WEBVTT
00:00:02.640 --> 00:00:11.119
One of the most challenging things about AI that no one talks about is not being able to talk about AI effectively.
00:00:11.279 --> 00:00:14.640
And I think that there's three main reasons for this.
00:00:14.880 --> 00:00:23.359
Number one, the tech changes faster than literally anyone can keep up with, even someone that talks about AI every day, like myself.
00:00:23.679 --> 00:00:26.559
Two, companies don't train their employees.
00:00:26.800 --> 00:00:29.839
And three, it's kind of the summation of those parts.
00:00:30.000 --> 00:00:38.159
There's a huge gap between what people know about AI and what the models are capable of and can actually accomplish.
00:00:38.479 --> 00:00:42.479
And so much of this can be solved by talking about it.
00:00:42.560 --> 00:00:44.640
But therein lies the problem.
00:00:45.200 --> 00:00:49.759
AI is a mystery wrapped up in ever-changing jargon.
00:00:50.000 --> 00:01:06.000
So we're kicking off volume two of our start here series, where we're gonna karate chop the AI jargon in the jugular and break down the basics of AI language and lingo that every business leader needs to know in 2026.
00:01:06.239 --> 00:01:07.760
All right, I'm excited.
00:01:07.920 --> 00:01:09.200
Let's get into it.
00:01:09.439 --> 00:01:13.920
Welcome to our start here series and everyday AI.
00:01:14.079 --> 00:01:23.280
So this brand new series is our essential beginner and advanced AI overviews to help you launch your year strong.
00:01:23.439 --> 00:01:31.680
So if one of your big goals is to double down or to just better learn AI, this new Start Here series is for you.
00:01:31.920 --> 00:01:32.319
All right.
00:01:32.400 --> 00:01:36.959
So if you haven't already, please go to starthearseries.com.
00:01:37.040 --> 00:01:38.560
Yeah, new URL.
00:01:38.799 --> 00:01:47.920
Uh so uh there you're gonna be forwarded straight to uh sign up page to get free access to our inner circle community.
00:01:48.079 --> 00:01:59.040
And not only will you get uh instant access to our Prime Prompt Polish Chat GPT prompt engineering course, but you will also get access to the Start Here series.
00:01:59.200 --> 00:02:13.759
So whether you're listening to this in mid-January or sometime in March, we're gonna be putting the recap of every single Start Here series um episode inside the community there, uh inside the community there, so you can listen to them all in order.
00:02:14.080 --> 00:02:14.560
All right.
00:02:14.800 --> 00:02:29.759
Uh, and if you are brand new, brand new, make sure you go listen to the first uh series of this, which was episode 691 or volume one in the start here series, which was generative AI, how it works and why it matters in 2026 more than ever.
00:02:30.000 --> 00:02:30.479
All right.
00:02:30.719 --> 00:02:36.080
If you're new, everyday AI, it's a daily live stream podcast and free daily newsletter.
00:02:36.240 --> 00:02:40.639
But I've done like 700 episodes, and the most common question I get is where do I start?
00:02:40.719 --> 00:02:43.759
That's the whole reason that we're doing this series.
00:02:43.919 --> 00:02:47.599
So let's cut the jargon down, shall we?
00:02:47.759 --> 00:02:48.960
Uh, we're gonna get straight into it.
00:02:49.039 --> 00:02:54.800
So don't worry, this isn't gonna be a super boring run through of random vocabulary words.
00:02:55.039 --> 00:03:03.039
Uh, one thing that I've realized that I kind of have a knack for uh over the years is almost being a translator, right?
00:03:03.199 --> 00:03:06.240
And uh, like literally, there is there's an instance.
00:03:06.479 --> 00:03:10.000
Um, I was working with a pretty big company, right?
00:03:10.240 --> 00:03:12.479
Doing like$20 billion in in revenue.
00:03:12.639 --> 00:03:16.400
And I was sitting with their C-suite and I was sitting with their their head of AI.
00:03:16.560 --> 00:03:21.360
And it seemed like sometimes these two sides weren't speaking the same language.
00:03:21.599 --> 00:03:23.520
And that's because there's so much jargon.
00:03:23.599 --> 00:03:26.960
And because, like I said, kind of that capabilities and understanding gap.
00:03:27.039 --> 00:03:32.400
And that's really what I've learned to do by doing this show over the last three years is connecting those two.
00:03:32.479 --> 00:03:42.319
So whether you are that highly technical person, you know, who understands all these terms, hopefully this will help you understand where the non-technical business leaders are coming from.
00:03:42.400 --> 00:03:52.639
And for our non-technical business leaders, hopefully this serves as a nice foundational look at the language and lingo you need to know to, you know, talk about it.
00:03:53.039 --> 00:03:54.400
Let's start.
00:03:55.039 --> 00:03:56.080
Start where we do.
00:03:56.240 --> 00:03:58.000
AI is not going anywhere, it's here to stay.
00:03:58.159 --> 00:03:58.479
All right.
00:03:58.560 --> 00:04:01.280
Uh, we went over that in the first episode, but it's worth repeating.
00:04:01.439 --> 00:04:06.719
ChatGPT, when it launched in November 2022, it was literally world-breaking technology.
00:04:06.879 --> 00:04:12.879
No one had ever seen anything like it before, uh, not just its capabilities, but also the sheer numbers.
00:04:13.039 --> 00:04:21.839
Uh, right, they had 100 million users within two months after launch, and ChatGPT now has nearly uh 900 million weekly active users.
00:04:21.920 --> 00:04:30.319
I think officially uh still at the 800 million weekly active users, and 40% of Americans use generative AI in just two years.
00:04:30.480 --> 00:04:38.879
Um, for comparison, it took the internet like five times as long to get that number of people actually using it.
00:04:41.120 --> 00:04:51.360
And I think honestly, ChatGPT's overwhelming popularity is one of the reasons that we have this huge capability divide, right?
00:04:51.600 --> 00:05:01.360
Do you know, as an example, you can give a single prompt to a large language model uh like Claude or uh ChatGPT, I think are probably the two best at it right now.
00:05:01.519 --> 00:05:08.560
And they can go research the web, they can uh, you know, personalize uh things for you.
00:05:08.720 --> 00:05:16.480
They can grab uh images, they can put together spreadsheets that you can actually download, PowerPoints that look really good, right?
00:05:16.879 --> 00:05:18.639
All in one prompt, right?
00:05:18.800 --> 00:05:20.879
Most people don't know that, right?
00:05:20.959 --> 00:05:22.879
Unless you listen to this show, then you probably do.
00:05:23.040 --> 00:05:24.959
Most people think AI is just a chat bot.
00:05:25.040 --> 00:05:31.439
It's not, it is something that is completing um economically valuable work at a rate higher than human experts.
00:05:31.519 --> 00:05:35.680
And recent benchmarks show that, according to blind studies.
00:05:36.560 --> 00:05:53.519
So essentially, yesterday's expert is today's beginner, and that is how quickly things change because your technical teams now are probably just spitting out things like tokens and you know, MCP and RAG and vector databases and Ralph Wiggins, right?
00:05:53.759 --> 00:05:58.480
Like it's this misunderstood jargon that just drives waste.
00:05:58.560 --> 00:06:04.720
It stalls pilots and the vocabulary shifts faster than anyone can track.
00:06:07.519 --> 00:06:10.800
To get it right, you require the right words.
00:06:10.959 --> 00:06:11.439
All right.
00:06:11.600 --> 00:06:17.279
There is a reason why this is volume two in our uh start here series.
00:06:17.439 --> 00:06:19.519
Because number one, we set the foundation.
00:06:19.680 --> 00:06:24.720
We said here is generative AI, here's why it's not going away, here's why you need to invest in it.
00:06:24.879 --> 00:06:29.120
And the next thing is you have to be able to communicate effectively about it.
00:06:29.279 --> 00:06:32.160
All right, because they require the right words.
00:06:32.319 --> 00:06:34.959
You can't evaluate vendors or manage risk alone.
00:06:35.120 --> 00:06:35.519
Uh, right.
00:06:35.600 --> 00:06:43.040
You need to bring your technical teams together with your people leading change management, you know, with your HR, with your marketing, with your sales.
00:06:43.199 --> 00:06:45.519
You all need to be able to speak the same language.
00:06:45.680 --> 00:06:50.079
Uh, and you, you know, this is the new business terms, right?
00:06:50.240 --> 00:06:56.480
Uh, you know, MCPs and and uh you know, A2As are the new KPIs.
00:06:56.800 --> 00:06:57.839
Uh, they are.
00:06:57.920 --> 00:07:03.279
Uh so you have to be able to translate AI phrases into things that move the needle for your company.
00:07:03.680 --> 00:07:07.920
This is ultimately, at least when we look at most large language models.
00:07:08.000 --> 00:07:13.759
And I'm assuming if you're listening to the show, you know, you're either using them on the front end or the back end or both.
00:07:14.000 --> 00:07:16.240
But this is all it comes down to.
00:07:16.480 --> 00:07:20.720
There's a prompt, there's an action, and there's an outcome.
00:07:20.879 --> 00:07:21.279
All right.
00:07:21.360 --> 00:07:23.360
So humans prompt a large language model.
00:07:23.519 --> 00:07:27.199
Hopefully, you're using our prime prompt polish context engineering way, right?
00:07:27.279 --> 00:07:36.079
So again, go to starthirseries.com, get free access to uh our new and updated PPP course that teaches you how to get the most out of a large language model.
00:07:36.160 --> 00:07:38.319
But it starts with the human prompting a large language model.
00:07:38.480 --> 00:07:49.759
Then the model will kind of go on with the task, uh, you know, it'll summarize, search, or draft, and the model will read the prompt inside of its context window and it works within the confines of its architecture, right?
00:07:50.000 --> 00:07:55.040
So much of the outcome is not even decided by the prompt.
00:07:55.199 --> 00:08:01.439
It's decided by you using the right model, the right mode for the right reason, for the right task, right?
