Introduction

Today I'm going to try in 15 minutes to share you what we have learned working over 200 people of different SMAs from Spain. What's the best way of implementing AI in their companies?

Understanding AI

1So first of all, we are going to talk from starting what is really AI, because we can see a lot of videos going to the hype. So let's try to see it really, really simple.

What is AI?

What is real AI?

Okay, so AI is not new. AI, we have been living with it years.

And here we can see some examples. Siri is AI. I'll explain it later. Google search, the algorithm for what result is higher or lower is also AI. and Tesla driving is also AI.

Examples of AI in Use

One question, does anyone here know what is AlphaGo? Hands up. Okay. Does someone want to explain it? For example, you.

Okay, so there is a deep mind of Google that is the laboratory that makes it, but AlphaGo is one AI that in 2016, win the best player of Go. That Go is a game really spread in Asia. And it's like a really of reasoning things and try to get to win the opponent. And it was like, okay, this is only for humans. And it was really mind blowing that an AI wins the best player.

But here I would like to do a really quick reflection that this game, the humans are keep playing, even though the AI is better. So that the AI do it better doesn't necessarily mean that humans are going to stop doing it.

And also, I think all of us here know, ChatGPT in 2022, that it has one technology that was GPT-3, but when GPT 3.5 appears, it was the platform that reached the billion of users in less time.

Types of AI

Also, later, two differentials of AI.

We have general AI, Narrow AI and general AI.

Narrow AI is also chagipity. It's also generative AI. It's a weak AI.

But the general AI, all of us have heard about it, but we have never used it because it's a fictitious one that only appears in the movies, and it might not be able to happen ever because of a limiting data.

We can speak about that later.

Machine Learning vs. Deep Learning

And machine learning versus deep learning. The difference really quick is machine learning, you have to be supervising the training really closely. So it limits how much data you can train.

But deep learning, you give it all the information and it doesn't need so much control of the human. 1And generative AI is a part of deep learning.

AI Technologies

Okay, now let's speak about, sorry, I'm a little bit ill. Let's speak about technologies of AI.

Popular AI Tools

Here I like to do a really important announcement. How many of you have you seen LinkedIn posts or Instagram posts or YouTube or whatever saying the 100 best AI tools? I have seen a lot.

You have really listened to it. Okay, these are tools. I can build a tool with my branding and say it's a new AI, but if we go to the models, to the base of AI, there are really, really deep, really few AIs that we can use.

So ChatGPT, we all know it, is the same technology that all the Microsoft and Windows services of AI use. That's important to know.

Later, we have Google Gemini, we are in Google campus, so Google, when ChagiPT appears, releases Bart, that later changes the name to Google Gemini. Cloude, how many of you have listened to Cloude? It's less known? Okay.

It's an AI that behind, with a lot of investment, is Amazon. And it's co-founder by ex-OpenAI employees. Well, I guess I don't have to explain that. OpenAI is the company behind ChatGPT, but just in case, that's the company.

Open Source AI Tools

Later, open source tools. We have two that they are... desire to talk about.

One is Meta, have JAMA 3.1, and right now is the best open source tool. And we also have Mistral, is maybe the less known of this leaderboard, and is an European AI company, at least the only one that speaks internationally.

And just as a fun fact, the first model they released, it was based in a YAMA. Now they do their own models.

Choosing the Right AI Model

This is something that we get asked a lot. Which AI model should I use?

You can see a lot of articles, research about our AI is the best of this research. But they are analysis made by their company, so sometimes it's not the most reliable.

So there is a hiring phase that has built this leaderboard that in a way that is trying to be not influenced by no one, selects which AI is the best. I'm going to do a quick example.

If you scan this QR, you will come here. So this is what I was showing to you and the way of working of this tool is, okay, users all around the world for free because they want to contribute to the community come here and say, hello, good morning.

As if it's ChagiPT, we write here the prompt and it generates two different answers. which we are going to do a test. Hands up, the one that thinks the best answer is the A. The B? A one? Okay, I'll type the B. Let's say that B is better. And now say, okay, A was work, B is Yama.

So they don't release the model till you vote. So it's more reliable. And that's what it's seen here.

Understanding Tokens

You also can, I invite you to enter here for selecting and say, okay, I want the best AI in Spanish in coding or in instruction following. And also there is something really important, that is the tokens.

