WEBVTT
00:00:01.520 --> 00:00:04.719
Have you ever felt overwhelmed by AI?
00:00:05.120 --> 00:00:16.239
Like there's certain aspects of artificial intelligence that you barely understand to begin with, yet you're expected to use it and it's changing every day.
00:00:16.480 --> 00:00:17.920
I understand where you're coming from.
00:00:18.160 --> 00:00:28.160
Chances are, if you're listening to this, you probably have a full-time job where you're expected to leverage AI, but you haven't had much formal training, if any, at all.
00:00:28.320 --> 00:00:32.240
And you definitely don't have extra hours in your day to learn.
00:00:32.479 --> 00:00:37.679
Oh, and yeah, I mentioned that thing where AI is changing literally every single day.
00:00:37.759 --> 00:00:41.920
And when I say I understand the challenge, I mean it.
00:00:42.240 --> 00:00:51.119
I mean literally, it's my only job to use, to build with, and to teach AI every single day for 10 to 12 hours.
00:00:51.200 --> 00:00:53.520
And that's all I've been doing for the past three years.
00:00:53.679 --> 00:00:56.399
And even I find it hard to keep up.
00:00:56.640 --> 00:00:57.759
But don't worry.
00:00:58.240 --> 00:01:01.359
That's where this new series comes into play.
00:01:01.520 --> 00:01:04.159
It's called the Start Here series.
00:01:04.319 --> 00:01:12.959
And whether you're a beginner, extremely confused, or you're someone that uses AI every single day, yet you're looking to double down.
00:01:13.280 --> 00:01:16.000
This new series is for you.
00:01:16.159 --> 00:01:22.799
Because one of the most common questions I get asked all the time is, where do I start?
00:01:23.040 --> 00:01:27.760
Or someone saying, Hey, I know a lot about large language models, but I want to know more about the creative side.
00:01:27.920 --> 00:01:28.640
Where do I start?
00:01:28.799 --> 00:01:34.400
To tell the truth, I haven't had a good answer until now.
00:01:34.560 --> 00:01:38.000
That's why we're kicking off this start here series.
00:01:38.159 --> 00:01:41.200
Uh so for those keeping up keeping up live, right?
00:01:41.359 --> 00:01:47.599
January is a time when I think most business leaders are setting new goals or trying to double down on good habits.
00:01:47.680 --> 00:01:55.599
So I know a lot of you are really trying to put in that extra effort here at the beginning of the beginning of the year to improve your understanding of AI.
00:01:55.760 --> 00:02:02.560
Uh, so that's why we're going to be releasing probably two of these episodes a week for the next five or so weeks.
00:02:02.719 --> 00:02:06.799
So, not necessarily abandoning our normal daily schedule.
00:02:06.879 --> 00:02:20.240
If you're an avid listener to the program, don't worry, we're just flexing a little bit here in the beginning of the year to try to help both beginners and advanced users alike get started and also get caught up.
00:02:20.400 --> 00:02:26.319
So we're going to be going over the basics like generative AI today, like what the heck is it?
00:02:26.479 --> 00:02:38.879
Um, to simplifying more advanced techniques like AI agents and expanding, uh explaining things like the model context protocol or what the health is a Ralph Wiggham loop, right?
00:02:38.960 --> 00:02:50.000
So whether it's concepts that we're zooming out and going back in time or things that are happening literally today and helping you put them in perspective, the start here series is gonna be for you.
00:02:50.159 --> 00:03:02.240
So we're gonna be going over, you know, the from the correct way to launch a successful AI pilot to how to measure ROI to how to read and understand benchmarks and choose which AI system is best for you or your company.
00:03:02.400 --> 00:03:06.479
We're gonna be tackling it all here in the Start Here series.
00:03:06.800 --> 00:03:07.599
You ready?
00:03:07.919 --> 00:03:09.439
All right, let's get into it.
00:03:09.599 --> 00:03:10.639
What's going on, y'all?
00:03:10.719 --> 00:03:12.719
If you're new here, my name's Jordan Wilson.
00:03:12.800 --> 00:03:14.960
In everyday AI, it's for you.
00:03:15.039 --> 00:03:17.360
And we've been doing it for a very long time.
00:03:17.520 --> 00:03:32.240
It is an unedited, unscripted daily live stream, podcast, and free daily newsletter, helping everyday business leaders like you and me not just keep up, but how we can leverage all the good stuff, make sense of the nonstop updates, and get ahead to grow our company and our career.
00:03:32.479 --> 00:03:37.759
So if that sounds like what you're trying to do, maybe this is episode number one for you.
00:03:37.840 --> 00:03:39.919
Maybe it's episode 700.
00:03:40.080 --> 00:03:40.960
It doesn't matter.
00:03:41.120 --> 00:03:43.360
We have something new and fresh for all of you.
00:03:43.520 --> 00:03:45.439
Yeah, a new URL to throw out there.
00:03:45.599 --> 00:03:45.919
Ready?
00:03:46.080 --> 00:03:50.560
So if you're interested, go to the go to starthearseries.com.
00:03:50.639 --> 00:03:52.719
That is starthearseries.com.
00:03:52.800 --> 00:03:56.080
So if you want to keep up with this particular series, go do that.
00:03:56.159 --> 00:04:05.520
You're gonna get a uh an invite link to sign up for our free community and you will be inserted directly into the Start Here Series uh kind of onboarding flow.
00:04:05.599 --> 00:04:13.360
So as we add to this, um there's gonna be more and more shows in there in that space, the dedicated space in our community.
00:04:13.520 --> 00:04:18.560
So if you want to keep up with all the shows in this series, uh connect with other leaders.
00:04:18.720 --> 00:04:24.079
I mean, literally, we have industry leaders in our free community, uh, that you can learn from.
00:04:24.240 --> 00:04:27.199
Go connect with, make sure you go to that uh website.
00:04:27.360 --> 00:04:32.399
All right, now we have that out of the way, just to let you know, these shows are gonna be a little quicker.
00:04:32.480 --> 00:04:37.600
All right, this one's uh, you know, might be 30 minutes, but most of these start here shows are gonna be about 20 minutes.
00:04:37.680 --> 00:04:42.639
I want to keep them very fast, very factual, which you know is gonna be a little hard for me.
00:04:42.800 --> 00:04:46.720
All right, so uh without further ado, let's get into it.
00:04:46.800 --> 00:04:54.959
Let's talk about generative AI, the basics, how it works, and why it matters in 2026 more than ever.
00:04:55.680 --> 00:04:59.519
So let's start with the pace in the reality.
00:04:59.759 --> 00:05:03.519
Well, nothing has spread this fast ever.
00:05:03.839 --> 00:05:05.920
Uh, nearly 900 million people.
00:05:06.079 --> 00:05:08.959
The last confirmed was 800 million, but I know the stat.
00:05:09.120 --> 00:05:15.040
It's nearly 900 million people are using Chat GPT weekly.
00:05:15.120 --> 00:05:17.600
I mean, that's more than the entire population of Europe.
00:05:17.759 --> 00:05:20.959
And Chat GPT, let's talk about an explosion.
00:05:21.120 --> 00:05:25.839
It reached 100 million users in its first two months.
00:05:26.480 --> 00:05:31.920
To get to that same number, the internet took eight years.
00:05:33.360 --> 00:05:40.240
So if you want to talk about a technology that you can't ignore, think about how commonplace the internet is in our day-to-day lives.
00:05:40.399 --> 00:05:43.680
You can't do too much without it, right?
00:05:43.920 --> 00:05:47.680
Especially if you're a knowledge worker, uh, you can't do too much without it.
00:05:47.920 --> 00:05:54.720
And just two years after launch, 40% of working age Americans are using generative AI.
00:05:55.040 --> 00:05:59.600
And the internet took seven years to reach that same level.
00:05:59.920 --> 00:06:13.600
So the generative AI technology, which is what large language models are kind of under the umbrella on, it is the most explosive growth of any technology ever.
00:06:16.319 --> 00:06:17.759
And companies are using it.
00:06:18.319 --> 00:06:21.040
Whether you have realized this or not, right?
00:06:21.360 --> 00:06:25.040
This conversation has changed a whole lot over the past few years.
00:06:25.279 --> 00:06:28.800
But nowadays, using AI is table stakes.
00:06:28.959 --> 00:06:30.160
Like you have to.
00:06:30.319 --> 00:06:31.839
You don't have a choice, right?
00:06:32.000 --> 00:06:34.879
Maybe three years ago it was a competitive advantage.
00:06:35.040 --> 00:06:45.680
You know, it was kind of novel, you know, to be using generative AI and large language models back in, you know, 21 uh or in 2022 when ChatGPT came out.
00:06:45.920 --> 00:06:48.800
It's not anything special today.
00:06:48.959 --> 00:06:51.439
You have to be using it today.
00:06:51.600 --> 00:06:59.120
Uh right now, about 80% of companies are even deploying AI agents that take actions, not just answer questions.
00:06:59.360 --> 00:07:08.399
So uh not only are 92% of Fortune 500 companies, you know, as an example, using open AI's technology, that's just one company, right?
00:07:08.720 --> 00:07:12.959
But 80% of companies are even deploying AI agents.
00:07:13.279 --> 00:07:13.680
All right.
00:07:13.839 --> 00:07:16.240
So yes, the space moves quickly.
00:07:16.399 --> 00:07:17.920
And yes, we are gonna zoom out.
00:07:18.000 --> 00:07:23.839
But I first just wanted to set the table, so to speak, and let you know where we are, just about right.
00:07:24.000 --> 00:07:26.720
And depending on where you're you're tuning in from, right?
00:07:26.800 --> 00:07:28.480
Yeah, we have listeners from all over the world.
00:07:28.560 --> 00:07:29.519
So thank you for that.
00:07:29.839 --> 00:07:33.040
But here in the US, that's usually the lens I'm talking through.
00:07:33.199 --> 00:07:34.399
I'm from Chicago.
00:07:34.639 --> 00:07:35.839
Hey, good to meet you.
00:07:35.920 --> 00:07:36.319
Uh, right.
00:07:36.480 --> 00:07:38.480
A lot of you are starting here with this episode.
00:07:38.720 --> 00:07:44.399
Um, everyone's using AI here in the US, every single company, right?
00:07:44.480 --> 00:07:45.759
You don't see a company anymore.
00:07:45.920 --> 00:07:50.879
I haven't met a company in a very long time that's not using whether officially or unofficially, right?
00:07:50.959 --> 00:07:57.279
Yeah, there's the whole shadow AI, uh, you know, in in some of those things, but every company is using AI.
00:07:57.439 --> 00:08:01.279
Um, so it's changed though.
00:08:01.600 --> 00:08:12.720
And I think, as strange as it is, I think a lot of people still have a very 2022 view of AI, right?
00:08:12.800 --> 00:08:18.959
Let's just use ChatGPT because that is the most widely uh used AI tool in the world, right?
00:08:19.040 --> 00:08:25.360
Obviously, uh Microsoft Copilot, very popular in the enterprise, Google Gemini, Anthropic Claude, right?
00:08:25.439 --> 00:08:27.839
Those are what I uh refer to as the big four.
00:08:27.920 --> 00:08:31.199
So if you ever hear me reference the big four, I'm not talking about consulting.
00:08:31.279 --> 00:08:33.840
I'm talking about those uh four companies.
00:08:34.000 --> 00:08:39.200
But uh now they're not just these friendly chatbots anymore.
00:08:39.519 --> 00:08:43.600
These AI systems are operating systems in and of themselves.
00:08:43.679 --> 00:08:44.879
I've been saying this for years.
00:08:45.120 --> 00:08:46.639
Companies need to make a decision.
00:08:46.720 --> 00:08:54.159
You need to move all your operating, your day-to-day operations into a large language model sooner rather than later, right?
00:08:54.240 --> 00:09:02.080
We'll probably explore that more in a later uh start here series, the exact process of uh choosing and setting up an AI operating system.
00:09:02.240 --> 00:09:08.000
It's not an official term, it's something I think I made it up a couple of years ago and I've been running with it ever since, the AI OS, right?
00:09:08.080 --> 00:09:14.240
But in the same way companies, you know, in the 90s or 80s or early 2000s depends, right?
00:09:14.399 --> 00:09:19.039
They gave all their employees computers at some point and they had to make a choice, right?
00:09:19.279 --> 00:09:21.279
Are we a Windows organization?
00:09:21.440 --> 00:09:23.120
Are we a Mac organization?
00:09:23.279 --> 00:09:24.559
Are we a Linux organization?
00:09:24.639 --> 00:09:26.159
You have to make the same choice.