00:08:01.519 --> 00:08:05.839
Uh, but essentially the model reads the prompt the human puts inside of its context window.
00:08:05.920 --> 00:08:10.000
Uh, you know, people, you need to understand things like context window.
00:08:10.079 --> 00:08:12.639
You need to understand tokens, and we're gonna be going over that today.
00:08:12.800 --> 00:08:19.680
Uh, then the model retrieves whatever data that it needs to be a helpful assistant and respond to your query, right?
00:08:19.839 --> 00:08:23.360
It's now models are agentic by nature, right?
00:08:23.680 --> 00:08:38.159
Agency, they make decisions on their own to accomplish a goal, which includes using different tools in their arsenal, uh, retrieving data on their own, you know, exploring multiple pathways and going back and saying, oh, this wasn't the right one.
00:08:38.240 --> 00:08:40.720
Um, and then they output a result.
00:08:40.879 --> 00:08:51.279
And then the humans hopefully observe what the model did, they trace uh the paths that it took, they verify the information, and then they use the outcomes.
00:08:51.519 --> 00:08:52.879
That's essentially right.
00:08:52.960 --> 00:08:55.039
If if you're like, well, how does AI even work?
00:08:55.120 --> 00:08:56.080
There you go, right?
00:08:56.240 --> 00:08:58.320
From prompt to action to outcome.
00:09:00.240 --> 00:09:06.480
And this is the foundation for every AI term out there.
00:09:06.799 --> 00:09:07.360
All right.
00:09:07.759 --> 00:09:09.519
And we did cover this last one.
00:09:09.600 --> 00:09:14.639
So this is the the 30-second uh refresher from volume one in this series.
00:09:14.799 --> 00:09:18.799
So generative AI systems, they generate something, right?
00:09:18.879 --> 00:09:22.639
Uh, whether that's text, images, code, video, music, right?
00:09:22.799 --> 00:09:25.840
Generative AI is not just chatbots.
00:09:25.919 --> 00:09:27.200
It is now everything.
00:09:27.360 --> 00:09:37.039
Uh large language models, which fall under uh, you know, kind of the generative AI umbrella, they generate responses by predicting the next word uh or token in a sequence.
00:09:37.120 --> 00:09:44.879
And they do that, you know, whether they're using uh the GPT technology, which is what a lot of today's most of today's large language models are built on top of.
00:09:45.039 --> 00:09:50.960
Uh, some of the you know, image models use diffusion technology, which is essentially they kind of fill in the dots, right?
00:09:51.120 --> 00:09:56.159
Where large language models use more kind of next uh next word or next token prediction.
00:09:56.720 --> 00:10:02.320
And most current AI products that you use, right, they're just wrapped around these other products, right?
00:10:02.399 --> 00:10:06.960
So essentially you have three big AI labs that dominate, right?
00:10:07.200 --> 00:10:18.960
You have Enthropic, you have Google, you have OpenAI and their models, Enthropic Claude, OpenAI's uh ChatGPT models, so the GPT models, and then you have the Google Gemini models.
00:10:19.039 --> 00:10:21.360
And you're like, oh wait, what about Microsoft Copilot?
00:10:21.440 --> 00:10:28.639
Uh well, they did just start making some of their own models, but for the most part, uh Copilot just uses uh the OpenAI models.
00:10:28.720 --> 00:10:34.320
They use a little bit of uh Anthropics models as well inside of their Microsoft, Microsoft Copilot 360.
00:10:35.279 --> 00:10:37.200
Uh or 365, sorry.
00:10:37.279 --> 00:10:38.559
Uh yeah, left out.
00:10:38.639 --> 00:10:41.440
That was Microsoft Xbox 360.
00:10:41.600 --> 00:10:48.799
Uh so the, you know, a lot of this other software that you use, most people don't know this, right?
00:10:48.960 --> 00:10:50.639
They're not out there creating their own models.
00:10:50.720 --> 00:10:52.480
They're not competing with OpenAI.
00:10:52.559 --> 00:10:59.120
For the most part, you know, if you're using any AI product and there's literally hundreds of thousands of AI products out there, right?
00:10:59.279 --> 00:11:03.200
Pieces of software, for the most part, they're using one of those three, probably.
00:11:03.279 --> 00:11:07.279
There's other smaller uh AI labs, but for the most part, those are the three that dominate.
00:11:07.440 --> 00:11:13.200
And essentially these models that we use, they don't technically understand our words, right?
00:11:13.440 --> 00:11:18.240
Um, the human who inputs the words, those words get translated to tokens.
00:11:18.399 --> 00:11:24.000
And the model itself, it doesn't think in words, it thinks and produces uh in tokens and then converts it back.
00:11:24.159 --> 00:11:28.799
So there's almost like a uh, you know, it's kind of like when you go into a new country, you have to convert your money.
00:11:28.879 --> 00:11:30.320
And then when you leave, you got to convert it back.
00:11:30.399 --> 00:11:31.519
It's the same thing, right?
00:11:31.600 --> 00:11:36.240
When I'm giving something to a large language model, it's can uh it's converting that to tokens, right?
00:11:36.320 --> 00:11:40.720
Which helps it truly understand what I'm trying to say because it doesn't understand words.
00:11:40.799 --> 00:11:42.720
It converts everything to tokens, right?
00:11:42.799 --> 00:11:51.919
Which is why sometimes, you know, how you communicate with a model in the same way of how you communicate with your teammates to get the most out of AI, how you communicate with the model is very important, right?
00:11:52.080 --> 00:11:56.080
Um, an example that I've given before, the word just, right?
00:11:56.480 --> 00:11:58.559
Um context matters.
00:11:58.720 --> 00:12:01.519
The word just can be tokenized by a model.
00:12:01.679 --> 00:12:05.039
Uh, just the single word just, seven different ways.
00:12:05.200 --> 00:12:06.639
The last time I checked, right?
00:12:06.799 --> 00:12:07.919
Something can be just.
00:12:08.080 --> 00:12:09.679
You can use just as a qualifier.
00:12:09.840 --> 00:12:10.960
I just want to talk.
00:12:11.120 --> 00:12:11.759
Um, right.
00:12:11.840 --> 00:12:14.240
There's like so many different meanings for words.
00:12:14.320 --> 00:12:21.679
Uh, in the same way that when you're young and you're learning whatever language that you learn, you learn, oh, different words can mean different things.
00:12:21.840 --> 00:12:26.080
And it's the words surrounding uh those words that give context and help us understand.
00:12:26.240 --> 00:12:32.399
So large language models, kind of the same thing, but uh instead of the words around, it's the tokens around.
00:12:33.200 --> 00:12:36.000
So let's talk about those tokens and context windows.
00:12:36.080 --> 00:12:36.960
This is important.
00:12:37.279 --> 00:12:40.639
Tokens are essentially text chunks, right?
00:12:40.720 --> 00:12:42.799
So they're about four characters each.
00:12:42.960 --> 00:12:51.519
So a word could be anywhere from one to, you know, uh up to a handful of tokens, and they're used to measure input and output length.
00:12:51.679 --> 00:12:57.519
So when you have a conversation with a large language model, like I said, everything gets put into tokens.
00:12:57.679 --> 00:13:02.720
And I like to think of, you know, an easy way to think about it is a hard drive, right?
00:13:03.200 --> 00:13:08.320
Um when you except let's think of a hard drive worked a different way, right?
00:13:08.480 --> 00:13:13.279
Right now, if you're trying to download a file and your hard drive's full, it just says, oh, it's full.
00:13:13.600 --> 00:13:15.600
A context window is like a hard drive.
00:13:15.679 --> 00:13:18.639
But the difference is it's automatically going to keep working.
00:13:18.720 --> 00:13:24.639
And instead of saying, hey, the context window is full, it's just gonna kick out the first thing that you put in, right?
00:13:24.799 --> 00:13:28.720
And a lot of times the first thing that you put into a large language model is the most important, right?
00:13:28.799 --> 00:13:31.120
It's saying, hey, here's the most important information.
00:13:31.200 --> 00:13:32.559
Here's what I want you to do for me.
00:13:32.799 --> 00:13:40.399
Then if you keep working with that, um, you know, keep spinning tokens back and forth with that large language model, eventually the hard drive is gonna get full.
00:13:40.480 --> 00:13:44.159
Except it's not gonna say hard drive's full, it's just gonna forget the first things you said.
00:13:44.240 --> 00:13:45.519
That's a context window.
00:13:45.600 --> 00:13:47.200
Um, and then parameters.
00:13:47.360 --> 00:13:50.399
Parameters are important to understand about large language models as well.
00:13:50.559 --> 00:13:55.679
Um, essentially, these are the neural network connections of a model.
00:13:55.919 --> 00:13:59.440
The more parameters it has, the more power, right?
00:13:59.519 --> 00:14:04.960
You can think of parameters uh kind of like horsepower of a car, uh right.
00:14:05.200 --> 00:14:13.840
There's more capability of these models that have more parameters, but there's higher computational costs in resource usage, right?
00:14:14.000 --> 00:14:18.720
So uh large, large language models, I think, are actually getting smaller.
00:14:18.879 --> 00:14:32.720
Uh, a lot of times companies don't say how many parameters there are, but you know, kind of the last widely reported uh one we knew about was the GPT-4 uh architecture from OpenAI that had about 2 trillion, uh reportedly 2 trillion parameters.
00:14:32.799 --> 00:14:33.440
So that's a lot.
00:14:33.600 --> 00:14:37.120
I've always said for a long time, large language models are going to get smaller.
00:14:37.360 --> 00:14:42.480
And in the long run, I think most people are gonna be using more small language models.
00:14:42.639 --> 00:14:45.519
We've even seen that a ton in the last week or two.
00:14:45.600 --> 00:14:54.000
Uh a lot of uh specialized models from the big companies, uh, right, for things like translation, uh, medical finance, right?
00:14:54.080 --> 00:14:57.360
The big AI labs are starting to put out very, very small models.
