Tejas Kumar

The Death of the Browser

2026 / InfoQ / 47:54

Watch this talk on YouTube

Transcript

53 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.

00:04Good morning InfoQ Dev Summit unique. I'm very happy to be here and this is so exciting and we have a lot of time together. I have like almost an hour so you have to sit here and listen to me yap for an hour. I'll make it worth your while, I promise. Hopefully and if it's not leave negative feedback in the app so they know, okay. I'm my name is Cage as he mentioned. That's pronounced like contagious to help you. Don't worry, I'm not contagious especially front row. If I spit on you, it's too close and I'm developer relations engineer at IBM and I work on AI. I work on a lot of AI solutions mostly AI research and development and how we can make AI more accessible to developers. As you know, IBM has a ton of AI offerings like Watson X. Like we actually have our own models like GPT-5 but like the ones that we make and and so how do we make this accessible to you to build things and and build revolutionary technology. Today we're not going to talk we're not going to talk about IBM at all actually because this talk is not like on behalf of IBM.

01:04It's not IBM speaking, it's me speaking and we're going to talk about AI innovation in 2025. To understand AI innovation in 2025, we need to sort of look at the the broad spectrum of time. Like how did we get to be where we are today and where are we going tomorrow? It's almost 2026, right? It's it's almost November and so before we do that, I need to give you a little bit of a disclaimer and tell you that there's there's just a lot of like hype around AI today, right? Like if you want to increase the valuation of your company, just put some AI feature on it and and watch the the sharks come to you. There's a lot of influencers, you know, saying things like hey, you got it you're going to miss out if you don't do this thing and there's a lot of stuff on LinkedIn and X that that can be kind of annoying. Anyone annoyed with AI here today? Yeah, almost everybody. See?

01:54And so this talk is not that. I just want to be very clear. We're going to look at um through the lens of like science and reason. Look at actually what AI is and how we can apply it. Um and I will also get into things like model context protocol MCP all of that. But but there is no hype here. There is no this is not like marketing. This is definitely not fear mongering. This is just logic, okay? And I hope we can um we came prepared for that. So, with that said, we're going to talk about AI. We're going to look at AI yesterday. We're going to look at AI today and we're we're going to look at AI tomorrow. Now, let's start right at the beginning uh and talk about AI from yes and I don't mean like yesterday, you know, like October 15th. I I mean like where did it start? And I think this is important because there's a lot of companies making a lot of money off the term AI and there's a lot of people saying a lot of things about AI as if it's new. And I'm here to tell you this morning AI is not new, right? If we think from first principles, artificial intelligence uh is is nothing new. Uh we've we've simulated intelligence from a long time ago. How how early did it start? It really started even with statistics uh in in like 1906.

03:03This guy named Andrey Markov um was a mathematician statistician um in university and was really interested in poetry, Russian poetry. Um and so he was like, "Hmm, interesting. There's some type of rhythm to the words in these poems." And through statistical modeling um came up with a way to predict with reasonable accuracy the next state of a system based on the current state of a system. What does that mean? Well, in the case of poems, the system is based on words. You have, you know, a word and then you can predict the next word based on the current word and he found success with this. Um this would eventually become what we call Markov chains and eventually grow into statistical Markov models named after him in 1906. Um the weather is actually predicted often times with a giant Markov chain. How does that work? Well, you know what the weather is today. I mean, look outside, right? And you know all the possible states of the system.

03:58It can be sunny, it can be cloudy, it can be rainy. If you're in Berlin, it can If you're in Germany, it can only be rainy. Um but it can be many states. And so, through Markov modeling, what you can do is say, "Okay, today is this. Therefore, [snorts] tomorrow the chance of that, aka the next state, is is this problem." Um we use Markov chains every day, uh especially if you have an iPhone. Uh who here has an iPhone? Do you have a Yeah, okay, cool. This The rest are not my friends. I'm joking. Uh If you have an iPhone and you pull up the keyboard, right? Above the keyboard, you will see these like word suggestions, you know, like you three of them, uh left, a middle, and a right. And if you I don't know if you've played this game, but I encourage you to try it. Like, pull up the keyboard, type one word, um just like I, the letter I, me, and then you get these three suggestions. If you tap on the middle suggestion repeatedly, uh over and over again, you get a sentence, maybe a paragraph, that doesn't make any sense.

04:55Okay? And that's because that's a giant Markov chain. They predict the next word based on the current one. Um but language doesn't work this way. Right? But this is a sort of the beginnings of AI. We predict the next state based on the current one. Fast forward a long time, and AI is still is is pretty old. Like, before I was born, we saw a lot of AI in the world. Uh for example, this Like, look at this. This is an intelligent system. Like, he eats the big ball, and then the ghosts are like, "Whoa!" and they start running away cuz then he can eat the ghosts, right? Um I would wager today that's like artificial intelligent It's literally like, "If something happens, then do that other thing." The difference between this and machine learning and AI we'll talk about, but this Do not be mistaken. This isn't a system that simulates intelligence, artificially intelligent.