00:09:26.399 --> 00:09:27.919
What is your company going to be?
00:09:28.000 --> 00:09:36.399
But now, in a few clicks, your entire organization's data can be accessed instantly in these AI operating systems, right?
00:09:36.559 --> 00:09:39.039
ChatGPT, they have a Teams and an enterprise plan.
00:09:39.120 --> 00:09:41.519
Gemini has a business and an enterprise plan.
00:09:41.679 --> 00:09:46.960
Obviously, Microsoft 365, uh, Copilot has an enterprise plan, Claude has an enterprise plan.
00:09:47.120 --> 00:09:52.240
These are for teams now, and they bring your data in instantly, and teams can collaborate, right?
00:09:52.399 --> 00:09:53.360
Seamlessly.
00:09:53.519 --> 00:09:56.159
So with no knowledge, even right now, right?
00:09:56.240 --> 00:09:58.720
Here's here's some stuff, so some new stuff, right?
00:09:58.960 --> 00:10:02.000
Uh yes, I know this is gonna age, right?
00:10:02.080 --> 00:10:10.240
If you're listening to this in June or July, but you know, Anthropic just came out with a tool for their desktop program called uh Claude Cowork.
00:10:10.559 --> 00:10:15.919
So with no knowledge at all, no tech know-how, you can use a desktop program.
00:10:16.080 --> 00:10:23.279
Let me say this again: you can use a desktop program uh using their Opus 4.5 model, one of the most powerful models in the world.
00:10:23.519 --> 00:10:32.960
It can control your computer, it can access your file system, it can it can browse the internet in your browser that's logged into everything.
00:10:34.879 --> 00:10:36.399
A large language model can do that.
00:10:36.480 --> 00:10:37.919
It's not just a friendly chat bot.
00:10:38.080 --> 00:10:39.759
That is the work us humans do.
00:10:39.919 --> 00:10:47.519
It can access my computer, it can access uh the terminal, it can code, it can access all my files on my local machine.
00:10:47.679 --> 00:10:53.360
It can go out and log into my email, log into any system, and access any data that I tell it to.
00:10:53.600 --> 00:10:54.960
That's where we are now.
00:10:55.120 --> 00:10:58.879
Today's generative AI is much more than friendly chatbots.
00:10:59.440 --> 00:11:01.279
But let's zoom off to understand.
00:11:01.519 --> 00:11:09.679
Because I think for whatever reason, people think of AI as a very new technology, which I think is one of the reasons why there's a high distress.
00:11:10.080 --> 00:11:12.559
The other reason why is most people don't know how to use it.
00:11:12.720 --> 00:11:25.440
And you see someone share some, you know, someone puts a screenshot of something from Chat GPT on on Facebook or LinkedIn, and someone usually that has no clue what they're doing, and it's uh an example of a large language model getting something very simple, very raw.
00:11:25.919 --> 00:11:29.360
FYI, anytime you see that, I've been doing this for a long time.
00:11:29.600 --> 00:11:32.720
99 times out of 100, it is human error.
00:11:32.799 --> 00:11:35.600
That human has no clue what they're doing, uh right.
00:11:35.840 --> 00:11:37.440
But AI is not new.
00:11:37.759 --> 00:11:41.919
AI has been around in some way, shape, or form for decades, up to 70 years.
00:11:42.000 --> 00:11:45.840
So the term artificial intelligence was actually coined in the 50s.
00:11:45.919 --> 00:11:49.759
And we've been kind of chasing this, you know, dream ever since.
00:11:49.840 --> 00:11:57.519
So there were early systems, uh, the expert systems in the 70s and the 80s, uh, right, used in a lot of different sectors.
00:11:57.679 --> 00:12:09.039
Uh, I think banking, uh, mortgages were some of the bigger ones, obviously in healthcare, but even early systems in the 70s uh were shown to be able to diagnose infections better than most doctors, right?
00:12:09.120 --> 00:12:13.679
And we obviously, with the advent of large language models, we have a lot more studies that show that.
00:12:13.840 --> 00:12:16.720
Um, so what is changed?
00:12:16.799 --> 00:12:18.720
It's not the ambition, right?
00:12:18.879 --> 00:12:23.600
It's finally there's the technology has evolved a lot.
00:12:23.759 --> 00:12:36.639
Uh, you know, there were a couple of breakthroughs in the 2010s, but obviously the big one that most people are familiar with is kind of the advent and the uh prolification of the transformer um that has led to Chat GPT.
00:12:37.279 --> 00:12:42.960
So the biggest thing that's changed is well, the technology and the compute needed, right?
00:12:43.039 --> 00:12:53.840
In the same way, how to, you know, access the internet, you know, in the early days, you had to have a computer bigger than a car, uh, right, and connect it up to all these cables.
00:12:53.919 --> 00:12:55.759
And now your cell phone obviously can.
00:12:55.840 --> 00:12:57.759
So the same thing is true with AI, right?
00:12:57.840 --> 00:13:02.480
These systems that used to be very slow and very big and very expensive, right?
00:13:02.639 --> 00:13:04.000
It's not like that anymore.
00:13:04.159 --> 00:13:10.559
The technology has changed, it's gotten exponentially uh faster, more affordable, and more powerful.
00:13:10.720 --> 00:13:17.120
Um, but probably the biggest breakthrough that led us to where we are today is in 2017.
00:13:17.279 --> 00:13:24.480
So uh that's when Google Researchers published the very popular research paper, Attention is all you need.
00:13:25.039 --> 00:13:33.600
And it quietly changed the trajectory, technically, not just of the artificial intelligence technology, but I would say of the world, right?
00:13:33.679 --> 00:13:41.200
The other thing I didn't really touch on artificial intelligence right now, and we maybe we'll tackle this in another start here series.
00:13:41.840 --> 00:13:46.159
It has become, I'd say, probably more important than oil.
00:13:46.480 --> 00:13:52.720
It's probably become more important than um countries military, uh, right.
00:13:52.879 --> 00:13:54.559
Whether you realize it or not, right?
00:13:54.720 --> 00:14:00.639
Maybe if you come back and listen to the show in 2027 or 2028, you'll be like, oh, okay, this weird guy was right.
00:14:00.799 --> 00:14:02.240
It is more important, right?
00:14:02.399 --> 00:14:10.320
That's why you're seeing uh a lot of geopolitical tensions heightened uh right now around the technology itself, uh, right.
00:14:10.559 --> 00:14:20.159
Essentially, you know, there's this thing called AI, there's this thing called AGI or artificial general intelligence, and then there's this thing called ASI, artificial superintelligence, right?
00:14:20.240 --> 00:14:22.320
That's when you start to get into the terminator thing.
00:14:22.480 --> 00:14:25.919
But uh we'll tackle this in a later start here series show.
00:14:26.000 --> 00:14:35.679
But essentially, the first uh, you know, country or company to develop AGI, artificial general intelligence, that is big, right?
00:14:35.840 --> 00:14:43.759
Anyways, it started with this paper, and that has kind of changed the course of the world and the business world, especially.
00:14:43.919 --> 00:14:49.679
But that paper uh essentially introduced the T in GPT, right?
00:14:49.759 --> 00:14:52.480
Uh uh generative preformed transformers.
00:14:52.559 --> 00:15:01.039
But the T in that paper was introducing the transformer architecture, which is the engine behind almost every single AI model today.
00:15:01.200 --> 00:15:04.639
Uh, and open AI was kind of the first to run with it, right?
00:15:04.799 --> 00:15:08.399
You know, a lot can be said with the current race between OpenAI and Google.
00:15:08.480 --> 00:15:11.200
They're definitely the front runners in the AI race.
00:15:11.360 --> 00:15:14.960
You know, Claude, uh Infrapex Claude is right there as well.
00:15:15.039 --> 00:15:19.919
Um, but OpenAI was the first to take this uh research and really productize it.
00:15:20.000 --> 00:15:32.559
So they built GPT one in 2018, GPT two in 2019, GPT-3 in 2020, which is actually when they started releasing it to the public uh with a lot of earlier software programs.
00:15:32.720 --> 00:15:39.919
That's when, you know, myself and my team, we started using it daily in 2020, uh, two years before ChatGPT came out.
00:15:40.000 --> 00:15:49.200
But most of the world kind of figured out about this big breakthrough that no one saw coming that originated in the 2017 Google Attention is all you need paper.
00:15:49.360 --> 00:15:51.600
It really came to life in November 2022.
00:15:51.679 --> 00:15:53.279
That was the Chat GPT moment.
00:15:53.440 --> 00:16:02.720
That is the line in the sand that I think has changed uh the course of not just the business world, but also the information world.
00:16:04.559 --> 00:16:08.080
So, how the heck do large language models work?
00:16:08.159 --> 00:16:13.039
All right, so you're like artificial intelligence has been around for a long time, right?
00:16:13.120 --> 00:16:15.600
And then there's this important thing called the transformer.
00:16:15.759 --> 00:16:20.320
Well, and then the transformer uh led to large language models.
00:16:20.480 --> 00:16:23.519
So here's kind of what they are and how they work, uh, right.
00:16:23.600 --> 00:16:30.080
So they predict essentially these large language models have been trained on the history of the internet, uh, right.
00:16:30.240 --> 00:16:32.960
There's a lot of lawsuits that are gonna come from that, right?
00:16:33.039 --> 00:16:35.440
Because people are like, well, what about copyrighted information?
00:16:35.600 --> 00:16:38.159
Yeah, uh it's gonna play out in the courts eventually.
00:16:38.320 --> 00:16:50.080
Uh, but these large language models essentially scrape not just everything on the internet that we use, the open web, but they scrape everything on the closed web, uh offline data sets, right?
00:16:50.240 --> 00:16:56.960
Essentially the entirety of human history and human knowledge has been scraped by these large language models.
00:16:57.120 --> 00:17:03.759
Then you have very smart humans uh at these big companies that then train the models and they go through a process.
00:17:03.840 --> 00:17:05.119
I'm not gonna get too dorky.
00:17:05.359 --> 00:17:09.279
It's called reinforcement learning with human feedback, but they train these models, right?
00:17:09.440 --> 00:17:17.759
So they say, hey, when you know someone asks a question about uh, you know, topic A, here's what a good output would be.
00:17:17.839 --> 00:17:20.880
Here's answer A is good, answer B is bad, right?
00:17:21.039 --> 00:17:24.240
And so they go through this reinforcement learning process to train the models.
00:17:24.319 --> 00:17:30.480
And these models are trained to be a helpful assistant, essentially with this, uh, their training data.
00:17:30.720 --> 00:17:31.200
All right.
00:17:31.359 --> 00:17:34.480
And this is how the earlier uh models worked.
00:17:34.640 --> 00:17:37.759
Today's models are much more sophisticated.
00:17:37.839 --> 00:17:52.480
Uh, right now, if you ever hear uh a term like scaffolding or uh agenc, uh right, today's large language models are night and day difference uh than the large language models that set off the Chat GPT uh craze, right?
00:17:52.559 --> 00:17:56.160
Those were kind of just, I call them old school transformers, right?
00:17:56.319 --> 00:17:59.599
Uh today's models, they're technically still transformers, right?
00:17:59.759 --> 00:18:01.680
But I call them uh reasoners, right?
00:18:01.759 --> 00:18:04.160
A lot of people call them reasoners or logic-based models.
00:18:04.240 --> 00:18:05.759
So they're much different now.
00:18:05.920 --> 00:18:10.559
Uh, there's a process that um they can kind of mimic human logic.
00:18:10.720 --> 00:18:22.720
So yes, they are still technically uh next token prediction machines, but there's uh essentially a uh step-by-step problem solving on top of that prediction engine that mimics human logic, right?
00:18:22.799 --> 00:18:29.839
And you can see it think uh step by step and go through and you know, use these different tools it has at its disposal.
00:18:30.000 --> 00:18:35.119
It might run code, it might go, you know, before it responds to you, these models that think now, right?
00:18:35.200 --> 00:18:37.680
They used to just spit out very quickly and answer, right?
00:18:37.759 --> 00:18:41.759
The earlier models, the GPT-2, the GPT-3 were really bad, uh right.
00:18:41.839 --> 00:18:49.279
The GPT-3-5 that launched with um, you know, Chad GPT a little bit better, it was at least coherent and usually accurate, right?
00:18:49.359 --> 00:18:52.640
But today's models, extremely impressive, right?
00:18:52.720 --> 00:18:55.519
They tackle problems like experienced human.
00:18:55.599 --> 00:19:02.400
Today's models score better on offline IQ tests than 99% of humans.