00:14:57.519 --> 00:15:07.440
I do think the future, uh, you know, there's gonna be, you know, open AI will have thousands or tens of thousands of models and they'll be spitting them out faster than we can even keep up with, right?
00:15:07.600 --> 00:15:11.759
Um, but anyways, parameters is capability.
00:15:13.039 --> 00:15:16.720
Now, let's talk about some dorky things.
00:15:16.879 --> 00:15:17.279
All right.
00:15:17.519 --> 00:15:20.159
And if you are a beginner, forgive me.
00:15:20.240 --> 00:15:28.879
Um, because for our more intermediate uh people or people who just need a reminder, um, we really have to understand data, right?
00:15:29.039 --> 00:15:33.279
So rag is probably a business term that you should know.
00:15:33.519 --> 00:15:39.279
And I think rag, traditional rag as we know it, is going to very much change, right?
00:15:39.360 --> 00:15:43.279
I'm not gonna say rag is going to go away, but rag is retrieval augmented generation.
00:15:43.440 --> 00:15:51.440
And that's where a model can retrieve relevant documents and then it grounds its responses using that retrieved content, right?
00:15:51.519 --> 00:15:59.919
So when, as an example, um, you know, before uh ChatGPT came out, the GPT technology was available and many businesses used it.
00:16:00.000 --> 00:16:08.960
And then they essentially uh used retrieval augmented generation uh to kind of wrap OpenAI's responses in their business data.
00:16:09.120 --> 00:16:12.639
Uh so embeddings are also something important to know.
00:16:12.799 --> 00:16:23.679
So these embeddings are it's kind of like group of data or turns uh text into vector embeddings, allowing for efficient similarity search in a vector database.
00:16:23.840 --> 00:16:24.240
All right.
00:16:24.480 --> 00:16:29.519
This one uh is a little bit more difficult to talk about, right?
00:16:29.600 --> 00:16:38.960
I'll probably uh in in the start here series uh in our community there, I'll probably put a couple graphics, uh, but it's it's a little hard to explain on a podcast without doing a deep dive, right?
00:16:39.039 --> 00:16:44.399
But essentially, text is split into smaller segments or chunks before the embeddings happen.
00:16:44.480 --> 00:16:47.919
And then the chunk size and the overlap settings control how it's done.
00:16:48.159 --> 00:16:50.559
So good chunking splits information logically.
00:16:50.639 --> 00:16:59.039
So answers include the right context and details, and bad chunking breaks up context or combines unrelated info, leading to wrong or incomplete answers.
00:16:59.200 --> 00:17:07.920
So this is pretty this is uh much more relevant for companies that are building with AP uh will building um with large language models on the back end, right?
00:17:08.000 --> 00:17:17.200
So when I say back end and front end, front end, if you're using a large language model on the front end, you're going to chat gpt.com and you're using that chat GPT interface.
00:17:17.279 --> 00:17:20.000
That's a front end user or claw.ai, right?
00:17:20.160 --> 00:17:30.720
A backend means that you maybe have a team of developers, or even now non-technical people are obviously using large language models on the back end, but you're using uh the company's API.
00:17:30.880 --> 00:17:33.440
And so essentially you're using their technology.
00:17:33.519 --> 00:17:40.720
And when you're doing that, a lot of times, then you might be working with RAG, uh, you know, these vector databases, chunking, all of that.
00:17:40.799 --> 00:17:45.200
So if you're a front-end user, you don't have to worry about that as much, but you should know what RAG is, right?
00:17:45.440 --> 00:17:52.640
RAG is what makes uh those queries uh kind of as they go off to the large language model, makes them much more accurate.
00:17:52.799 --> 00:18:05.039
It reduces the the likelihood and the prevalence of hallucinations when you are first interjecting um your company's data, your secret sauce, uh, how your company thinks, right?
00:18:05.119 --> 00:18:11.680
All that important stuff that helps new hires and uh, you know, people that have been at a company for a very long time make good decisions.
00:18:11.759 --> 00:18:14.720
It also helps large language models make good decisions.
00:18:15.440 --> 00:18:18.160
And large language models are obviously changing.
00:18:18.400 --> 00:18:25.440
Um, you know, they started as a helpful assistant in that little chat window, and there is nothing else they can really do.
00:18:25.759 --> 00:18:29.279
Now, models are agentic by nature.
00:18:29.440 --> 00:18:30.000
All right.
00:18:30.160 --> 00:18:44.960
Uh, I like to say we have our old school transformer models, and then we have our newer reasoning models, even though these reasoning models are still built for the most part on the transformer technology, the generated pre-trained transformer GVT technology, right?
00:18:45.039 --> 00:18:51.279
I like to say we had more of these, you know, old school transformer, helpful assistant chatbots.
00:18:51.440 --> 00:18:54.799
Now we have more autonomous workers, right?
00:18:54.960 --> 00:19:01.279
These models now, specifically, you know, anything from Claude in the 4.5 um family.
00:19:01.359 --> 00:19:05.680
So OBIS 4.5, Sonnet 4.5, Haiku 4.5.
00:19:05.920 --> 00:19:12.160
Um, and then with Google Gemini, all the Gemini's new models uh, you know, are hybrid or can think, right?
00:19:12.240 --> 00:19:16.880
So they can either give you that kind of next token prediction or they can think about it, right?
00:19:16.960 --> 00:19:24.319
They can be agentic in nature, or sometimes they think if they need to think about it, or they just think and they're like, all right, I'm gonna spit on an answer.
00:19:24.400 --> 00:19:29.119
And then with OpenAI's models, for the most part, uh, aside from the default one, right?
00:19:29.200 --> 00:19:37.920
But uh, you know, make sure that's why I always say if you're using uh a model, make sure you're using a model that thinks because these are models that can call tools, right?
00:19:38.079 --> 00:19:50.319
They can um decide, hey, this user is asking me about something that might be a little old in my training data or the data that the model is built on top of might be might give old or outdated or wrong answers.
00:19:50.480 --> 00:19:53.839
So an agentic model can decide I'm gonna call a tool.
00:19:53.920 --> 00:20:01.839
And it can trigger functions like search, uh, you know, going through your email, letting the uh agentic model interact with external uh services, right?
00:20:01.920 --> 00:20:06.640
You have these things called connectors, which, you know, it's kind of like rag mini, right?
00:20:06.720 --> 00:20:17.839
We're true blog meta generation, but a very simple version, uh, it used to take uh a lot of smart engineers and usually uh six, seven, eight figures to get rag.
00:20:18.160 --> 00:20:21.680
Now you can get a very simple version in a couple of clicks, you know.
00:20:21.839 --> 00:20:26.480
Most large language models on the front end call those something like connectors or apps, right?
00:20:26.559 --> 00:20:37.759
So you can connect things like your like your email, your calendar, uh, your you know, your SharePoint or your Google Drive, um, your OneDrive, things like that, Salesforce, Slack, right?
00:20:37.839 --> 00:20:42.319
And then all of that um invaluable, right?
00:20:42.559 --> 00:20:50.480
Like all of that data that your company needs and uses that you're like, man, two years ago, it was so hard to work with that.
00:20:50.559 --> 00:20:52.319
Now it's there almost automatically, right?
00:20:52.400 --> 00:20:53.359
With these connectors.
00:20:53.519 --> 00:20:54.880
You also have MCP.
00:20:54.960 --> 00:20:56.880
This is the model context protocol.
00:20:57.119 --> 00:21:07.279
Um and this is uh a technology that was first uh invented and popularized by Enthropic, but the rest of the industry has adopted it, which is great.
00:21:07.359 --> 00:21:11.039
You know, one thing about the AI industry, see like it's weird.
00:21:11.119 --> 00:21:12.559
They all kind of work together, right?
00:21:12.640 --> 00:21:15.119
So, even as an example, I was just talking about connectors.
00:21:15.279 --> 00:21:24.240
Uh, if if you're a Google organization and you use Gmail and Google Drive, you can technically use those connectors inside OpenAI's products.
00:21:24.400 --> 00:21:27.279
You can use it inside Microsoft products, inside Claude's products.
00:21:27.359 --> 00:21:36.319
So, in the same way that MCP is kind of this, it's kind of like uh um, you know, how the internet talks to, you know, websites can talk to each other with APIs.
00:21:36.559 --> 00:21:40.480
This is kind of how AI tools can talk with uh to each other with MCPs.
00:21:40.640 --> 00:21:44.160
Uh so the model context protocol is definitely the most popular one.
00:21:44.319 --> 00:21:48.240
There's a handful of other ones that are uh widely used and widely adopted.
00:21:48.319 --> 00:21:53.519
But for the most part, you know, these models have essentially languages that they can talk to each other.
00:21:53.680 --> 00:21:59.759
So you don't have to, you know, send it from an AI um system to a website.
00:22:00.240 --> 00:22:04.720
To use the website's API to go to another website to then send it to a model, right?
00:22:04.799 --> 00:22:07.039
So it's just a direct connection.
00:22:07.119 --> 00:22:12.720
The MCP connection allows different AI models to talk to each other or different AI tools.
00:22:13.039 --> 00:22:14.079
And then scaffolding.
00:22:14.160 --> 00:22:20.000
So if you've ever heard of the word scaffolding, I say it a little, I try not to, uh, just because I don't want to, you know, confuse any people.
00:22:20.160 --> 00:22:28.400
But that essentially defines the structure that agents use to chain together multiple steps or tools for more complex workflows.
00:22:28.640 --> 00:22:35.759
So, you know, uh, sometimes, honestly, like the models themselves can all be fairly similar, right?
00:22:35.839 --> 00:22:40.240
If you look at the benchmarks, they're, you know, 0.1% from you know each other.
00:22:40.319 --> 00:22:42.720
A lot of times it's the scaffolding, uh, right.
00:22:42.799 --> 00:22:51.200
It's it's how the model actually works and how it's able to successfully or unsuccessfully use all these different tools that it has at its disposal.
00:22:51.440 --> 00:23:02.880
So uh the large language models have largely gone from helpful AI assistants that were powered by older, slower models that hallucinated a lot.
