What Do Madison Fellows Do? | Building AI-Powered Marketing Intelligence Systems | Professor Bear

[0:02]Hey, Professor Bear here. Uh, I'm going to talk a little bit about if you join the Madison project as a humanitarians volunteer, what exactly do you do? I'm going to do these videos for all of the projects. We're just starting with Madison. Madison is just tools for marketing, branding, um, market research, surveys, that kind of thing. Primarily like a lot of the projects, these are angantic AI pro uh projects. The main tool we build at Gantic workflows in is something called N8N. It's a not quite open source, but it's free for education. You can download it locally and we're going to extend NAN quite a bit. So what N8N is a visual

[0:46]Tool and why that works well for us is it allows our domain experts. You're going to if you're in medicine, you're going to be working with people, you know, actual people in the industry doing marketing, branding, etc. who have a lot of experience in that. And they need to be able to build workflows in order to test them. So, and everyone that we work with can create little visual things. If we used a more Pythonic tool like Crew AI or whatever, we just can't get our domain experts to really do that. Um, what what this is is just visual programming. So they have a bazillion components built into NAM. Do

[1:23]Things like, you know, fill out a Google form or connect to Discord or whatever. You build them as a fellow. You're also going to build custom components that we're going to contribute back to one the Madison project which is open source and to N and N. So if somebody is building an end workflow and they want to use one of the components that are built in medicine, it's open source, they're free to do that. The other thing we're going to do quite a bit where Ed is a great framework, but where it may be a little bit weak is in sort of the mathematics around reinforcement

[1:59]Learning. So these agents like for example self-driving car or something that um plays Pac-Man or Pong, these are agents. Asia has been around for a long time and reinforcement learning theory has been around for a long time. What this means here, for example, if we were in a self-driving car, your actions would be turning the steering wheel a particular degree to the right or left. It would be pushing on the gas a little bit or pushing on it hard. It'd be pushing on the brake a little bit, pushing on it hard. These are the actions. In pong, it might be moving the paddle up or moving

[2:35]The pedal down. The environment is what you see in Pong. And that would be sets of, you know, three or four screens from a game, the actual pixels on the screen. In a car, it would be the cameras and it may be LAR as well if if the self-driving car uses LAR, but it' be its input. The state is a particular action like this is what is on my Pong game right now. This is what is in my cameras on the car right now. There's a red light. there's, you know, a dog walking across the street, whatever. Your agent is what learns. And so, angenic AI and agents are kind of the

[3:14]Same thing. What's a little bit different though in these angenic frameworks now, and this is where things like N8N come into play is we have triggers. What trigger means is we don't necessarily use it all the time. So, a trigger for a car would be when you start your car. Trigger might be when I want to do some stock analysis or branding analysis whatever or could be a market event. If we see something some sentiment happening about a company that might be a trigger big difference between sort of anenic frameworks and conventional reinforcement learning is the language models themselves. Some people and it's probably most people

[3:53]Rely on the LLM to basically be the agent. That is the entire decisions is done by the LLM. probably not the best idea. Probably a better idea would be to have the LLM be an input to your agent. And when I start talking about reinforcement layering theory, uh you'll sort of figure out why. Uh but this allows us to use, you know, multi-armed bandits, uh you know, Q-learning, policy grains, whatever. And but this is very powerful. the the the the language models controlling the agents is a huge thing but it shouldn't be the only thing in an angentic system. The other thing that these are memory so we can store things

[4:40]And use those as input particularly personalized context what context is important like who are you what language you speak etc and access to tools tools are kind of tools in terms of getting information there also tools in action we can go fetch some information if the agent thinks we need to fetch the information so there's going to be a lot of extending nan and reinforcement learning in general to adapt reinforcement learning to angenic frameworks. But you don't do that initially. The first thing you do if you're a volunteer is you build a tool. So you pick a tool, for example, a something to get some

[5:18]Information about the sentiment of a company. So some tools you might think about are, you know, research agents that research a company, content agents that either check content or create content. Uh be performance agents, you know, are you measuring how how good is your your uh ad campaign on Instagram versus Facebook versus this versus that? Um what you won't do initially is be part of the orchestration layer which is this which is the actual you know mathematics about reinforcement learning. You can sort of earn that privilege by doing a good job on your first agent. So what you do is you build an agent to

[6:01]Do something um and most of the project we have are angentic. They're just in different domains like Dehoff is in biology. Uh Microoft is in finance. Popper is in skepticism and how do we trust AI? Those kinds of things. But if you specifically geared towards medicine, the agency you'll build will be focused on marketing, branding, advertising. It's great if you're looking for a job in marketing, branding, or advertising because you'll also work with experts in the field who know about marketing, branding, and exercising. which is also important as a developer as somebody who's developing agents and software for marketing branding and agents it helps to learn something about the field

[6:45]Itself. So you'll do that and so that's it. So to summarize uh Madison is an aenic framework for marketing, branding and advertising. Our policy is we want you to work on something you're interested in. Our experience is that if you don't work on something you're interested and you don't do a good job, everything we build is open source. So you can take this and include it in your job or start a company or do whatever you want to do. This is primarily for your learning and you'll learn a lot about angentic AI in medicine and in other projects which do angentic AI. But why pick medicine is

[7:28]Because you're you're either interested in also more generative like the MOV project doesn't make TV commercials. We don't really generate much media. Instead, that's more analyzing financial data. First, gathering financial data and analyzing it. Madison can be more creative. So if you like angic AI but also have sort of a creative bent, you maybe want to make some short films or commercials or you know ads or whatever. But also what's also interesting is the synthetic personas. So what a synthetic persona is is when we ask a language model to act like somebody act like a northeastern student who just graduated and then start asking them questions.

[8:13]Uh, another thing that we do in the Madison project which we're not doing in any other projects and that's critically important because more and more often people are asking language models for advice. What restaurant should I go to? Well, they go to claude or chat TV and ask them what should I order? What should I buy? This is already common. Maybe our last research we did was maybe 30% but is growing very quickly. And so the ability to not only interview real people but inter interview what we call synthetic personas is sort of the future of advertising. So that's a big part of what we do as well. So to

[8:51]Summarize if you join the Madison project you're going to be working with real world experts who have a lot of experience in branding marketing advertising and you're initially start building an angic tool. some useful little tool, you know, up here some basically a component here that you can drag into 8 and N8M and it does something useful. And then if you do a good job on that, you can either continue to do that if that's what you like doing building the level tools or sort of more talk about reinforcement learning and learning in angentic AI in general which is what we call our orchestration layer. So

[9:31]Hopefully this gives you a sense of what Madison does, I'm going to do the same thing for the rest of our projects. So people sort of know if I join medicine, what am I doing? Um so hopefully this tells you if I join medicine what am I doing? You are going to build angentic tools for marketing, branding and advertising. Uh we just want you to have a good faith. So if you have a technical background, we also look for project managers, people who want to manage other people doing this. You'll learn a lot about Agenda AI. If you join this, we look for just a good

[10:08]Faith effort that you're going to do the work. You're going to put in the work. Put in the work, you'll learn. If you don't put in the work, you won't learn anything. And we'll write you a letter saying, "Thank you, but you didn't put in the work." So hopefully that explains things. Uh if it doesn't, I guess I'm making more videos on this. And again for the humanitarians it helps support us. If you like share and do that we are a nonprofit. So if you don't want to donate but you want to support the nonprofit like it, share it, do something like that. That helps us. So please do that if you're

[10:43]Inclined to do that. And I guess pretty soon I'll make another video on or other projects. So you can look at those as well. Okay. Take care. Um see you soon I guess.