00:19:02.480 --> 00:19:05.680
They are literally in the top 1%, uh uh, right?
00:19:05.839 --> 00:19:12.799
They are at genius level scores, uh, taking offline uh IQ tests that they have not been trained on.
00:19:12.960 --> 00:19:13.279
All right.
00:19:13.440 --> 00:19:19.039
So that's where uh kind of large language models are today, how they kind of work, uh right.
00:19:19.200 --> 00:19:26.559
So scrape data uh from data sources, humans train them uh to be helpful assistance for users.
00:19:26.799 --> 00:19:29.039
A user asks a model a query.
00:19:29.119 --> 00:19:35.920
Uh, the model might just pull that out from its training data, or it might decide, hey, I need to use some tools at my disposal.
00:19:36.079 --> 00:19:38.160
I should probably go fetch something on the web.
00:19:38.400 --> 00:19:41.519
It seems like the user is asking about something very current.
00:19:41.599 --> 00:19:48.640
Um, maybe I should use a data analysis, maybe I should run some code under the hood to get to this, uh, to get to this answer.
00:19:48.799 --> 00:20:02.799
So, yes, uh this uh kind of this scaffolding or the tools that models use uh help it hallucinate less and just provide much more robust and impressive outputs.
00:20:03.279 --> 00:20:06.720
So uh the scale is hard to comprehend, right?
00:20:06.799 --> 00:20:09.200
We're talking about billions and trillions of parameters.
00:20:09.359 --> 00:20:14.960
So a parameter is essentially a learned pattern that these models have gone through in their training data.
00:20:15.039 --> 00:20:19.759
So more parameters just means that there's more patterns that the models can uh recognize and use.
00:20:19.920 --> 00:20:24.160
I like to think of it's it's a connection in its big neural network brain.
00:20:24.240 --> 00:20:34.000
Uh so GPT 3 had 175 billion parameters, GPT-4 reportedly had over uh about almost two trillion parameters.
00:20:34.079 --> 00:20:41.119
And a lot of today's newer models, we don't know how many, but we've seen reports that they're anywhere from two to four trillion parameters.
00:20:41.279 --> 00:20:41.599
All right.
00:20:41.759 --> 00:20:45.519
So it's just uh an insane amount of data.
00:20:45.680 --> 00:20:49.200
And then there's also something called the context window, and even that has exploded.
00:20:49.279 --> 00:20:54.640
And that's essentially how much a large language model can remember before it starts to forget, uh, right.
00:20:54.720 --> 00:21:02.160
And we do in our in our free, you know, uh Prime Prompt Polish course uh that you can access in our community, by the way.
00:21:02.240 --> 00:21:14.079
So you can just go to starthirseries.com, uh, sign up for our community, and you will also get instant access to that free uh prompt engineering, context engineering course that's been taken by more than 15,000 people.
00:21:14.319 --> 00:21:18.079
Anyways, a tote uh context window is extremely important.
00:21:18.160 --> 00:21:25.519
And that's another thing that has scaled alongside with models because the earlier versions, they would forget things almost right away, right?
00:21:25.759 --> 00:21:29.839
Today's you can work with them sometimes, depending on how you have it set up.
00:21:30.000 --> 00:21:35.680
You can work with them for hours or days until they start to forget things, right?
00:21:35.759 --> 00:21:38.880
And there's obviously uh some um autocompaction.
00:21:39.039 --> 00:21:44.400
Uh, you know, uh these models are starting to work uh kind of automatically compacting this information.
00:21:44.480 --> 00:21:47.759
So it keeps a longer uh kind of memory going.
00:21:48.640 --> 00:21:52.720
So generative AI isn't just text anymore, right?
00:21:52.880 --> 00:21:59.039
So large language models started as mainly text based models, uh, but now they're multimodal by default.
00:21:59.119 --> 00:22:06.240
And there's a lot of different Uh kind of techniques, but everything kind of now falls under under this generative AI umbrella, right?
00:22:06.319 --> 00:22:10.960
Where traditional uh artificial intelligence was more deterministic, right?
00:22:11.039 --> 00:22:14.160
It was based on more uh decision trees, right?
00:22:14.319 --> 00:22:17.519
If else or if this then logic, uh right.
00:22:17.680 --> 00:22:19.519
So uh traditional AI, right?
00:22:19.599 --> 00:22:27.119
The expert systems in the in the 80s and used through the 90s, et cetera, uh, a lot of it was more deterministic, right?
00:22:27.200 --> 00:22:28.720
It was very rule-based.
00:22:28.880 --> 00:22:31.279
That's not large language models, right?
00:22:31.440 --> 00:22:37.599
Uh, you know, people say hallucinations uh or creativity is a feature, not a bug.
00:22:37.759 --> 00:22:38.319
Um, right.
00:22:38.480 --> 00:22:40.400
That's what this next token prediction.
00:22:40.559 --> 00:22:47.200
And what that means is, right, like if think if a little kid goes to touch a stove, what do you say?
00:22:47.599 --> 00:22:49.279
You say, oh, be careful.
00:22:49.440 --> 00:22:49.920
That's hot.
00:22:50.079 --> 00:22:50.720
Oh, that's hot.
00:22:50.799 --> 00:22:51.440
Don't touch that.
00:22:51.599 --> 00:22:52.640
Be careful, right?
00:22:52.880 --> 00:22:55.119
That's what you would naturally say, right?
00:22:55.359 --> 00:22:58.880
Humans, I think, very much are like large language models.
00:22:58.960 --> 00:23:08.799
So when people say, oh, large language models, they can't, they can't reason, well, they they kind of can because it's based on the reasoning of millions of humans in theory.
00:23:08.960 --> 00:23:09.519
Uh, right.
00:23:09.680 --> 00:23:12.559
But the same thing can be said for different types of models.
00:23:12.640 --> 00:23:18.799
Uh, on the image side, you have very popular uh, you know, AI image generators as well.
00:23:18.960 --> 00:23:21.839
You know, three, four years ago, they were very bad.
00:23:21.920 --> 00:23:24.880
They didn't even look like you couldn't even tell what something was.
00:23:25.279 --> 00:23:34.480
You know, some of the earlier versions, like, you know, Dolly 2 or some of the earlier versions of Midjourney, today's AI models, you can't tell the difference, right?
00:23:34.640 --> 00:23:35.920
I've mentioned this before.
00:23:36.079 --> 00:23:37.839
Uh, I used to do a lot of photography.
00:23:37.920 --> 00:23:39.599
I've taken more than a million photos.
00:23:39.680 --> 00:23:45.200
Uh, I've owned like, I don't know, eight different DSLRs, you know, professional cameras.
00:23:45.359 --> 00:23:46.720
I can't tell the difference.
00:23:46.960 --> 00:23:52.640
And if and if anyone tells you today, I think maybe three months ago, you could tell the difference.
00:23:52.799 --> 00:23:54.000
Today, no one can.
00:23:54.160 --> 00:23:54.400
Right.
00:23:54.480 --> 00:23:56.480
So, yes, there's different types of AI.
00:23:56.720 --> 00:23:58.000
It's not just text.
00:23:58.160 --> 00:24:02.880
Uh, there's, you know, text to music that's really good, you know, Suno V5.
00:24:03.200 --> 00:24:13.440
Uh, there's text to video, great companies out there, runway, uh, their Gen 4.5, Google VO3.1, uh, Sora 2, a lot of Chinese models, right?
00:24:13.519 --> 00:24:15.759
Uh, so it's not just text.
00:24:16.000 --> 00:24:17.920
It is multimodal.
00:24:18.160 --> 00:24:20.240
These models are multimodal by default.
00:24:20.400 --> 00:24:28.640
So uh, you know, text, images, music, soundscapes, sound effects, writing code, uh, it is all over the place.
00:24:29.680 --> 00:24:32.240
And it works, right?
00:24:32.640 --> 00:24:38.160
Yeah, maybe you read a bad headline recently that said, you know, 95% of AI pilots failed.
00:24:38.240 --> 00:24:38.720
They don't.
00:24:38.799 --> 00:24:40.000
That was uh marketing.
00:24:40.079 --> 00:24:41.200
Uh the ROI is true.
00:24:41.279 --> 00:24:54.000
If you look at real studies, right, that talk to thousands or tens of thousands uh of business leaders, almost every single reputable study in the world shows that the ROI return on investment of AI is exponential.
00:24:54.240 --> 00:25:03.119
So, as an example, the International Data Corporation found that companies get$3.70 back for every$1 invested in generative AI.
00:25:03.279 --> 00:25:08.400
And top performers are seeing a$10 plus uh dollar return per dollar invested.
00:25:08.480 --> 00:25:20.880
Uh Snowflake and uh ESG Enterprise Strategy Group survey uh of more than of almost 2,000 business leaders show that 92% of early adopters say their AI investments are already paying for themselves.
00:25:21.039 --> 00:25:24.400
So yeah, already positive uh ROI.
00:25:24.559 --> 00:25:38.160
And then uh similarly, other studies show that up to 98% of the same companies uh are that have previously invested in AI are looking to increase their investment.
00:25:38.640 --> 00:25:42.240
98% are increasing their investment, right?
00:25:42.319 --> 00:25:46.480
So it's no longer experimentation like it was in 2023.
00:25:46.640 --> 00:25:47.920
Now it's all about scale.
00:25:48.079 --> 00:25:52.079
Now it's all about, oh, wait, we can crush our competitors with AI.
00:25:52.240 --> 00:25:54.960
Unfortunately, it's about reducing headcount.
00:25:55.039 --> 00:25:56.319
We'll get to that in a little bit.
00:25:56.640 --> 00:25:59.519
But the returns are there and they are undeniable.
00:25:59.599 --> 00:26:04.079
And the economic skit stakes as well are massive, right?
00:26:04.319 --> 00:26:09.759
Uh, like I said, never trust small-scale studies or studies that have uh uh an agenda.
00:26:10.160 --> 00:26:16.640
But um, when you talk about does AI work, I mean, look at Anthropic's uh November study.
00:26:16.799 --> 00:26:19.119
Uh 100,000 real conversations.
00:26:19.200 --> 00:26:23.519
They found that AI reduces task completion time by 80%.
00:26:24.079 --> 00:26:33.839
Uh there was a McKinsey Digital study a couple of years ago, same thing, uh, said between 75 to 80% time savings on standard knowledge work tasks.
00:26:34.079 --> 00:26:42.559
Uh PWC study of 50,000 workers globally said 92% of daily AI users report productivity gains.
00:26:43.039 --> 00:26:43.440
Right?
00:26:43.599 --> 00:26:46.160
92% are reporting productivity gains.
00:26:46.400 --> 00:26:51.839
And I want to talk to the other 8% and teach them a couple of things, uh, but these stats are undeniable.
00:26:52.079 --> 00:26:57.519
These large language models do the tasks that used to, right?
00:26:57.680 --> 00:27:01.839
I remember tasks I used to do 15 years ago, a lot of researching, right?
00:27:02.000 --> 00:27:14.079
Looking, having 20 tabs open, you know, reading these PDFs, grabbing information out, personalizing it, synthesizing it, uh, putting it in spreadsheets, maybe then eventually turning those spreadsheets into a PowerPoint, something like that.
00:27:14.240 --> 00:27:18.880
Those are projects that would take 40, 50, 60, 100 hours.
00:27:19.200 --> 00:27:21.759
I literally timed this the other day.
00:27:22.240 --> 00:27:24.400
Uh takes like 10 minutes now.
00:27:24.640 --> 00:27:30.079
One model, one prompt, if you know what you're doing, can do that entire process in 10 minutes.
00:27:30.319 --> 00:27:35.599
And it's about 99.7% factual if you know what you're doing, right?
00:27:35.839 --> 00:27:40.720
Today's models are completely unrecognizable from the Chat GPT of 2022.
00:27:40.880 --> 00:27:48.960
And you might be thinking, wait, if all these AI models are that good, what's happening to jobs?
00:27:49.279 --> 00:27:49.920
Right.
00:27:50.480 --> 00:27:51.599
Let me just tell you this.
00:27:51.759 --> 00:27:55.839
If you're brand new to the show, I'm a realist, right?
00:27:56.079 --> 00:28:04.799
I'm not someone drinking this AI Kool-Aid and being like, oh, you know, none of us are gonna have to work and we're all gonna live, you know, this utopian AI dream.
00:28:04.880 --> 00:28:05.920
No, I don't think so.
00:28:06.000 --> 00:28:08.079
I think it it might get a little bad.
00:28:08.240 --> 00:28:09.039
Um, right.
00:28:09.119 --> 00:28:10.799
And we're already starting to see that.