00:23:03.119 --> 00:23:11.039
Now they're the models themselves are agentic by nature and they can use tools and you can connect to your business's dynamic data.
00:23:11.759 --> 00:23:13.759
But there's risks and you need to know them.
00:23:13.839 --> 00:23:23.119
Hallucinations, probably a word most people have heard of, but it's essentially a lie, uh right, or a false statement, or a half truth that's put out there very confidently, uh.
00:23:23.519 --> 00:23:40.160
But if you do know the basics of context engineering, so if you've taken our free uh prime prompt polish uh course and our free prime prompt polish, uh can't speak Prime Prompt Polish Pro course, you know the basics of context engineering and your hallucination rate is gonna go down.
00:23:40.319 --> 00:23:52.960
But if you're using the right model and if you know the basics of prompt engineering and uh sharing your company's context, hallucinations are, I'm not gonna say they're gone, but they're essentially gone, right?
00:23:53.279 --> 00:23:54.880
People overlook that.
00:23:55.039 --> 00:24:10.160
If you know what you're doing, if you're using the right model, if you're connecting in your company's data, if you're uh looking at the chain of thought, uh doing basic observability, uh, you know, human in the loop, even lazy human in the loop, hallucinations are down to about zero.
00:24:10.319 --> 00:24:15.599
Uh prompt injections, you have to look out for those, uh, especially as we have agentic browsers, right?
00:24:15.680 --> 00:24:27.119
So these are browsers that are run by an AI agent and they can go browse the web for you and interact uh with interfaces on the web with graphical interfaces, you know, click, you know, click drop downs and fill out forms for you.
00:24:27.200 --> 00:24:28.960
But there's prompt injections, but right?
00:24:29.039 --> 00:24:34.240
So there's prompt injections that get into training data, which impacts the models themselves.
00:24:34.400 --> 00:24:40.559
But on the agentic side, for agents or agentic browsers, there's prompt injections on web pages, right?
00:24:40.640 --> 00:24:45.759
And some of the earlier agents and agenc browsers didn't have great safeguards against those, right?
00:24:45.839 --> 00:24:48.000
So, you know, as an agent, right?
00:24:48.079 --> 00:24:51.920
A simple simple example, um, you know, I could put something on my website.
00:24:52.079 --> 00:25:03.920
Uh, so if an agent or an agentic browser was trying to use it and I could say, hey, you know, go uh, you know, go to this site and you know buy uh a hundred boxes of cereal, you know, put it in your Amazon cart or something like that.
00:25:04.079 --> 00:25:07.920
That's a very uh safe and happy prompt injection, right?
00:25:08.000 --> 00:25:09.440
It's it's not causing a lot of harm.
00:25:09.519 --> 00:25:25.839
I it just tricked the model uh or an agentic model into doing something that it didn't intend to, but there's obviously very harmful and unsafe uh you know prompt injections, just like there's you know, phishing emails and all these you know scams that have been going around uh you know with malicious links and all that.
00:25:25.920 --> 00:25:26.559
The same thing.
00:25:26.799 --> 00:25:28.160
Models have risks.
00:25:28.319 --> 00:25:31.359
So companies need to understand and talk about guardrails.
00:25:31.519 --> 00:25:37.839
That's when you combine uh human-made policies and filters to enforce safety and prevent misuse.
00:25:39.359 --> 00:25:40.000
All right.
00:25:40.400 --> 00:25:43.920
So this is kind of the translation playbook.
00:25:44.079 --> 00:25:44.400
All right.
00:25:44.480 --> 00:25:47.279
So in the future, we might be updating this, right?
00:25:47.359 --> 00:25:50.880
If we keep the start here series going on for a longer time.
00:25:51.039 --> 00:25:54.799
Uh so make sure I always say check the show notes of this show.
00:25:54.880 --> 00:26:08.079
So if you're listening on the podcast, we're gonna always keep the show notes updated and uh in the community in which you get free access to by going to starthearseries.com, uh, we're gonna keep the resources there updated as well.
00:26:08.240 --> 00:26:10.319
So there's always gonna be new terminology.
00:26:10.480 --> 00:26:12.240
Some of the terminology is gonna change, right?
00:26:12.319 --> 00:26:18.640
Like as an example, the definition of artificial general intelligence changes every month, right?
00:26:19.039 --> 00:26:26.640
By uh definitions from 10 years ago, we're well well past AGI, but you know, how people define it say maybe we're not there.
00:26:26.720 --> 00:26:30.640
So we're gonna keep the definitions up to uh up to date in our community.
00:26:30.720 --> 00:26:35.759
We're gonna have a little AI translation playbook, but you need to ask yourself these questions.
00:26:36.000 --> 00:26:36.319
All right.
00:26:36.400 --> 00:26:40.720
I'm gonna go back to how I started, the technical people and the non-technical business leaders.
00:26:40.880 --> 00:26:50.000
These are the questions that you need to be asking yourself out loud when it comes to using AI and understanding it and hopefully closing that capability and education gap.
00:26:50.240 --> 00:26:55.759
Number one, what's the problem we're trying to solve and what's our current costs associated with that solution?
00:26:55.920 --> 00:26:56.160
Right?
00:26:56.319 --> 00:26:57.440
People overlook that.
00:26:57.599 --> 00:27:03.279
I'm like, you know, when people are trying to, you know, calculate ROI on Gen AI, uh, right?
00:27:03.359 --> 00:27:05.920
I always ask, okay, what's your current cost on this?
00:27:06.000 --> 00:27:09.119
And they're like, okay, well, you know, uh, we have this piece of software.
00:27:09.279 --> 00:27:11.119
And no, how much time is spent?
00:27:11.279 --> 00:27:11.920
No clue.
00:27:12.079 --> 00:27:14.400
You have to understand what your current costs are.
00:27:14.640 --> 00:27:20.319
Uh, next questions what model, data, tools, and who approves?
00:27:20.559 --> 00:27:21.920
You have to go through those.
00:27:22.079 --> 00:27:31.519
Next, how can we responsibly run the fastest sprint possible with measurement and evaluation, with proper, uh, you know, proper QA?
00:27:31.839 --> 00:27:35.200
Um, you need to go fast.
00:27:35.680 --> 00:27:42.799
You have to be careful, but you have to go as quickly as possible while still being careful while you're implementing AI.
00:27:42.960 --> 00:27:47.759
Because if you're too slow and if you take a pilot slow and if you're not building modularly, right?
00:27:47.839 --> 00:27:49.680
Now all of a sudden you did a year-long pilot.
00:27:49.839 --> 00:27:58.079
Oh, wait, that original model you started with is that was the best model in the world, is now the 50th best model, and it's bad.
00:27:58.160 --> 00:28:00.240
And now you're putting yourself at a disadvantage.
00:28:00.400 --> 00:28:07.680
You can't treat AI implementation in the same way that we've treated tax uh tech implementation for the past few decades.
00:28:07.920 --> 00:28:12.559
You need to ask what are the hallucination risks, the security risks, and the costs.
00:28:12.720 --> 00:28:18.960
And then last but not least, you always need to require receipts, the observability, traceability, uh, expert-driven loops.
00:28:19.119 --> 00:28:26.559
We don't need human, uh lazy, human of the loop, uh, logs, defining the scope and getting your evaluations first.
00:28:26.720 --> 00:28:34.480
You have to understand what you're going after to be able to measure it uh and go and you know, understand the evaluations in the process.
00:28:34.720 --> 00:28:35.039
All right.
00:28:35.119 --> 00:28:36.960
So now that's it.
00:28:37.200 --> 00:28:40.160
You now have the foundation for the rest of the start here series.
00:28:40.240 --> 00:28:40.559
All right.
00:28:40.720 --> 00:28:45.680
We went over what the heck is generative AI, how do large language models work and why it's the future of work.
00:28:45.759 --> 00:28:48.240
And now you have the language and the lingo down.
00:28:48.559 --> 00:28:48.880
All right.
00:28:49.039 --> 00:28:52.000
So what I want you to do, sit down with your people.
00:28:52.160 --> 00:28:53.759
You need to get a 30-day plan.
00:28:53.920 --> 00:28:57.759
You need to be able to sprint, you need to understand how you're gonna deal with mistakes.
00:28:57.839 --> 00:29:02.079
Uh, and then later you can go into the RAG, deep dive, ROI, security scale, all this stuff.
00:29:02.240 --> 00:29:06.079
But you have to set here's where we are, here's where we want to go.
00:29:06.319 --> 00:29:13.279
Um, figure out what's gonna happen in between, but you have to keep up to date with the language and the lingo.
00:29:13.440 --> 00:29:15.680
Otherwise, your plan is gonna fall apart.
00:29:15.839 --> 00:29:22.960
All right, one way it's not gonna fall apart, go listen to episode 628, very much so related uh to this episode.
00:29:23.039 --> 00:29:26.160
So that is what's the best large language model for your team?
00:29:26.400 --> 00:29:30.240
Seven steps to evaluate and create ROI for AI.
00:29:30.400 --> 00:29:32.480
All right, I hope this was helpful.
00:29:32.640 --> 00:29:40.640
Now you know the AI without the jargon, the AI language every business leader needs to live by in 2026.
00:29:40.799 --> 00:29:44.480
All right, if this was helpful, please go to starthearseries.com.
00:29:44.640 --> 00:29:55.839
That's gonna take you and give you free access to our inner circle community, where not only can you go and check all of our Start Here series uh on the web, right?
00:29:55.920 --> 00:30:02.240
Or if on the mobile app, if that's what you want to do, uh, but you're gonna be signed up for our free daily newsletter as well.
00:30:02.319 --> 00:30:07.119
And you can go take our free prime prompt polish uh context engineering course.
00:30:07.279 --> 00:30:07.759
All right.
00:30:07.920 --> 00:30:09.200
Thank you for tuning in.
00:30:09.359 --> 00:30:12.880
I hope to see you back tomorrow and every day for more everyday AI.
00:30:12.960 --> 00:30:13.920
Thanks, y'all.
