the chips, TSMC manufactures them, and ASML in the Netherlands makes the machines that make them, so the real bottleneck sits three layers below the model labs.
Anthropic made a single bet that programming is the shortest path to AGI, while OpenAI spread across video, images, audio and consumer products.
xAI could not win on model talent, so it bet on compute, and now rents Colossus 2 capacity in Memphis to labs that cannot build their own.
Google DeepMind owns the full stack, from chips to cloud to twenty years of search data, and its only real handicap has been bureaucratic inertia.
Once AI can run its own research, the only remaining bottlenecks are compute and energy, because you can spawn a hundred thousand researchers in parallel.
A year of recursive self-improvement could compress ten to twenty years of normal progress, which changes how fast an opportunity window closes.
Delegating beats prompting: hand over context, the goal, what great looks like, and how the result gets verified.
Most people skip the verification step, which is the one that decides whether the output is usable.
AI did not make new things possible. It made old processes profitable enough that they look new.
The test for any AI business idea is whether the underlying process was profitable before AI. If it wasn't, it probably still isn't.
Businesses pay for repeatable procedural pipelines that turn a raw input into a finished output, so find the pipeline before you find the model.
First customers come from unscalable manual work, and only then do you figure out which part of it an AI can compress.
Opportunities are getting bigger and more frequent, but the window to act on each one is getting shorter.
Takeaway
Learn the stack before you learn the tool
WHAT TO LEARN
Understanding where models came from, what they physically run on, and which processes already made money is what turns AI from a set of tricks into judgment.
02AI history, 1950s to now
Terms like weights and activation functions stop being intimidating once you know they were named in the 1950s to describe a machine that sums weighted inputs and fires.
The first AI winter was a compute failure, not a theory failure, which is why a 1950s architecture suddenly worked in 2012 when AlexNet won on image classification.
What was called artificial intelligence for thirty years, the expert system, is what we now just call a program, so a rules engine is not a model.
03How LLMs actually work
A token is a fragment of a word rather than a word, which explains why models stumble on spelling, counting and unfamiliar names.
A model samples from a distribution of likely next tokens instead of always picking the top one, so some unpredictability is designed in rather than broken.
Pre-training compresses enormous amounts of text into weights, and the later human-feedback stage rewards whatever raters liked best, which is where sycophancy comes from.
Hallucination follows from learning the statistics of language rather than a record of facts, because the text a model trained on was never required to be true.
04The AI stack
The stack runs chips, data centers, energy, labs, APIs, harnesses and implementation, and a newcomer starts at the top layer with no view of the six below.
NVIDIA designs chips, TSMC manufactures them, and ASML builds the machines TSMC needs, so a handful of suppliers constrain the entire industry.
Data centers depend on utilities, grids and municipal approval, which is why energy policy has become an AI story.
05The major AI labs
OpenAI spread into video, images and audio while Anthropic bet everything on programming as the path to general intelligence, and that split shows in both product lines.
Google DeepMind owns chips, cloud and two decades of search data, so its slow start reflects bureaucracy rather than capability.
xAI could not compete on model talent so it competed on compute, and now rents Colossus 2 capacity in Memphis to labs that cannot build their own.
06Recursive self-improvement
Recursive self-improvement means AI running its own research, at which point compute and energy are the only limits because researcher count becomes arbitrary.
Read the optimistic, cautious and skeptical forecasts side by side rather than picking the one that matches your mood.
07The tools to learn
Pick one chat assistant, one harness and one automation tool and commit, because the leading options are close enough that comparison shopping is the expensive mistake.
Treat automation tooling as possibly temporary scaffolding rather than a permanent skill investment.
08How to prompt
Delegate instead of prompting: hand over context, the goal, what great looks like, and how the result will be verified.
Verification is the step almost everyone drops, and naming the check up front is what makes output usable without rereading every line.
09Turning AI into money
AI did not make new things possible, it made old processes profitable enough to look new, so start from a process that already earned.
Run every idea through one filter: was this profitable before AI? If it wasn't, compressing it with a model rarely changes the answer.
What businesses pay for is a repeatable procedural pipeline that converts a raw input into a finished output, so find the pipeline before you pick the model.
First customers come from unscalable manual work, and only after doing it by hand can you see which part a model should compress.
10AI opportunities are coming
Opportunities are getting bigger and arriving more often while the window to act on each one shrinks, which is the whole argument for doing the foundations now.
Glossary
Terms worth knowing.
Perceptron
The earliest artificial neuron, designed by Frank Rosenblatt in the 1950s. It takes several inputs, multiplies each by a weight, sums them, and fires an output. Every modern neural network is a scaled-up version of this idea.
AI winter
A period of collapsed funding and interest in artificial intelligence that followed the unmet promises of the 1950s and 1960s. The theory held up, but the hardware of the era could not run it.
Expert system
A rule-based program from the 1970s through the 2000s that encoded human knowledge as large if-then decision trees. It was called AI at the time, but contains no learning component.
AlexNet
The 2012 deep learning model that won an image classification contest by a wide margin and restarted mainstream interest in neural networks.
Transformer
The 2017 neural network architecture built around an attention mechanism. It underpins GPT, Claude, Gemini and nearly every current large language model.
Token
The unit of text a model reads and writes. Usually a fragment of a word rather than a whole word, which is why a seven-character word may be split into two or three pieces.
Next-token prediction
The core operation of a language model. It computes a probability distribution over possible next tokens and samples one, repeating the process to generate text.
Pre-training
The first and largest training stage, where a model learns statistical structure by reading enormous amounts of text. A purely pre-trained model can complete text but is not yet a usable assistant.
Fine-tuning from human feedback
A second training stage where people rate competing model outputs and the model is adjusted toward the preferred ones. It makes models conversational, and also makes them agreeable.
Sycophancy
A model's tendency to tell the user what they want to hear. It emerges because human raters in the feedback stage consistently preferred flattering answers.
Chain of thought
A technique where a model writes out intermediate reasoning steps and feeds them back to itself before answering. It improves hard-task accuracy and gives humans something readable to inspect.
