Skip to content

AI Yesterday, Today, and Tomorrow: Extending AI Systems with Model Context Protocol

2026 / How to Web 2025 / 20:28

Watch this talk on YouTube

Transcript

20 paragraphs

This is an automatic transcript of the recording above. It is published in full and unedited, apart from correcting names the recogniser reliably mishears. It will contain mistakes.

00:15Good afternoon, Bucharest. It's so good to see you today. How are you asleep? Great. Um, my name is Tis. That's pronounced like contagious. Don't worry, I'm not contagious. It's just some people have trouble with this name. Even the moderator asked had to ask me and she thought it was teas and so this should just help you uh pronounce this. I'm an AI developer relations engineer at IBM. Uh so two IBM talks in a row. I I need to say I'm not here speaking on behalf of IBM so to say. I just work there um on AI and research and developer communities. Okay. Today unfortunately or fortunately our talk is going to be about AI. And I say unfortunately because I think a lot of us are just tired of AI, right? There there's some AI fatigue. She's nodding.

01:00She's like, "Yeah, I hate that shit." And I I I I I'm with you. And so what I'm trying to do today is give you a talk that is not um he's leaving. That's how like tired of AI they are. I'm trying to give you a talk today that is not um based in hype or or based in um high valuations for the sake of like you're going to miss out, but to give you a view on AI sort of through time yesterday, today and tomorrow uh with nuance based in science and engineering and and and in a way that will make it actionable for you. You'll be able to actually go to work or go home and use this in your life to add value because I I I believe it is useful if we remove all the marketing. Okay. And so that's our plan for today. To get started, we don't have a lot of time. So I want to start by talking about AI yesterday. And I don't mean like actually yesterday where um OpenAI announced this like AI slop video app called Sora 2 that happened yesterday. I'm not talking about yesterday. Um Claude 4.5 sonnet new model dropped yesterday. I'm not talking about yesterday. We're going to go back a few years uh to where AI began, specifically generative AI and um think about go from there to today to the future. Okay. When generative AI was new, nobody cared. Absolutely nobody cared. GPD2, have you used it ever? No.

02:18Oh, okay. Like two of you. GPD2 really? Um she's now she's like no. So GPD2 um nobody used it. GPD3 came out and still nobody cared until OpenAI put a beautiful user interface on it, a chat interface called a chat GPT and then everybody was like oh my god and and the the main thesis I want to share with you today is you can build the greatest things but if the user interface or the human interface is not there nobody cares. Okay. Um so we'll start with that. Chat GPT was new in 2022 November and we started using it but when we used it some limitations became very clear. Uh let's travel through time and use sort of early chat GPT um to understand and and and and understand the limitations. Now when we do this we're going to use the model GPT 3.5 um and I'm going to use a tool that I work on called Langflow. It's open source. It's free. There's no commerce to be had here. Um, but it's it's a great way to visualize AI workflows and then execute them and then host them on the internet and that's why I'm using it. It's this beautiful like one sizefits-all thing.

03:20Okay, so this is Langflow and we're going to build like a version of chat GPT here. So chat input, chat output, kind of the essentials and we have a language model in between in between the chat input and the chat output something like this, right? And so this is kind of what chat GPT was except the model was GPT3 3 actually. So this is like a very early version of Chad GPT. And if we use it um we can test it. We'd say hi and if the internet works it says cool hi how can I assist you today? Fantastic. The problem with early like chat GPT of yesterday was it was it what underlied it was a machine learning model whose weights were frozen. It had no access to like real-time information. For example, we can do something like um what movies are playing in the theater today, right?