05:44The The problem, or rather the limiting factor here is that the intelligence is just a bunch of rules that we know ahead of time, right? Like people wrote this code and they're like, "Okay, so if this then that." And that's how they simulated that's how they created like artificial intelligence. Rule-based the technical term for this as many of you know is is rule-based AI. It's it's rule-based where the programmers of in this case Pac-Man knew the rules ahead of time. I grew up playing Mortal Kombat and Prince of Persia and a bunch of games like Super Mario, right? Where do you have like enemies coming at you and you do stuff and stuff. That's all the examples of rule-based AI. It's it's old.

06:24That was around 1980 that Pac-Man came out, but a few years before this we also started to see what would become generative AI. So so the way that we reason about AI today um there was a paper published that highlighted something called backpropagation which is which is an interesting algorithmic construct which set the stage for deep learning and the next sort of evolution in AI. Um for those of you who don't know very much about machine learning, I'm here to tell you that you don't actually have to know machine learning to be an AI engineer. We'll talk about that uh later, but how machine learning models work is you have an input and they predict an output with some level of you know certainty [snorts] or confidence, right? Um between the inputs and the outputs is a bunch of layers they're called, but I don't like this term.

07:12If you've ever seen a soundboard, you know like a there there's one in the back there like a big mixing board an audio console that that audio engineers use, you'll usually see a lot of strips with a lot of little knobs, you know, like many many knobs. Um that's kind of what a machine learning model looks like. You you've got an input layer many many hidden hidden layers of of knobs that you can turn. And you turn these knobs and you tweak them so that the output is what you want. Backpropagation algorithmically says you do one pass through all your layers of probabilities and then you arrive at an outcome. And then the algorithm goes sort of literally back propagates, looks at the weights or the knobs, and says, "Okay, I need to adjust these to get a higher confidence outcome." And it's just a big loop where you go and turn knobs a bunch of times until at the end you get what is an actually a machine learning model. Um each knob you may call a parameter. So, model like GPT-4 today has 600 billion parameters. Think of it exactly that way. 600 billion little knobs you can turn to increase the probability of something useful, okay? That's a lot of knobs, but that's in essence a machine learning model. Started in 1974 with this paper on backpropagation and deep learning. That was the way until for the next few decades, that's kind of how it worked. And there were many new architectures

08:32that used different configurations of these layers. Okay? There were things like recur- recurrent, excuse me, neural networks and GANs gen- generative adversarial networks, etc. Fast forward to 2017 um and the the team of Google Brain scientists mostly published a paper. I'm sure many of you are familiar with this paper. It's called Attention Is All You Need um that highlighted a new architecture for um models, specifically language models. An architecture with machine learning models is nothing more than just a configuration of layers. You have this layer that does this, that layer, and you just arrange them in an order to get the output you want, okay? Um this paper, I mean, was foundational.

09:12This is the basis, the beginnings of ChatGPT. Um it was published in 2017. Look at those names. They're all Google except like one dude from the University of Toronto and Ilya who was no attribution. Um but this paper outlined a mechanism through code where we can model attention very similar to human attention. Where if you think about how we reason about language, we don't pay attention to single words. Usually, if you're reading a book, if you're reading a blog post, if you're reading this, um you don't focus on one word and then think of the next. You focus on many words in parallel. And that's exactly what they model here. It's called multi-headed attention, where in code, you model attention on one word and its surrounding words. They call them tokens, which is roughly one English word. And you they pay attention to these in code in parallel. Now, this paper is 11 pages long. It's not very uh it's not very challenging to read, I should say. I'd encourage you to read it yourself. It's public. It's free online.

10:09This would set the basis >> [snorts] >> for what would become ChatGPT. Okay? Um but the point I'm about to make here is very, very, very important, uh which is from this paper, the implementation of it was a model called GPT-1, then GPT-2, and then GPT-3. And all these models had great results, because this is a very good architecture. Um but they released GPT-3, and nobody cared. Nobody cared. Like, ChatGPT blew up not because of the model, but in 2022, um OpenAI, they had this great model. It was doing good work, but nobody knew or cared. Until what? Until they put a chat user interface on top of it. Until they literally did some UI work. Okay, a little text box, a little UI. Okay, and now we stream in some words. Oh, wow, what's this? And then people started to care. Right? This is a very important point that I'm making here, which is you can innovate something great and and and wonderful, but if the user experience is just not there, nobody's going to care.