Context window
The total number of tokens a model can hold at once, covering the prompt, any attached material and its own output. Roughly analogous to working memory.
Hallucination
Confident output that is factually wrong. It arises because a model learned the statistical shape of text rather than a grounded record of the world.
Harness
The software layer wrapped around a raw model that gives it tools, file access and a loop in which to act. Claude Code, Codex and Cursor are harnesses.
Recursive self-improvement
The point at which AI conducts its own research and development without human involvement, so each generation of models designs a faster and cheaper next generation.
Open-weight model
A model whose trained parameters are published, so anyone can download and run it locally rather than calling a hosted API.
Fab
A semiconductor fabrication plant. TSMC in Taiwan runs the fabs that physically manufacture the chip designs companies like NVIDIA create.
Resources
Things they pointed at.
04:51channelBut what is a neural network? (3Blue1Brown)
05:13channelThe Moment We Stopped Understanding AI (Welch Labs)
05:23channelThe 35-Year History of ChatGPT (Art of the Problem)
05:32linkThe Bitter Lesson by Rich Sutton
11:07channelLarge Language Models Explained Briefly (3Blue1Brown)
11:26channelIntro to Large Language Models (Andrej Karpathy)
11:56channelDeep Dive into LLMs like ChatGPT (Andrej Karpathy)
12:19linkTracing the Thoughts of a Large Language Model (Anthropic)
12:33channelTransformers, the tech behind LLMs (3Blue1Brown)
17:56channelAI Harness 101: An In-Depth Explanation (the host's own video)
18:08channelHow I Use LLMs (Andrej Karpathy)
18:23linkState of AI Report 2026 (Air Street Capital)
18:39toolEpoch AI Trends
25:28linkThe Gentle Singularity by Sam Altman
25:39linkWe Must Pace the Frontier by Dario Amodei
26:00linkMachines of Loving Grace by Dario Amodei
26:08linkTime Horizon 1.1 (METR)
26:14linkAI 2027
26:19linkAI's recursive self-improvement might not come so quickly after all (MIT)
33:55channelHow I'd Start AI Consulting in 2026 If I Could Start Over (the host's own video)
34:06channelWhat I'd Learn Instead of AI Automation in 2027 (the host's own video)
34:16linkDo Things That Don't Scale by Paul Graham
34:35linkRequests for Startups
29:46productMaker School (90-day paid accountability program)
35:41productMaker Zero (free resource vault with every link from the video)
Quotables
Lines you could clip.
00:55
“You'll know more about artificial intelligence than likely 99% of all humans that have ever lived.”
outsized promise stated in one sentence, no setup required→ TikTok hook↗ Tweet quote
01:49
“Anytime anything looks really complicated, if you just go back to when it was invented, and put yourself in the minds of the people that were dealing with the problems that necessitated those inventions, it often becomes really straightforward.”
a general learning principle that travels far outside AI→ newsletter pull-quote↗ Tweet quote
03:46
“The technology had just finally caught up to the theory.”
compresses seventy years of AI history into eight words→ IG reel cold open↗ Tweet quote
10:33
“What we're doing is more akin to dropping a bunch of different types of solutions in a petri dish and then slowly monitoring as it grows. We're just doing it in silicon, not biology.”
vivid, counterintuitive, and resets how most people picture AI research→ TikTok hook↗ Tweet quote
20:40
“Anthropic was literally like, we are text in and text out, baby. The only thing that matters is we're going to make our models the smartest programmers on planet Earth.”
names a strategic difference people argue about, in quotable form→ IG reel cold open↗ Tweet quote
25:09
“This means that in one year, you might see approximately 10 to 20 years of progress.”
a single startling number that works as a standalone claim→ TikTok hook↗ Tweet quote
29:02
“Delegate. Don't prompt.”
three words, instantly memorable, carries a whole framework behind it→ IG reel cold open↗ Tweet quote
31:20
“AI hasn't made anything new possible. What it has done is it has made old processes much more profitable.”
contrarian against the standard AI pitch and immediately actionable→ newsletter pull-quote↗ Tweet quote
35:02
“AI opportunities are getting a lot bigger, they're also getting a lot more frequent, but the time window that you can capitalize on them is a lot shorter.”
urgency without hype, good closer for a short→ TikTok hook↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphoranalogystory
So I spent the last seven years building and operating AI businesses all the way back from GPT -2 in 2019 to intermediate image generators in 2021 and 2022 to modern cloud code and codex based systems. This has been very financially lucrative and quite interesting, but with AI now moving into real world knowledge based roles, I know a lot of people are wondering.
How do I actually learn AI? And not just an entry -level understanding of AI, not just how to use the tools, but how to gain a broader appreciation and a more foundational kind of comprehension of what exactly is going on and why. So in this video I wanted to put together a straightforward one -week curriculum designed to take you from somebody that knows very little about AI or maybe you've just experimented with one or two tools to actually having a very strong understanding of AI, both its underpinnings and more broadly how you can use AI in a variety of ways, both for knowledge work, your business, your hobbies, and your personal life.
Ultimately by the end of this video you'll feel a lot more confident about how to use AI and you'll know more about artificial intelligence than likely 99 % of all humans that have ever lived. This is a broad roadmap, and so I'm going to link a bunch of resources down in the description below that you guys could use in order to, you know, actually watch the videos and read the references that I am going to be including.
I'm also going to provide a brief high -level description of all the various concepts that I wish that I had started on would have saved me so much time, probably years of my life, back at the beginning of my own journey. So in order to really get this, we're going to start back in the 1950s with AI history, where Frank Rosenblatt initially invented the idea of a modern neural network, what he called a perceptron.
It's important to understand the history behind artificial neural networks and later on, you know, machine learning and deep learning and now modern artificial intelligence, because a lot of what we're dealing with today seems really complicated. And if you're just airdropped into it, you're going to think, why the hell did they do this?