00:00:02.640 --> 00:00:11.119
One of the most challenging things about AI that no one talks about is not being able to talk about AI effectively.
00:00:11.279 --> 00:00:14.640
And I think that there's three main reasons for this.
00:00:14.880 --> 00:00:23.359
Number one, the tech changes faster than literally anyone can keep up with, even someone that talks about AI every day, like myself.
00:00:23.679 --> 00:00:26.559
Two, companies don't train their employees.
00:00:26.800 --> 00:00:29.839
And three, it's kind of the summation of those parts.
00:00:30.000 --> 00:00:38.159
There's a huge gap between what people know about AI and what the models are capable of and can actually accomplish.
00:00:38.479 --> 00:00:42.479
And so much of this can be solved by talking about it.
00:00:42.560 --> 00:00:44.640
But therein lies the problem.
00:00:45.200 --> 00:00:49.759
AI is a mystery wrapped up in ever-changing jargon.
00:00:50.000 --> 00:01:06.000
So we're kicking off volume two of our start here series, where we're gonna karate chop the AI jargon in the jugular and break down the basics of AI language and lingo that every business leader needs to know in 2026.
00:01:06.239 --> 00:01:07.760
All right, I'm excited.
00:01:07.920 --> 00:01:09.200
Let's get into it.
00:01:09.439 --> 00:01:13.920
Welcome to our start here series and everyday AI.
00:01:14.079 --> 00:01:23.280
So this brand new series is our essential beginner and advanced AI overviews to help you launch your year strong.
00:01:23.439 --> 00:01:31.680
So if one of your big goals is to double down or to just better learn AI, this new Start Here series is for you.
00:01:31.920 --> 00:01:32.319
All right.
00:01:32.400 --> 00:01:36.959
So if you haven't already, please go to starthearseries.com.
00:01:37.040 --> 00:01:38.560
Yeah, new URL.
00:01:38.799 --> 00:01:47.920
Uh so uh there you're gonna be forwarded straight to uh sign up page to get free access to our inner circle community.
00:01:48.079 --> 00:01:59.040
And not only will you get uh instant access to our Prime Prompt Polish Chat GPT prompt engineering course, but you will also get access to the Start Here series.
00:01:59.200 --> 00:02:13.759
So whether you're listening to this in mid-January or sometime in March, we're gonna be putting the recap of every single Start Here series um episode inside the community there, uh inside the community there, so you can listen to them all in order.
00:02:14.080 --> 00:02:14.560
All right.
00:02:14.800 --> 00:02:29.759
Uh, and if you are brand new, brand new, make sure you go listen to the first uh series of this, which was episode 691 or volume one in the start here series, which was generative AI, how it works and why it matters in 2026 more than ever.
00:02:30.000 --> 00:02:30.479
All right.
00:02:30.719 --> 00:02:36.080
If you're new, everyday AI, it's a daily live stream podcast and free daily newsletter.
00:02:36.240 --> 00:02:40.639
But I've done like 700 episodes, and the most common question I get is where do I start?
00:02:40.719 --> 00:02:43.759
That's the whole reason that we're doing this series.
00:02:43.919 --> 00:02:47.599
So let's cut the jargon down, shall we?
00:02:47.759 --> 00:02:48.960
Uh, we're gonna get straight into it.
00:02:49.039 --> 00:02:54.800
So don't worry, this isn't gonna be a super boring run through of random vocabulary words.
00:02:55.039 --> 00:03:03.039
Uh, one thing that I've realized that I kind of have a knack for uh over the years is almost being a translator, right?
00:03:03.199 --> 00:03:06.240
And uh, like literally, there is there's an instance.
00:03:06.479 --> 00:03:10.000
Um, I was working with a pretty big company, right?
00:03:10.240 --> 00:03:12.479
Doing like$20 billion in in revenue.
00:03:12.639 --> 00:03:16.400
And I was sitting with their C-suite and I was sitting with their their head of AI.
00:03:16.560 --> 00:03:21.360
And it seemed like sometimes these two sides weren't speaking the same language.
00:03:21.599 --> 00:03:23.520
And that's because there's so much jargon.
00:03:23.599 --> 00:03:26.960
And because, like I said, kind of that capabilities and understanding gap.
00:03:27.039 --> 00:03:32.400
And that's really what I've learned to do by doing this show over the last three years is connecting those two.
00:03:32.479 --> 00:03:42.319
So whether you are that highly technical person, you know, who understands all these terms, hopefully this will help you understand where the non-technical business leaders are coming from.
00:03:42.400 --> 00:03:52.639
And for our non-technical business leaders, hopefully this serves as a nice foundational look at the language and lingo you need to know to, you know, talk about it.
00:03:53.039 --> 00:03:54.400
Let's start.
00:03:55.039 --> 00:03:56.080
Start where we do.
00:03:56.240 --> 00:03:58.000
AI is not going anywhere, it's here to stay.
00:03:58.159 --> 00:03:58.479
All right.
00:03:58.560 --> 00:04:01.280
Uh, we went over that in the first episode, but it's worth repeating.
00:04:01.439 --> 00:04:06.719
ChatGPT, when it launched in November 2022, it was literally world-breaking technology.
00:04:06.879 --> 00:04:12.879
No one had ever seen anything like it before, uh, not just its capabilities, but also the sheer numbers.
00:04:13.039 --> 00:04:21.839
Uh, right, they had 100 million users within two months after launch, and ChatGPT now has nearly uh 900 million weekly active users.
00:04:21.920 --> 00:04:30.319
I think officially uh still at the 800 million weekly active users, and 40% of Americans use generative AI in just two years.
00:04:30.480 --> 00:04:38.879
Um, for comparison, it took the internet like five times as long to get that number of people actually using it.
00:04:41.120 --> 00:04:51.360
And I think honestly, ChatGPT's overwhelming popularity is one of the reasons that we have this huge capability divide, right?
00:04:51.600 --> 00:05:01.360
Do you know, as an example, you can give a single prompt to a large language model uh like Claude or uh ChatGPT, I think are probably the two best at it right now.
00:05:01.519 --> 00:05:08.560
And they can go research the web, they can uh, you know, personalize uh things for you.
00:05:08.720 --> 00:05:16.480
They can grab uh images, they can put together spreadsheets that you can actually download, PowerPoints that look really good, right?
00:05:16.879 --> 00:05:18.639
All in one prompt, right?
00:05:18.800 --> 00:05:20.879
Most people don't know that, right?
00:05:20.959 --> 00:05:22.879
Unless you listen to this show, then you probably do.
00:05:23.040 --> 00:05:24.959
Most people think AI is just a chat bot.
00:05:25.040 --> 00:05:31.439
It's not, it is something that is completing um economically valuable work at a rate higher than human experts.
00:05:31.519 --> 00:05:35.680
And recent benchmarks show that, according to blind studies.
00:05:36.560 --> 00:05:53.519
So essentially, yesterday's expert is today's beginner, and that is how quickly things change because your technical teams now are probably just spitting out things like tokens and you know, MCP and RAG and vector databases and Ralph Wiggins, right?
00:05:53.759 --> 00:05:58.480
Like it's this misunderstood jargon that just drives waste.
00:05:58.560 --> 00:06:04.720
It stalls pilots and the vocabulary shifts faster than anyone can track.
00:06:07.519 --> 00:06:10.800
To get it right, you require the right words.
00:06:10.959 --> 00:06:11.439
All right.
00:06:11.600 --> 00:06:17.279
There is a reason why this is volume two in our uh start here series.
00:06:17.439 --> 00:06:19.519
Because number one, we set the foundation.
00:06:19.680 --> 00:06:24.720
We said here is generative AI, here's why it's not going away, here's why you need to invest in it.
00:06:24.879 --> 00:06:29.120
And the next thing is you have to be able to communicate effectively about it.
00:06:29.279 --> 00:06:32.160
All right, because they require the right words.
00:06:32.319 --> 00:06:34.959
You can't evaluate vendors or manage risk alone.
00:06:35.120 --> 00:06:35.519
Uh, right.
00:06:35.600 --> 00:06:43.040
You need to bring your technical teams together with your people leading change management, you know, with your HR, with your marketing, with your sales.
00:06:43.199 --> 00:06:45.519
You all need to be able to speak the same language.
00:06:45.680 --> 00:06:50.079
Uh, and you, you know, this is the new business terms, right?
00:06:50.240 --> 00:06:56.480
Uh, you know, MCPs and and uh you know, A2As are the new KPIs.
00:06:56.800 --> 00:06:57.839
Uh, they are.
00:06:57.920 --> 00:07:03.279
Uh so you have to be able to translate AI phrases into things that move the needle for your company.
00:07:03.680 --> 00:07:07.920
This is ultimately, at least when we look at most large language models.
00:07:08.000 --> 00:07:13.759
And I'm assuming if you're listening to the show, you know, you're either using them on the front end or the back end or both.
00:07:14.000 --> 00:07:16.240
But this is all it comes down to.
00:07:16.480 --> 00:07:20.720
There's a prompt, there's an action, and there's an outcome.
00:07:20.879 --> 00:07:21.279
All right.
00:07:21.360 --> 00:07:23.360
So humans prompt a large language model.
00:07:23.519 --> 00:07:27.199
Hopefully, you're using our prime prompt polish context engineering way, right?
00:07:27.279 --> 00:07:36.079
So again, go to starthirseries.com, get free access to uh our new and updated PPP course that teaches you how to get the most out of a large language model.
00:07:36.160 --> 00:07:38.319
But it starts with the human prompting a large language model.
00:07:38.480 --> 00:07:49.759
Then the model will kind of go on with the task, uh, you know, it'll summarize, search, or draft, and the model will read the prompt inside of its context window and it works within the confines of its architecture, right?
00:07:50.000 --> 00:07:55.040
So much of the outcome is not even decided by the prompt.
00:07:55.199 --> 00:08:01.439
It's decided by you using the right model, the right mode for the right reason, for the right task, right?