11:17Uh that that rhymes unintentionally. If the user experience is not there, nobody's going to care. Okay? And And that's exactly the state of I'm not a rapper or poet, uh I try to be. Um but that's that's kind of what made ChatGPT successful. It's it's Sure, the model is great, but the the product UI and UX is ultimately what got them to 100 million plus weekly users, okay? It's a very important point for us. So, ChatGPT came out 2022. Who used ChatGPT in 2022 for the first time? Okay, all of you. Yeah, it blew up. It blew up. Um, I'm privileged and thankful to say I've been working in AI since before this. Uh, but ChatGPT came out 2022, revolutionary, undoubtedly reached uh many many millions of users very rapidly. But, there were problems with it very early on. Um, specifically, there were three problems with it. What were those problems? Problem number one, hallucination, right? Like it would just make up nonsense. Uh, and for us, for some reason, we we give it like authority. And so, when it makes up nonsense or hallucinates, we think, "Oh, that's that's uh that must be right." I don't I don't know what it is with us, but we we see something and we think, "Oh, that must be true." Um, unless it's very clearly false. So, Google tried rolling out AI overviews, right? Some of you may have seen this, the memes of um somebody typed into Google, "How many uh cigarettes should a pregnant woman smoke

12:40every day?" Right? Um, and just so you know, the answer is zero. I just want to be clear about that. Um, but the the Google AI overview said, you know, "Between three and five every day." Right? That That's It's uh hallucinations are a real problem. So, that's probably the biggest. Problem number two was knowledge cutoff. So, the model The first of all, GPT-3 finished its training. What does "finished its training" even mean? It just means the bunch of knobs and weights were like sealed. The model was ready in October of 2021 ish, okay? Um, from there, the model is not going to be like retrained. It's it it takes a lot of compute and a lot of resources and a lot of GPUs to do this. So, they they have a trained model up until a certain date, and then its knowledge is kind of just frozen there.

13:25So, if you ask a question like, "Hey, what's What movie should I go see today? It will say I can't help you. Right? That's that's kind of a problem. There is no access to real-time information. So, if you want to use this in your company and you have new users who signed up and you're maybe doing a prompt like hey, what ads should I show my new users based on their preferences, the model will say oh no, I was trained in 2021. I don't even know who you are. Right? I'm kind of useless. And then the third problem was the problem of limited context. So, these models have you can say kind of memory. Um that's an oversimplification, no doubt. But they can hold in context like a finite number of tokens. And a token is roughly one English word. So, today most context windows are like 100 to 200,000 rough English words. Any more than that and you run into problems. The the model will just start over as if it's a new conversation. Now, there are some models with very large context windows. Gemini, Google's models have like you could actually put the entire Harry Potter book series into Gemini.

14:28Uh and it will that's actually cool, but it's rare. Finite context is a problem. Okay, so we have three big problems. Problem number one is hallucination. Number two is knowledge cut off. Number three is finite context. I'd like to actually show you these problems so we can explore together how you solve them. And and you can solve them and you the solution from 2022 is still the solution today, okay? And so I I think this is very important to kind of spend some time demoing. So, someone's notebook is here. Uh I don't know who's this is, but thanks. Let's read it. It's uh it's a list of upcoming talks anyway. Um Whose is yours? >> [laughter] >> Okay, so uh this is the problem. So, I'm going I was watching a cool video on YouTube.

15:08So, this is I'm going to use a tool called LangFlow. LangFlow is it's it's free, it's open source. I love this conference that they say, you know, no hidden marketing. Um fantastic. This is open source tool. There's no marketing or money to be made. It's great for diagramming though. So, I'm going to show you this problem. I'm going to make a flow. It's visually a great way to understand how these things work. So, I have a chat input and a chat output. You can think of this as like a chat user interface. And I'm going to get a language model from OpenAI and I'm going to choose like an old model GPT-3.5, okay? Um now, I'm going to connect the chat input to the chat output right here. Um and we should that's basically it. We have a chatbot now that uses OpenAI's older model. So, I'm going to ask something real time, right? Like um what movies are playing in the theater today? Right? Um and as you can expect, uh I'm sorry, but I'm unable to provide real-time information. This is the classic experience of early ChatGPT.

16:06This doesn't know. Um there's also knowledge cut off and so on. Um how do we solve this? We solve this with a very popular technique that's kind of not cool anymore cuz we're still in yesterday. Uh it's called rag. Anyone heard of rag? Uh it's not Yeah, it's not what I'm wearing. Um it's it's a it's it stands for retrieval augmented generation. Retrieval meaning you retrieve accurate up-to-date information. You get it. You do like fetch. You do a network request. You get the information. And you give it to the language model. So, you get the answer, put it in the prompt. Let's call it rag, okay? Retrieve Let's look at how rag actually works uh here. And so, I'm just going to zoom out. We'll implement rag here. So, what we need is we need to retrieve we need to get movies somehow from the internet. Let's assume we have this in this information available cuz we do. So, I'm going to leave the chat input right there, but I need to get movies somehow. So, if I go to um Rotten Tomatoes, right? These are a list of movies in the theater.

17:04Fantastic. So, what I'm going to do is this is my big database. I'm going to get this data. So, I'm going to get a URL component here. And I'm going to paste this. And what this will do is it will get me the text content of this website. This is equivalent to getting some stuff out of your database. Okay, I hope that's clear. So So now we have the answer. We have the question from the user. I'm going to do some prompt engineering. This is literally what that term means. I have a prompt template. And here I'm going to say, this is the question from the user. And I'll do question. Um here are some data from my database, right? Data. And so you just have these two things, the question and data.

