YT YouTube2Text

YouTube2Text → Transcripts → The Last AI Roadmap Video You’ll Ever Watch

The Last AI Roadmap Video You’ll Ever Watch — Transcript

by Nick Saraev · 8,202 words · 1,220 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00So, I spent the last 7 years building
0:01and operating AI businesses all the way
0:03back from GPT-2 in 2019 to intermediate
0:06image generators in 2021 and 2022 to
0:09modern cloud code and Codex based
0:11systems. This has been very financially
0:12lucrative and quite interesting, but
0:14with AI now moving into real-world
0:16knowledge-based roles, I know a lot of
0:17people are wondering, how do I actually
0:19learn AI? And not just an entry-level
0:21understanding of AI, not just how to use
0:23the tools, but how to gain a broader
0:24appreciation and a more foundational
0:27kind of comprehension of what exactly is
0:29going on and why. So, in this video, I
0:31wanted to put together a straightforward
0:33one-week curriculum designed to take you
0:35from somebody that knows very little
0:36about AI or maybe you've just
0:37experimented with one or two tools to
0:39actually having a very strong
0:41understanding of AI, both its
0:43underpinnings and more broadly how you
0:45can use AI in a variety of ways, both
0:47for knowledge work, your business, your
0:49hobbies, and your personal life.
0:50Ultimately, by the end of this video,
0:51you'll feel a lot more confident about
0:53how to use AI, and you'll know more
0:55about artificial intelligence than
0:56likely 99% of all humans that have ever
0:58lived. This is a broad road map, and so
1:01I'm going to link a bunch of resources
1:03down in the description below that you
1:04guys could use in order to, you know,
1:05actually watch the videos and read the
1:08references that I am going to be
1:09including. I'm also going to provide a
1:11brief, high-level description of all the
1:13various concepts that I wish that I had
1:15started on. Would have saved me so much
1:17time, probably years of my life, um back
1:18at the beginning of my own journey. So,
1:20in order to really get this, we're going
1:21to start back in the 1950s with AI
1:23history, uh where Frank Rosenblatt
1:25initially invented the idea of a modern
1:28neural network, what he called a
1:29perceptron. It's important to understand
1:31the history behind artificial neural
1:33networks and later on, you know, machine
1:35learning and deep learning and now
1:37modern artificial intelligence because a
1:39lot of what we're dealing with today
1:40seems really complicated. And if you
1:42just air-dropped into it, you're going
1:43to think, why the hell did they do this?
1:45And what the hell does that term mean?
1:46And what was the motivation behind this?
1:48The reality is, anytime anything looks
1:49really complicated, if you just go back
1:51to when it was invented and put yourself
1:53in the minds of the people that were
1:54dealing with the the that necessitated
1:56those inventions, it often becomes
1:57really straightforward. And so, with
1:59Frank Rosenblatt back in the 1950s, this
2:01whole idea of a perceptron, when you
2:02understand the history behind it, sort
2:04of what he was trying to solve, a lot of
2:06the modern terms of, you know,
2:07artificial neural networks and
2:09regression models and weighted functions
2:11and so on and so forth, all that stuff
2:12becomes really clear. At a high level,
2:14what ended up happening was the
2:15invention of this artificial neural
2:16network led to a tremendous amount of
2:18excitement around artificial
2:20intelligence in general. So, people
2:22started talking about how, "Hey, in the
2:23next 5 to 10 years, to have, you know,
2:26AI models that are going to run
2:27everything for us. Machines will think
2:29and they'll govern our economy and
2:30they'll do really, really crazy things."
2:32But, obviously, at the time, they were
2:33pretty compute bottlenecks. They didn't
2:34really have anything in anywhere near
2:36close to what we have now. And so, you
2:38know, most of these things did not come
2:40to pass. And that led to a massive,
2:42widespread disillusionment with AI,
2:44known as an AI winter. Um after the AI
2:46winter, okay, there became renewed
2:48interest in designing what are called
2:49expert systems. These expert systems, at
2:52a high level, are just like computer
2:53programs like we're dealing with now.
2:55They were really complicated if-then or,
2:57you know, binary search tree-based
2:59systems that tried to deal with problems
3:01on a procedural and deterministic basis.
3:03And they worked pretty well for some
3:05problems in their fields. Not at all
3:06based on neural networks and really
3:08nowhere near what we consider AI today.
3:10Uh they were still quite useful.
3:11Nowadays, that's just our modern
3:12conception of a computer program. But,
3:14back then, these expert systems were
3:16still called artificial intelligence.
3:18People used expert systems all the way
3:20up until maybe 2011, 2012, when AlexNet
3:23broke onto the deep learning scene, um
3:25absolutely crushing on an image
3:28classification task at the time. Um it
3:30performed so much better than virtually
3:32all other approaches that people started
3:33to say, "Wait a second. Are these guys
3:35using that crazy architecture back from
3:37like the 1950s and 1960s? Why the hell
3:39is it working now?" It led to renewed
3:41interest in artificial neural networks
3:43and sort of what they were capable of.
3:44And really what had happened there, if
3:45you think about it, is the technology
3:47had just finally caught up to the
3:48theory. You know, we we've had the
3:50theoretical underpinnings behind the
3:52technology we're using today, ChatGPT,
3:54Codex, for the most part for the last 70
3:55years or so, but we needed the actual
3:57hardware and computes to catch up to,
4:00you know, where theoreticians were back
4:01then in order for us to actually be able
4:03to use them to do cool things like image
4:05classification and so on. That takes us
4:07to 2017 where the transformer was first
4:09released. This is the same architecture
4:11that underlies ChatGPT, Claude, Gemini,
4:14and really most successful large
4:15language models today. That basis became
4:17the foundation of the generative
4:19pre-trained transformer GPT-1,
4:21GPT-2, GPT-3, later ChatGPT, which is
4:25based on GPT-3's architecture, and now
4:28the explosion and sort of, you know,
4:30extraordinary diversity of LLMs
4:32currently available on the scene. Then
4:342024 hit, we invented reasoning models
4:36that also added significantly more
4:38complexity to how,
4:40uh, you know, large language models
4:41solve tasks. And really the rest is
4:43history, kind of like our current date.