I'm going to explain this with an example, and it's the long-term memory and short-term memory. Okay, long-term memory, it could be if you ask ChatGPT or Gemini, what's the capital of Spain? They will answer Madrid because they have been trained with a lot of data and they can answer from there. But if you attach a PDF, a file, speaking about your company and you ask them a question about that, that's going to be the certain memory.

And that's the token. That is an equivalent in words. We are going to set another example here.

Because when someone say, I asked to chatGPT and it doesn't remember, it's because of that. Okay, chatGPT doesn't understand about words. It understand about numbers and probability.

So when you tell chatGPT this, he read these numbers. and is the limit it has. You can see that 250 characters is 53 tokens.

So, if you attack a higher PDA for text and it doesn't know how to answer, it might be because of this. So, depending on what you want to do, you can use Gemini that is 2 million tokens of context, Cloudy, GPT-4.0 or YAML, depending on what you want. It's not the only thing to see, but it's something really important.

Creating an AI Assistant

Okay, now let's go to create or see how to create an AI assistant. First of all, when I speak about AI assistant, I speak about if you generate an assistant from API of OpenAI, from Google API, from cloud projects, from GPTs, whatever.

Defining the Purpose

First of all, we have to know for what we want to use. With our clients, most of the times, they don't know for what purpose they want to use AI and if it's going to be relevant or is going to work bad in this use case.

So we also use this matrix for say, what you should use. If we are going to use AI for first time and we want to use something really difficult, now okay, this is impact effort and we try to go always to quit wins.

Okay, so quit wins is something that with more or less low effort, I'm going to have a big impact in my productivity. And if you want a help for finding this use case, there you have a GPT that you can scan and talk with him, tell him, and we are going to see it, tell him a lot about how you do, what you work, and it will generate a specific and personalized use case for your work.

Main Tools for AI Assistants

Okay, this is the main tools that we use, Cloudy or GPT. So, GPT, they have name, description, instructions, all these parts. The most relevant is instructions.

In instructions, we have to write all how we want to behave. AI, we sometimes say to our clients, is like a junior, that it can know a lot, but you have to take it really, really clear. And there you can upload notelets in documents.

The best file, it would be TXT or PDF. Functions, the idea here is, let's activate what you are not going to use, and be more productive. Actions, if you are going to use API for an external tool, and a cloud is much simple.

You can only upload a description, instructions, and knowledge. Sometimes less is more.

Okay, here.

Effective Communication with AI

How to Talk to an LLM

How to talk to an LLM.

It's really important that you know that sometimes we ask something and we think we are giving enough context or enough information, but it's not the case.

If I tell you to imagine a dog playing with a ball in a park, try to imagine it. All of us have a different image in our mind. So we have to be really, really specific.

First, in text LLMs, we have to say, okay, how we want to perform as an expert in medicine, engineering, or maybe a marketing consultant. The context, okay, what's my background, what's my company, what's specifically, if I want to generate, for example, an article for the blog of the web, what exactly I want, I want one each one, two, three, how many loop of storytelling I want to use, whatever.

Sometimes you would be surprised that some client has come and said it doesn't work properly and it's because they have explained a lot of his life and their work but they haven't said do this really specifically. And details.

It's really important to say what tone, what delivery format, and the tone, it can also, just for example, it can also replicate this type of speaking. For example, if you want to make a tweet assistant, you upload a lot of tweets and it will create this style of tweet.

I'm being forced to go ahead, so I really want to make you an example really quick of all of GPT's I might have. This is the prompt that it used for generating articles of ECO. I have difference, this I don't know if it's the more update, but that's the idea.

You have to be really specific on what you want.

Conclusion and Q&A

And now I think it's going to be better to go to Q&A. And if you have any questions, we can do it.

Okay, the interface is have this option that I don't really recommend it. Maybe it's if you're first time and you know where to begin, you can come here, but all this part of configure is enough here.

So this is built specifically for one type of company with all the data of that company and also internet access to them. So it generates a list of ideas and we can tell Okay, we want this blog post.

So also sometimes we try to generate AI or generate assistance of AI of themes or topics that we don't know. So the best thing is know what you are talking about so you can be really specific or contact maybe in this case marketing specialist so it can help you in exactly, okay, I asked to a marketing specialist, what should I use for Instagram post and planning of them?

If he gives the keys, you can give these keys to the AI, so it's gonna be more specific. And here we can be a good time speaking with the AI, but that's the point.

And also we are generating in deep a guide of prompting. If you scan this QR in the next weeks, a mate of mine would send you a PDF with more in deep that we don't have today.

Thank you.