17:47So the question goes here. The data goes here. And and we give this prompt to the language model. So notice the language model itself has not changed. What has changed is the way I build my app. Okay, I just get data from my database and so on. So let's now go try that same prompt here in the playground. Um what movies are playing in the theater today? And so the language model is still GP3, but now it has it just just like that knows stuff. This is rag, and this is how we got around rag a lot of times. This is how companies still do it today. It's a really good thing. There's a movie titled If I Had Legs, I'd Kick You. I don't even know what that's about, but um Yeah, who names a movie? Is that true?

18:33Like is that a Oh, it's this one. 95 Anyway, okay, so um Maybe the movie's AI generated. Okay, so this is This is working. Uh but but it's If you want to see rag in the wild today, where can you go look? Well, I have news for you. I don't know if you've noticed this, ChatGPT itself performs rag. It's it's it's such a it's a first-class technique when working with AI. In fact, if you don't believe me, uh let's go to chat.openai.com and I ask the same prompt, right? Um and check it out. What you will see is as soon as they're they're thinking. I don't want you to think skip. And so it's going to it's going to search the web. It's literally doing that Rotten Tomatoes thing. And there you go. It depends on what you mean, but in Munich, um these are the movies, right? There's no if I had legs I'd kick you. But, um that's rag. And it literally it shows you the sources from where it retrieved, okay? So rag is the de facto standard technique. It works great. Um and that's how you use it if you live in 2024.

19:41Um this was like the year 2024, so last year was the year of rag. Everyone was performing rag. Um especially they were performing rag this way to get real-time information. But this solves two of three problems, right? It solves problem number one, which is hallucinations. There's As you can see, there's no hallucinations here. Another way to remove hallucinations is to lower the temperature of your model. So if you go to the the controls of your language model, you see temperature right here. Something like 0.1 is a very good value. The This rule applies to humans and LLMs. Um the higher the temperature, the more you hallucinate. It's true. You ever had a fever and you start seeing some stuff? Like it's it's the same. And so, um you want to keep the temperature low and this is how you do rag. So that solves two of three problems. It solves hallucinations, no doubt. The The movies are not hallucinated. It solves um the the knowledge cut off. It got movies from 2025. This model was fitted in 2021. Uh 2020, yeah, something like that. And so, okay, but what about context windows?

20:46How do we solve that? Another solution for this is you would use something called vector search. Anyone familiar with vector search? Yeah, vector search. It's like It's like a search engine, but it searches not on keyword density, but instead on semantic meaning. So it associates words like animal and pet with dog, even though the word dog may not match. does this make sense? And so, similarity search allows you to store bodies, like little pockets of text in your vector database, and based on a user's incoming query, you find the closest linguistic result to what the user wants, and you put that in your prompt. Okay, and that really helps long context because you only get relevant context, and you're selective with what context you put in your context window.

21:30Does that make sense? And so, I'm I'm using the word context a lot, but that's kind of how it works. RAG was a great solution, and it worked well almost every single time. Um and this was 2024. So, we started in 1906, then 1974, 1980, we're now uh 2017 with the transformers, we're now in 2024. But, let's talk about today. Uh what is 2025 characterized by in the world of AI? Um Anyone been to San Francisco? If you Okay, yeah, good, don't. Um if you >> [laughter] >> If you go to San Francisco, uh I'm not even making this up, people start conversations by saying 2025 is the year of agents. Like, this actually is a thing. Like, you go sit down to have coffee with someone, and before they say hi, they're like, "Hey, 2025 is the year of agents." Like, you know, it's it's a different culture, man. Um and so, uh so, that's characteristic of 2025.

22:27That's a cool picture. Um that He's like taking a photo in the mirror. Um 2025 is the year of AI agents. What are agents? Well, now we opened the we opened a can of worms, we have to talk about agents, and we have to talk about agents in detail. That's the only way we talk about things. Um Let's uh Let's go back to our slides here, and I'm going to skip forward cuz I talked about RAG and stuff. Let's Yeah, let's today, 2025. Um agents. What are What are agents? Uh What are agents? Specifically, what what is an agent? Um I think it's it's it's a very important to reason from first principles. That's how we all go there together. And I'm not some guy saying big words to you, okay? An agent is one, a person, an entity that has agency. That's all it is, okay?

23:18But then what is agency? Well, agency, classically defined, is a person or thing, meaning a computer, so organic or artificial, through which power is exerted or an end is achieved. Meaning, if you if you're a human being here and you can like make a choice, I'm choosing to come on stage and talk to you, that's me using my agency. I'm doing effort to do a thing for you today, okay? That's So, human agency is the ability to make decisions to do tasks. Artificial intelligence agency or AI agents is the same thing, it's just in code. Okay, it's the ability to make choices to do tasks. So, let's walk through a really practical version of what agency actually looks like. If you come to me and you say, "Hey Tejas, multiply the first 10 prime numbers."