4:45So, that's a really broad overview, but
4:46there's actually way more underneath the
4:48scenes. Some fantastic resources for
4:50this are 3Blue1Brown's What Is a Neural
4:53Network series. This actually walks you
4:54through the foundational mathematics
4:56behind how a neural network processes
4:58and computes. I know that can be kind of
5:00scary to people that aren't big fans of
5:01math, myself included, but um, it is
5:04really cool to just sort of like grok it
5:06and sort of understand that flash of
5:08insight behind, wait a second, this is
5:09something really similar to what my
5:10brain is doing. Aside from that, a
5:12really good resource is by Welch Labs,
5:14The Moment We Stopped Understanding AI.
5:16That details what occurred around 2012
5:18or so with, um, you know, AlexNet and
5:20sort of the deep learning revolution.
5:22And then finally, The Art of the Problem
5:23has an incredible video called The
5:2535-Year History of ChatGPT. Once you're
5:27done with those three videos, again,
5:29which you can find down below in the
5:30description, I'd highly recommend
5:31reading The Bitter Lesson by Rich
5:33Sutton. Um, fantastic work that I think
5:35really grounds you in what most
5:37researchers believe to be the path to
5:39artificial general intelligence and
5:41ultimately superintelligence. Next, you
5:43need to understand how large language
5:44models actually work. And so, there are
5:46a lot of resources that I'm going to
5:47throw at you here, uh, and it is a very
5:49broad topic, but I'm going to try, you
5:51know, just giving you the 80/20 up front
5:53right now. Hopefully, this sparks some
5:54of your neurons, which then can make you
5:56a little more excited to learn about the
5:57resources that I'm linking below. Uh the
5:59first is there eight key ideas to
6:01understand in how modern AI works. First
6:04is tokens. Now, token is essentially a
6:07portion of a word. You know, word might
6:09be seven characters long. A token might
6:11be four or five characters. And so, you
6:13know, what large language models will do
6:15is they will split text into these
6:17chunks of tokens, and then the model
6:19will read them one token at a time, or
6:22write them one token at a time. And so,
6:24this is a little bit alien for people
6:25that like to think in words. For all
6:27intents and purposes, most people
6:28believe, you know, an artificial
6:30intelligence model just generates the
6:32most likely next word, but it's a lot
6:33more complicated than that because of
6:35this token concept. It's not really a
6:36word, it's sort of like a
6:38a fraction of a word, so to speak. And
6:40the reason why is because there are just
6:41a lot of different fractions of words
6:43that you can rinse, repeat, recombine in
6:46a variety of different ways to make
6:47different words. There's also next token
6:49prediction, which is the idea that at
6:50its core, what's going on with these
6:52neural networks is we're running a bunch
6:53of like very complicated matrix math
6:55using the same things you will have
6:56learned in the three blue three brown
6:58one blue video, um in order to, you
7:00know, generate the most statistically
7:03likely next token. That's not exactly
7:05true. It's a little bit more complicated
7:07than that. Really, what you're doing is
7:08you're generating a distribution of the
7:10most likely next tokens, and then you're
7:11sampling from, or you're taking from
7:14that distribution, maybe like the top
7:16five or whatever, and then iterating on
7:18that randomly. Cuz if you always
7:19selected the statistically most likely
7:21next token, you know, you'd basically
7:22always know what you were going to say
7:24ahead of time. You need a little bit of
7:26variance. You need a little bit of
7:27jitter. And that's a very important
7:28concept to understand in next token
7:30prediction for large language models.
7:32Afterwards, this concept of pre-training
7:33is also crucial. And pre-training is the
7:36initial kind of steps involved in
7:38getting a model off the the and having
7:40it be able to reason over things. Uh
7:42pre-training is basically what like
7:44GPT-3 was. It was basically a purely
7:46pre-trained model. I mean, I actually
7:48still used that model and sort of its
7:49fledgling alien-like intelligence uh to
7:52build a business that did over a million
7:53dollars. And so, that's how I made my
7:54first million with GPT-3, which most
7:57people today would look at and be like,
7:58"That thing, you know, that thing has
8:00less neurons than a freaking beetle. How
8:02the hell are you going to make a
8:03million-dollar business with that?"
8:04Well, the way I was able to do so was
8:06really understanding just how much data
8:08and information was compressed via
8:10pre-training into the model's
8:12intelligence and ultimately what
8:13patterns I could extract from that to do
8:15things like, you know, write really
8:17high-quality articles by pretending that
8:19I was actually an award-winning
8:20journalist, feeding that in as my system
8:22prompt, and then having it iterate over
8:23and over and over again. Nowadays,
8:25models go a little further than just
8:26pre-training. We actually fine-tune them
8:28through human feedback. That's another
8:30important concept you guys are going to
8:31learn a lot about with these resources
8:32where basically not only do we pre-train
8:35and sort of like instantiate this brain,
8:37but then we we coax and shape and evolve
8:39this brain into uh something that can
8:42actively produce productive
8:44conversations with human beings. And so,
8:46this is where human beings are shown a
8:47bunch of different outputs and they're
8:48asked to rate the the best output
8:51according to them. And usually what ends
8:53up happening is the human beings rate
8:54the output that sort of glazes them the
8:56most, makes them feel the best. And
8:58because of that, you also end up with
8:59model sycophancy and a bunch of other
9:01things. Again, nothing that you'd really
9:02understand on a deep level unless you
9:04did this foundational sort of learning.