00:08:01.519 --> 00:08:05.839
Uh, but essentially the model reads the prompt the human puts inside of its context window.
00:08:05.920 --> 00:08:10.000
Uh, you know, people, you need to understand things like context window.
00:08:10.079 --> 00:08:12.639
You need to understand tokens, and we're gonna be going over that today.
00:08:12.800 --> 00:08:19.680
Uh, then the model retrieves whatever data that it needs to be a helpful assistant and respond to your query, right?
00:08:19.839 --> 00:08:23.360
It's now models are agentic by nature, right?
00:08:23.680 --> 00:08:38.159
Agency, they make decisions on their own to accomplish a goal, which includes using different tools in their arsenal, uh, retrieving data on their own, you know, exploring multiple pathways and going back and saying, oh, this wasn't the right one.
00:08:38.240 --> 00:08:40.720
Um, and then they output a result.
00:08:40.879 --> 00:08:51.279
And then the humans hopefully observe what the model did, they trace uh the paths that it took, they verify the information, and then they use the outcomes.
00:08:51.519 --> 00:08:52.879
That's essentially right.
00:08:52.960 --> 00:08:55.039
If if you're like, well, how does AI even work?
00:08:55.120 --> 00:08:56.080
There you go, right?
00:08:56.240 --> 00:08:58.320
From prompt to action to outcome.
00:09:00.240 --> 00:09:06.480
And this is the foundation for every AI term out there.
00:09:06.799 --> 00:09:07.360
All right.
00:09:07.759 --> 00:09:09.519
And we did cover this last one.
00:09:09.600 --> 00:09:14.639
So this is the the 30-second uh refresher from volume one in this series.
00:09:14.799 --> 00:09:18.799
So generative AI systems, they generate something, right?
00:09:18.879 --> 00:09:22.639
Uh, whether that's text, images, code, video, music, right?
00:09:22.799 --> 00:09:25.840
Generative AI is not just chatbots.
00:09:25.919 --> 00:09:27.200
It is now everything.
00:09:27.360 --> 00:09:37.039
Uh large language models, which fall under uh, you know, kind of the generative AI umbrella, they generate responses by predicting the next word uh or token in a sequence.
00:09:37.120 --> 00:09:44.879
And they do that, you know, whether they're using uh the GPT technology, which is what a lot of today's most of today's large language models are built on top of.
00:09:45.039 --> 00:09:50.960
Uh, some of the you know, image models use diffusion technology, which is essentially they kind of fill in the dots, right?
00:09:51.120 --> 00:09:56.159
Where large language models use more kind of next uh next word or next token prediction.
00:09:56.720 --> 00:10:02.320
And most current AI products that you use, right, they're just wrapped around these other products, right?
00:10:02.399 --> 00:10:06.960
So essentially you have three big AI labs that dominate, right?
00:10:07.200 --> 00:10:18.960
You have Enthropic, you have Google, you have OpenAI and their models, Enthropic Claude, OpenAI's uh ChatGPT models, so the GPT models, and then you have the Google Gemini models.
00:10:19.039 --> 00:10:21.360
And you're like, oh wait, what about Microsoft Copilot?
00:10:21.440 --> 00:10:28.639
Uh well, they did just start making some of their own models, but for the most part, uh Copilot just uses uh the OpenAI models.
00:10:28.720 --> 00:10:34.320
They use a little bit of uh Anthropics models as well inside of their Microsoft, Microsoft Copilot 360.
00:10:35.279 --> 00:10:37.200
Uh or 365, sorry.
00:10:37.279 --> 00:10:38.559
Uh yeah, left out.
00:10:38.639 --> 00:10:41.440
That was Microsoft Xbox 360.
00:10:41.600 --> 00:10:48.799
Uh so the, you know, a lot of this other software that you use, most people don't know this, right?
00:10:48.960 --> 00:10:50.639
They're not out there creating their own models.
00:10:50.720 --> 00:10:52.480
They're not competing with OpenAI.
00:10:52.559 --> 00:10:59.120
For the most part, you know, if you're using any AI product and there's literally hundreds of thousands of AI products out there, right?
00:10:59.279 --> 00:11:03.200
Pieces of software, for the most part, they're using one of those three, probably.
00:11:03.279 --> 00:11:07.279
There's other smaller uh AI labs, but for the most part, those are the three that dominate.
00:11:07.440 --> 00:11:13.200
And essentially these models that we use, they don't technically understand our words, right?
00:11:13.440 --> 00:11:18.240
Um, the human who inputs the words, those words get translated to tokens.
00:11:18.399 --> 00:11:24.000
And the model itself, it doesn't think in words, it thinks and produces uh in tokens and then converts it back.
00:11:24.159 --> 00:11:28.799
So there's almost like a uh, you know, it's kind of like when you go into a new country, you have to convert your money.
00:11:28.879 --> 00:11:30.320
And then when you leave, you got to convert it back.
00:11:30.399 --> 00:11:31.519
It's the same thing, right?
00:11:31.600 --> 00:11:36.240
When I'm giving something to a large language model, it's can uh it's converting that to tokens, right?
00:11:36.320 --> 00:11:40.720
Which helps it truly understand what I'm trying to say because it doesn't understand words.
00:11:40.799 --> 00:11:42.720
It converts everything to tokens, right?
00:11:42.799 --> 00:11:51.919
Which is why sometimes, you know, how you communicate with a model in the same way of how you communicate with your teammates to get the most out of AI, how you communicate with the model is very important, right?
00:11:52.080 --> 00:11:56.080
Um, an example that I've given before, the word just, right?
00:11:56.480 --> 00:11:58.559
Um context matters.
00:11:58.720 --> 00:12:01.519
The word just can be tokenized by a model.
00:12:01.679 --> 00:12:05.039
Uh, just the single word just, seven different ways.
00:12:05.200 --> 00:12:06.639
The last time I checked, right?
00:12:06.799 --> 00:12:07.919
Something can be just.
00:12:08.080 --> 00:12:09.679
You can use just as a qualifier.
00:12:09.840 --> 00:12:10.960
I just want to talk.
00:12:11.120 --> 00:12:11.759
Um, right.
00:12:11.840 --> 00:12:14.240
There's like so many different meanings for words.
00:12:14.320 --> 00:12:21.679
Uh, in the same way that when you're young and you're learning whatever language that you learn, you learn, oh, different words can mean different things.
00:12:21.840 --> 00:12:26.080
And it's the words surrounding uh those words that give context and help us understand.
00:12:26.240 --> 00:12:32.399
So large language models, kind of the same thing, but uh instead of the words around, it's the tokens around.
00:12:33.200 --> 00:12:36.000
So let's talk about those tokens and context windows.
00:12:36.080 --> 00:12:36.960
This is important.
00:12:37.279 --> 00:12:40.639
Tokens are essentially text chunks, right?
00:12:40.720 --> 00:12:42.799
So they're about four characters each.
00:12:42.960 --> 00:12:51.519
So a word could be anywhere from one to, you know, uh up to a handful of tokens, and they're used to measure input and output length.
00:12:51.679 --> 00:12:57.519
So when you have a conversation with a large language model, like I said, everything gets put into tokens.
00:12:57.679 --> 00:13:02.720
And I like to think of, you know, an easy way to think about it is a hard drive, right?
00:13:03.200 --> 00:13:08.320
Um when you except let's think of a hard drive worked a different way, right?
00:13:08.480 --> 00:13:13.279
Right now, if you're trying to download a file and your hard drive's full, it just says, oh, it's full.
00:13:13.600 --> 00:13:15.600
A context window is like a hard drive.
00:13:15.679 --> 00:13:18.639
But the difference is it's automatically going to keep working.
00:13:18.720 --> 00:13:24.639
And instead of saying, hey, the context window is full, it's just gonna kick out the first thing that you put in, right?
00:13:24.799 --> 00:13:28.720
And a lot of times the first thing that you put into a large language model is the most important, right?
00:13:28.799 --> 00:13:31.120
It's saying, hey, here's the most important information.
00:13:31.200 --> 00:13:32.559
Here's what I want you to do for me.
00:13:32.799 --> 00:13:40.399
Then if you keep working with that, um, you know, keep spinning tokens back and forth with that large language model, eventually the hard drive is gonna get full.
00:13:40.480 --> 00:13:44.159
Except it's not gonna say hard drive's full, it's just gonna forget the first things you said.
00:13:44.240 --> 00:13:45.519
That's a context window.
00:13:45.600 --> 00:13:47.200
Um, and then parameters.
00:13:47.360 --> 00:13:50.399
Parameters are important to understand about large language models as well.
00:13:50.559 --> 00:13:55.679
Um, essentially, these are the neural network connections of a model.
00:13:55.919 --> 00:13:59.440
The more parameters it has, the more power, right?
00:13:59.519 --> 00:14:04.960
You can think of parameters uh kind of like horsepower of a car, uh right.
00:14:05.200 --> 00:14:13.840
There's more capability of these models that have more parameters, but there's higher computational costs in resource usage, right?
00:14:14.000 --> 00:14:18.720
So uh large, large language models, I think, are actually getting smaller.
00:14:18.879 --> 00:14:32.720
Uh, a lot of times companies don't say how many parameters there are, but you know, kind of the last widely reported uh one we knew about was the GPT-4 uh architecture from OpenAI that had about 2 trillion, uh reportedly 2 trillion parameters.
00:14:32.799 --> 00:14:33.440
So that's a lot.
00:14:33.600 --> 00:14:37.120
I've always said for a long time, large language models are going to get smaller.
00:14:37.360 --> 00:14:42.480
And in the long run, I think most people are gonna be using more small language models.
00:14:42.639 --> 00:14:45.519
We've even seen that a ton in the last week or two.
00:14:45.600 --> 00:14:54.000
Uh a lot of uh specialized models from the big companies, uh, right, for things like translation, uh, medical finance, right?
00:14:54.080 --> 00:14:57.360
The big AI labs are starting to put out very, very small models.