00:28:10.880 --> 00:28:12.079
Talk to recent grads.
00:28:12.160 --> 00:28:13.680
How many recent grads do you know?
00:28:13.839 --> 00:28:19.839
It's hard to get a job because companies just aren't hiring anymore, especially the companies that have figured out AI, right?
00:28:19.920 --> 00:28:21.279
So uh some stats here.
00:28:21.359 --> 00:28:28.640
Uh so only 30% of graduates in 2025 secured a job in their field, and that's down from 41% the year before.
00:28:28.799 --> 00:28:31.039
That is a huge year-over-year drop.
00:28:31.119 --> 00:28:36.720
And entry level hiring is down 44% from its peak three years ago.
00:28:36.960 --> 00:28:43.599
So entry-level hiring going down nearly 50% is catastrophic.
00:28:43.759 --> 00:28:50.960
And just the number of uh graduates not being able to secure a job in their field, that's uh worrisome as well.
00:28:51.039 --> 00:29:00.480
Uh, right now, 62% of employers say that candidates should have AI knowledge, but 55% of graduates say their degree programs didn't prepare them.
00:29:00.640 --> 00:29:12.880
So, yeah, everyone wants AI experience, but unfortunately, a lot of colleges and universities, you know, from 2022 to 2024 or 2025, just banned AI.
00:29:13.119 --> 00:29:18.079
So it has created, especially in the US, kind of a crisis, right?
00:29:18.240 --> 00:29:21.680
Uh companies can't find the experience that they need.
00:29:21.839 --> 00:29:31.039
So instead, they're just doubling, tripling down their AI investments as these models get more and more smart, or as these uh models get smarter, and they're like, wait, maybe we don't need all these people.
00:29:31.200 --> 00:29:40.000
Also, 51% of recent grads are second-guessing their career choice due to AI, up from 33% the year prior.
00:29:40.079 --> 00:29:42.160
That is a huge jump, right?
00:29:42.319 --> 00:29:45.039
Anyone of you study statistics and you see something like that.
00:29:45.119 --> 00:29:48.160
Uh, this is from a study I covered this in an earlier episode.
00:29:48.319 --> 00:29:52.319
Uh, it had always been around that like 25 to 30% mark.
00:29:52.400 --> 00:29:55.119
And then it just skyrocketed to more than 50%.
00:29:55.599 --> 00:30:00.079
And I do assume probably in three years, that number is gonna be more than three-fourths.
00:30:00.240 --> 00:30:08.240
I would assume that the overwhelming majority, probably three and four college grads, are gonna be like, what did I just go to school for?
00:30:09.440 --> 00:30:10.079
Right?
00:30:10.400 --> 00:30:14.160
Unless you're in something related to AI.
00:30:14.480 --> 00:30:16.400
It's like, what did I go to school for?
00:30:16.559 --> 00:30:16.960
Yeah.
00:30:17.279 --> 00:30:17.920
All right.
00:30:18.400 --> 00:30:20.400
And companies are betting big on it.
00:30:20.559 --> 00:30:30.319
So global AI spending on AI has already is expected to hit$2 trillion this year, and that's a 37% jump from 2025.
00:30:30.480 --> 00:30:36.079
Um, and Gartner predicts that 40% of enterprise apps will include AI agents by year end.
00:30:36.160 --> 00:30:49.599
So essentially, everything age uh sorry, companies and enterprises are spending more and more money on AI in the common software that most companies use, right?
00:30:49.680 --> 00:30:52.720
I'll just throw some basic examples, right?
00:30:52.960 --> 00:30:59.359
Salesforce is a very um, you know, popular CRM, the most popular CRM in the world.
00:30:59.519 --> 00:31:02.160
Uh, and then, you know, click up, let's use that as an example.
00:31:02.319 --> 00:31:05.039
Smaller company, but they're a you know, pretty big CRM company.
00:31:05.119 --> 00:31:07.920
You know, you can say the same thing for you know, HubSpot, right?
00:31:08.160 --> 00:31:15.680
Those companies, basic software that tens of millions of businesses use, everything's an agent now.
00:31:16.160 --> 00:31:16.480
Right?
00:31:16.720 --> 00:31:19.920
Everything is being agentified before our very eyes.
00:31:20.079 --> 00:31:26.480
The very tasks that humans would do inside of this software, it's all just becoming an agent now.
00:31:28.880 --> 00:31:32.720
So as we wrap up here, what does all this mean?
00:31:33.200 --> 00:31:33.599
All right.
00:31:33.759 --> 00:31:38.799
So now you know a little bit about the history of generative AI, uh, right?
00:31:39.039 --> 00:31:46.720
AI is not new, but the transformer technology that led to large language models is relatively new.
00:31:46.799 --> 00:31:54.799
And the adoption of large language models has been unlike anything we've ever seen, and it's not even close.
00:31:55.839 --> 00:31:58.480
So, why does it matter this year more than ever?
00:32:00.240 --> 00:32:13.519
Well, I've been the crazy guy yelling at you all for many years to not wait any longer to train your employees, uh, to invest in the future of work, right?
00:32:16.799 --> 00:32:19.119
Just using AI is not gonna do anything.
00:32:19.279 --> 00:32:25.759
It's not gonna do anything for your department, it's not gonna do anything for your career, it's not gonna do anything for your company, right?
00:32:25.839 --> 00:32:37.599
I think a lot of uh, you know, a lot of enterprises thought, like, okay, well, yeah, we'll just buy some co-pilot seats, we'll buy some, you know, Chat GPT enterprise seats, and you know, that'll make the board happy.
00:32:37.680 --> 00:32:43.599
And, you know, maybe our people are a little more productive and we'll be able to quietly reduce headcount and everything will be good.
00:32:43.839 --> 00:32:45.440
It's not like that anymore.
00:32:45.680 --> 00:32:58.720
Companies that adopted AI early are now three times more likely to see operating profit impact uh up to 5% than those companies that are still in the experimentation phase.
00:32:59.039 --> 00:32:59.680
Right?
00:33:00.000 --> 00:33:07.680
You can't be experimenting anymore, both as individuals, uh as departments, as organizations.
00:33:09.440 --> 00:33:13.519
You can't treat this as, you know, we're gonna pilot this AI thing.
00:33:13.839 --> 00:33:14.160
No.
00:33:14.880 --> 00:33:16.480
You have to hit the ground running.
00:33:16.559 --> 00:33:29.920
You have to hit the ground learning, you have to hit the ground experimenting, you have to hit the ground measuring, you have to hit the ground scoping, you have to hit the ground running your own internal benchmarks, but you have to hit the ground running as fast as you can, right?
00:33:30.559 --> 00:33:34.640
You have to go slow, but you have to go fast, right?
00:33:35.039 --> 00:33:44.559
You have to be methodical, you have to measure, you have to uh know, you have to be able to quantify, but you have to do it as fast as possible, right?
00:33:44.799 --> 00:33:57.200
Because the gap between AI fluent workers and AI fluent companies and everyone else is widening every single month, every single week, every single day.
00:33:57.839 --> 00:34:07.039
So by companies sitting there and saying, Yeah, we're gonna do a year-long pilot, you know, we're gonna make sure we get this AI thing right, we're gonna do a slow rollout.
00:34:07.839 --> 00:34:10.079
It's not gonna work, right?
00:34:10.960 --> 00:34:19.119
Even Fortune 500 companies that have that mentality, they're gonna get eaten up.
00:34:19.280 --> 00:34:21.599
You're already starting to see stories of it, right?
00:34:21.760 --> 00:34:24.400
Companies that thought they were too big for AI, right?
00:34:24.559 --> 00:34:28.239
You saw a lot of companies, FYI, doing about face, uh, right?
00:34:28.320 --> 00:34:32.880
Some of the consulting companies, uh, you know, now investing billions of dollars.
00:34:32.960 --> 00:34:37.440
Uh, you know, some of the banks said, Oh, no, we're large language models never touching that.
00:34:37.760 --> 00:34:38.719
Let's laugh at that.
00:34:39.280 --> 00:34:41.760
No, now they're investing billions of dollars.
00:34:42.159 --> 00:34:46.719
You can't not play the game because you don't have a choice.
00:34:48.159 --> 00:34:54.480
The future of work is generative AI is large language models.
00:34:54.960 --> 00:34:57.599
So the window is closing, y'all.
00:34:57.679 --> 00:35:04.079
And that is why generative AI matters in 2026 more than ever.
00:35:04.400 --> 00:35:06.800
This is the year to make your move.
00:35:06.960 --> 00:35:08.880
This is the year to level up.
00:35:09.039 --> 00:35:10.000
And guess what?
00:35:10.239 --> 00:35:12.400
It starts here with the start here series.
00:35:12.559 --> 00:35:19.199
All right, thank you for tuning in to the first uh volume of the start here series.
00:35:19.280 --> 00:35:22.480
Like I said, future ones are gonna be much, much faster.
00:35:22.639 --> 00:35:25.119
Uh, about 20 to 25 minutes.
00:35:25.199 --> 00:35:29.440
And make sure if you're still listening, check the show notes of this episode.
00:35:29.599 --> 00:35:35.360
So uh, if you're listening on January 15th, right, that's when uh you know this show is debuting.
00:35:35.440 --> 00:35:36.320
Nothing's gonna be there.
00:35:36.400 --> 00:35:39.119
But in the future, we're gonna go and update the show notes.
00:35:39.199 --> 00:35:44.320
So if you're listening on the podcast as we release new episodes, uh, we will make sure to link them there.
00:35:44.480 --> 00:35:47.920
But more importantly, just go to starthirseries.com.
00:35:48.079 --> 00:35:52.880
There you can uh sign up for our community for free.
00:35:53.039 --> 00:36:01.920
Uh, you can go follow, it's gonna send you straight uh to the space that we have set up that's gonna have just the start here series in there, nothing else.
00:36:02.000 --> 00:36:02.800
So you can focus.
00:36:02.960 --> 00:36:09.760
So uh whether you're hearing this in mid-January, mid-February, uh, late 2027, it doesn't matter.
00:36:09.920 --> 00:36:10.719
It's gonna be there.
00:36:10.800 --> 00:36:14.239
You can get caught up uh and instantly level up.
00:36:14.400 --> 00:36:14.800
All right.
00:36:14.880 --> 00:36:16.480
Uh so here's the final take.
00:36:16.639 --> 00:36:20.320
AI isn't new, but generative AI's compounding impact is.
00:36:20.480 --> 00:36:31.679
And today's large language models move faster than anyone can track, and they can even outperform humans in blind tests at producing economically viable work and valuable work.
00:36:31.840 --> 00:36:35.360
So don't think about AI upskilling or AI reskilling.
00:36:35.519 --> 00:36:37.280
If you do that, you're gonna fail.
00:36:37.440 --> 00:36:39.760
That is the wrong way to approach AI.
00:36:39.920 --> 00:36:47.039
You have to unlearn, you have to unlearn good habits and you have to build a solid foundation from scratch.
00:36:47.360 --> 00:36:49.440
AI first, AI native.
00:36:49.519 --> 00:36:51.599
You don't get to sprinkle AI on the top.
00:36:51.760 --> 00:36:52.559
It's not gonna work.
00:36:52.719 --> 00:36:53.039
All right.
00:36:53.199 --> 00:36:54.400
Uh, thank you for tuning in.
00:36:54.559 --> 00:36:58.320
Like I said, please go to starthearseries.com.
00:36:58.480 --> 00:37:00.320
If this was helpful, tell someone about it.
00:37:00.400 --> 00:37:01.599
Please subscribe to the podcast.
00:37:01.760 --> 00:37:07.199
Share this if you're listening uh on social media, on LinkedIn, tag someone who needs to hear this.
00:37:07.440 --> 00:37:09.920
We all need to start somewhere.
00:37:10.079 --> 00:37:10.480
All right.
00:37:10.639 --> 00:37:13.760
Don't let the rapid pace of AI confuse you.
00:37:13.840 --> 00:37:16.480
Don't let it slow yourself or your company down.
00:37:16.639 --> 00:37:17.760
That is my job.
00:37:18.000 --> 00:37:18.880
I work for you.
00:37:18.960 --> 00:37:21.199
You don't have to spend hours every single day.
00:37:21.440 --> 00:37:23.440
I do it, I cut it to you straight.
00:37:23.599 --> 00:37:23.920
All right.
00:37:24.079 --> 00:37:25.519
So thank you for tuning in.
00:37:25.679 --> 00:37:29.280
Hope to see you back tomorrow and every day for more everyday AI.
00:37:29.360 --> 00:37:30.400
Thanks, y'all.