9:06This might only take you a few hours,
9:07but very, very important. After that,
9:08with reasoning models and chain of
9:10thought, you actually feed an output
9:12back in as input and then simulate the
9:14steps required in order to make a
9:16decision. Chain of thought is why models
9:18can do extraordinary things today, and
9:20it's also a good monitor tool. It's a
9:22good observability tool for human beings
9:24to sort of understand what exactly is
9:26going on in this model's weird
9:27alien-like brain when I give it a task
9:29and ask it to complete it. A really
9:31important concept you're going to learn
9:32a lot about is this idea of a context
9:34window, which is the total number of
9:35tokens that you can actually fit into a
9:37model at one time. You know, the longer
9:39the token window, the context window
9:40rather, the more tokens you can fit in.
9:43And ultimately speaking, the more a
9:44model can reason over a bunch of
9:45different things to kind of stick
9:46together. Very analogous to human
9:48working memory, but you can go up super
9:50deep there, too. Hallucination is
9:51another really important bit. You know,
9:53because the statistical nature of how
9:54models think and you know, they sample
9:56from that distribution, they are also
9:59from time to time going to fall prey to
10:01biases and errors like coming up with
10:03things that don't actually make any
10:04sense. Why? Because the way that the
10:06model initially sort of gained its
10:08ability to to talk was not through, you
10:11know, grounding it in real-world facts.
10:13The way it gained the ability to talk
10:15and to reason and think was just looking
10:17at a bunch of text and then analyzing
10:18the statistical correlations between
10:19them. And it turns out text isn't always
10:22the real. We don't always just talk
10:23about things that are real. And so
10:24because we don't always talk about
10:25things that are actually real and
10:26grounded in the real world, the model
10:28will, like that, learn that, "Oh, okay,
10:30I can write about things that don't
10:31necessarily exist." Finally, the concept
10:33of it being grown and not built. You
10:35know, a lot of people talk about AI
10:37researchers building these really smart
10:39intelligence and and useful models. The
10:41reality is they don't actually really
10:43build them. What we're doing is more
10:44akin to dropping a bunch of different
10:46types of solutions in a Petri dish and
10:48then slowly monitoring as it grows.
10:50We're just doing it in silica, not
10:52biology, so you don't have Lovecraftian
10:53horror novels written about it. Um but
10:56really, like we are growing alien-style
10:58intelligence as through mathematics. And
11:00I think that's important to get. Couple
11:01of the resources I'm going to show you
11:02are going to go really deep into that
11:03and then you'll learn a lot more about
11:04what it means to be grown and not built.
11:06So, some really good ones. Large
11:08language models explained briefly, again
11:10by Three Blue One Brown. That's uh 5 or
11:1210 minutes and it's super
11:13straightforward. It walks you through
11:14essentially how large language models
11:16work on a foundational levels. Then two
11:19really good video resources from Andrej
11:20Karpathy, who's widely considered to be
11:22one of the best educators in AI. I look
11:24up to that guy quite a bit. Um first is
11:26intro to large language models. It's a
11:281-hour explanation for a general
11:30audience. It occurred before reasoning
11:32models have sort of hit the scene back
11:33in 2023. Because of that, there a bunch
11:36of things that he says that now I look
11:38at and I say, "Wow, if I were a
11:39researcher back in 2023, that's probably
11:41what I'd believe as well." And you also
11:43see how we're solving the problems that
11:45he talks about in that video through
11:47things like, you know, chain of thought
11:49and reasoning, which were developed
11:50immediately after. So, it's very cool
11:51because this is in effect also a short
11:53history. After that, he has a great
11:55video called deep dive into LLMs like
11:57ChatGPT. It's about 3 or 4 hours or so,
12:00and it's sort of it's almost like a
12:01novel, essentially. But, it gives you
12:03everything from pre-training to
12:04fine-tuning and reasoning models. And
12:07ultimately, what you learn in that video
12:08is going to be most of today. Probably
12:10going to take you a couple sittings to
12:11make through, but still worth it. Then,
12:13in terms of reading, would highly
12:14recommend you guys check out a primary
12:16source from Anthropic. They wrote a
12:18wonderful article called Tracing the
12:20Thoughts of a Large Language Model. This
12:22goes deep into what the inside of a
12:24model's brain looks like. Ultimately,
12:26why it does things like hallucinate.
12:28It's also fairly accessible to a modern
12:30audience. If you wanted to go a little
12:31bit deeper and learn a little bit more
12:33about how transformers, the specific
12:35architecture that underlies ChatGPT,
12:37Codex, other models, and stuff like that
12:38work, um I'd also recommend watching
12:40Transformers: The Tech Behind LLMs by
12:433Blue1Brown. This is about half an hour
12:45or so. It's more technical, for sure,
12:46but it is also extraordinarily powerful.
12:48Once you're done with that, it's
12:49important to understand the stack, okay?
12:51A lot of people talk about artificial
12:53intelligence as if it's some
12:54abstraction, just like a software tool
12:56or even a mathematical idea. But, it's a
12:58lot more than that. You know, there's
13:00hardware that has to run these models.
13:02There are chips that have to run, you
13:04know, the the the mathematical
13:05operations that are occurring. And they
13:06also have to be fabricated as well, and
13:08there are very few fabricators. And so,
13:09you know, this sort of model production
13:11also starts impacting politics and sort
13:14of what various government leaders are
13:15doing in various regions of the world.
13:17And ultimately speaking, like why
13:18certain pressures have changed models,
13:20let's say, from the East and from China
13:23uh in a slightly different direction
13:24than models from the West because of
13:25computer availability. You know, this
13:27stuff is not obviously crucial to the
13:29utility there, but again, gaining a
13:30strong foundational understanding of the
13:32context behind this will make you a far
13:33better operator later on. And so, first,
13:35you're going to start with the chain.