24:07Right? I would look at you and say, "I don't even know what a prime number is." Okay? And three, I get I Anyway, so um I I would say I don't know how to do that, but I have a language model in my brain, literally, to parse language. We all do. It's called your name GPT. No, and so I have a language model in my brain. And so you're going to tell me to do some complex arithmetic and I'm going to do the language model in my brain is going to say, "Hey, you can't do that, but there's a tool you can use based on its description that was taught to you in fourth grade math class." Okay? Some of you this may be resonating. The tool's called a calculator. Okay? And and I learned how to use this tool and I learned more importantly, I went to school and I learned its description. Use this for arithmetic.

24:53So, now I know, okay, there's a tool. And so what's going to happen is I'm going to pull up the tool with my agency. So, you understand, my agency is going to give me this tool, right? So, I'm going to use this calculator, and [snorts] then the language model in my mind is going to generate the inputs to the tool. So, you say multiply the first 10 prime numbers. Then the language model in my mind generates, okay, 2 * 3 * 7, right? And it's just like the the inputs come from my brain. Does it make sense? And so, then I do the final thing, which is I hit the equal button. And then the tool gives me an output. It shows me the value, and then it goes straight to my language model, and then I tell you the answer. Does it make sense? This whole thing that I identified is literally what AI agents are and how they work, but that button code. Okay? Let me show you a demo of agency. Um which here, so you said multiply the first whatever prime number. This is me This is the demo. Check it out. So, 1 No, 2 * 3 * 7 * I don't know the other prime number, 11 I don't know. And so, there. That's That's me using agency, you know? So, how then do we do this, but with AI? How do we do this artificially? Um let's take a look.

26:01So, if we come back to LangChain, and I'm using LangChain again to show you kind of the flow of things. That's kind of the point, okay? Um we're going to get rid of all of this. Um and we're going to perform rag, but this is going to be agentic rag. What's the difference between classic rag and agentic rag? In agentic rag, the machine makes the decisions. I don't make decisions. So, um I'm going to delete this prompt template. I'm going to give my Leave my chat input, and the URL I'm going to turn into a tool that the agent can use, okay? I turn on tool mode. We're going to get rid of this language model, and instead of a language model, we're going to actually use an agent. You may ask, what is an agent? An agent is a language model plus plus. Like you, you're a language model, but you have more, you know? That's exactly what an agent is, okay? So, we'll give this tool the URL tool to the agent. We give the chat input to the agent. And just like that, we've got a cool agent that can browse the internet. And so, let's try this again. We're going to check for rag. This is not going to work. I'm going to spoiler you. Um let's choose a better model. This is not going to work for reasons I will tell you in a second.

27:04But, if I send this prompt now, you can watch the agent use its agency. And And this is that moment where you go like, "Wow, that's so cool." This is not working because I didn't plug in the chat output. you didn't see any output, you know. Um so, let's try that again. And we'll send this. And now you can see the agent it got the current date, and it said to find out go check the theaters' websites, you know. So, it didn't do this for a few reasons. One, it didn't know what tool to use. So, this the description of this tool is not clear enough for the language model to understand. So, I'll say, "Use this tool to get movie listings from Rotten Tomatoes." And I can be as prescriptive of as I want in the description. But, now I've clarified what the tool does, and so I can, you know, do it again. Um and now because of the language model connecting the agent to the tool, um suddenly it knows Look at this, it's accessing fetch content, going to Rotten Tomatoes cuz I described the tool, and it performed agentic rag.

28:01You see that? So, the cool thing about this is I didn't go I didn't tell it where to get the data from. This whole thing was generated by the agent itself. It knew to go to this URL. I didn't copy paste this. This was the agent thinking. Does that make sense? It was literally agency, choosing which website to go to. In fact, we can do better. We can say, we'll just change the description, "Use this tool to browse the internet for answers." How cool is that? So, now we're getting more general. So, now we can do something like, "When is Tejas's talk at InfoQ Dev Summit 2025?" Let's do Munich just for safety. Um Right? And so, now the agent with its agency can go browse the internet, fetch content, and it got a 404 page um cuz this website is poorly built uh No no disrespect, but uh wow, that's unfortunate. Um W- Why is this a 404? Okay, so um let's uh let's go back to that. Uh no, but let's actually use a different tool.

29:11Instead of a URL because the URL tool allows the LLM to think of some URL. And the URL may not exist, this is feedback for you, but instead of thinking about a URL, wouldn't it make more sense to just think about a search query? Right? Instead of like thinking about because URLs are kind of hit or miss. So, let's do this. Let's go web search. This makes more sense. And we'll do tool mode. We'll give this tool and let's just check that tool description. Performs a basic DuckDuckGo search. Great, that's actually what I want. So, let's try that again. This time different tool, but same task, right? And now it's using the different tool, performing rag, and in a couple seconds uh my you know, you'll you'll decide how you're going to rate this talk, you know. That's uh that's going to happen.