And what the hell does that term mean? And what was the motivation behind this? The reality is anytime anything looks really complicated, if you just go back to when it was invented, and put yourself in the minds of the people that were dealing with the problems that necessitated those inventions, it often becomes really straightforward.
And so with Frank Rosenblatt back in the 1950s and this whole idea of a perceptron, when you understand the history behind it, sort of what he was trying to solve, a lot of the modern terms of artificial neural networks and regression models and weighted functions and so on and so forth, all of that stuff becomes really clear.
At a high level, what ended up happening was the invention of this artificial neural network led to a tremendous amount of excitement around artificial intelligence in general. So people started talking about how, hey, in the next five to 10 years, we're going to have AI models that are going to run everything for us. machines will think and they'll govern our economy and they'll do really really crazy things but obviously at the time they were pretty compute bottlenecks they didn't really have anything in anywhere near close to what we have now and so you know most of these things did not come to pass and that led to a massive widespread disillusionment with ai known as an ai winter After the AI winter, there became renewed interest in designing what are called expert systems.
These expert systems at a high level are just like computer programs like we're dealing with now. They were really complicated if -then or binary search tree based systems that tried to deal with problems on a procedural and deterministic basis. And they worked pretty well for some problems in their fields, not at all based on neural networks and really nowhere near what we consider AI today.
They were still quite useful. Nowadays, that's just our modern conception of a computer program. But back then, these expert systems were still called artificial intelligence.
People used expert systems all the way up until maybe 2011, 2012, when AlexNet broke onto the deep learning scene, absolutely crushing on an image classification task at the time. performed so much better than virtually all other approaches that people started to say, wait a second, are these guys using that crazy architecture back from like the 1950s and 1960s?
Why the hell is it working now? It led to renewed interest in artificial neural networks and sort of what they were capable of. And really what had happened there, if you think about it, is the technology had just finally caught up to the theory.
You know, we've had the theoretical underpinnings behind the technology that we're using today, chat GPT codecs, for the most part, for the last 70 years or so. But we needed the actual hardware and compute to catch up to where theoreticians were back then in order for us to actually be able to use them to do cool things like image classification and so on.
That takes us to 2017, where the Transformer was first released. This is the same architecture that underlies ChatGPT, Claude, Gemini, and really most successful large language models today. That basis became the foundation of the generative pre -trained Transformer, GPT -1, GPT -2, GPT -3, later ChatGPT, which is based on GPT -3's architecture, and now the explosion and sort of, you know, extraordinary diversity of LLMs currently available on the scene.
Then 2024 hit, we invented reasoning models that also added significantly more complexity to how, you know, large language models solve tasks. And really the rest is history, kind of like our current date. So that's a really broad overview, but there's actually way more underneath the scenes.
Some fantastic resources for this are 3Blue1Brown's What is a Neural Network series. This actually walks you through the foundational mathematics behind how a neural network processes and computes. I know that can be kind of scary to people that aren't big fans of math, myself included, but it is really cool to just sort of like grok it and sort of understand that flash of insight behind, wait a second, this is something really similar to what my brain is doing.
Aside from that, a really good resource is by Welch Labs, The Moment We Stopped Understanding AI. That details what occurred around 2012 or so with, you know, AlexNet and sort of the deep learning revolution. And then finally, The Art of the Problem has an incredible video called The 35 -Year History of ChatGPT.
Once you're done with those three videos, again, which you can find down below in the description, I'd highly recommend reading The Bitter Lesson by Rich Sutton. Fantastic work that I think really grounds you in what most researchers believe to be the path to artificial general intelligence. and ultimately superintelligence.
Next, you need to understand how large language models actually work. And so there are a lot of resources that I'm going to throw at you here, and it is a very broad topic, but I'm going to try, you know, just giving you the 80 -20 upfront right now. Hopefully this sparks some of your neurons, which then can make you a little more excited to learn about the resources that I'm linking below.
The first is there are eight key ideas to understand in how modern AI works. First is tokens. Now, token is essentially a portion of a word.
A word might be seven characters long. A token might be four or five characters. And so what large language models will do is they will split text into these chunks of tokens, and then the model will read them one token at a time or write them one token at a time.
And so this is a little bit alien for people that like to think in words. For all intents and purposes, most people believe an artificial intelligence model just generates the most likely next word. But it's a lot more complicated than that because of this token concept.
It's not really a word. It's sort of like a fraction of a word, so to speak. And the reason why is because there are just a lot of different fractions of words that you can rinse, repeat, and recombine in a variety of different ways to make different words.
There's also next token prediction, which is the idea that at its core, what's going on with these neural networks is we're running a bunch of like very complicated matrix math using the same things you will have learned in the three blue, three brown, one blue video in order to generate the most statistically likely next token.
That's not exactly true. It's a little bit more complicated than that. Really, what you're doing is you're generating a distribution of the most likely next tokens, and then you're sampling from or you're taking from that distribution, maybe like the top five or whatever, and then iterating on that randomly.
Because if you always selected the statistically most likely next token, you'd basically always know what you were going to say ahead of time. You need a little bit of variance. You need a little bit of jitter.
And that's a very important concept to understand in next token prediction for large language models. Afterwards, this concept of pre -training is also crucial. And pre -training is the initial kind of steps involved in getting a model off the ground and having to be able to reason over things.
Pre -training is basically what like GPT -3 was. It was basically a purely pre -trained model and actually still used that model and sort of its fledgling alien -like intelligence to build a business that did over a million dollars. And so that's how I made my first million with GPT -3, which most people today would look at and be like, that thing, you know, that thing has less neurons than a freaking beetle.
How the hell are you going to make a million dollar business with that? Well, the way I was able to do so was really understanding just how much data and information was compressed via pre -training into the model's intelligence, and ultimately what patterns I could extract from that to do things like, you know, write really high -quality articles by pretending that I was actually an award -winning journalist, feeding that in as my system prompt, and then having to iterate over and over and over again.