04:10Uh and it will just not know. I'm sorry, but I'm unable to get rid. This was a big limitation with G Chad GPD yesterday. How does Chad GPD solve it today? Let's go use chat GPD right now. And I'll paste the same prompt. And what you will see is some interesting developments to today. It searches the web. You see that it searched the web and it got answers. that searching the web is a technique called rag. Who here has heard of rag? Who you? Yeah, that's yesterday. You're outdated. Rag. Rag stands for retrieval augmented generation. You retrieve some authoritative data and you use it to augment the generated output. So, okay, let's travel through time and let's add rag here. So, I'm going to add a URL just here. And this URL will go to the internet, will go to this website that has the latest movies and theaters and it will give it to the model. Okay. How does rag work though? Well, we add an intermediary layer between the chat input and the model. And this layer is a prompt composition. This is the basis of all rags. So how rag works is I take I write a prompt that is a superset of the user's prompt. So I'll write this prompt. The user asked question which is a variable as you can see here. And the data is and I'll put another variable data. Now answer the user. So this is an actual rag prompt. Okay. And so what we'll do is we'll take the question from the user and plug it in here. We'll take

05:35the data from the authoritative data source, the website, plug it in here. Right? And now we have rag. We have we're retrieving movies from the internet. We're getting the user's query. We're giving it to the so in the prompt we include the answer to the question. Does that make sense? This is rag. Okay. So how does this work? Let's try. So I'm going to open the playground. And notice it's the same prompt. I'll paste it and I'll send um and now like suddenly it has all this big list of movies right this is how we got around that limitation yesterday last year this is most of 2024 was doing rag and rag is still a technique today but it has changed it has changed if we fast forward to today to be agentic 2025 I I hate myself for saying this because it's one of those cliche sentences 2025 is the year of AI AI agents. Yuck. Okay.

06:27Anyway, um, what does that mean? Rag has become a little bit more agentic. What are agents? Well, I'm an agent. You're an agent. We're all agents. Agents are entities, human or machine, that have the ability to make choices. I chose to come be here today. I chose to drink water. Right? If you ask me to do a task that I don't know how to do, I my language model in my brain will remind me of a task by description. So, hey Tis, multiply the first 10 prime numbers. I I'm like, I don't even know what a prime number is. But if I did, um, I would use a tool described to multiply prime numbers. You know what it is? It's calculator. So, I would like naturally go reach for the tool, generate inputs to the tool from my language model in my head and then, you know, two one times three. I don't know the numbers. I'm bad at math. Anyway, um, and then I would hit equals, and the tool would give me an output to my lang.

07:20Does this make sense? So agency allows us to tool to choose tools based on their description. Rag is now agentic and it looks like this. So instead of this like prompt, template, business, whatever, we take the language model and swap it with an agent, an actual agent. And on the surface, an agent really doesn't do anything. An agent is kind of just like a language model except an agent can use tools. There's this little input here. So this URL tool I'm going to actually turn into a tool. You see tool mode. Tool mode on. And what you'll notice is the actual URL has disappeared. I don't paste a URL anymore. The agent uses its agency to know which website to go to it when it gives the does it make sense? It gives an input to the URL tool. Okay. So we'll pass the chat input raw from the user and we'll pass the tool just like this.

08:11This is how rag has changed. And now using agency, the agent can just go browse the internet using this tool. So let's take a look. So I'll say go to this website. And now it just knows how to go to the website, you know, because it has the tool and find movies playing in the theater. Okay. And we and now you'll actually watch it. We'll actually use the use the tool uh if the internet allows. We did talk about internet yesterday uh during the rehearsal and I was told it would be better but here we are. Let me just quickly check this. And also just for posterity let's use a better model. Okay, there we go. So it's doing something and we should have a I can actually do this better by just doing this whole thing offline with local models. Um, what network am I on?

09:15Thank you. Yeah, that's actually Wow, this guy knows Langlow better than me. It give him a round of applause. That's fantastic. Thank you. Uh, let's try that. And now we plugged in the output. Perfect. And so we have fetch content. It used the tool. It got the answer. Fantastic. You see, and you can actually see the inputs it gave the tool. Okay, let's do better. I used a calculator example. So we'll take a calculator and we'll do tool mode. And now I have two tools. Look at that. And so playground, right? Let's do multiply the first 10 prime 10 10 prime numbers, right? And with agency, my agent evaluates expression. It gives these are prime numbers. Cool. I didn't even know that. And and it gets a result. So it used a tool. Agents are cool because they can use tools. That rhymes. You can get a tattoo of it if you want. Um, but also tools make our lives work and tools and code are just functions. Functions that receive an input, return an output. The LLM generates the input and reads the outputs. Does this make sense? Now that tools are functions, they can connect with APIs. And this is where we get to sort of tomorrow territory. Let I really suck in general, but I also really suck at at managing my calendar. I'm not good at this. Ask any organizer of the conference, they will tell you. And so what I built was an agent to manage my calendar. And it just works. And you can do this yourself using tools. So tools,