29:55So, here we go. It it searched the web. This is the search query. You can kind of inspect. It found these results. Good start. And now let's check Tejas Kumar's talk is scheduled for Thursday at 9:00 a.m. CEST and the talk is titled So, it did that. The agent used rag itself, right? And this is kind of where we are in 2025. Agents just have a mind of their own. They can do things. from multi- Let's actually do the multiplication thing as well because feel like it's parallel. There's a calculator here. Calc- you later. Um you use a tool mode and you give the tool. And agents can use as many tools as you want. And so, we'll now do, you know, multiply the first 10 prime numbers, right? Um and it'll just do the job. It's using evaluate expression. It did that. It got some 6 billion something Uh and it did it by generating literally generating the inputs and reading the output as I did. It's me but fake or or another way of doing it. It's it's a it's a human agent but artificial. You get it? So that's this is how AI agents work. 2025 this year was characterized by AI agents. Everyone's doing agents always everywhere to the point where it's kind of irritating. Like there's a new coding agent every week.

31:10If you If you want to know which is the best one, use cursor. It's the best I don't have any partnership with them or anything. I just think it's it's a tremendous tool. It's it's I've actually built a similar tool so I understand the complexity and cursor is just a class of its own, okay? So this is agency and this entire year AI in 2025 has been about agents. There are limitations with agents as well, specifically around long-running tasks. And the biggest problem today with agents is context management, actually. Um what when you reach the limits of a model's context, even with the vector store or whatever it may be, um an agent can't perform tasks that take hours and hours, multi-step tasks, yet.

31:52But that's where we're going. Okay. So this is the state today. This is agentic rag um and agents in general. The cool thing about agents is a language model is just a piece of the puzzle, right? Um let's explore how we can make this even more practical to enable an agent to actually not just read but also write, to do stuff for me. I suck at managing a calendar. Uh just ask the organizers of the conference. They're like, "Are you sure you're going to be here?" Um and so let me show you an agent sort of in its prime, not not just that reads but also writes. I I built this actually use this myself, an agent to manage my calendar. So if I pull up Google Calendar, um this is like not fake, I promise you.

32:37And I turn on tool mode. Um I'm just going to go here, do this. And I can choose what it can do or can't do on my calendar. So, should I allow it to delete events? I don't know. Uh I'm just going to say do everything. It's fine. You So, sometimes you need to run before you can walk, you know? Um And so, this is my calendar. Welcome. And so, what I'm going to do is I'm going to open my calendar here uh in split view. Actually, that's not a good idea. Um we'll we'll do that. So, I'm going to now come here. And it can do this on my calendar. So, what I'm going to do is I can say, you know, make a lunch appointment for me today at 1:00 p.m. Europe Berlin time. Right? And I just, you know, some some models struggle with dates, so I'll say today October 16th, 2025, just in case. Um but if you're building an application, you can, you know, inject this in your prompt yourself. Um this looks good. So, I'm going to send this off.

33:30Uh and I'm going to now open my calendar. Hopefully, I open it fast enough so we can see that somewhere here it should appear. Uh open in split view. So, it's creating a Google Calendar event. And your lunch appointment has successfully been scheduled for at 1:00 p.m. in the Europe Berlin. So cool, it did it. Um and we can see the event here. So, it totally Here we go. So, lunch appointment. You see that? It totally did it. It did exactly This didn't exist before. You saw my calendar. Um and it did it. And it decided I need 1 and 1/2 hours for lunch. Thank you. Uh That's the agent working. Uh but you might ask, "Okay, um what if I want to use this, but I don't want to use it um in this LangFlow UI, right? Like, you can't build a business and expect your customers to like go use LangFlow, right? Um No, you want them to use your application. Ideally, you want them to use no application, just to do a job.

34:31And so, how can we do that? Well, LangFlow, thankfully, is is kind of cool because this diagram that I drew here with my calendar and stuff, I can expose this whole thing um over HTTP. So I this this whole flow now becomes an API, and it's open source. So it's running on localhost, but I can you know, it's it's it's a Docker container. I can put it wherever I want. And if I do a network fetch to this endpoint, I will just like trigger this entire flow and give the output to a user. So any front end can just talk to this as is, okay? Um but I'm I want to I have like 15 minutes left. I can't believe it's been so long. It feels like 5 minutes, honestly. Um but this would make this calendar agent accessible from any user interface that you build for your company or whatever.

35:15Um I want to spend the last few minutes talking about the future, 2026 and beyond, uh where I believe the web and software in general is going, and I think it's important to prepare us all for that. Um I think people are not going to care about your product UI. Uh that's probably the best way to say it. Um there's an there's a new standard in the world of AI called MCP, model context protocol. Anyone familiar with MCP? Okay, everybody, cool. Um it's a it's a it's Some would say, "Oh, it's too early. It's too experimental." It's not. Uh Stripe is using it. Um Neon is using Databricks is using a number Microsoft. There's a rumor that the next version of macOS at the OS level will have MCP support. So believe me, it's not Wild West. It's not too early. MCP allows you to use agent agentic tools from anywhere um without any UI, so to speak. Uh and so what if I wanted to use this calendar agent in ChatGPT? Can I do that? I can over MCP.