Nowadays models go a little further than just pre -training. We actually fine -tune them through human feedback. That's another important concept because you're going to learn a lot about with these resources where basically not only do we pre -train and sort of like instantiate this brain, but then we coax and shape and evolve this brain into something that can actively produce productive conversations with human beings.
And so this is where human beings are shown a bunch of different outputs, and they're asked to rate the best output according to them. And usually what ends up happening is the human beings rate the output that sort of glazes them the most, makes them feel the best. And because of that, you also end up with model sycophancy and a bunch of other things.
Again, nothing that you'd really understand on a deep level unless you did this foundational sort of learning. This might only take you a few hours, but very, very important. After that, with reasoning models and chain of thought, you actually feed an output back in as input and then simulate the steps required in order to make a decision.
Chain of thought is why models can do extraordinary things today. And it's also a good monitor tool. It's a good observability tool for human beings to sort of understand what exactly is going on in this model's weird alien -like brain when I give it a task and ask it to complete it.
A really important concept that you're going to learn a lot about is this idea of a context window, which is the total number of tokens that you can actually fit into a model at one time. You know, the longer the token window, the context window rather, the more tokens you can fit in. And ultimately speaking, the more a model can reason over a bunch of different things kind of stick together.
Very analogous to human working memory, but you can go up super deep there too. Hallucination is another really important bit, you know, because the statistical nature of how models think and, you know, they sample from that distribution. They are also from time to time going to fall prey to biases and errors like coming up with things that don't actually make any sense.
Why? Because the way that the model initially sort of gained its ability to to talk was not through, you know, grounding it in real world facts. The way.
gain the ability to talk and reason and think, was just looking at a bunch of text and then analyzing the statistical correlations between them. And it turns out text isn't always real. We don't always just talk about things that are real.
And so because we don't always talk about things that are actually real and grounded in the real world, the model will, like that, learn that, oh, okay, I could write about things that don't necessarily exist. Finally, the concept of it being grown and not built. You know, a lot of people talk about AI researchers building these really smart, intelligent, and useful models.
The reality is they don't actually really build them. What we're doing is more akin to dropping a bunch of different types of solutions in a petri dish and then slowly monitoring as it grows. We're just doing it in silicon, not biology, so you don't have Lovecraftian horror novels written about it.
But really, we are growing alien -style intelligences through mathematics, and I think that's important to get. A couple of the resources I'm going to show you are going to go really deep into that, and then you'll learn a lot more about what it means to be grown and not built. So some really good ones.
Large Language Models Explained Briefly, again by 3Blue1Brown. That's five or ten minutes and it's super straightforward. It walks you through essentially how large language models work on foundational levels.
Then two really good video resources from Andrej Karpathy, who's widely considered to be one of the best educators in AI. I look up to that guy quite a bit. First is Intro to Large Language Models.
It's a one -hour explanation for a general audience. It occurred before reasoning models sort of hit the scene back in 2023. Because of that, there are a bunch of things that he says that now I look at and I say, wow, if I were a researcher back in 2023, that's probably what I'd believe as well.
And you also see how we're solving the problems that he talks about in that video through things like, you know, chain of thought and reasoning, which were developed immediately after. So it's very cool because this is in effect also a short history. After that, he has a great video called Deep Dive into LLMs.
like ChatGPT. It's about three or four hours or so, and it's almost like a novel, essentially. But it gives you everything from pre -training to fine -tuning and reasoning models.
And ultimately, what you learn in that video is going to be most of today. Probably going to take you a couple settings to make through, but still worth it. Then, in terms of reading, I would highly recommend you guys check out a primary source from Anthropic.
They wrote a wonderful article called Tracing the Thoughts of a Large Language Model. This goes deep into what the inside of a model's brain looks like, ultimately why it does things like hallucinate. It's also fairly accessible to a modern audience.
If you wanted to go a little bit deeper and learn a little bit more about how Transformers, the specific architecture that underlies ChatGPT, Codex, other models, and stuff like that work, I'd also recommend watching Transformers, the tech behind LLMs by 3Blue1Brown. This is about half an hour or so.
It's more technical for sure, but it is also extraordinarily powerful. Once you're done with that, it's important to understand the stack. A lot of people talk about artificial intelligence as if it's some abstraction, just like a software tool or even a mathematical idea.
But it's a lot more than that. You know, there's hardware that has to run these models. There are chips that have to run, you know, the mathematical operations that are occurring.
And they also have to be fabricated as well. And there are very few fabricators. And so, you know, this sort of model production also starts impacting politics and sort of what various government leaders are doing in various regions of the world.
And ultimately speaking, like why certain pressures have changed model. let's say, from the East and from China in a slightly different direction than models from the West because of compute availability. You know, this stuff is not obviously crucial to the utility of AI.
But again, gaining a strong foundational understanding of the context behind this will make you a far better operator later on. And so first, you're going to start with the chain. And so there are a variety of layers kind of stacked like an onion that ultimately result in the artificial intelligences that we know and love today.
They start with chips. Then they moved to data centers. Then they moved to energy.
Then they moved to model labs. Then they moved to APIs. then apps and harnesses, and finally actual implementation.
So on the chip level, you have big companies like NVIDIA, some of the most valued companies on planet Earth that actually design said chip architectures. You know, they don't actually manufacture them. They send them off to a company like TSMC or the Taiwan Semiconductor Manufacturing Company.
They make the chips. And you know, what's interesting is TSMC itself, the company that you know, makes the chips.
Well, somebody needs to make the machines that make the chips. Turns out that's another company, ASML, which is in the Netherlands. And so now you're starting to get into like supply chains of the things that manufacture this stuff and the hardware and the infrastructure that ultimately, you know, most modern machine learning, artificial intelligence is built off of.
Most people don't understand that. Most people don't get that there's actually a complicated supply chain at the very base of AI. And that might be even more valuable and more powerful than the stuff that model labs like philanthropic and open AI are doing.
On the data center level, you have Microsoft, you have Google. You have Amazon, you have Meta, you have, I don't know, Oracle, you have X, you have these large companies that are now pouring billions, sometimes even on the plan order of trillions of dollars into building plants where you will just line the walls with the chips that are, you know, invented by NVIDIA, made by TSMC.