00:00:01.520 --> 00:00:04.719
Have you ever felt overwhelmed by AI?
00:00:05.120 --> 00:00:16.239
Like there's certain aspects of artificial intelligence that you barely understand to begin with, yet you're expected to use it and it's changing every day.
00:00:16.480 --> 00:00:17.920
I understand where you're coming from.
00:00:18.160 --> 00:00:28.160
Chances are, if you're listening to this, you probably have a full-time job where you're expected to leverage AI, but you haven't had much formal training, if any, at all.
00:00:28.320 --> 00:00:32.240
And you definitely don't have extra hours in your day to learn.
00:00:32.479 --> 00:00:37.679
Oh, and yeah, I mentioned that thing where AI is changing literally every single day.
00:00:37.759 --> 00:00:41.920
And when I say I understand the challenge, I mean it.
00:00:42.240 --> 00:00:51.119
I mean literally, it's my only job to use, to build with, and to teach AI every single day for 10 to 12 hours.
00:00:51.200 --> 00:00:53.520
And that's all I've been doing for the past three years.
00:00:53.679 --> 00:00:56.399
And even I find it hard to keep up.
00:00:56.640 --> 00:00:57.759
But don't worry.
00:00:58.240 --> 00:01:01.359
That's where this new series comes into play.
00:01:01.520 --> 00:01:04.159
It's called the Start Here series.
00:01:04.319 --> 00:01:12.959
And whether you're a beginner, extremely confused, or you're someone that uses AI every single day, yet you're looking to double down.
00:01:13.280 --> 00:01:16.000
This new series is for you.
00:01:16.159 --> 00:01:22.799
Because one of the most common questions I get asked all the time is, where do I start?
00:01:23.040 --> 00:01:27.760
Or someone saying, Hey, I know a lot about large language models, but I want to know more about the creative side.
00:01:27.920 --> 00:01:28.640
Where do I start?
00:01:28.799 --> 00:01:34.400
To tell the truth, I haven't had a good answer until now.
00:01:34.560 --> 00:01:38.000
That's why we're kicking off this start here series.
00:01:38.159 --> 00:01:41.200
Uh so for those keeping up keeping up live, right?
00:01:41.359 --> 00:01:47.599
January is a time when I think most business leaders are setting new goals or trying to double down on good habits.
00:01:47.680 --> 00:01:55.599
So I know a lot of you are really trying to put in that extra effort here at the beginning of the beginning of the year to improve your understanding of AI.
00:01:55.760 --> 00:02:02.560
Uh, so that's why we're going to be releasing probably two of these episodes a week for the next five or so weeks.
00:02:02.719 --> 00:02:06.799
So, not necessarily abandoning our normal daily schedule.
00:02:06.879 --> 00:02:20.240
If you're an avid listener to the program, don't worry, we're just flexing a little bit here in the beginning of the year to try to help both beginners and advanced users alike get started and also get caught up.
00:02:20.400 --> 00:02:26.319
So we're going to be going over the basics like generative AI today, like what the heck is it?
00:02:26.479 --> 00:02:38.879
Um, to simplifying more advanced techniques like AI agents and expanding, uh explaining things like the model context protocol or what the health is a Ralph Wiggham loop, right?
00:02:38.960 --> 00:02:50.000
So whether it's concepts that we're zooming out and going back in time or things that are happening literally today and helping you put them in perspective, the start here series is gonna be for you.
00:02:50.159 --> 00:03:02.240
So we're gonna be going over, you know, the from the correct way to launch a successful AI pilot to how to measure ROI to how to read and understand benchmarks and choose which AI system is best for you or your company.
00:03:02.400 --> 00:03:06.479
We're gonna be tackling it all here in the Start Here series.
00:03:06.800 --> 00:03:07.599
You ready?
00:03:07.919 --> 00:03:09.439
All right, let's get into it.
00:03:09.599 --> 00:03:10.639
What's going on, y'all?
00:03:10.719 --> 00:03:12.719
If you're new here, my name's Jordan Wilson.
00:03:12.800 --> 00:03:14.960
In everyday AI, it's for you.
00:03:15.039 --> 00:03:17.360
And we've been doing it for a very long time.
00:03:17.520 --> 00:03:32.240
It is an unedited, unscripted daily live stream, podcast, and free daily newsletter, helping everyday business leaders like you and me not just keep up, but how we can leverage all the good stuff, make sense of the nonstop updates, and get ahead to grow our company and our career.
00:03:32.479 --> 00:03:37.759
So if that sounds like what you're trying to do, maybe this is episode number one for you.
00:03:37.840 --> 00:03:39.919
Maybe it's episode 700.
00:03:40.080 --> 00:03:40.960
It doesn't matter.
00:03:41.120 --> 00:03:43.360
We have something new and fresh for all of you.
00:03:43.520 --> 00:03:45.439
Yeah, a new URL to throw out there.
00:03:45.599 --> 00:03:45.919
Ready?
00:03:46.080 --> 00:03:50.560
So if you're interested, go to the go to starthearseries.com.
00:03:50.639 --> 00:03:52.719
That is starthearseries.com.
00:03:52.800 --> 00:03:56.080
So if you want to keep up with this particular series, go do that.
00:03:56.159 --> 00:04:05.520
You're gonna get a uh an invite link to sign up for our free community and you will be inserted directly into the Start Here Series uh kind of onboarding flow.
00:04:05.599 --> 00:04:13.360
So as we add to this, um there's gonna be more and more shows in there in that space, the dedicated space in our community.
00:04:13.520 --> 00:04:18.560
So if you want to keep up with all the shows in this series, uh connect with other leaders.
00:04:18.720 --> 00:04:24.079
I mean, literally, we have industry leaders in our free community, uh, that you can learn from.
00:04:24.240 --> 00:04:27.199
Go connect with, make sure you go to that uh website.
00:04:27.360 --> 00:04:32.399
All right, now we have that out of the way, just to let you know, these shows are gonna be a little quicker.
00:04:32.480 --> 00:04:37.600
All right, this one's uh, you know, might be 30 minutes, but most of these start here shows are gonna be about 20 minutes.
00:04:37.680 --> 00:04:42.639
I want to keep them very fast, very factual, which you know is gonna be a little hard for me.
00:04:42.800 --> 00:04:46.720
All right, so uh without further ado, let's get into it.
00:04:46.800 --> 00:04:54.959
Let's talk about generative AI, the basics, how it works, and why it matters in 2026 more than ever.
00:04:55.680 --> 00:04:59.519
So let's start with the pace in the reality.
00:04:59.759 --> 00:05:03.519
Well, nothing has spread this fast ever.
00:05:03.839 --> 00:05:05.920
Uh, nearly 900 million people.
00:05:06.079 --> 00:05:08.959
The last confirmed was 800 million, but I know the stat.
00:05:09.120 --> 00:05:15.040
It's nearly 900 million people are using Chat GPT weekly.
00:05:15.120 --> 00:05:17.600
I mean, that's more than the entire population of Europe.
00:05:17.759 --> 00:05:20.959
And Chat GPT, let's talk about an explosion.
00:05:21.120 --> 00:05:25.839
It reached 100 million users in its first two months.
00:05:26.480 --> 00:05:31.920
To get to that same number, the internet took eight years.
00:05:33.360 --> 00:05:40.240
So if you want to talk about a technology that you can't ignore, think about how commonplace the internet is in our day-to-day lives.
00:05:40.399 --> 00:05:43.680
You can't do too much without it, right?
00:05:43.920 --> 00:05:47.680
Especially if you're a knowledge worker, uh, you can't do too much without it.
00:05:47.920 --> 00:05:54.720
And just two years after launch, 40% of working age Americans are using generative AI.
00:05:55.040 --> 00:05:59.600
And the internet took seven years to reach that same level.
00:05:59.920 --> 00:06:13.600
So the generative AI technology, which is what large language models are kind of under the umbrella on, it is the most explosive growth of any technology ever.
00:06:16.319 --> 00:06:17.759
And companies are using it.
00:06:18.319 --> 00:06:21.040
Whether you have realized this or not, right?
00:06:21.360 --> 00:06:25.040
This conversation has changed a whole lot over the past few years.
00:06:25.279 --> 00:06:28.800
But nowadays, using AI is table stakes.
00:06:28.959 --> 00:06:30.160
Like you have to.
00:06:30.319 --> 00:06:31.839
You don't have a choice, right?
00:06:32.000 --> 00:06:34.879
Maybe three years ago it was a competitive advantage.
00:06:35.040 --> 00:06:45.680
You know, it was kind of novel, you know, to be using generative AI and large language models back in, you know, 21 uh or in 2022 when ChatGPT came out.
00:06:45.920 --> 00:06:48.800
It's not anything special today.
00:06:48.959 --> 00:06:51.439
You have to be using it today.
00:06:51.600 --> 00:06:59.120
Uh right now, about 80% of companies are even deploying AI agents that take actions, not just answer questions.
00:06:59.360 --> 00:07:08.399
So uh not only are 92% of Fortune 500 companies, you know, as an example, using open AI's technology, that's just one company, right?
00:07:08.720 --> 00:07:12.959
But 80% of companies are even deploying AI agents.
00:07:13.279 --> 00:07:13.680
All right.
00:07:13.839 --> 00:07:16.240
So yes, the space moves quickly.
00:07:16.399 --> 00:07:17.920
And yes, we are gonna zoom out.
00:07:18.000 --> 00:07:23.839
But I first just wanted to set the table, so to speak, and let you know where we are, just about right.
00:07:24.000 --> 00:07:26.720
And depending on where you're you're tuning in from, right?
00:07:26.800 --> 00:07:28.480
Yeah, we have listeners from all over the world.
00:07:28.560 --> 00:07:29.519
So thank you for that.
00:07:29.839 --> 00:07:33.040
But here in the US, that's usually the lens I'm talking through.
00:07:33.199 --> 00:07:34.399
I'm from Chicago.
00:07:34.639 --> 00:07:35.839
Hey, good to meet you.
00:07:35.920 --> 00:07:36.319
Uh, right.
00:07:36.480 --> 00:07:38.480
A lot of you are starting here with this episode.
00:07:38.720 --> 00:07:44.399
Um, everyone's using AI here in the US, every single company, right?
00:07:44.480 --> 00:07:45.759
You don't see a company anymore.
00:07:45.920 --> 00:07:50.879
I haven't met a company in a very long time that's not using whether officially or unofficially, right?
00:07:50.959 --> 00:07:57.279
Yeah, there's the whole shadow AI, uh, you know, in in some of those things, but every company is using AI.
00:07:57.439 --> 00:08:01.279
Um, so it's changed though.
00:08:01.600 --> 00:08:12.720
And I think, as strange as it is, I think a lot of people still have a very 2022 view of AI, right?
00:08:12.800 --> 00:08:18.959
Let's just use ChatGPT because that is the most widely uh used AI tool in the world, right?
00:08:19.040 --> 00:08:25.360
Obviously, uh Microsoft Copilot, very popular in the enterprise, Google Gemini, Anthropic Claude, right?
00:08:25.439 --> 00:08:27.839
Those are what I uh refer to as the big four.
00:08:27.920 --> 00:08:31.199
So if you ever hear me reference the big four, I'm not talking about consulting.
00:08:31.279 --> 00:08:33.840
I'm talking about those uh four companies.
00:08:34.000 --> 00:08:39.200
But uh now they're not just these friendly chatbots anymore.
00:08:39.519 --> 00:08:43.600
These AI systems are operating systems in and of themselves.
00:08:43.679 --> 00:08:44.879
I've been saying this for years.
00:08:45.120 --> 00:08:46.639
Companies need to make a decision.
00:08:46.720 --> 00:08:54.159
You need to move all your operating, your day-to-day operations into a large language model sooner rather than later, right?
00:08:54.240 --> 00:09:02.080
We'll probably explore that more in a later uh start here series, the exact process of uh choosing and setting up an AI operating system.
00:09:02.240 --> 00:09:08.000
It's not an official term, it's something I think I made it up a couple of years ago and I've been running with it ever since, the AI OS, right?
00:09:08.080 --> 00:09:14.240
But in the same way companies, you know, in the 90s or 80s or early 2000s depends, right?
00:09:14.399 --> 00:09:19.039
They gave all their employees computers at some point and they had to make a choice, right?
00:09:19.279 --> 00:09:21.279
Are we a Windows organization?
00:09:21.440 --> 00:09:23.120
Are we a Mac organization?
00:09:23.279 --> 00:09:24.559
Are we a Linux organization?
00:09:24.639 --> 00:09:26.159
You have to make the same choice.