13:36And so, there are a variety of layers
13:38kind of stacked like an onion that
13:40ultimately result in the artificial
13:42intelligences that we know and love
13:43today. They start with chips, then they
13:45move to data centers, then they move to
13:47energy, then they move to model labs,
13:49then they move to APIs, then apps and
13:51harnesses, and finally actual
13:53implementation. So, on the chip level,
13:55uh you have big companies like Nvidia,
13:58some of the most value companies on
13:59planet Earth that actually design said
14:02chip architectures. You know, they don't
14:03actually manufacture them. They send
14:05them off to company like TSMC or the
14:07Taiwan Semiconductor Manufacturing
14:09Company. Um they they make the chips.
14:12And you know, what's interesting is TSMC
14:14itself, the company that, you know,
14:16makes the chips, well, somebody needs to
14:18make the machines that make the chips.
14:20Turns out that's another company, um
14:22ASML, which is in the Netherlands. And
14:24so, now you're starting to get into like
14:26supply chains of the things that
14:27manufacture this stuff and the hardware
14:30and the infrastructure that ultimately,
14:31you know, most modern machine learning
14:33artificial intelligence is built off of.
14:35Most people don't understand that. Most
14:36people don't get that there's actually a
14:37complicated supply chain at the very
14:40base of AI, and that might be even more
14:42valuable and more powerful than the
14:43stuff that model labs like Anthropic
14:45and, you know, OpenAI are doing. On the
14:47data center level, you know, you have
14:48Microsoft, you have Google, you have
14:50Amazon, you have Meta, you have, uh I
14:53don't know, Oracle, you have X. You have
14:55these large companies that are now
14:57pouring billions, sometimes even on the
15:00planned order of trillions of dollars,
15:02into building plants where you will just
15:05line the walls with the chips that are,
15:08you know, invented by Nvidia, made by
15:10TSMC, and then, um you know, actually
15:13one step up the chain, machines produced
15:16by ASML. And so, the idea is to have
15:18rows and rows and rows of these, you
15:21know, chips basically running in perfect
15:23unison. Um, you know, the the the
15:25outputs of TSMC, for instance, um, to do
15:28the the processing in such ways that
15:30Nvidia prescribes. And so, you know, big
15:33companies are pouring, again, hundreds
15:34of billions of dollars into this because
15:36they know that that is where the
15:36calculations are actually going to be
15:38occurring. But those big data centers,
15:39they also don't exist in isolation,
15:41right? Because they need something to
15:43power them. Every single one of these
15:44operations requires an extraordinary
15:47amount of power in order to ultimately
15:49work at the scale and scope of which we
15:51want them to. And so, you actually have
15:53utilities and power grids. You have, you
15:55know, municipal governments that are
15:56responsible for
15:58saying yes or no to where these big data
16:00centers are going to go. And so, you
16:02know, we're just working our way through
16:04that onion right now, but hopefully you
16:05understand that's a little bit more
16:06complicated than just like, "Oh,
16:07Anthropic and OpenAI, they do
16:08everything." Speaking of, then you have
16:10the model apps. Okay, at that layer, you
16:12have companies like OpenAI, Anthropic.
16:14You have Google. You have X as well, you
16:16know, some of these companies are
16:17vertically integrated. So, they're
16:18trying to work at multiple levels of
16:19this layer. Um, you have DeepMind. You
16:21also have a couple of open-weight labs.
16:23Now, these sorts of companies, they're
16:24actually responsible for the software
16:26implementation. They're responsible for
16:27coming up with almost like the the the
16:29mathematics behind what computations
16:31will be done on the chips. Then, one
16:33step up from them, you also have, you
16:35know, the labs themselves, some of which
16:37offer APIs, but now tools like
16:39OpenRouter, which are basically
16:41intermediary layers where, you know, you
16:43make a request to them and then they
16:44dole out the requests to whatever
16:46provider makes the most financial sense
16:47at that moment in time. But what's
16:49ultimately sending most of those
16:50requests? It's not human beings just at
16:52our computers anymore. Really, it's one
16:53level up to harnesses, to automated
16:56agent frameworks that do these sorts of
16:58requests on mass. And so, these are, you
17:00know, Codexes. These are your Claude
17:02codes. These are your Cursors. And then
17:03finally, you actually have
17:04implementation, which is the people,
17:07like myself, probably like most of you,
17:10that use all of this technology, okay,
17:12in the in the AI vertical, and then take
17:15it and then apply it to other verticals
17:17out there. You know, like your your your
17:19local plumber, or like some big
17:21consulting company whose operations you
17:23are optimizing, or something like that.
17:25Maybe, you know, you're implementing
17:26into your own business, or your own
17:27life, or your hobbies, your personal,
17:28like whatever. Ultimately speaking, the
17:30implementation layer is the one where we
17:32sort of go, "So what?" And take all of
17:34the transformative outputs of these
17:36industries, and then apply it to
17:38something different. So, as somebody
17:40that, you know, has probably recently
17:41been introduced to artificial
17:42intelligence technology, you are
17:44starting at the implementation layer,
17:45which is all the way at the very top of
17:47that onion. But, you have to realize
17:48just how deep everything goes if you
17:50really want to gain an appreciation for
17:51what exactly is going on and why. Some
17:53great resources here to watch. Um on the
17:55AI harnesses specifically, I have a
17:57great video called AI Harness: An
17:59In-depth Explanation, um what a harness
18:01is, why it matters just as much as the
18:02actual underlying model. There's a
18:04wonderful, um
18:06uh video resources by Andrej Karpathy as
18:08well, called How I Use LLMs. It's
18:10basically like a deep walk-through of
18:12how a power user, like him, you know,
18:14somebody that now works at least, um as
18:15of the time of this recording, at
18:16Anthropic, uh uses actual underlying
18:18artificial intelligence.