00:14:57.519 --> 00:15:07.440
I do think the future, uh, you know, there's gonna be, you know, open AI will have thousands or tens of thousands of models and they'll be spitting them out faster than we can even keep up with, right?
00:15:07.600 --> 00:15:11.759
Um, but anyways, parameters is capability.
00:15:13.039 --> 00:15:16.720
Now, let's talk about some dorky things.
00:15:16.879 --> 00:15:17.279
All right.
00:15:17.519 --> 00:15:20.159
And if you are a beginner, forgive me.
00:15:20.240 --> 00:15:28.879
Um, because for our more intermediate uh people or people who just need a reminder, um, we really have to understand data, right?
00:15:29.039 --> 00:15:33.279
So rag is probably a business term that you should know.
00:15:33.519 --> 00:15:39.279
And I think rag, traditional rag as we know it, is going to very much change, right?
00:15:39.360 --> 00:15:43.279
I'm not gonna say rag is going to go away, but rag is retrieval augmented generation.
00:15:43.440 --> 00:15:51.440
And that's where a model can retrieve relevant documents and then it grounds its responses using that retrieved content, right?
00:15:51.519 --> 00:15:59.919
So when, as an example, um, you know, before uh ChatGPT came out, the GPT technology was available and many businesses used it.
00:16:00.000 --> 00:16:08.960
And then they essentially uh used retrieval augmented generation uh to kind of wrap OpenAI's responses in their business data.
00:16:09.120 --> 00:16:12.639
Uh so embeddings are also something important to know.
00:16:12.799 --> 00:16:23.679
So these embeddings are it's kind of like group of data or turns uh text into vector embeddings, allowing for efficient similarity search in a vector database.
00:16:23.840 --> 00:16:24.240
All right.
00:16:24.480 --> 00:16:29.519
This one uh is a little bit more difficult to talk about, right?
00:16:29.600 --> 00:16:38.960
I'll probably uh in in the start here series uh in our community there, I'll probably put a couple graphics, uh, but it's it's a little hard to explain on a podcast without doing a deep dive, right?
00:16:39.039 --> 00:16:44.399
But essentially, text is split into smaller segments or chunks before the embeddings happen.
00:16:44.480 --> 00:16:47.919
And then the chunk size and the overlap settings control how it's done.
00:16:48.159 --> 00:16:50.559
So good chunking splits information logically.
00:16:50.639 --> 00:16:59.039
So answers include the right context and details, and bad chunking breaks up context or combines unrelated info, leading to wrong or incomplete answers.
00:16:59.200 --> 00:17:07.920
So this is pretty this is uh much more relevant for companies that are building with AP uh will building um with large language models on the back end, right?
00:17:08.000 --> 00:17:17.200
So when I say back end and front end, front end, if you're using a large language model on the front end, you're going to chat gpt.com and you're using that chat GPT interface.
00:17:17.279 --> 00:17:20.000
That's a front end user or claw.ai, right?
00:17:20.160 --> 00:17:30.720
A backend means that you maybe have a team of developers, or even now non-technical people are obviously using large language models on the back end, but you're using uh the company's API.
00:17:30.880 --> 00:17:33.440
And so essentially you're using their technology.
00:17:33.519 --> 00:17:40.720
And when you're doing that, a lot of times, then you might be working with RAG, uh, you know, these vector databases, chunking, all of that.
00:17:40.799 --> 00:17:45.200
So if you're a front-end user, you don't have to worry about that as much, but you should know what RAG is, right?
00:17:45.440 --> 00:17:52.640
RAG is what makes uh those queries uh kind of as they go off to the large language model, makes them much more accurate.
00:17:52.799 --> 00:18:05.039
It reduces the the likelihood and the prevalence of hallucinations when you are first interjecting um your company's data, your secret sauce, uh, how your company thinks, right?
00:18:05.119 --> 00:18:11.680
All that important stuff that helps new hires and uh, you know, people that have been at a company for a very long time make good decisions.
00:18:11.759 --> 00:18:14.720
It also helps large language models make good decisions.
00:18:15.440 --> 00:18:18.160
And large language models are obviously changing.
00:18:18.400 --> 00:18:25.440
Um, you know, they started as a helpful assistant in that little chat window, and there is nothing else they can really do.
00:18:25.759 --> 00:18:29.279
Now, models are agentic by nature.
00:18:29.440 --> 00:18:30.000
All right.
00:18:30.160 --> 00:18:44.960
Uh, I like to say we have our old school transformer models, and then we have our newer reasoning models, even though these reasoning models are still built for the most part on the transformer technology, the generated pre-trained transformer GVT technology, right?
00:18:45.039 --> 00:18:51.279
I like to say we had more of these, you know, old school transformer, helpful assistant chatbots.
00:18:51.440 --> 00:18:54.799
Now we have more autonomous workers, right?
00:18:54.960 --> 00:19:01.279
These models now, specifically, you know, anything from Claude in the 4.5 um family.
00:19:01.359 --> 00:19:05.680
So OBIS 4.5, Sonnet 4.5, Haiku 4.5.
00:19:05.920 --> 00:19:12.160
Um, and then with Google Gemini, all the Gemini's new models uh, you know, are hybrid or can think, right?
00:19:12.240 --> 00:19:16.880
So they can either give you that kind of next token prediction or they can think about it, right?
00:19:16.960 --> 00:19:24.319
They can be agentic in nature, or sometimes they think if they need to think about it, or they just think and they're like, all right, I'm gonna spit on an answer.
00:19:24.400 --> 00:19:29.119
And then with OpenAI's models, for the most part, uh, aside from the default one, right?
00:19:29.200 --> 00:19:37.920
But uh, you know, make sure that's why I always say if you're using uh a model, make sure you're using a model that thinks because these are models that can call tools, right?
00:19:38.079 --> 00:19:50.319
They can um decide, hey, this user is asking me about something that might be a little old in my training data or the data that the model is built on top of might be might give old or outdated or wrong answers.
00:19:50.480 --> 00:19:53.839
So an agentic model can decide I'm gonna call a tool.
00:19:53.920 --> 00:20:01.839
And it can trigger functions like search, uh, you know, going through your email, letting the uh agentic model interact with external uh services, right?
00:20:01.920 --> 00:20:06.640
You have these things called connectors, which, you know, it's kind of like rag mini, right?
00:20:06.720 --> 00:20:17.839
We're true blog meta generation, but a very simple version, uh, it used to take uh a lot of smart engineers and usually uh six, seven, eight figures to get rag.
00:20:18.160 --> 00:20:21.680
Now you can get a very simple version in a couple of clicks, you know.
00:20:21.839 --> 00:20:26.480
Most large language models on the front end call those something like connectors or apps, right?
00:20:26.559 --> 00:20:37.759
So you can connect things like your like your email, your calendar, uh, your you know, your SharePoint or your Google Drive, um, your OneDrive, things like that, Salesforce, Slack, right?
00:20:37.839 --> 00:20:42.319
And then all of that um invaluable, right?
00:20:42.559 --> 00:20:50.480
Like all of that data that your company needs and uses that you're like, man, two years ago, it was so hard to work with that.
00:20:50.559 --> 00:20:52.319
Now it's there almost automatically, right?
00:20:52.400 --> 00:20:53.359
With these connectors.
00:20:53.519 --> 00:20:54.880
You also have MCP.
00:20:54.960 --> 00:20:56.880
This is the model context protocol.
00:20:57.119 --> 00:21:07.279
Um and this is uh a technology that was first uh invented and popularized by Enthropic, but the rest of the industry has adopted it, which is great.
00:21:07.359 --> 00:21:11.039
You know, one thing about the AI industry, see like it's weird.
00:21:11.119 --> 00:21:12.559
They all kind of work together, right?
00:21:12.640 --> 00:21:15.119
So, even as an example, I was just talking about connectors.
00:21:15.279 --> 00:21:24.240
Uh, if if you're a Google organization and you use Gmail and Google Drive, you can technically use those connectors inside OpenAI's products.
00:21:24.400 --> 00:21:27.279
You can use it inside Microsoft products, inside Claude's products.
00:21:27.359 --> 00:21:36.319
So, in the same way that MCP is kind of this, it's kind of like uh um, you know, how the internet talks to, you know, websites can talk to each other with APIs.
00:21:36.559 --> 00:21:40.480
This is kind of how AI tools can talk with uh to each other with MCPs.
00:21:40.640 --> 00:21:44.160
Uh so the model context protocol is definitely the most popular one.
00:21:44.319 --> 00:21:48.240
There's a handful of other ones that are uh widely used and widely adopted.
00:21:48.319 --> 00:21:53.519
But for the most part, you know, these models have essentially languages that they can talk to each other.
00:21:53.680 --> 00:21:59.759
So you don't have to, you know, send it from an AI um system to a website.
00:22:00.240 --> 00:22:04.720
To use the website's API to go to another website to then send it to a model, right?
00:22:04.799 --> 00:22:07.039
So it's just a direct connection.
00:22:07.119 --> 00:22:12.720
The MCP connection allows different AI models to talk to each other or different AI tools.
00:22:13.039 --> 00:22:14.079
And then scaffolding.
00:22:14.160 --> 00:22:20.000
So if you've ever heard of the word scaffolding, I say it a little, I try not to, uh, just because I don't want to, you know, confuse any people.
00:22:20.160 --> 00:22:28.400
But that essentially defines the structure that agents use to chain together multiple steps or tools for more complex workflows.
00:22:28.640 --> 00:22:35.759
So, you know, uh, sometimes, honestly, like the models themselves can all be fairly similar, right?
00:22:35.839 --> 00:22:40.240
If you look at the benchmarks, they're, you know, 0.1% from you know each other.
00:22:40.319 --> 00:22:42.720
A lot of times it's the scaffolding, uh, right.
00:22:42.799 --> 00:22:51.200
It's it's how the model actually works and how it's able to successfully or unsuccessfully use all these different tools that it has at its disposal.
00:22:51.440 --> 00:23:02.880
So uh the large language models have largely gone from helpful AI assistants that were powered by older, slower models that hallucinated a lot.