00:09:26.399 --> 00:09:27.919
What is your company going to be?
00:09:28.000 --> 00:09:36.399
But now, in a few clicks, your entire organization's data can be accessed instantly in these AI operating systems, right?
00:09:36.559 --> 00:09:39.039
ChatGPT, they have a Teams and an enterprise plan.
00:09:39.120 --> 00:09:41.519
Gemini has a business and an enterprise plan.
00:09:41.679 --> 00:09:46.960
Obviously, Microsoft 365, uh, Copilot has an enterprise plan, Claude has an enterprise plan.
00:09:47.120 --> 00:09:52.240
These are for teams now, and they bring your data in instantly, and teams can collaborate, right?
00:09:52.399 --> 00:09:53.360
Seamlessly.
00:09:53.519 --> 00:09:56.159
So with no knowledge, even right now, right?
00:09:56.240 --> 00:09:58.720
Here's here's some stuff, so some new stuff, right?
00:09:58.960 --> 00:10:02.000
Uh yes, I know this is gonna age, right?
00:10:02.080 --> 00:10:10.240
If you're listening to this in June or July, but you know, Anthropic just came out with a tool for their desktop program called uh Claude Cowork.
00:10:10.559 --> 00:10:15.919
So with no knowledge at all, no tech know-how, you can use a desktop program.
00:10:16.080 --> 00:10:23.279
Let me say this again: you can use a desktop program uh using their Opus 4.5 model, one of the most powerful models in the world.
00:10:23.519 --> 00:10:32.960
It can control your computer, it can access your file system, it can it can browse the internet in your browser that's logged into everything.
00:10:34.879 --> 00:10:36.399
A large language model can do that.
00:10:36.480 --> 00:10:37.919
It's not just a friendly chat bot.
00:10:38.080 --> 00:10:39.759
That is the work us humans do.
00:10:39.919 --> 00:10:47.519
It can access my computer, it can access uh the terminal, it can code, it can access all my files on my local machine.
00:10:47.679 --> 00:10:53.360
It can go out and log into my email, log into any system, and access any data that I tell it to.
00:10:53.600 --> 00:10:54.960
That's where we are now.
00:10:55.120 --> 00:10:58.879
Today's generative AI is much more than friendly chatbots.
00:10:59.440 --> 00:11:01.279
But let's zoom off to understand.
00:11:01.519 --> 00:11:09.679
Because I think for whatever reason, people think of AI as a very new technology, which I think is one of the reasons why there's a high distress.
00:11:10.080 --> 00:11:12.559
The other reason why is most people don't know how to use it.
00:11:12.720 --> 00:11:25.440
And you see someone share some, you know, someone puts a screenshot of something from Chat GPT on on Facebook or LinkedIn, and someone usually that has no clue what they're doing, and it's uh an example of a large language model getting something very simple, very raw.
00:11:25.919 --> 00:11:29.360
FYI, anytime you see that, I've been doing this for a long time.
00:11:29.600 --> 00:11:32.720
99 times out of 100, it is human error.
00:11:32.799 --> 00:11:35.600
That human has no clue what they're doing, uh right.
00:11:35.840 --> 00:11:37.440
But AI is not new.
00:11:37.759 --> 00:11:41.919
AI has been around in some way, shape, or form for decades, up to 70 years.
00:11:42.000 --> 00:11:45.840
So the term artificial intelligence was actually coined in the 50s.
00:11:45.919 --> 00:11:49.759
And we've been kind of chasing this, you know, dream ever since.
00:11:49.840 --> 00:11:57.519
So there were early systems, uh, the expert systems in the 70s and the 80s, uh, right, used in a lot of different sectors.
00:11:57.679 --> 00:12:09.039
Uh, I think banking, uh, mortgages were some of the bigger ones, obviously in healthcare, but even early systems in the 70s uh were shown to be able to diagnose infections better than most doctors, right?
00:12:09.120 --> 00:12:13.679
And we obviously, with the advent of large language models, we have a lot more studies that show that.
00:12:13.840 --> 00:12:16.720
Um, so what is changed?
00:12:16.799 --> 00:12:18.720
It's not the ambition, right?
00:12:18.879 --> 00:12:23.600
It's finally there's the technology has evolved a lot.
00:12:23.759 --> 00:12:36.639
Uh, you know, there were a couple of breakthroughs in the 2010s, but obviously the big one that most people are familiar with is kind of the advent and the uh prolification of the transformer um that has led to Chat GPT.
00:12:37.279 --> 00:12:42.960
So the biggest thing that's changed is well, the technology and the compute needed, right?
00:12:43.039 --> 00:12:53.840
In the same way, how to, you know, access the internet, you know, in the early days, you had to have a computer bigger than a car, uh, right, and connect it up to all these cables.
00:12:53.919 --> 00:12:55.759
And now your cell phone obviously can.
00:12:55.840 --> 00:12:57.759
So the same thing is true with AI, right?
00:12:57.840 --> 00:13:02.480
These systems that used to be very slow and very big and very expensive, right?
00:13:02.639 --> 00:13:04.000
It's not like that anymore.
00:13:04.159 --> 00:13:10.559
The technology has changed, it's gotten exponentially uh faster, more affordable, and more powerful.
00:13:10.720 --> 00:13:17.120
Um, but probably the biggest breakthrough that led us to where we are today is in 2017.
00:13:17.279 --> 00:13:24.480
So uh that's when Google Researchers published the very popular research paper, Attention is all you need.
00:13:25.039 --> 00:13:33.600
And it quietly changed the trajectory, technically, not just of the artificial intelligence technology, but I would say of the world, right?
00:13:33.679 --> 00:13:41.200
The other thing I didn't really touch on artificial intelligence right now, and we maybe we'll tackle this in another start here series.
00:13:41.840 --> 00:13:46.159
It has become, I'd say, probably more important than oil.
00:13:46.480 --> 00:13:52.720
It's probably become more important than um countries military, uh, right.
00:13:52.879 --> 00:13:54.559
Whether you realize it or not, right?
00:13:54.720 --> 00:14:00.639
Maybe if you come back and listen to the show in 2027 or 2028, you'll be like, oh, okay, this weird guy was right.
00:14:00.799 --> 00:14:02.240
It is more important, right?
00:14:02.399 --> 00:14:10.320
That's why you're seeing uh a lot of geopolitical tensions heightened uh right now around the technology itself, uh, right.
00:14:10.559 --> 00:14:20.159
Essentially, you know, there's this thing called AI, there's this thing called AGI or artificial general intelligence, and then there's this thing called ASI, artificial superintelligence, right?
00:14:20.240 --> 00:14:22.320
That's when you start to get into the terminator thing.
00:14:22.480 --> 00:14:25.919
But uh we'll tackle this in a later start here series show.
00:14:26.000 --> 00:14:35.679
But essentially, the first uh, you know, country or company to develop AGI, artificial general intelligence, that is big, right?
00:14:35.840 --> 00:14:43.759
Anyways, it started with this paper, and that has kind of changed the course of the world and the business world, especially.
00:14:43.919 --> 00:14:49.679
But that paper uh essentially introduced the T in GPT, right?
00:14:49.759 --> 00:14:52.480
Uh uh generative preformed transformers.
00:14:52.559 --> 00:15:01.039
But the T in that paper was introducing the transformer architecture, which is the engine behind almost every single AI model today.
00:15:01.200 --> 00:15:04.639
Uh, and open AI was kind of the first to run with it, right?
00:15:04.799 --> 00:15:08.399
You know, a lot can be said with the current race between OpenAI and Google.
00:15:08.480 --> 00:15:11.200
They're definitely the front runners in the AI race.
00:15:11.360 --> 00:15:14.960
You know, Claude, uh Infrapex Claude is right there as well.
00:15:15.039 --> 00:15:19.919
Um, but OpenAI was the first to take this uh research and really productize it.
00:15:20.000 --> 00:15:32.559
So they built GPT one in 2018, GPT two in 2019, GPT-3 in 2020, which is actually when they started releasing it to the public uh with a lot of earlier software programs.
00:15:32.720 --> 00:15:39.919
That's when, you know, myself and my team, we started using it daily in 2020, uh, two years before ChatGPT came out.
00:15:40.000 --> 00:15:49.200
But most of the world kind of figured out about this big breakthrough that no one saw coming that originated in the 2017 Google Attention is all you need paper.
00:15:49.360 --> 00:15:51.600
It really came to life in November 2022.
00:15:51.679 --> 00:15:53.279
That was the Chat GPT moment.
00:15:53.440 --> 00:16:02.720
That is the line in the sand that I think has changed uh the course of not just the business world, but also the information world.
00:16:04.559 --> 00:16:08.080
So, how the heck do large language models work?
00:16:08.159 --> 00:16:13.039
All right, so you're like artificial intelligence has been around for a long time, right?
00:16:13.120 --> 00:16:15.600
And then there's this important thing called the transformer.
00:16:15.759 --> 00:16:20.320
Well, and then the transformer uh led to large language models.
00:16:20.480 --> 00:16:23.519
So here's kind of what they are and how they work, uh, right.
00:16:23.600 --> 00:16:30.080
So they predict essentially these large language models have been trained on the history of the internet, uh, right.
00:16:30.240 --> 00:16:32.960
There's a lot of lawsuits that are gonna come from that, right?
00:16:33.039 --> 00:16:35.440
Because people are like, well, what about copyrighted information?
00:16:35.600 --> 00:16:38.159
Yeah, uh it's gonna play out in the courts eventually.
00:16:38.320 --> 00:16:50.080
Uh, but these large language models essentially scrape not just everything on the internet that we use, the open web, but they scrape everything on the closed web, uh offline data sets, right?
00:16:50.240 --> 00:16:56.960
Essentially the entirety of human history and human knowledge has been scraped by these large language models.
00:16:57.120 --> 00:17:03.759
Then you have very smart humans uh at these big companies that then train the models and they go through a process.
00:17:03.840 --> 00:17:05.119
I'm not gonna get too dorky.
00:17:05.359 --> 00:17:09.279
It's called reinforcement learning with human feedback, but they train these models, right?
00:17:09.440 --> 00:17:17.759
So they say, hey, when you know someone asks a question about uh, you know, topic A, here's what a good output would be.
00:17:17.839 --> 00:17:20.880
Here's answer A is good, answer B is bad, right?
00:17:21.039 --> 00:17:24.240
And so they go through this reinforcement learning process to train the models.
00:17:24.319 --> 00:17:30.480
And these models are trained to be a helpful assistant, essentially with this, uh, their training data.
00:17:30.720 --> 00:17:31.200
All right.
00:17:31.359 --> 00:17:34.480
And this is how the earlier uh models worked.
00:17:34.640 --> 00:17:37.759
Today's models are much more sophisticated.
00:17:37.839 --> 00:17:52.480
Uh, right now, if you ever hear uh a term like scaffolding or uh agenc, uh right, today's large language models are night and day difference uh than the large language models that set off the Chat GPT uh craze, right?
00:17:52.559 --> 00:17:56.160
Those were kind of just, I call them old school transformers, right?
00:17:56.319 --> 00:17:59.599
Uh today's models, they're technically still transformers, right?
00:17:59.759 --> 00:18:01.680
But I call them uh reasoners, right?
00:18:01.759 --> 00:18:04.160
A lot of people call them reasoners or logic-based models.
00:18:04.240 --> 00:18:05.759
So they're much different now.
00:18:05.920 --> 00:18:10.559
Uh, there's a process that um they can kind of mimic human logic.
00:18:10.720 --> 00:18:22.720
So yes, they are still technically uh next token prediction machines, but there's uh essentially a uh step-by-step problem solving on top of that prediction engine that mimics human logic, right?
00:18:22.799 --> 00:18:29.839
And you can see it think uh step by step and go through and you know, use these different tools it has at its disposal.
00:18:30.000 --> 00:18:35.119
It might run code, it might go, you know, before it responds to you, these models that think now, right?
00:18:35.200 --> 00:18:37.680
They used to just spit out very quickly and answer, right?
00:18:37.759 --> 00:18:41.759
The earlier models, the GPT-2, the GPT-3 were really bad, uh right.
00:18:41.839 --> 00:18:49.279
The GPT-3-5 that launched with um, you know, Chad GPT a little bit better, it was at least coherent and usually accurate, right?
00:18:49.359 --> 00:18:52.640
But today's models, extremely impressive, right?
00:18:52.720 --> 00:18:55.519
They tackle problems like experienced human.
00:18:55.599 --> 00:19:02.400
Today's models score better on offline IQ tests than 99% of humans.