18:20Um I'd also read a little bit about the
18:22State of AI Report. There's a lovely one
18:25called the State of AI Report 2026 by a
18:27company called Air Street Capital. It's
18:28about 45 minutes or so, almost 250
18:30slides. So, you're going to have to
18:31spend a fair amount of time on this.
18:33But, if you skim that, you'll understand
18:34a lot about what I talked about here at
18:35the deeper levels of the onion. There's
18:37also, um a great piece by Epoch, Epoch
18:40AI Trends. These are numbers on compute,
18:42how much money it costs to do things
18:43like, you know, train models, how much
18:45money is in data centers. After this,
18:46I'd spend some time learning about the
18:48specific players in the game. I mean, I
18:49touched on a few of them earlier, right?
18:51You need to know who's building frontier
18:52AI, and how their philosophies are a
18:54little bit different. You know, every
18:56one of these companies has a different
18:57philosophy. That's why they have
18:58different companies. The two main ones
19:00in the lead, at at least at the time of
19:02this recording, um OpenAI and Anthropic,
19:05themselves sort of spun off of a bunch
19:07of work that other researchers are doing
19:09at Google. You You they sort of had like
19:10ideological issues with some of the way
19:12that the research was occurring at
19:13Google. And so, that spun OpenAI. Uh a
19:16few generations into the models of
19:17OpenAI, Anthropic had some ideological
19:19issues that was going on with OpenAI, so
19:20then they spun off. And so, everything
19:23is still connected sort of on a
19:24foundational underpinning sort of level.
19:26You have to know that. And so, you know,
19:28the labs as of right now, uh you have
19:30OpenAI, massive, massive consumer
19:32company, probably one of the biggest and
19:35most lucrative artificial intelligence
19:36companies ever. And obviously the ones
19:38to first bring to light ChatGPT and make
19:40it a consumer product. I'd still say
19:42they own a fair amount of the market,
19:43and I don't want to date this video's,
19:45you know,
19:46education potential by putting specific
19:48numbers on it. But next to OpenAI, you
19:50have Anthropic. The whole idea about
19:52OpenAI is, you know, they want to get to
19:54an automated research intern. However,
19:56along the way, they've taken some
19:58detours. They've made consumer apps and
19:59consumer products, things like Sora.
20:01They've made, you know, image models.
20:03They've made like GPT image, for
20:05instance. They've um done a bunch of
20:06things in both video, now audio, voice
20:09transcription. this is sort of like your
20:10general purpose AI company. They were
20:12the first to the scene, and they
20:13obviously know that there's a large
20:14financial incentive to getting this
20:16right. Um but, you know, they're not
20:17exactly like super all in on any one
20:19thing. Speaking of companies that are
20:20all in on one thing, Anthropic. The key
20:22differentiator between Anthropic and
20:24OpenAI that most people don't realize is
20:25Anthropic is 1,000,000%
20:28all in on computer programming as a
20:30means to get to AGI and ultimately ASI.
20:33Whereas OpenAI was releasing all these
20:35cool consumer apps, they were dabbling
20:37in video, they're dabbling in photo,
20:39Anthropic was literally like, "We are
20:40text in and text out, baby. The only
20:42thing that matters is we're going to
20:43make our models the smartest programmers
20:45on planet Earth. Then we're going to
20:46make them program the next generation of
20:48the the models themselves. If wasting
20:50even 1% of those resources that, you
20:52know, we're we're spending on video
20:54takes away from our ability to reach
20:56that AGI or ASI, it will all have not
20:59have been worth it." And so, their whole
21:01sort of modus operandi is to build this
21:02better future by racing ahead to, you
21:05know, recursive self-improvement, a
21:06concept we'll talk about later, and then
21:08using that, and once they're at RSI,
21:10they'll kind of do everything else.
21:11There's also Google DeepMind. Now, this
21:13is probably one of the oldest labs out
21:15there. Um, you know, this is the one
21:17that made AlphaGo way back in the day.
21:19It's the one that, you know, sort of
21:21took the world by storm, I think,
21:22probably 7, 8, or 9 years ago, before
21:25OpenAI or anything like that was even
21:27created or invented. It was really the
21:29thing that sort of defined a generation
21:31of AI researchers, and obviously Google
21:33DeepMind is part of Google. DeepMind was
21:35its own separate lab before its
21:37acquisition. Google obviously had a lot
21:38of financial incentives to acquiring it,
21:39but um, these guys are sort of like
21:42vertically integrated into everything.
21:44They own the full stack, cuz Google also
21:45manufactures chips. It also
21:47manufactures, or rather runs, um, cloud
21:50infrastructure, so think Google Search.
21:51It has probably more data than any other
21:53one of these companies, simply because
21:55it's been a search engine for the better
21:56part of the last 15 or 20 years. A lot
21:58of people refer to them as sort of a
21:59slow giant, where, you know, when all
22:01this AI stuff initially kicked off, um,
22:04it was Google DeepMind that people were
22:06looking at and being like, "Oh my god,
22:07these guys are probably going to take
22:08over, right?" But then they just didn't
22:09do anything for a really long time. And
22:11the reality is, the bigger a company you
22:12are, the more bureaucratic your
22:14processes are, but also the more things
22:16that you have to move around in order to
22:17like actually kick-start momentum. That
22:19said, the bigger a company is, the more
22:21inertia it has. Just like a really heavy
22:24freight train, when it starts moving in
22:25a direction, although it can take some
22:26time to pick up, once it does start
22:28moving and eventually keep moving, it's
22:30very difficult to stop. There's also
22:32Meta, okay? And Meta, by Mark
22:34Zuckerberg, has been well behind the
22:35model race for quite a while, only very
22:37recently caught up. But they've
22:38championed themselves essentially being
22:40like open-source, you know, models for
22:42the people. We want everybody to have
22:44their own personal superintelligence,
22:45and so they dropped a bunch of products
22:47to that end, like Muse. And they're not
22:48afraid to say why. They want to make a
22:50ton of money with this stuff, and they
22:51think that by empowering all humans with
22:53their own superintelligences, they can
22:54create markets that have never been
22:55created before. You also have xAI with
22:58Elon Musk. Now, their big bet really is
23:00compute. Their huge bet is we have the
23:03most money ever. We also have what we
23:06need for fabrication of large data
23:08centers and compute farms um because of
23:10SpaceX, because of Tesla, and the fact
23:13that we've essentially been heading, you
23:14know, a robotics manufacturing, um
23:17specifically spaceship uh manufacturing,
23:19but also things like solar panels for
23:20the better part of the last like 10 or
23:2215 years. And so, despite the fact that
23:24they weren't really capable of competing
23:25in the model race initially, they just
23:27didn't have the, you know, model talent
23:29or the know-how, um they're now making
23:31massive strides in compute build-outs.