36:15How MCP works is it's a standard server client protocol, like HTTP, you know? Um when you open an MCP client, let's talk about HTTP, actually. A browser. A browser is an HTTP client, right? Your website has an HTTP server. So, the browser says, "Hey server, at this domain, give me the webpage." And the server gives it. Client-server, okay? MCP works exactly the same way. We're an MCP client, like ChatGPT. ChatGPT is an MCP client. An MCP client opens up, registers with an external MCP server, and says, "Hey MCP server, what context do you have for me?" Model context protocol. "What context do you have for me?" And context can be tools, like the calendar tool, the calculator tool, the web search tool. It can be tools.

37:03Or it can be not tools. It can be um prompts. It can be conversation history. It can be a number of things, okay? And so, "What what context do you have for me?" And so, what you can do with this is you can share tools with you can extend the capabilities of just like ChatGPT with your own custom tools, like with your calendar manager. Do you understand? And so, this means you never actually have to leave ChatGPT to do work. Uh let me show let me show you this. I think practically this makes more sense. So, let's use my calendar thing, but not use it via LangFlow, and not use it via not use it via some web interface, but use it over MCP. So, LangFlow exposes also an MCP server. So, I'm going to come in here, and instead of API, I'm going to share this as an MCP server.

37:51Uh and now, I can edit my tools. This doesn't make any sense, this description. So, I'm going to say, "Use this tool to manage my calendar, right? Calendar, like that, okay?" And I'll call this I'll give it the name calendar. Calendar tool. So, now I can expose this flow over MCP, and quickly add it to any of these MCP clients. These are all MCP clients. So, I'll add it maybe to Claude by clicking here. And now, if I restart Claude, so I'm going to quit Claude, I'm going to open it again. Uh we have Claude for those of you who aren't familiar with Claude, it's exactly like ChatGPT. Uh, it's it's just a competitor. But, look at this. The U The user interface is kind of the same. And if I open Claude and I go to my tools, you can see LF starter project.

38:33That's the thing that I just made, LangFlow starter project. And so, now I can say, "What's on my calendar today?" Right? Send it off. And now I'm not using LangFlow, but I'm just using some generic MCP client, something like Claude. And check it out. It's using the calendar tool that I just made over MCP. Because when I opened Claude, it said, "Hey, MCP server, what tools do you have?" And it gave it the calendar tool. So, now factory Claude has its capabilities extended. And it has these items on your calendar today, indeed. And there's the lunch appointment that I just built for 90 minutes. You see that? And so, Claude can do this on my behalf. So, let's talk about UX. Uh, previously, if I wanted to manage my calendar, I would have to go on some type of calendar app, Google Calendar, and then I have to look, first of all. But, but there's cookie banners.

39:26Uh, there's there's Google saying your Google Workspace is about to expire. There's all kinds of notifications. Um, and then, okay, I reach there, and but then I have no mouse. Uh, but I need to create a new event. But, how do you create an calendar? You click and drag, right? Like in the You know what I mean? Like if if we're here on Google Calendar, check out this UX, fantastic UX. It's so great. Um, if I want to make an event, I do this. Um, this. You see that? Um, but what if I have no mouse? That's kind of hard. Do I go here? Maybe. But, what if the page doesn't have JavaScript? What if this screen reader is not accessible? Uh, what do The web is just a combination of decisions that were made, sometimes bad ones. Okay? And I say this as a web engineer. Like, we build things like carousels, you know, that are not really accessible. We build accordions. Like, the UI is challenging.

40:18And so, traditionally, I would have to go use some web interface and click on things that maybe I can't click on. Click on things that maybe need JavaScript. And if there's no What if I'm in a country with poor network connectivity, right? The web is hard. Uh but this is how it used to be. Okay, so how do we fix this? Well, maybe some new company comes along, invests millions in a great UI team, and they do great UI work. Great. What if we don't need UI at all? Wouldn't that be something? What if I didn't Look at this Exactly, you see that? You see that? That's exactly what I mean. Like, what if we didn't need any of this, but instead of all of this, I could just from here do everything I need. Do you Do you understand? What if this is my home, and I never need to leave my house and go venture across the wild wide web, you know? And and and like accessibility and and and and cookie banners and all of What if I didn't need to navigate that, but I just spoke, and my calendar was managed, my email was managed, I could buy stuff on Amazon by just saying buy it here, right? What if that was the future? And I think that is the future.