00:19:02.480 --> 00:19:05.680
They are literally in the top 1%, uh uh, right?
00:19:05.839 --> 00:19:12.799
They are at genius level scores, uh, taking offline uh IQ tests that they have not been trained on.
00:19:12.960 --> 00:19:13.279
All right.
00:19:13.440 --> 00:19:19.039
So that's where uh kind of large language models are today, how they kind of work, uh right.
00:19:19.200 --> 00:19:26.559
So scrape data uh from data sources, humans train them uh to be helpful assistance for users.
00:19:26.799 --> 00:19:29.039
A user asks a model a query.
00:19:29.119 --> 00:19:35.920
Uh, the model might just pull that out from its training data, or it might decide, hey, I need to use some tools at my disposal.
00:19:36.079 --> 00:19:38.160
I should probably go fetch something on the web.
00:19:38.400 --> 00:19:41.519
It seems like the user is asking about something very current.
00:19:41.599 --> 00:19:48.640
Um, maybe I should use a data analysis, maybe I should run some code under the hood to get to this, uh, to get to this answer.
00:19:48.799 --> 00:20:02.799
So, yes, uh this uh kind of this scaffolding or the tools that models use uh help it hallucinate less and just provide much more robust and impressive outputs.
00:20:03.279 --> 00:20:06.720
So uh the scale is hard to comprehend, right?
00:20:06.799 --> 00:20:09.200
We're talking about billions and trillions of parameters.
00:20:09.359 --> 00:20:14.960
So a parameter is essentially a learned pattern that these models have gone through in their training data.
00:20:15.039 --> 00:20:19.759
So more parameters just means that there's more patterns that the models can uh recognize and use.
00:20:19.920 --> 00:20:24.160
I like to think of it's it's a connection in its big neural network brain.
00:20:24.240 --> 00:20:34.000
Uh so GPT 3 had 175 billion parameters, GPT-4 reportedly had over uh about almost two trillion parameters.
00:20:34.079 --> 00:20:41.119
And a lot of today's newer models, we don't know how many, but we've seen reports that they're anywhere from two to four trillion parameters.
00:20:41.279 --> 00:20:41.599
All right.
00:20:41.759 --> 00:20:45.519
So it's just uh an insane amount of data.
00:20:45.680 --> 00:20:49.200
And then there's also something called the context window, and even that has exploded.
00:20:49.279 --> 00:20:54.640
And that's essentially how much a large language model can remember before it starts to forget, uh, right.
00:20:54.720 --> 00:21:02.160
And we do in our in our free, you know, uh Prime Prompt Polish course uh that you can access in our community, by the way.
00:21:02.240 --> 00:21:14.079
So you can just go to starthirseries.com, uh, sign up for our community, and you will also get instant access to that free uh prompt engineering, context engineering course that's been taken by more than 15,000 people.
00:21:14.319 --> 00:21:18.079
Anyways, a tote uh context window is extremely important.
00:21:18.160 --> 00:21:25.519
And that's another thing that has scaled alongside with models because the earlier versions, they would forget things almost right away, right?
00:21:25.759 --> 00:21:29.839
Today's you can work with them sometimes, depending on how you have it set up.
00:21:30.000 --> 00:21:35.680
You can work with them for hours or days until they start to forget things, right?
00:21:35.759 --> 00:21:38.880
And there's obviously uh some um autocompaction.
00:21:39.039 --> 00:21:44.400
Uh, you know, uh these models are starting to work uh kind of automatically compacting this information.
00:21:44.480 --> 00:21:47.759
So it keeps a longer uh kind of memory going.
00:21:48.640 --> 00:21:52.720
So generative AI isn't just text anymore, right?
00:21:52.880 --> 00:21:59.039
So large language models started as mainly text based models, uh, but now they're multimodal by default.
00:21:59.119 --> 00:22:06.240
And there's a lot of different Uh kind of techniques, but everything kind of now falls under under this generative AI umbrella, right?
00:22:06.319 --> 00:22:10.960
Where traditional uh artificial intelligence was more deterministic, right?
00:22:11.039 --> 00:22:14.160
It was based on more uh decision trees, right?
00:22:14.319 --> 00:22:17.519
If else or if this then logic, uh right.
00:22:17.680 --> 00:22:19.519
So uh traditional AI, right?
00:22:19.599 --> 00:22:27.119
The expert systems in the in the 80s and used through the 90s, et cetera, uh, a lot of it was more deterministic, right?
00:22:27.200 --> 00:22:28.720
It was very rule-based.
00:22:28.880 --> 00:22:31.279
That's not large language models, right?
00:22:31.440 --> 00:22:37.599
Uh, you know, people say hallucinations uh or creativity is a feature, not a bug.
00:22:37.759 --> 00:22:38.319
Um, right.
00:22:38.480 --> 00:22:40.400
That's what this next token prediction.
00:22:40.559 --> 00:22:47.200
And what that means is, right, like if think if a little kid goes to touch a stove, what do you say?
00:22:47.599 --> 00:22:49.279
You say, oh, be careful.
00:22:49.440 --> 00:22:49.920
That's hot.
00:22:50.079 --> 00:22:50.720
Oh, that's hot.
00:22:50.799 --> 00:22:51.440
Don't touch that.
00:22:51.599 --> 00:22:52.640
Be careful, right?
00:22:52.880 --> 00:22:55.119
That's what you would naturally say, right?
00:22:55.359 --> 00:22:58.880
Humans, I think, very much are like large language models.
00:22:58.960 --> 00:23:08.799
So when people say, oh, large language models, they can't, they can't reason, well, they they kind of can because it's based on the reasoning of millions of humans in theory.
00:23:08.960 --> 00:23:09.519
Uh, right.
00:23:09.680 --> 00:23:12.559
But the same thing can be said for different types of models.
00:23:12.640 --> 00:23:18.799
Uh, on the image side, you have very popular uh, you know, AI image generators as well.
00:23:18.960 --> 00:23:21.839
You know, three, four years ago, they were very bad.
00:23:21.920 --> 00:23:24.880
They didn't even look like you couldn't even tell what something was.
00:23:25.279 --> 00:23:34.480
You know, some of the earlier versions, like, you know, Dolly 2 or some of the earlier versions of Midjourney, today's AI models, you can't tell the difference, right?
00:23:34.640 --> 00:23:35.920
I've mentioned this before.
00:23:36.079 --> 00:23:37.839
Uh, I used to do a lot of photography.
00:23:37.920 --> 00:23:39.599
I've taken more than a million photos.
00:23:39.680 --> 00:23:45.200
Uh, I've owned like, I don't know, eight different DSLRs, you know, professional cameras.
00:23:45.359 --> 00:23:46.720
I can't tell the difference.
00:23:46.960 --> 00:23:52.640
And if and if anyone tells you today, I think maybe three months ago, you could tell the difference.
00:23:52.799 --> 00:23:54.000
Today, no one can.
00:23:54.160 --> 00:23:54.400
Right.
00:23:54.480 --> 00:23:56.480
So, yes, there's different types of AI.
00:23:56.720 --> 00:23:58.000
It's not just text.
00:23:58.160 --> 00:24:02.880
Uh, there's, you know, text to music that's really good, you know, Suno V5.
00:24:03.200 --> 00:24:13.440
Uh, there's text to video, great companies out there, runway, uh, their Gen 4.5, Google VO3.1, uh, Sora 2, a lot of Chinese models, right?
00:24:13.519 --> 00:24:15.759
Uh, so it's not just text.
00:24:16.000 --> 00:24:17.920
It is multimodal.
00:24:18.160 --> 00:24:20.240
These models are multimodal by default.
00:24:20.400 --> 00:24:28.640
So uh, you know, text, images, music, soundscapes, sound effects, writing code, uh, it is all over the place.
00:24:29.680 --> 00:24:32.240
And it works, right?
00:24:32.640 --> 00:24:38.160
Yeah, maybe you read a bad headline recently that said, you know, 95% of AI pilots failed.
00:24:38.240 --> 00:24:38.720
They don't.
00:24:38.799 --> 00:24:40.000
That was uh marketing.
00:24:40.079 --> 00:24:41.200
Uh the ROI is true.
00:24:41.279 --> 00:24:54.000
If you look at real studies, right, that talk to thousands or tens of thousands uh of business leaders, almost every single reputable study in the world shows that the ROI return on investment of AI is exponential.
00:24:54.240 --> 00:25:03.119
So, as an example, the International Data Corporation found that companies get$3.70 back for every$1 invested in generative AI.
00:25:03.279 --> 00:25:08.400
And top performers are seeing a$10 plus uh dollar return per dollar invested.
00:25:08.480 --> 00:25:20.880
Uh Snowflake and uh ESG Enterprise Strategy Group survey uh of more than of almost 2,000 business leaders show that 92% of early adopters say their AI investments are already paying for themselves.
00:25:21.039 --> 00:25:24.400
So yeah, already positive uh ROI.
00:25:24.559 --> 00:25:38.160
And then uh similarly, other studies show that up to 98% of the same companies uh are that have previously invested in AI are looking to increase their investment.
00:25:38.640 --> 00:25:42.240
98% are increasing their investment, right?
00:25:42.319 --> 00:25:46.480
So it's no longer experimentation like it was in 2023.
00:25:46.640 --> 00:25:47.920
Now it's all about scale.
00:25:48.079 --> 00:25:52.079
Now it's all about, oh, wait, we can crush our competitors with AI.
00:25:52.240 --> 00:25:54.960
Unfortunately, it's about reducing headcount.
00:25:55.039 --> 00:25:56.319
We'll get to that in a little bit.
00:25:56.640 --> 00:25:59.519
But the returns are there and they are undeniable.
00:25:59.599 --> 00:26:04.079
And the economic skit stakes as well are massive, right?
00:26:04.319 --> 00:26:09.759
Uh, like I said, never trust small-scale studies or studies that have uh uh an agenda.
00:26:10.160 --> 00:26:16.640
But um, when you talk about does AI work, I mean, look at Anthropic's uh November study.
00:26:16.799 --> 00:26:19.119
Uh 100,000 real conversations.
00:26:19.200 --> 00:26:23.519
They found that AI reduces task completion time by 80%.
00:26:24.079 --> 00:26:33.839
Uh there was a McKinsey Digital study a couple of years ago, same thing, uh, said between 75 to 80% time savings on standard knowledge work tasks.
00:26:34.079 --> 00:26:42.559
Uh PWC study of 50,000 workers globally said 92% of daily AI users report productivity gains.
00:26:43.039 --> 00:26:43.440
Right?
00:26:43.599 --> 00:26:46.160
92% are reporting productivity gains.
00:26:46.400 --> 00:26:51.839
And I want to talk to the other 8% and teach them a couple of things, uh, but these stats are undeniable.
00:26:52.079 --> 00:26:57.519
These large language models do the tasks that used to, right?
00:26:57.680 --> 00:27:01.839
I remember tasks I used to do 15 years ago, a lot of researching, right?
00:27:02.000 --> 00:27:14.079
Looking, having 20 tabs open, you know, reading these PDFs, grabbing information out, personalizing it, synthesizing it, uh, putting it in spreadsheets, maybe then eventually turning those spreadsheets into a PowerPoint, something like that.
00:27:14.240 --> 00:27:18.880
Those are projects that would take 40, 50, 60, 100 hours.
00:27:19.200 --> 00:27:21.759
I literally timed this the other day.
00:27:22.240 --> 00:27:24.400
Uh takes like 10 minutes now.
00:27:24.640 --> 00:27:30.079
One model, one prompt, if you know what you're doing, can do that entire process in 10 minutes.
00:27:30.319 --> 00:27:35.599
And it's about 99.7% factual if you know what you're doing, right?
00:27:35.839 --> 00:27:40.720
Today's models are completely unrecognizable from the Chat GPT of 2022.
00:27:40.880 --> 00:27:48.960
And you might be thinking, wait, if all these AI models are that good, what's happening to jobs?
00:27:49.279 --> 00:27:49.920
Right.
00:27:50.480 --> 00:27:51.599
Let me just tell you this.
00:27:51.759 --> 00:27:55.839
If you're brand new to the show, I'm a realist, right?
00:27:56.079 --> 00:28:04.799
I'm not someone drinking this AI Kool-Aid and being like, oh, you know, none of us are gonna have to work and we're all gonna live, you know, this utopian AI dream.
00:28:04.880 --> 00:28:05.920
No, I don't think so.
00:28:06.000 --> 00:28:08.079
I think it it might get a little bad.
00:28:08.240 --> 00:28:09.039
Um, right.
00:28:09.119 --> 00:28:10.799
And we're already starting to see that.