23:33And they're gotten to the point now
23:34where their data centers, like Colossus
23:362 in Memphis, are literally being rented
23:38out to other model labs that don't have
23:40anywhere near that much capacity. So,
23:42they're performing strategic
23:43partnerships where it matters and
23:44ultimately leading the model race in in
23:46that hardware space, at least as of the
23:48time of this recording. Now, all these
23:49labs are racing towards this idea of
23:51recursive self-improvement. In plain
23:53words, that just means that an AI is
23:55doing the research and development of
23:57AI. And it does so completely
23:59independently and autonomously without
24:01any human involvement. Whether or not
24:02it's slower or faster than human beings'
24:05own approaches, the benefit to RSI, at
24:07least from an AI research perspective,
24:09is when you have AI that is capable of
24:11actually figuring this stuff out, you
24:13can just run it arbitrarily fast and
24:14then arbitrarily scoped. So, you can
24:16spawn, you know, 100,000 of these
24:18researchers, a million of these
24:19researchers. All you really need is
24:21compute and energy. Those two are the
24:22main bottlenecks.
24:24The idea is once you have RSI, the
24:25models can begin actually optimizing
24:27their own inference and also optimizing
24:29their own training. And then the next
24:30generation of the models will be, you
24:32know, an exponentially uh greater degree
24:34faster and then cheaper to run. So,
24:36those models will then optimize the next
24:38generation of exponentially faster and
24:40cheaper to run until, I mean, honestly,
24:42people are talking about these models
24:43essentially attaining omniscience, you
24:45know, knowing everything about
24:46everything, simply because, you know,
24:47they can compute and they can think so
24:49many orders of magnitude faster than
24:51people can. And we're getting pretty
24:52close. I'm not going to date this, but
24:54suffice to say, recursive
24:55self-improvement is something that I
24:57think is probably starting to occur
24:58already, if not is in full swing right
25:00now, which is one of the main
25:01motivations behind me making this video,
25:03by the way. I thought it was important
25:04everybody understand just what the
25:06hell's going on and how quickly all of
25:07it is happening. This means that in 1
25:09year, you might see approximately 10 to
25:1220 years of progress. Try and wrap your
25:14mind around how much progress you've had
25:15in the last year, and then how much
25:16progress we would have in an
25:17exponentially increasing situation. We
25:19could be at the point where models don't
25:21drop once a month. We could be at the
25:22point where models drop literally every
25:235 minutes. And I don't say that
25:25facetiously. Anyway, some great
25:26resources to read here. The Gentle
25:28Singularity by Sam Altman. Uh it's maybe
25:3115 or 20 minutes or so. It's a very
25:33optimistic case, and ultimately
25:34speaking, it's probably a little bit
25:36biased for investment reasons, but it's
25:37an awesome piece. Um We Must Pace the
25:40Frontier by Dario Amodei. He's the CEO
25:42of um Anthropic. Uh you know, this is a
25:45cautious sort of case, whereas OpenAI is
25:47like, "Hey, everything's going to be
25:48awesome. Everything's going to be
25:49amazing." You know, Anthropic's like,
25:50"No, no, no, we got to be a little more
25:51careful." And that's the point that he's
25:53making in that piece. If you read that
25:55piece start to finish, and it can be
25:56kind of a mouthful, but it is quite
25:57important and valuable. Um alongside his
26:00other piece, Machines of Loving Grace, I
26:02think you start to understand just how
26:04maniacally focused on achieving that RSI
26:07a lot of these companies are. Time
26:08Horizon 1.1 by METR. It's a company
26:11that's heavily involved in tracking what
26:13is going on in these AI models and how
26:14quickly they are progressing. I also
26:16think it's important to understand the
26:17skeptic's case. Um here's one from MIT
26:19called AI's Recursive Self-Improvement
26:21Might Not Come So Quickly After All. And
26:23then, if you really want to understand
26:25not just the model dynamics, but the
26:28geopolitical dynamics, like I was
26:30telling you guys about earlier, read AI
26:322027. It's a detailed forecast of what
26:35is occurring over the course of the last
26:36few years and will occur over the course
26:38of the next few years, maybe into the
26:392030s or so with artificial intelligence
26:41technology. Next up, I wish I had done
26:44all of this ahead of time before I
26:45learned the tools, because I think this
26:47would have had quite the impact on, you
26:48know, what I did and why. But,
26:50eventually, when push comes to shove,
26:51you can't spend all your time in a la-la
26:53land obsessing over geopolitics or
26:55whatever, you actually got to, you know,
26:57when push comes to shove, learn a damn
26:58tool. And so, this is all about the
27:00tools. And as of the time of this
27:01recording, you have a few major ones.
27:03You have Claude, you have ChatGPT, um
27:05you have Gemini, you know, you have X's
27:08Grok, you have a few open weights models
27:11coming out of China that are competing
27:12as well. Uh but these are really the
27:14main contenders, and these are probably
27:15going to be the ones that are going to
27:16proceed. I guess you could have Muse as
27:18well, depending on how things go over
27:19the course of the next few months.