10:39as I mentioned, they're just code. They're just functions. I have this Apple calendar tool that I made. By the way, everything here is code. All these if I click the code button, it's all just code. It's and and you there's documentation. You can everything is code including tools. So this tool Apple calendar, I will turn it into tool mode. And I will pass this as a tool, right? And that's it. I have this agent. This is actually crazy because the Apple calendar lives on my computer. This is the Apple calendar app right here. Um, and as you can see, my calendar is kind of empty, but today on October 1st, I have to give a talk right now. So, if I go to the playground and I can I can literally just here, let's do this side by side, right? I can say um, you know, make a calendar event for 1:30 p.m. Bucharest time to give a talk, right? Um and the agent will use the so it gets the current date there that that exists now it it just did the job right um how how fast was that you know what I mean and so a this is the age of agents that's today this is not tomorrow that's today things like this exist you can build them with langlow or others okay but where are we going tomorrow now that things like this exist what if I don't want to use the langflow interface to build tools like this but I want to use the tools from anywhere, right? I want to pick up my phone and use a voice assistant. Hey,

12:05book me a flight or I want to use my laptop, but I don't want to use Langflow. Yuck. I want to use like chat GPT, which I already use for everything else. I want to use Claude. I want to use something else. This is where MCP or model context protocol comes into effect. Okay, what is MCP? Similar to HTTP, hypertext transfer protocol. Model context protocol is a protocol. It's a it's a client server protocol. You have a server somewhere. You have a client somewhere. The client when it opens talks to the server. It says, "Hey, what context do you have for me?" And then the server says, "Well, I have these tools. A tool for the calendar, a tool for this, a tool for that." And then the client can call those tools remotely.

12:47Does that make sense? That's so you're just passing functions over the network. It's There's a lot of people who will use big words to confuse and gatekeep you. Don't listen to them. It's all functions. So we can we can do this. Okay. And so with MCP any client can talk to any server. So Langflow is an MCP server. It gives context or tools to MCP clients like chat GPT like claude which is AI app and like cursor. Anyone using cursor? Cursor is a way. Yeah. So cursor is an MCP client. Um Langflow can work as an MCP server. So what I the point I'm trying to make is I can now manage my calendar from anywhere. Let's look at that. Um so I will open actually I need to expose this MCP server first. So what I'm going to do is share and I can share this as an MCP server. And so now I've got to edit my tools. So this thing is called new flow. I need to give a description. This is how the agent knows when to call the tool. So I'll say use this tool to manage my calendar. I'll call it calendar tool. I don't know. And then we'll close it. And now what I can do is I can just auto install this to any MCP client like cursor or claude. So I'll install it to both claude and cursor.

14:00And that's it. It's installed. So now if I open claude um I can just like manage my calendar from this. I don't have to use langlow. So this is claude. This is my calendar, right? Um why does it okay today? I want today. There we go. make uh 2 3 p.m. Bucharest time appointment in my calendar for lunch. Right? I'm just going to check that it's actually configured here. Perfect. And now what's going to happen? This is not Langflow. This is just an app like chat GPT that happens to be an MCP client. Right? And it will just use my tool. Look at that. It's it's tremendous. And so now what it'll do is it will ask for permission because it it it won't just indiscriminately use. Okay. Well, I I guess it didn't ask for permission because I already gave it permission.