And then, you know, actually one step up the chain machines produced by ASML. And so the idea is to have rows and rows and rows. of these, you know, chips basically running in perfect unison, you know, the outputs of TSM MC, for instance, to do the processing in such ways that NVIDIA prescribes.
And so, you know, big companies are pouring, again, hundreds of billions of dollars into this because they know that that is where the calculations are actually going to be occurring. But those big data centers, they also don't exist in isolation, right? Because they need something to power them.
Every single one of these operations requires an extraordinary amount of power in order to ultimately work at the scale and scope of which we want them to. And so you actually have utilities and power grids. You have, you know, municipal governments that are responsible for saying yes or no to where these big data centers are going to go.
And so, you know, we're just working our way through that onion right now, but hopefully you understand it's a little bit more complicated than just like, oh, Anthropic and OpenAI, they do everything. Speaking of, then you have the model labs. At that layer, you have companies like OpenAI, Anthropic.
You have Google. You have X as well. Some of these companies are vertically integrated, so they're trying to work at multiple levels of this layer.
You have DeepMind. You also have a couple of open -weight labs. These sorts of companies, they're actually responsible for the software implementation.
They're responsible for coming up with almost like the mathematics behind what computations will be done on the chips. One step up from them, you also have the labs themselves, some of which offer APIs. Now tools like OpenRouter, which are basically intermediary layers where, you know, you make a request to them and then they dole out the requests to whatever provider makes the most financial sense at that moment in time.
But what's ultimately sending most of those requests? It's not human beings just at our computers anymore. Really, it's one level up to harnesses.
to automated agent frameworks that do these sorts of requests en masse. And so these are, you know, codexes, these are your cloud codes, these are your cursors. And then finally, you actually have implementation, which is the people like myself, probably like most of you that use all of this technology, okay, in the in the AI vertical, and then take it and then apply it to other verticals out there, like your local plumber, or like some big consulting company whose operations you are optimizing, or something like that.
Maybe you're implementing into your own business, or your own life, or your hobbies, your personal, like whatever. Ultimately speaking, the implementation layer is the one where we sort of go, so what, and take all of the transformative outputs of these industries, and then apply it to something different. So as somebody that has probably recently been introduced to artificial intelligence technology, you are starting at the implementation layer, which is all the way at the very top of that onion.
But you have to realize just how deep everything goes if you really want to gain an appreciation for what exactly is going on and why. Some great resources here to watch on the AI harness. So specifically, I have a great video called AI Harness, an in -depth explanation, what a harness is, why it matters just as much as the actual underlying model.
There's a wonderful video resources by Andre. carpathy as well called how i use llms it's basically like a deep walkthrough of how a power user like him you know somebody that now works at least as of the time this recording at anthropic uses actual underlying artificial intelligence i'd also read a little bit about the state of AI report.
There's a lovely one called the state of AI report 2026 by a company called Airstreet Capital. It's about 45 minutes or so almost 250 slides, you are going to have to spend a fair amount of time on this. But if you skim that you'll understand a lot about what I talked about here at the deeper levels of the onion.
There's also a great piece by Epoch, Epoch AI trends. These are numbers on compute, how much money it costs to do things like, you know, train models, how much money is in data centers. After this, I'd spend some time learning about the specific players in the game.
I mean, I touched on a few of them earlier, right? You need to know who's building frontier AI and how their philosophies are a little bit different.
You know, every one of these companies has a different philosophy. That's why they have different companies. The two main ones in the lead, at least at the time of this recording, OpenAI and Anthropic themselves sort of spun off of a bunch of work that other researchers are doing at Google.
You know, they sort of had like ideological issues with some of the way that the research was occurring at Google. And so that spawned OpenAI. A few generations into the models of OpenAI, Anthropic had some ideological issues that was going on with OpenAI.
they spun off. And so everything is still connected sort of on a foundational underpinning sort of level, you have to know that. And so you know, the labs as of right now, you have open AI.
massive, massive consumer company, probably one of the biggest and most lucrative artificial intelligence companies ever, and obviously the ones to first bring to light ChatGPT and make it a consumer product. I'd still say they own a fair amount of the market, and I don't want to date this video's, you know, education potential by putting specific numbers on it.
But next to OpenAI, you have Anthropic. The whole idea about OpenAI is, you know, they want to get to an automated research intern. However, along the way, they've taken some detours.
They've made consumer apps and consumer products, things like Sora. They've made, you know, image models. They've made like GPT image, for instance.
They've done a bunch of things in both video, now audio, voice transcription. And so this is sort of like your general purpose AI company. They were the first to the scene, and they obviously know that there's a large financial incentive to getting this right.
But, you know, they're not exactly like super all in on any one thing. Speaking of companies that are all in on one thing, Anthropic. The key differentiator between Anthropic and OpenAI that most people don't realize is Anthropic is 1 million percent all in on computer programming as a means to get to AGI and ultimately ASI.
Whereas OpenAI was releasing all these cool consumer apps, they were dabbling in video, they're dabbling in photo. Anthropic was literally like, we are text in and text out, baby. The only thing that matters is we're going to make our models the smartest programmers on planet Earth.
And then we're going to make them program the next generation of... the models themselves. If wasting even 1 % of those resources that we're spending on video takes away from our ability to reach that.
AGI or ASI, it will all not have been worth it. And so their whole sort of modus operandi is to build this better future by racing ahead to, you know, recursive self -improvement, a concept we'll talk about later, and then using that. And once they're at RSI, they'll kind of do everything else.
There's also Google DeepMind. Now, this is probably one of the oldest labs out there. You know, this is the one that made AlphaGo way back in the day.
It's the one that, you know, sort of took the world by storm, I think, probably seven, eight or nine years ago before OpenAI or anything like that. was even created or invented. It was really the thing that sort of defined a generation of AI researchers.