27:20Reality is, like, most of these are
27:22going to be quite similar, but you do
27:23just have to pick one and then start
27:24talking to it. In addition to the chat
27:26assistant layer, again, this is kind of
27:27like an onion, one level up, you have
27:29the agent layer, you have the harness
27:31layer. And so, the variety of different
27:33harnesses, you have Claude Code for
27:35instance, you have Codex, you have Kimmy
27:37Code, right? There are a variety of
27:38these different harnesses, which are
27:40sort of software implementations and
27:41approaches to uh these these models that
27:44sort of constrain them and control them
27:45in various directions. You'll know a lot
27:47more about this than what I'm explaining
27:48right now, because you'll have actually
27:50gone through the whole harness section
27:52of our curriculum. Finally, from there,
27:54I'd recommend learning some sort of
27:55automation tool. Um I don't mean
27:57automation tool in the traditional
27:58sense, but I think it's important to
28:00understand that the way that most people
28:02are going to be applying these models is
28:04not really just to ask them how much
28:06protein is in their yogurt, as my
28:07co-host Jack Roberts typically says. It
28:09is to ultimately make money or do some
28:12sort of economic end, you know, achieve
28:14something um in a business-type
28:16scenario. And so, the way that you do
28:18that typically is by automating
28:19processes that currently produce revenue
28:21or produce some sort of positive value.
28:23And so, automation tools typically hook
28:25up artificial intelligence models and
28:27then empower them with access to other
28:29tools to allow them to do things within
28:30a business in a reliable and predictable
28:32way. And so, personally, what I would do
28:33is I would pick a chat assistant, so
28:35maybe Claude, um a harness or an agent
28:38an agentic layer on top of that,
28:39something like Claude Code, and then
28:41finally, I'd pick an automation tool,
28:42something like, I don't know, N8N.
28:44Although these automation tools as is
28:46may not really be required over the
28:47course of next years depending on how
28:49quickly AI moves to to take over all
28:51that stuff. Some simple and easy ways to
28:53wrap your mind around prompting, which
28:55is the process of putting, you know, a
28:57request into an AI model to kick off
28:59some sort of prompt or action chain, um
29:01is delegate. Don't prompt. At the end of
29:03the day, these models actually work
29:04better if you anthropomorphize them. For
29:06those of you guys that don't know, that
29:07word just means treat them like you
29:08would treat a person. And so typically
29:10when delegating to other people, you
29:12always want to have four things. Um
29:13first is context. So who you are, what
29:15the work is for, you know, why you want
29:17to do it and so on. The second is goal,
29:19which is what the result that you want
29:21out of this. The third is what great
29:23looks like. So some sort of example or
29:25criteria that you are explicitly stating
29:27means the job is done. And then finally,
29:29how to verify that the job has been done
29:32to your, uh you know, specifications.
29:34And so maybe you verify it with like a
29:36screenshot, or maybe you verify it by
29:37comparing the pixel-perfect accuracy of
29:39the thing that it Claude generated
29:41versus the thing that, you know, you
29:42generated. Something like that. This is
29:44most of what I teach on my channel and
29:46also inside of Maker School, which is my
29:4890-day accountability program that
29:49guarantees you your first customer for a
29:51paying AI or automation service in 90
29:53days, or I'll give you all your money
29:55back. The reason why is because we walk
29:56you through how to do things like prompt
29:58engineer, how to give agents strong
30:00tasks, and so on. Anyway, ad over, you
30:02should check it out. Some things that
30:04you guys could learn or uh uh read or
30:06watch to really embody, you know, what
30:08it is to be effective with these models.
30:10First, AI Fluency uh Framework and
30:13Foundations. This is a brief 4-hour mini
30:15course by Anthropic Academy that
30:17actually walks you through everything to
30:18do with their version of their large
30:21language model, Claude, but also, you
30:23know, Claude code, what all these
30:24different tools are for, and ultimately
30:26how you can apply them to business use
30:27cases. Next up, uh 2.4 million dollars
30:30of prompt engineering hacks in 53
30:32minutes. That's by me. I spent 53
30:34minutes basically just ringing out my
30:36brain like a wet towel trying to give
30:37you guys as much value as possible. So
30:38definitely check that out below, too.
30:40And I know I'm probably going to come
30:41across as you know Anthropic uh uh a
30:44script kitty or super biased or
30:45something. But if you wanted to go a
30:47little bit deeper into this, I would
30:48also read through the interactive prompt
30:50engineering tutorial by Anthropic. It's
30:52hands-on. They actually have agents just
30:54built in directly to your your
30:55structure. You sort of like quizzes. All
30:57this stuff was built by AI. And so it's
30:59sort of like a meta example of what AI
31:01can be used for. Both education, both
31:04upskilling, and both teaching people and
31:06onboarding them to your product. All
31:07things that you can learn how to do with
31:09AI as well. Finally, how to turn all of
31:11this stuff into money. I mean it's one
31:12thing to learn how to use a tool, but
31:14how do you actually apply it to real
31:15businesses? The whole idea here behind
31:18this part of the roadmap is that AI
31:20hasn't made anything new possible. What
31:23it has done it is it has made old
31:24processes much more profitable. And it
31:27is in the making of those old processes
31:29more profitable that some new things
31:30have seemed like they become possible.