41:29That's That's literally what companies like OpenAI are going for. They want ChatGPT to replace the browser and to replace your web browsing experience. They want you to live inside ChatGPT. Again, to to just make sure this metaphor lands, how it works pre-AI, pre-MCP, is um we open the web, and we open we open a browser, and we leave our house, so to speak. We go to amazon.com. We go to Zillow. We go to these other websites. We leave our house. Uh and the websites have different standards for accessibility. Some of them don't work. There's cookie banners, whatever. Um what if we never needed to leave our house? Everything is in one app, which is Claude or ChatGPT or some MCP client. Um that's kind of the future. Because here's the deal. What if we could even use a computer through this? In fact, Cursor just released support for this, which I think is very awesome. In Cursor agent, you can just open a browser and say, you know, go on amazon.de and find clothes, right? And Cursor, by the way, is also an MCP client. And so, that's what it's going to do now. It's literally like navigating to this website in a browser um and telling you what it sees. And so, I don't really need to go do this myself. And the benefit here is that I can just speak things into existence. So here, I see the amazon.de homepage is successfully loaded. Now I need to accept the cookies dialogue.

42:49Um but the best thing about this is I have an agent to do all this. And it's as you can see those three dots on the bottom, it's generating. But then the question becomes, what do I do while it generates? What do I do while this thing cool, he accepted the cookies. What do I do while my agent does my shopping? What do I do while my agent books my aircraft tickets? What do I do while my agent um manages my calendar? And the answer is I live my life. I don't I don't spend hours looking through emails only to delete spam after spam after spam. I don't like manage a calendar and do all the No, I actually have a family that I like. Uh and I I have friends. I like doing sports. I A lot of us live today wanting to get more time back.

43:36Uh I think that's true. I think a lot of people we we spend a lot of time doing nonsense. Um I have to get my boarding pass for my flight tomorrow out of Munich. Um how cool would it be if an agent did it for me while I'm here talking to you? In fact, that's exactly what happened. So, full disclosure, I have a team of agents working for me all the time, and it's great. I I have so much time to do the things I actually want. And this is the vision I have for all of us. That's where indeed I believe AI is going in 2020 uh six plus. I'll finish with this. There's just a couple minutes left, but OpenAI recently had a dev day where they showcased some of this uh where they also agree with me that they want the web to go.

44:13I will say consolidating on a platform like Open AI is probably not where I wanted to go. I want this to be distributed. But this is what they released in case you missed it. I think this is very important to show you. I'm going to come over here and skip forward to this right here. This is their dev day YouTube event. I'm not going to show you all of it, but there's this really awesome demo this guy did around 11 minutes where this is so cool. Um And some of you may have seen it if you have. So what he's doing is he's at-mentioning Zillow, which is like a home search platform for property. And what it this does is it connects to Zillow and it embeds this actual interactive map of properties inside chat GP. This is chat GPT.

44:55And then you can kind of find your apartment, find your houses. You can expand the map fully. Whoa, that was not um Let's let's go back real quick. Because here you can So when you click on a house, you can expand the map and scroll scroll scroll. And then finally, you can say I would like to buy this one or I would like to rent this one. And all this experience doesn't happen on some separate website with its own cookie banners. It happens inside chat GPT or Claude or any MCP client. Cursor, Windsurf, whatever, right? And so MCP allows this beautiful experience that I think is for me a lot better than varied experiences across the web. Um This is this is And you can actually build this today. There's an SDK to get your apps into chat GPT that I think is is worth taking advantage of. In fact, if you have something customer or external facing and you're not here um you you may be missing customers. With that this is what I'm going to say. This this talk started about UX. We talked about GPT-1, GPT-2, GPT-3 and how nobody cared until there was a chat UI on top of it.

46:05That's where we started an hour ago. Okay? We'll finish with UX. Um UX is changing. UX is used to be go to a bunch of websites. It's now just prompt things into existence. What does that look like for you? I think it's worth considering and this is indeed where AI is going in the future. Let's wrap up. Let's talk about next steps. Uh this has been the better portion of an hour. Um he's telling me I have 5 minutes. There's a clock here that says I have 1 minute. Who wins? Guess we'll never know. Um the I I This has been the better portion of an hour and and frankly I have enjoyed this very much. Uh as you can tell I I probably I like talking a lot. Uh I actually This isn't enough and so I [snorts] I do have a podcast uh where I talk about these things longer. There's like each episode is like 2 hours or so.

46:53>> [laughter] >> Uh which I don't know if you want to do that to yourself. Um but it is actually it's very helpful. It It It does give you the latest updates and so on, but also there's guest episodes I think which is more important. We talk to people who lead companies working with AI, the CEOs of Vercel and Netlify, there's people from, you know, Cloudflare, whatever. Uh and so it's it's just worth listening to at least to me. Um second, the tool that I used to build those workflows and then expose MCP servers and APIs, it's called LangFlow. Uh it's free, it's open source, and it's uh it's Yeah, it's free. It's open source and there's no money to be made from it, but I I think it's it's just a great tool. And so, if you want to check it out, it's langflow.org/desktop.

47:33That QR code will take you there as well. Um and then if you have any questions, I'm here uh of course. I'm happy to answer your questions after somewhere. But if you want to find us on socials, that's @TejasKumar_ and @LangFlow_AI. All right. Hey, thank you so much. This has been a real pleasure to be with you today.

More talks

All 23 transcripts

Elsewhere

There is every talk I have given, all 80 of them, ConTejas Code, the podcast, and Fluent React, the O'Reilly book on how React works inside.