And obviously Google DeepMind is part of Google. DeepMind was its own separate lab before its acquisition. Google obviously had a lot of financial incentives to acquiring it, but these guys are sort of like vertically integrated into everything.
They own the full stack. Because Google also manufactures chips. It also manufactures or rather runs cloud infrastructure.
So think Google search. It has probably more data than any other one of these companies simply because it's been a search engine for the better part of the last 15 or 20 years. A lot of people refer to them as sort of a slow giant where, you know, when all this AI stuff initially kicked off, it was Google deep mind that people were looking at being like, oh my God, these guys are probably going to take over, right?
But then they just didn't do anything for a really long time. And the reality is the bigger a company you are, the more bureaucratic your process is. are, but also the more things that you have to move around in order to like actually kickstart momentum.
That said, the bigger a company is, the more inertia it has. Just like a really heavy freight train, when it starts moving in a direction, although it can take some time to pick up, once it does start moving and eventually keep moving, it's very difficult to stop. There's also Meta, okay?
And Meta by Mark Zuckerberg has been well behind the model race for quite a while, only very recently caught up. But they've championed themselves essentially being like open source, you know, models for the people. We want everybody to have their own personal super intelligence.
And so they've dropped a bunch of products to the end like Muse, and they're not afraid to say why. They want to make a ton of money with this stuff. And they think that by empowering all humans with their own super intelligences, they can create markets that have never been created before.
You also have XAI with Elon Musk. Now, their big bet really is compute. Their huge bet is we have the most money ever.
We also have what we need for fabrication of large data centers and compute farms because of SpaceX. because of Tesla, and the fact that we've essentially been heading, you know, robotics manufacturing, specifically spaceship manufacturing, but also things like solar panels for the better part of the last like 10 or 15 years.
And so despite the fact that they weren't really capable of competing in the model race, initially, they just didn't have the, you know, model talent or the know how, they're now making massive strides in compute build outs. And they're gotten to the point now where their data centers like Colossus two in Memphis are literally being rented out to other model labs that don't have anywhere.
near that much capacity. So they're performing strategic partnerships where it matters and ultimately leading the model race in that hardware space, at least as the time of this recording. Now, all these labs are racing towards this idea of recursive self -improvement.
In plain words, that just means that an AI is doing the research and development of AI, and it does so completely independently and autonomously without any human involvement. Whether or not it's slower or faster than human beings' own approaches, the benefit to RSI, at least from an AI research perspective, is when you have AI that is capable of actually figuring this stuff out, you can just run it arbitrarily fast and then arbitrarily scoped.
So you can spawn 100 ,000 of these researchers, a million of these researchers. All you really need is compute and energy. Those two are the main bottlenecks.
The idea is once you have RSI, the models can begin actually optimizing their own inference and also optimizing their own training. And then the next generation of the models will be, you know, an exponentially greater degree faster and then cheaper to run. So those models will then optimize the next generation of exponentially faster and cheaper to run until, I mean, honestly, people are talking about these models essentially attaining omniscience, you know, knowing everything about everything, simply because, you know, they can compute and they can think so many orders of magnitude faster than people can.
And we're getting pretty close. I'm not going to date this, but suffice to say, recursive self -improvement is something that I think is probably starting to occur already, if not is in full swing right now, which is one of the main motivations behind me making this video. By the way, I thought it was important everybody understand just what the hell is going on and how quickly all of it is happening.
This means that in one year, you might see approximately 10 to 20 years of progress. Try and wrap your mind around how much progress you've had in the last year and then how much progress we would have in an exponentially increasing situation. We could be at the point where models don't drop.
once a month. It could be at the point where models drop literally every five minutes, and I don't say that facetiously. Anyway, some great resources to read here.
The Gentle Singularity by Sam Altman. It's maybe 15 or 20 minutes or so. It's a very optimistic case, and ultimately speaking, it's probably a little bit biased for investment reasons, but it's an awesome piece.
We Must Pace the Frontier by Dario Amode. He's the CEO of Anthropic. You know, this is a cautious sort of case, whereas OpenAI is like, hey, everything's going to be awesome.
Everything's going to be amazing. You know, Anthropic's like, no, no, no, we got to be a little more careful. And that's the point that he's making in that piece.
If you read that piece start to finish, and it can be kind of a mouthful, but it is quite important and valuable. Alongside his other piece, Machines of Loving Grace, I think you start to understand just how maniacally focused on achieving that RSI a lot of these companies are. Time Horizon 1 .1 by METR.
It's a company that's heavily involved in track. what is going on in these AI models and how quickly they are progressing. I also think it's important to understand the skeptics case.
Here's one from MIT called AI's recursive self -improvement might not come so quickly after all. And then if you really want to understand not just the model dynamics, but the geopolitical dynamics, like I was telling you guys about earlier, read AI 2027. It's a detailed forecast of what is occurring over the course of the last few years and will occur over the course of the next few years, maybe into the 2030s or so with artificial intelligence technology.
Next up, I wish I had done all of this ahead of time before I learned the tools, because I think this would have had quite the impact on, you know, what I did and why. But eventually, when push comes to shove, you can't spend all your time in a la -la land obsessing over geopolitics or whatever. You actually got to, you know, when push comes to shove, learn a damn tool.
And so this is all about the tools. And as of the time of this recording, you have a few major ones. You have Claude, you have ChachiPT, you have Gemini, you know, you have X's Grok, you have a few open weights models coming out of China that are competing as well.
But these are really the main contenders. And these are probably going to be the ones that are going to proceed. I guess you could have Muse as well, depending on how things go over the course of the next few months.
Reality is like most of these are going to be quite similar, but you do just have to pick one and then start talking to it. In addition to the chat assistant layer, again, this is kind of like an onion. One level up, you have the agent layer, you have the harness layer.
And so the variety of different harnesses, you have Cloud Code, for instance, you have Codex. You have Kimi code, right? There are a variety of these different harnesses, which are sort of software implementations and approaches to these models that sort of constrain them and control them in various directions.