31:33But really these things were always
31:34possible. The whole idea about
31:35artificial general intelligence is that
31:37it is the same intelligence as a human
31:39being approximately in all domains,
31:40right? And so, you know, all the things
31:42an AGI can do a human being can do. It's
31:44just about the time scale. And so when
31:46people talk about AGI and like
31:48artificial intelligence producing all
31:49these new incredible things, I'm always
31:51like, "Well, the question is more was
31:53that thing profitable before AI?" You
31:55know, was it profitable before the next
31:57cool new model drop? If the answer to
31:58that question was no, it's probably
32:00still not profitable now. But if it is
32:02profitable now and you apply AI to it,
32:04now we're talking. And so, the ability
32:06to turn these wonderful developments
32:07into some sort of monetary outcome
32:09typically relies on your ability to
32:11understand what a good monetary process
32:13actually looks like. And in general, for
32:15most business, that looks like a very
32:17structured step-by-step repeatable and
32:20procedural workflow that, you know,
32:22takes some raw good. Might be, I don't
32:24know, like a person that just doesn't
32:26know anything about whatever you're
32:27selling. If you're selling education,
32:28which might be, you know, I don't know,
32:30some sort of like raw thing, maybe an
32:32iron ingot or something. And then it
32:34goes through a pipeline filled with, I
32:36don't know, a bunch of processes to turn
32:38that iron ingot into something else. And
32:39then, you know, outcomes, I don't know,
32:41some sort of like working hammer or
32:43working tool that you can use on a farm.
32:45It is in that pipeline, it is in that
32:47process, that ultimately value is
32:49delivered and you are paid. And so, you
32:51know, what you need to do in order to
32:52become really good at turning these
32:54wonderful technological advances into
32:55money, is you just need to understand
32:57what pre-existing processes made
32:58businesses money and then how can we
33:01make that a 100x faster in order to
33:03generate much more leverage with
33:05artificial intelligence. And so, the
33:06simplest and easiest way to really
33:07understand all this stuff is actually
33:08just to go out there and and try selling
33:10it to people. Ask them, "So, like, what
33:11exactly keeps you up at night? So, what
33:13exactly is your business currently doing
33:15in order to make revenue?" Wow, really?
33:17Hmm, that's interesting. I wonder if
33:18that's something an AI could do faster.
33:20Let me think. It's not just to like, you
33:22know, come up with a bunch of great
33:23ideas using artificial intelligence for
33:25the sake of great ideas, but it's
33:27actually to look at like pre-existing
33:28processes and then just implement
33:30artificial intelligence to iterate on
33:31them and make them a little bit better,
33:32a little bit faster. The cool thing is,
33:34you know, if you grow 1% better every
33:35day, you grow, I don't know, 2 or 300%
33:37better by the end of the year. AI can
33:39compress that whole year-long process
33:41into one that that legitimately only
33:42takes a few days. And as somebody that's
33:44now made, you know, over $10 million
33:46using artificial intelligence tools, it
33:48is shocking what that difference can do
33:50for, you know, a company's wallet. Some
33:52very simple and straightforward guides
33:54to this, how I would start AI consulting
33:56in 2026 if I could start over, that is
33:58by yours truly, took about half an hour
34:00for that video start to finish. Uh, if
34:02you guys like that sort of thing, then
34:04another really good one is what I'd
34:06learn instead of AI automation in 2027.
34:08That's about 20 to 25 minutes. Um, on
34:11the reading side, probably my favorite
34:12piece ever throughout this entire video
34:15is probably this one, Do things that
34:17don't scale by Paul Graham. Uh, it's
34:18about, I don't know, 10 or 20 minutes or
34:20so. Basically, it's why your first
34:21customers always come from doing
34:23unscalable things by hand, aka the sorts
34:26of things that you were always doing
34:27whether or not AI was around. And then
34:29ultimately speaking, when you identify
34:30what those things are, how you can
34:32convert that into some sort of highly
34:33leveraged AI process. Next is request
34:35for startups. This will take you just a
34:36few minutes to read, but basically it's
34:37like a place where people make requests
34:39for companies to build solutions to
34:41problems other companies are having. So,
34:42this can help put you in the mindset of
34:44somebody that's like looking for ways
34:46that you can apply artificial
34:47intelligence to things that people are
34:50currently seeing detract from their
34:51ability to make money or ultimately
34:53speaking just solve problems for them.
34:54Okay, ultimately AI opportunities are
34:56coming. In general, they're getting a
34:58lot bigger. They're also getting a lot
35:00more frequent, but the time window that
35:02you can capitalize on them is a lot
35:04shorter. And so, I wanted to make this
35:05video because I feel like by
35:06understanding the context behind all
35:08this, and really by just spending, you
35:10know, a few days if not a week, you can
35:12achieve such a well-rounded
35:14understanding of what exactly is going
35:15on here that, you know, you ultimately
35:17can actually be the person that knows
35:18whether or not an opportunity is good,
35:20whether or not these new developments
35:21are real or they're total fluff, whether
35:23or not you need to be concerned about
35:24certain impacts geopolitically, whether
35:26or not you need to worry about the job
35:27market or what have you. This is what I
35:29wish I'd learned when I had started this
35:30AI journey forever ago. And hopefully,
35:32my video alongside all the resources
35:34that I've linked down below can help
35:36take you guys to that level as well. If
35:37you guys want these resources, um just
35:39click the second link in the description
35:41to head over to Maker Zero. It's where I
35:42give everything away completely for
35:44free. This will be up at the top of the
35:45recently uploaded module inside of
35:48classroom. If you guys also want to
35:49learn how to do this for a business
35:50purpose, as mentioned, check out Maker
35:52School. That's the first link in the
35:53description. Aside from that, guys,
35:54thank you very much for watching this
35:55video all the way to the end. I really
35:57appreciate every single one of you,
35:58especially the ones that do make it this
35:59far. If you guys have any suggestions
36:01for future videos for me to make, just
36:03let me know. I'd be more than happy to.
36:05I get most of my ideas from my audience
36:06at this point. Um please do me a solid,
36:08like, subscribe, and I love every single
36:11one of you. Have a lovely rest of the
36:12day, and I'll catch all y'all in the
36:13next one.

About this transcript

This page contains the full transcript of The Last AI Roadmap Video You’ll Ever Watch by Nick Saraev, generated from the public captions YouTube serves with the video. The transcript has 8,202 words across 1,220 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.