00:23:03.119 --> 00:23:11.039
Now they're the models themselves are agentic by nature and they can use tools and you can connect to your business's dynamic data.
00:23:11.759 --> 00:23:13.759
But there's risks and you need to know them.
00:23:13.839 --> 00:23:23.119
Hallucinations, probably a word most people have heard of, but it's essentially a lie, uh right, or a false statement, or a half truth that's put out there very confidently, uh.
00:23:23.519 --> 00:23:40.160
But if you do know the basics of context engineering, so if you've taken our free uh prime prompt polish uh course and our free prime prompt polish, uh can't speak Prime Prompt Polish Pro course, you know the basics of context engineering and your hallucination rate is gonna go down.
00:23:40.319 --> 00:23:52.960
But if you're using the right model and if you know the basics of prompt engineering and uh sharing your company's context, hallucinations are, I'm not gonna say they're gone, but they're essentially gone, right?
00:23:53.279 --> 00:23:54.880
People overlook that.
00:23:55.039 --> 00:24:10.160
If you know what you're doing, if you're using the right model, if you're connecting in your company's data, if you're uh looking at the chain of thought, uh doing basic observability, uh, you know, human in the loop, even lazy human in the loop, hallucinations are down to about zero.
00:24:10.319 --> 00:24:15.599
Uh prompt injections, you have to look out for those, uh, especially as we have agentic browsers, right?
00:24:15.680 --> 00:24:27.119
So these are browsers that are run by an AI agent and they can go browse the web for you and interact uh with interfaces on the web with graphical interfaces, you know, click, you know, click drop downs and fill out forms for you.
00:24:27.200 --> 00:24:28.960
But there's prompt injections, but right?
00:24:29.039 --> 00:24:34.240
So there's prompt injections that get into training data, which impacts the models themselves.
00:24:34.400 --> 00:24:40.559
But on the agentic side, for agents or agentic browsers, there's prompt injections on web pages, right?
00:24:40.640 --> 00:24:45.759
And some of the earlier agents and agenc browsers didn't have great safeguards against those, right?
00:24:45.839 --> 00:24:48.000
So, you know, as an agent, right?
00:24:48.079 --> 00:24:51.920
A simple simple example, um, you know, I could put something on my website.
00:24:52.079 --> 00:25:03.920
Uh, so if an agent or an agentic browser was trying to use it and I could say, hey, you know, go uh, you know, go to this site and you know buy uh a hundred boxes of cereal, you know, put it in your Amazon cart or something like that.
00:25:04.079 --> 00:25:07.920
That's a very uh safe and happy prompt injection, right?
00:25:08.000 --> 00:25:09.440
It's it's not causing a lot of harm.
00:25:09.519 --> 00:25:25.839
I it just tricked the model uh or an agentic model into doing something that it didn't intend to, but there's obviously very harmful and unsafe uh you know prompt injections, just like there's you know, phishing emails and all these you know scams that have been going around uh you know with malicious links and all that.
00:25:25.920 --> 00:25:26.559
The same thing.
00:25:26.799 --> 00:25:28.160
Models have risks.
00:25:28.319 --> 00:25:31.359
So companies need to understand and talk about guardrails.
00:25:31.519 --> 00:25:37.839
That's when you combine uh human-made policies and filters to enforce safety and prevent misuse.
00:25:39.359 --> 00:25:40.000
All right.
00:25:40.400 --> 00:25:43.920
So this is kind of the translation playbook.
00:25:44.079 --> 00:25:44.400
All right.
00:25:44.480 --> 00:25:47.279
So in the future, we might be updating this, right?
00:25:47.359 --> 00:25:50.880
If we keep the start here series going on for a longer time.
00:25:51.039 --> 00:25:54.799
Uh so make sure I always say check the show notes of this show.
00:25:54.880 --> 00:26:08.079
So if you're listening on the podcast, we're gonna always keep the show notes updated and uh in the community in which you get free access to by going to starthearseries.com, uh, we're gonna keep the resources there updated as well.
00:26:08.240 --> 00:26:10.319
So there's always gonna be new terminology.
00:26:10.480 --> 00:26:12.240
Some of the terminology is gonna change, right?
00:26:12.319 --> 00:26:18.640
Like as an example, the definition of artificial general intelligence changes every month, right?
00:26:19.039 --> 00:26:26.640
By uh definitions from 10 years ago, we're well well past AGI, but you know, how people define it say maybe we're not there.
00:26:26.720 --> 00:26:30.640
So we're gonna keep the definitions up to uh up to date in our community.
00:26:30.720 --> 00:26:35.759
We're gonna have a little AI translation playbook, but you need to ask yourself these questions.
00:26:36.000 --> 00:26:36.319
All right.
00:26:36.400 --> 00:26:40.720
I'm gonna go back to how I started, the technical people and the non-technical business leaders.
00:26:40.880 --> 00:26:50.000
These are the questions that you need to be asking yourself out loud when it comes to using AI and understanding it and hopefully closing that capability and education gap.
00:26:50.240 --> 00:26:55.759
Number one, what's the problem we're trying to solve and what's our current costs associated with that solution?
00:26:55.920 --> 00:26:56.160
Right?
00:26:56.319 --> 00:26:57.440
People overlook that.
00:26:57.599 --> 00:27:03.279
I'm like, you know, when people are trying to, you know, calculate ROI on Gen AI, uh, right?
00:27:03.359 --> 00:27:05.920
I always ask, okay, what's your current cost on this?
00:27:06.000 --> 00:27:09.119
And they're like, okay, well, you know, uh, we have this piece of software.
00:27:09.279 --> 00:27:11.119
And no, how much time is spent?
00:27:11.279 --> 00:27:11.920
No clue.
00:27:12.079 --> 00:27:14.400
You have to understand what your current costs are.
00:27:14.640 --> 00:27:20.319
Uh, next questions what model, data, tools, and who approves?
00:27:20.559 --> 00:27:21.920
You have to go through those.
00:27:22.079 --> 00:27:31.519
Next, how can we responsibly run the fastest sprint possible with measurement and evaluation, with proper, uh, you know, proper QA?
00:27:31.839 --> 00:27:35.200
Um, you need to go fast.
00:27:35.680 --> 00:27:42.799
You have to be careful, but you have to go as quickly as possible while still being careful while you're implementing AI.
00:27:42.960 --> 00:27:47.759
Because if you're too slow and if you take a pilot slow and if you're not building modularly, right?
00:27:47.839 --> 00:27:49.680
Now all of a sudden you did a year-long pilot.
00:27:49.839 --> 00:27:58.079
Oh, wait, that original model you started with is that was the best model in the world, is now the 50th best model, and it's bad.
00:27:58.160 --> 00:28:00.240
And now you're putting yourself at a disadvantage.
00:28:00.400 --> 00:28:07.680
You can't treat AI implementation in the same way that we've treated tax uh tech implementation for the past few decades.
00:28:07.920 --> 00:28:12.559
You need to ask what are the hallucination risks, the security risks, and the costs.
00:28:12.720 --> 00:28:18.960
And then last but not least, you always need to require receipts, the observability, traceability, uh, expert-driven loops.
00:28:19.119 --> 00:28:26.559
We don't need human, uh lazy, human of the loop, uh, logs, defining the scope and getting your evaluations first.
00:28:26.720 --> 00:28:34.480
You have to understand what you're going after to be able to measure it uh and go and you know, understand the evaluations in the process.
00:28:34.720 --> 00:28:35.039
All right.
00:28:35.119 --> 00:28:36.960
So now that's it.
00:28:37.200 --> 00:28:40.160
You now have the foundation for the rest of the start here series.
00:28:40.240 --> 00:28:40.559
All right.
00:28:40.720 --> 00:28:45.680
We went over what the heck is generative AI, how do large language models work and why it's the future of work.
00:28:45.759 --> 00:28:48.240
And now you have the language and the lingo down.
00:28:48.559 --> 00:28:48.880
All right.
00:28:49.039 --> 00:28:52.000
So what I want you to do, sit down with your people.
00:28:52.160 --> 00:28:53.759
You need to get a 30-day plan.
00:28:53.920 --> 00:28:57.759
You need to be able to sprint, you need to understand how you're gonna deal with mistakes.
00:28:57.839 --> 00:29:02.079
Uh, and then later you can go into the RAG, deep dive, ROI, security scale, all this stuff.
00:29:02.240 --> 00:29:06.079
But you have to set here's where we are, here's where we want to go.
00:29:06.319 --> 00:29:13.279
Um, figure out what's gonna happen in between, but you have to keep up to date with the language and the lingo.
00:29:13.440 --> 00:29:15.680
Otherwise, your plan is gonna fall apart.
00:29:15.839 --> 00:29:22.960
All right, one way it's not gonna fall apart, go listen to episode 628, very much so related uh to this episode.
00:29:23.039 --> 00:29:26.160
So that is what's the best large language model for your team?
00:29:26.400 --> 00:29:30.240
Seven steps to evaluate and create ROI for AI.
00:29:30.400 --> 00:29:32.480
All right, I hope this was helpful.
00:29:32.640 --> 00:29:40.640
Now you know the AI without the jargon, the AI language every business leader needs to live by in 2026.
00:29:40.799 --> 00:29:44.480
All right, if this was helpful, please go to starthearseries.com.
00:29:44.640 --> 00:29:55.839
That's gonna take you and give you free access to our inner circle community, where not only can you go and check all of our Start Here series uh on the web, right?
00:29:55.920 --> 00:30:02.240
Or if on the mobile app, if that's what you want to do, uh, but you're gonna be signed up for our free daily newsletter as well.
00:30:02.319 --> 00:30:07.119
And you can go take our free prime prompt polish uh context engineering course.
00:30:07.279 --> 00:30:07.759
All right.
00:30:07.920 --> 00:30:09.200
Thank you for tuning in.
00:30:09.359 --> 00:30:12.880
I hope to see you back tomorrow and every day for more everyday AI.
00:30:12.960 --> 00:30:13.920
Thanks, y'all.