You'll know a lot more about this than what I'm explaining right now, because you'll have actually gone through the whole harness section of our curriculum. Finally, from there, I'd recommend learning some sort of automation tool. I don't mean automation tool in the traditional sense, but.
I think it's important to understand that the way that most people are going to be applying these models is not really just to ask them how much protein is in their yogurt, as my co -host Jack Roberts typically says. It is to ultimately make money. or do some sort of economic end, you know, achieve something in a business type scenario.
And so the way that you do that typically is by automating processes that currently produce revenue or produce some sort of positive value. And so automation tools typically hook up artificial intelligence models and then empower them with access to other tools to allow them to do things within a business in a reliable and predictable way.
And so personally, what I would do is I would pick a chat assistant, so maybe Claude, a harness or an agent. an agentic layer on top of that, something like Cloud Code. And then finally, I'd pick an automation tool, something like, I don't know, N8n.
Although these automation tools as is may not really be required over the course of the next few years, depending on how quickly AI moves to take over all that stuff. Some simple and easy ways to wrap your mind around prompting, which is the process of putting a request into an AI model to kick off some sort of prompt or action chain, is delegate.
Don't prompt. At the end of the day, these models actually work better if you anthropomorphize them. For those of you guys that don't know, that word just means treat them like you would treat a person.
And so typically when delegating to other people, you always want to have four things. First is context. So who you are, what the work is for, why you want to do it, and so on.
The second is goal, which is the result that you want out of this. The third is what great looks like. So some sort of example or criteria that you are explicitly stating means the job is done.
And then finally, how to verify that the job has been done to your, you know, specifications. And so maybe you verify it with like a screenshot, or maybe you verify it by comparing the pixel perfect accuracy of the thing that Claude generated versus the thing that, you know, you generated, something like that. This is most of what I teach on my channel and also inside of Maker School, which is my 90 -day accountability program that guarantees you your first customer for a paying AI or automation service in 90 days, or I give you all your money back.
The reason why is because we walk you through how to do things like prompt engineer, how to give agents strong tasks, and so on. Anyway, add over, you should check it out. Some things that you guys could learn or read or watch to really embody what it is to be effective with these models.
First, AI fluency, framework, and foundations. This is a brief four -hour mini course by Anthropic Academy that actually walks you through everything to do with their version of their large language model, Claude, but also, you know, Claude code, what all these different tools are for, and ultimately how you can apply them to business use cases.
Next up, $2 .4 million of prompt engineering hacks in 53 minutes. That's by me. I spent 53 minutes basically just wringing out my brain like a wet towel, trying to give you guys as much value as possible.
So definitely check that out below too. And I know I'm probably going to come across as, you know, anthropic, a script kitty or super biased or something. But if you wanted to go a little bit deeper into this, I would also read through the interactive prompt engineering tutorial by Anthropic.
It's hands on. They actually have agents just built in directly to your structure. You sort of like quizzes.
All this stuff was built by AI. And so it's sort of like a meta example of what I can be used for both education, both upskilling and both teaching. people and onboarding them to your product.
All things that you can learn how to do with AI as well. Finally, how to turn all of this stuff into money. I mean, it's one thing to learn how to use a tool, but how do you actually apply it to real businesses?
The whole idea here behind this part of the roadmap is that AI hasn't made anything new possible. What it has done is it has made old processes much more profitable. And it is in the making of those old processes more profitable that some new things have seemed like they've become possible.
But really, these things were always possible. The whole idea about artificial general intelligence is that it is the same intelligence as a human being approximately in all domains, right? And so, you know, all the things an AGI can do, a human being can do.
It's just about the timescale. And so when people talk about AGI and like artificial intelligence producing all these new incredible things, I'm always like, well, the question is more, was that thing profitable before AI? Was it profitable before the next cool new model drop?
If the answer to that question was no, it's probably still not profitable now. But if it is profitable now and you apply AI to it, now we're talking. And so the ability to turn these wonderful developments into some sort of monetary outcome typically relies on your ability to understand what a good monetary process actually looks like.
And in general, for most business, that looks like a very structured, step -by -step, repeatable and procedural workflow that takes some raw good. might be, I don't know, like a person that just doesn't know anything about whatever you're selling if you're selling education, which might be, you know, I don't know, some sort of like raw thing, maybe an iron ingot or something.
And then it goes through a pipeline filled with, I don't know, a bunch of processes to turn that iron ingot into something else. And then, you know, out comes, I don't know, some sort of like working hammer or working tool that you can use on a farm. It is in that pipeline, it is in that process that ultimately value is delivered and you are paid.
And so, you know, what you need to do in order to become really good at turning these wonderful technological advances into money is you just need to understand what pre -existing processes made businesses money and then how can we make that 100x faster in order to... generate much more leverage with artificial intelligence.
And so the simplest and easiest way to really understand all this stuff is actually just to go out there and try selling it to people, ask them. So like, what exactly keeps you up at night? So what exactly is your business currently doing in order to make revenue?
Wow, really? Hmm. That's interesting.
I wonder if that's something an AI could do faster. Let me think. It's not just to like, you know, come up with a bunch of great ideas using artificial intelligence for the sake of great ideas, but it's actually to look at like preexisting processes and then just implement artificial intelligence to iterate on them and make them a little bit better, a little bit faster.
The cool thing is, you know, if you grow 1 % better every day, you grow, I don't know, two or 300 % better by the end of the year. AI can compress that whole year long process into one that legitimately only takes a few days. And as somebody that's now made, you know, over $10 million using artificial intelligence tools, it is shocking what that difference can do for, you know, a company's wallet.
Some very simple and straightforward guides to this, how I would start AI consulting in 2026, if I could start over, that is by yours truly took about half an hour for that video start to finish. If you guys like that sort of thing, then another really good one is what I'd learn instead of AI automation in 2027. That's about 20 to 25 minutes.
On the reading side, probably my favorite piece ever throughout this entire video is probably this one. Do Things That Don't Scale by Paul Graham. That's about, I don't know, 10 or 20 minutes or so.