14:50Okay, the machines are not evil and alive. Um, and so it totally added this appointment right here. You see that? And so I can do this from I can do it from cursor. So as I'm on, let's open cursor. This is cursor. I was actually coding an SVG. This is actually true. Um, and now that I have it in cursor, let's go confirm that I have it in cursor in under tools. Perfect. I have it. I can do something like um find out when Tis's talk is at how to web 2025 and add it to my calendar and it will use the combination of all the tools right like it will use the web search the URL it composes tools together while I'm coding I'm not even leaving my context and it just does a job um so it's searching for this stuff um the web search didn't give information because the website doesn't render without JavaScript. So I can interrupt it if I need to and say here. So now it's using a browser. It's like actually and and I can actually watch this brow. Let's go let's go watch this browser. Look at this. It's like clicking here on this website going to the agenda to try and find like where my talk is, right? This kind of stuff is wild. So anyway, I found one page and I can see TIS is a speaker, but now I need to find where the and it's just going and it's composing my tools. Um, and you can actually look it it it brought up that that program dropdown. Are you seeing this? That's today. And so I I have like

16:244 minutes left, man. But I hope it can add it to my calendar because that'd be crazy. There you go. So like this whole thing is being navigated by an agent, right? I'm not doing anything. Um, and at some point it will add my talk. I need to really manage. Perfect. Okay, it found it. I'm on the agenda page. Let me check day two. The problem with the agenda is my name is covered by another element on the page. Okay. And so it's going to struggle a little bit, but I'm making a very strong point here for my conclusion. So stay with me. Um, I'm on day two, but I still don't see TIS. um let me see if I can find it. And so now it's writing some JavaScript to run in the browser and my name is not and so it's just really going and going. Um and so it will continue this but ultimately it will succeed. I need to manage time.

17:15We can look at it at the end. But um the point I'm trying to make, let's skip all this stuff and just talk about tomorrow. Um that was the demo for tomorrow by the way. Agents controlling browsers on our behalf so that we don't have to. I'll wrap up with this. It's all about UX. That's what I'm going to say. Um, I used to be a front-end developer. I actually wrote the book on React. Uh, and then I moved from front end to AI because AI is just a different vector for UX. What is UX? It's the experience we have as users on the web. And I tell you what, the web sucks. I hate the web. Um, sometimes because I tried to book a flight today and the my browser was had trouble loading JavaScript and I couldn't do anything.

17:56And sometimes there's like accordians, you know, or carousels and you can't navigate that And then there's people who are blind and who are and it's it's it's the web is like a composition of bad decisions made by people sometimes, okay? It's just hard to navigate. And so what if I didn't have to navigate it, you know? What if one of my agents could do that for me and get a job done, right? What if I could just type a prompt into like chat GPT and be like, "Hey, please book me a flight to Taiwan. This is my budget. I want to arrive at night. I want to leave in the morning. Here's my frequent flyer numbers. Have fun." Right? That's it. And then you know what I do? I I hit send. It does whatever it needs to do. I go spend time with my family and friends. And then I get a confirmation email. That is where AI is going. It's all about the UX. Okay. Um, I I genuinely believe that. In fact, try it.

18:56Like I actually have some some applications that I'm using. My social media is largely just run by a team of agents. I host a podcast. It's run by a team of agents. You can tell because m dashes. Anyway, um, what does this mean? Does this mean our jobs are going away? Like where is this leading? you know, um, and I genuinely believe that if we get AI to do the busy work like booking the flights and filling the forms, then we get a lot of time back to do what really matters, spending time with our loved ones, touching grass, and enjoying life. Uh, with that, let me give you some next steps as I wrap up. As you can tell, I like talking. I do this a lot. Uh, 20 minutes is challenging, man, for real. And so, I have a podcast where I talk way more, okay? like like there are episodes that are very long, some of them two hours plus about AI and where we're going and how to build with AI also about fitness and finance and it's just uh I would invite you to listen to that if if you think there's value there. Number two, the tool that I used was called Langflow. It's free and open source.

20:00It's a desktop application you can download to build agents yourself. I work on it. It's open source and it's honestly run by a small team. So if you have if something breaks, open an issue on GitHub. We want to fix it. And then finally, here's me on socials. If you have questions or things I can clarify, I would love to connect with you. Uh, thank you so much for your time and attention.

More talks

All 19 transcripts

Elsewhere

There is every talk I have given, all 69 of them, ConTejas Code, the podcast, and Fluent React, the O'Reilly book on how React works inside.