00:28:10.880 --> 00:28:12.079
Talk to recent grads.
00:28:12.160 --> 00:28:13.680
How many recent grads do you know?
00:28:13.839 --> 00:28:19.839
It's hard to get a job because companies just aren't hiring anymore, especially the companies that have figured out AI, right?
00:28:19.920 --> 00:28:21.279
So uh some stats here.
00:28:21.359 --> 00:28:28.640
Uh so only 30% of graduates in 2025 secured a job in their field, and that's down from 41% the year before.
00:28:28.799 --> 00:28:31.039
That is a huge year-over-year drop.
00:28:31.119 --> 00:28:36.720
And entry level hiring is down 44% from its peak three years ago.
00:28:36.960 --> 00:28:43.599
So entry-level hiring going down nearly 50% is catastrophic.
00:28:43.759 --> 00:28:50.960
And just the number of uh graduates not being able to secure a job in their field, that's uh worrisome as well.
00:28:51.039 --> 00:29:00.480
Uh, right now, 62% of employers say that candidates should have AI knowledge, but 55% of graduates say their degree programs didn't prepare them.
00:29:00.640 --> 00:29:12.880
So, yeah, everyone wants AI experience, but unfortunately, a lot of colleges and universities, you know, from 2022 to 2024 or 2025, just banned AI.
00:29:13.119 --> 00:29:18.079
So it has created, especially in the US, kind of a crisis, right?
00:29:18.240 --> 00:29:21.680
Uh companies can't find the experience that they need.
00:29:21.839 --> 00:29:31.039
So instead, they're just doubling, tripling down their AI investments as these models get more and more smart, or as these uh models get smarter, and they're like, wait, maybe we don't need all these people.
00:29:31.200 --> 00:29:40.000
Also, 51% of recent grads are second-guessing their career choice due to AI, up from 33% the year prior.
00:29:40.079 --> 00:29:42.160
That is a huge jump, right?
00:29:42.319 --> 00:29:45.039
Anyone of you study statistics and you see something like that.
00:29:45.119 --> 00:29:48.160
Uh, this is from a study I covered this in an earlier episode.
00:29:48.319 --> 00:29:52.319
Uh, it had always been around that like 25 to 30% mark.
00:29:52.400 --> 00:29:55.119
And then it just skyrocketed to more than 50%.
00:29:55.599 --> 00:30:00.079
And I do assume probably in three years, that number is gonna be more than three-fourths.
00:30:00.240 --> 00:30:08.240
I would assume that the overwhelming majority, probably three and four college grads, are gonna be like, what did I just go to school for?
00:30:09.440 --> 00:30:10.079
Right?
00:30:10.400 --> 00:30:14.160
Unless you're in something related to AI.
00:30:14.480 --> 00:30:16.400
It's like, what did I go to school for?
00:30:16.559 --> 00:30:16.960
Yeah.
00:30:17.279 --> 00:30:17.920
All right.
00:30:18.400 --> 00:30:20.400
And companies are betting big on it.
00:30:20.559 --> 00:30:30.319
So global AI spending on AI has already is expected to hit$2 trillion this year, and that's a 37% jump from 2025.
00:30:30.480 --> 00:30:36.079
Um, and Gartner predicts that 40% of enterprise apps will include AI agents by year end.
00:30:36.160 --> 00:30:49.599
So essentially, everything age uh sorry, companies and enterprises are spending more and more money on AI in the common software that most companies use, right?
00:30:49.680 --> 00:30:52.720
I'll just throw some basic examples, right?
00:30:52.960 --> 00:30:59.359
Salesforce is a very um, you know, popular CRM, the most popular CRM in the world.
00:30:59.519 --> 00:31:02.160
Uh, and then, you know, click up, let's use that as an example.
00:31:02.319 --> 00:31:05.039
Smaller company, but they're a you know, pretty big CRM company.
00:31:05.119 --> 00:31:07.920
You know, you can say the same thing for you know, HubSpot, right?
00:31:08.160 --> 00:31:15.680
Those companies, basic software that tens of millions of businesses use, everything's an agent now.
00:31:16.160 --> 00:31:16.480
Right?
00:31:16.720 --> 00:31:19.920
Everything is being agentified before our very eyes.
00:31:20.079 --> 00:31:26.480
The very tasks that humans would do inside of this software, it's all just becoming an agent now.
00:31:28.880 --> 00:31:32.720
So as we wrap up here, what does all this mean?
00:31:33.200 --> 00:31:33.599
All right.
00:31:33.759 --> 00:31:38.799
So now you know a little bit about the history of generative AI, uh, right?
00:31:39.039 --> 00:31:46.720
AI is not new, but the transformer technology that led to large language models is relatively new.
00:31:46.799 --> 00:31:54.799
And the adoption of large language models has been unlike anything we've ever seen, and it's not even close.
00:31:55.839 --> 00:31:58.480
So, why does it matter this year more than ever?
00:32:00.240 --> 00:32:13.519
Well, I've been the crazy guy yelling at you all for many years to not wait any longer to train your employees, uh, to invest in the future of work, right?
00:32:16.799 --> 00:32:19.119
Just using AI is not gonna do anything.
00:32:19.279 --> 00:32:25.759
It's not gonna do anything for your department, it's not gonna do anything for your career, it's not gonna do anything for your company, right?
00:32:25.839 --> 00:32:37.599
I think a lot of uh, you know, a lot of enterprises thought, like, okay, well, yeah, we'll just buy some co-pilot seats, we'll buy some, you know, Chat GPT enterprise seats, and you know, that'll make the board happy.
00:32:37.680 --> 00:32:43.599
And, you know, maybe our people are a little more productive and we'll be able to quietly reduce headcount and everything will be good.
00:32:43.839 --> 00:32:45.440
It's not like that anymore.
00:32:45.680 --> 00:32:58.720
Companies that adopted AI early are now three times more likely to see operating profit impact uh up to 5% than those companies that are still in the experimentation phase.
00:32:59.039 --> 00:32:59.680
Right?
00:33:00.000 --> 00:33:07.680
You can't be experimenting anymore, both as individuals, uh as departments, as organizations.
00:33:09.440 --> 00:33:13.519
You can't treat this as, you know, we're gonna pilot this AI thing.
00:33:13.839 --> 00:33:14.160
No.
00:33:14.880 --> 00:33:16.480
You have to hit the ground running.
00:33:16.559 --> 00:33:29.920
You have to hit the ground learning, you have to hit the ground experimenting, you have to hit the ground measuring, you have to hit the ground scoping, you have to hit the ground running your own internal benchmarks, but you have to hit the ground running as fast as you can, right?
00:33:30.559 --> 00:33:34.640
You have to go slow, but you have to go fast, right?
00:33:35.039 --> 00:33:44.559
You have to be methodical, you have to measure, you have to uh know, you have to be able to quantify, but you have to do it as fast as possible, right?
00:33:44.799 --> 00:33:57.200
Because the gap between AI fluent workers and AI fluent companies and everyone else is widening every single month, every single week, every single day.
00:33:57.839 --> 00:34:07.039
So by companies sitting there and saying, Yeah, we're gonna do a year-long pilot, you know, we're gonna make sure we get this AI thing right, we're gonna do a slow rollout.
00:34:07.839 --> 00:34:10.079
It's not gonna work, right?
00:34:10.960 --> 00:34:19.119
Even Fortune 500 companies that have that mentality, they're gonna get eaten up.
00:34:19.280 --> 00:34:21.599
You're already starting to see stories of it, right?
00:34:21.760 --> 00:34:24.400
Companies that thought they were too big for AI, right?
00:34:24.559 --> 00:34:28.239
You saw a lot of companies, FYI, doing about face, uh, right?
00:34:28.320 --> 00:34:32.880
Some of the consulting companies, uh, you know, now investing billions of dollars.
00:34:32.960 --> 00:34:37.440
Uh, you know, some of the banks said, Oh, no, we're large language models never touching that.
00:34:37.760 --> 00:34:38.719
Let's laugh at that.
00:34:39.280 --> 00:34:41.760
No, now they're investing billions of dollars.
00:34:42.159 --> 00:34:46.719
You can't not play the game because you don't have a choice.
00:34:48.159 --> 00:34:54.480
The future of work is generative AI is large language models.
00:34:54.960 --> 00:34:57.599
So the window is closing, y'all.
00:34:57.679 --> 00:35:04.079
And that is why generative AI matters in 2026 more than ever.
00:35:04.400 --> 00:35:06.800
This is the year to make your move.
00:35:06.960 --> 00:35:08.880
This is the year to level up.
00:35:09.039 --> 00:35:10.000
And guess what?
00:35:10.239 --> 00:35:12.400
It starts here with the start here series.
00:35:12.559 --> 00:35:19.199
All right, thank you for tuning in to the first uh volume of the start here series.
00:35:19.280 --> 00:35:22.480
Like I said, future ones are gonna be much, much faster.
00:35:22.639 --> 00:35:25.119
Uh, about 20 to 25 minutes.
00:35:25.199 --> 00:35:29.440
And make sure if you're still listening, check the show notes of this episode.
00:35:29.599 --> 00:35:35.360
So uh, if you're listening on January 15th, right, that's when uh you know this show is debuting.
00:35:35.440 --> 00:35:36.320
Nothing's gonna be there.
00:35:36.400 --> 00:35:39.119
But in the future, we're gonna go and update the show notes.
00:35:39.199 --> 00:35:44.320
So if you're listening on the podcast as we release new episodes, uh, we will make sure to link them there.
00:35:44.480 --> 00:35:47.920
But more importantly, just go to starthirseries.com.
00:35:48.079 --> 00:35:52.880
There you can uh sign up for our community for free.
00:35:53.039 --> 00:36:01.920
Uh, you can go follow, it's gonna send you straight uh to the space that we have set up that's gonna have just the start here series in there, nothing else.
00:36:02.000 --> 00:36:02.800
So you can focus.
00:36:02.960 --> 00:36:09.760
So uh whether you're hearing this in mid-January, mid-February, uh, late 2027, it doesn't matter.
00:36:09.920 --> 00:36:10.719
It's gonna be there.
00:36:10.800 --> 00:36:14.239
You can get caught up uh and instantly level up.
00:36:14.400 --> 00:36:14.800
All right.
00:36:14.880 --> 00:36:16.480
Uh so here's the final take.
00:36:16.639 --> 00:36:20.320
AI isn't new, but generative AI's compounding impact is.
00:36:20.480 --> 00:36:31.679
And today's large language models move faster than anyone can track, and they can even outperform humans in blind tests at producing economically viable work and valuable work.
00:36:31.840 --> 00:36:35.360
So don't think about AI upskilling or AI reskilling.
00:36:35.519 --> 00:36:37.280
If you do that, you're gonna fail.
00:36:37.440 --> 00:36:39.760
That is the wrong way to approach AI.
00:36:39.920 --> 00:36:47.039
You have to unlearn, you have to unlearn good habits and you have to build a solid foundation from scratch.
00:36:47.360 --> 00:36:49.440
AI first, AI native.
00:36:49.519 --> 00:36:51.599
You don't get to sprinkle AI on the top.
00:36:51.760 --> 00:36:52.559
It's not gonna work.
00:36:52.719 --> 00:36:53.039
All right.
00:36:53.199 --> 00:36:54.400
Uh, thank you for tuning in.
00:36:54.559 --> 00:36:58.320
Like I said, please go to starthearseries.com.
00:36:58.480 --> 00:37:00.320
If this was helpful, tell someone about it.
00:37:00.400 --> 00:37:01.599
Please subscribe to the podcast.
00:37:01.760 --> 00:37:07.199
Share this if you're listening uh on social media, on LinkedIn, tag someone who needs to hear this.
00:37:07.440 --> 00:37:09.920
We all need to start somewhere.
00:37:10.079 --> 00:37:10.480
All right.
00:37:10.639 --> 00:37:13.760
Don't let the rapid pace of AI confuse you.
00:37:13.840 --> 00:37:16.480
Don't let it slow yourself or your company down.
00:37:16.639 --> 00:37:17.760
That is my job.
00:37:18.000 --> 00:37:18.880
I work for you.
00:37:18.960 --> 00:37:21.199
You don't have to spend hours every single day.
00:37:21.440 --> 00:37:23.440
I do it, I cut it to you straight.
00:37:23.599 --> 00:37:23.920
All right.
00:37:24.079 --> 00:37:25.519
So thank you for tuning in.
00:37:25.679 --> 00:37:29.280
Hope to see you back tomorrow and every day for more everyday AI.
00:37:29.360 --> 00:37:30.400
Thanks, y'all.