Basically, it's why your first customers always come from doing unscalable things by hand, aka the sorts of things that you were always doing, whether or not AI was around. And then ultimately speaking, when you identify what those things are, how you can convert that into some sort of highly leveraged AI process. Next is Requests for Startups.
This will take you just a few minutes to read, but basically it's like a place where people make requests for companies to build solutions to problems other companies are having. So this can help put you in the mindset of somebody that's like looking for ways that you can apply artificial intelligence to things that people are currently seeing detract from their ability to make money, or ultimately speaking, just solve problems for them.
Okay, ultimately, AI opportunities are coming. In general, they're getting a lot bigger. They're also getting a lot more frequent, but the time window that you can capitalize on them is a lot shorter.
And so I wanted to make this video because I feel like by understanding the context behind all this, and really by just spending a few days, if not a week, you can achieve such a well -rounded understanding of what exactly is going on here that you ultimately can actually be the person that knows whether or not an opportunity is good, whether or not these new developments are real or they're total fluff, whether or not you need to be concerned about certain impacts geopolitically, whether or not you need to worry about the job market or what have you.
This is what I wish I'd learned when I had started this AI journey forever ago, and hopefully my video alongside all the resources that I've linked down below can help take you guys to that level as well. If you guys want these resources, just click the second link in the description to head over to Maker Zero. It's where I give everything away completely for free.
This will be up at the top of the recently uploaded module inside of Classroom. If you guys also want to learn how to do this for a business purpose, as mentioned, check out Maker School. That's the first link in the description.
Aside from that, guys, thank you very much for watching this video all the way to the end. I really appreciate every single one of you, especially the ones that do make it this far. If you guys have any suggestions for future videos for me to make, just let me know.
I'd be more than happy to. I get most of my ideas from my audience at this point. Please do me a solid like, subscribe, and I love every single one of you.
Have a lovely rest of the day, and I'll catch all y 'all in the next one.
The Hook
The bait, then the rug-pull.
Seven years of building AI businesses, compressed into a week of homework. The promise lands fast: finish the curriculum and you'll know more about artificial intelligence than almost anyone who has ever lived. What follows is not a tool tour. It's a sequence, and the sequence is the point.
Frameworks
Named ideas worth stealing.
31:17list
The seven-stage AI roadmap
AI history
How LLMs work
The AI stack
The players
The race to RSI
Learn the tools
Turn it into money
The spine of the whole video, shown on screen as a checklist that fills in stage by stage. The order is deliberate: context first, mechanics second, industry third, tools fourth, money last.
Steal forany curriculum where learners currently start at the tool and never get the context
06:00list
The eight key ideas behind modern AI
Tokens
Next-token prediction
Pre-training
Fine-tuning from human feedback
Reasoning and chain of thought
Context window
Hallucination
Grown, not built
A minimum viable mental model of a language model. Each idea maps to a behavior you've already noticed, which is why learning them retroactively explains months of confusing output.
Steal foronboarding anyone non-technical onto an AI product
13:44model
The AI stack, layered like an onion
Chips
Data centers
Energy
Model labs
APIs
Apps and harnesses
Implementation
Seven layers from silicon to the person applying it in a plumbing business. NVIDIA designs, TSMC manufactures, ASML makes the machines TSMC needs, hyperscalers build the data centers, utilities and municipalities gate the power, labs write the math, harnesses drive the requests, and implementers decide what it's for.
Steal forexplaining why a supply-chain story is also a geopolitics story
29:02list
Delegate, don't prompt: hand over four things
Context
Goal
What great looks like
How to verify
Treat the model like a capable new hire rather than a search box. The fourth item is the one most people drop, and it is the one that decides whether you can trust the output without reading every line.
Steal forany task brief, human or model
31:53concept
The pre-AI profitability test
Before building anything, ask whether the process was profitable before AI existed. If it wasn't, AI probably hasn't fixed that. If it was, compressing it with AI is where the margin is.
Steal forfiltering a list of AI business ideas down to the ones worth building
CTA Breakdown
How they asked for the click.
VERBAL ASK
35:39link
“If you guys want these resources, just click the second link in the description to head over to Maker Zero. It's where I give everything away completely for free.”
Two-tier close. The free resource vault is named first and carries all the links the video promised, so the ask feels like completing the lesson rather than interrupting it. The paid program is mentioned once mid-video at the prompting section and again briefly here, both times attached to a specific capability just taught.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Nick Saraev breaks down seven concrete habit swaps that separate the highest-output AI users from everyone else: talk instead of type, keep AI listening through every meeting, show instead of tell, let AI interview you, demand flaws instead of opinions, and just try the thing yourself before assuming it's impossible.
Nick Saraev races a vanilla Claude agent against REA, the trending GitHub tool for AI-powered app cloning. He finds the underlying tech is decades old, and the agents themselves are already commoditizing the software built around them.
A behavioral-neuroscience grad turned AI-agency founder breaks down ten ways to use AI as a learning tool, and the one way people default to that quietly works against them.
An AI automation agency owner making seven figures a year argues the easy money in his own industry is closing fast, then lays out what's actually worth building toward next.
A creator who says he's burned $31,141 on Claude Code turns that spend into sixteen concrete habits, from running evals on prompts to keeping a backup model on standby.
00:40
01:14
01:46
02:29
02:56
03:31
03:41
04:18
04:45
05:12
05:43
06:34
07:01
07:56
08:22
08:54
09:17
09:44
10:11
10:38
11:06
11:33
12:00
12:27
12:54
13:22
14:16
14:43
15:10
15:37
16:03
16:32
16:59
17:31
17:53
18:21
18:46
19:15
19:42
20:22
21:04
21:31
22:08
22:25
22:52
23:20
23:49
24:52
25:04
25:25
26:03
26:30
26:49
28:05
28:09
28:46
29:13
30:03
30:40
31:11
31:17
32:23
32:51
33:18
33:45
34:12
34:39
35:29
35:55
07:00
07:30
10:46
29:36
23:59
25:59