Summary
MCP (Model Context Protocol)
Privacy and Security Concerns
Personal Assistant Use Cases
Legacy Software Integration
Cost and Infrastructure
Scaling from Prototype to Production
Notes
Transcript
Whenever that's going to settle, like packing up your stuff. Yeah. I got it.
Okay, action. use cloud to create the form and knowing Like it lives on my-The form. The form, yeah, the Google form. And it lives in my-browser The MCP. Okay, yeah, now it's 14 responses, so it should be A couple more people. Um...
Yeah, so the purpose of this first session is a chat about like between Also, how do we want to organize around this? But we want to create a pretty good session.
We just wanted to be mostly demand-driven in terms of like, that's why I think the first session we just want to gauge what you guys know about it, what you don't know, what you're interested in, and then hopefully come up with probably many tutorials or... small projects that we can walk through and start to understand different uses. Yeah.
We were hesitant to send it in the free care first. Because I think first session wanted it to feel more like a focus group where we're just talking about what we're interested in. Yeah. Feel free to obviously send it around.
I'm also hoping at some point to get Chris Blattman to come and talk through one of his Claude Blattman tutorials, if you've heard. Have you heard his song? He's like, completely in the deep end with Claude now, and he has a website called claudeflatman.com, and he's used it to transform every single research workflow he has. He's evangelizing Claude all across the internet. So, Ricardo Hausman sends his tutorials and all of your.
Maybe that's a good place to start. So like what else have you guys heard about agentic AI? What do you know already? Or what do you not know?
So we were thinking maybe kind of like at the beginning, like some people showing how they use them or how personal experiences. Bye. We're just starting. So maybe we're talking, we're thinking, I mean just the first part of a session being, hey, let's explore these use cases or someone can show us, I'm using them to make my life easier with these, I don't know, with these steps, right? And then at the end, maybe have time to actually do it yourself, right?
And that brain starts, like we're taking notes and transcribing them with AI also to create basically a knowledge base of this. for us to share resources. We're thinking, hello.
So maybe to have like a document, this is well documented, right? So all of these things probably relies on having the correct knowledge, right? The correct MP file or the text file that creates like all your requirements specifically, right? So we want to create also like a database for us from these sessions, from good resources, from what works for us. We've seen a lot of resources out there for free and trying to curate some of those into notes that we can share and contribute to together to make it bigger.
Okay, so it's kind of like an open source version you can think of like an open source plotter chat GPT but it's like totally customizable so you can essentially build out without a subscription and the agent for yourself so that's like one of the tools that we could also eventually talk about my friend has connected it to his whatsapp and he runs code by texting it like on whatsapp I don't know how that works but pretty pretty crazy yeah one one fun thing that we can create is like a file with all the tools that we find like I know I'm a big user of Cursor, for example, and I use it a lot, and I told other people to use it.
It's also for most of us our last quarter, so we want to both have better experiences with this, at least know what is out there and how can we use them. We also have this kind of Don't worry. reactions about ethics and also what should be done and what should not and how can we actually use them better for ourselves. And I think that's an ongoing discussion, but we also been talking to James on the CAD program.
To actually have those discussions here, Oswald, so maybe one of the sessions could be, okay, let's talk about the limits of this, or even if this is a language of core values and principles, and how can you make it safer? So I don't know if you heard, but Anthropy, for example, has this constitutional AI, so it's basically what is the constitution behind what they believe, okay, they have trained the model to be or to behave, or the rules that it should never or you don't go further than that.
And I think that's super interesting. And I have actually one resource from a class that I took where you can create your own constitutional. and and and tree not free more alone on a set of rules and like they the principles behind what you want to create. And we can make one session about that. I think that's also super cool. So the idea is basically we can lead all of these things together and make sense of it together.
And we're not here to teach you how to do it or teach, you know, like, not a TA role or anything like that, but just let's build this together. So anyone can participate, anyone can do the session, anyone can just join us also in leading the organization of this, the logistics. So that's our vision of this.
So what do you guys currently know, or what do you not know when you're interested in learning related to agentic AI? I think we should have the session on MCP and How to make your own tools. Thank you. That's how you do it.
MCP is Model Context Protocol, and so that's sort of the standard way that you can make your own sort of functions or like data resources and have a LLM be able to communicate with them. And so any LLM should be able to, it's like, an API point specific to that LLM. And so that's how you make your own tools and have it do very particular things.
And you can create your own data. I can never give cloud, well, any AI more access than literally just one folder. So I'm really interested, I'm scared, on giving it access to anything else, like more, even like cloud.
So I want to say, like, hey, look at all my emails. Like, literally manage my calendar. Like, you know, if it will be like a secretary of some sort, like, I just want to tell them, like, I got to do this and this. Prioritize. But that wouldn't mind that I have to give it access to my whole like-Totally. And I don't even trust Google that already has it.
Yeah. So yeah. I'm also super interested in that. And something that I was just mentioning this before that I want to go through is Chris Blattman has made an intro tutorial for anyone who-and I guess it's more research bent, but it's anyone who wants to-use agents for more organization on research workflows, team management. So one of the first tutorials that I wanted to go through was that one which he connects to his email and his calendar and he just like, creates a system to better basically communicate across apps and figure out your workflows and organization better.
I think he did a triage version of the emails to see what is important and highly, triage them, they're only able to answer important things faster. But I also have this, I'm also worried and scared about giving access to one of my, you know, the computer. I specifically ask to, ask me always, like, what are you doing? And ask my permission before running something. Because if you run some Python or whatever, install this thing, and you really don't know what is behind, it can be malicious code that impersonates your, steal all your information and you never know about it.
So I think that's the risk with, like autonomous agents, like law and all of this, if you start using computers. And we've been thinking, How to do this? Thank you. more secure And there's a lot of also solutions that want to charge you for doing that. Like I keep, I think it's one of the tools that as well as all there to, to basically what it is, they deploy like an isolated system where like an instance and IBS or whatever, where you can deploy your, like you can leave there without access to your personal information, but you know, like configuring their own, their own environment.
Like a virtual machine. Yeah. Something like that. And I think probably that's like the best, at this point, but we can have one session about how best practice for doing that, or even try, and I also wanted to do it, I haven't done it before, so I think it's cool to do.
What else? Yeah. Um. So I don't know very much about Agenda K.I. I've been very interested in it. In my business role where I work, we use a lot of legacy software.
And I wanted to explore the idea of building some kind of agent that could interface with legacy software. Like I just Googled it, 'cause I can remember it was called like the Clogs computer use I don't know. It's the one that can basically see your screen and interact with your computer. Something like that. I don't know.
That's the biggest pain point in my job. We have accounting software that is 20 years old. And Hundreds of people use it every day, and they spend hours and hours and hours doing it. The dumbest things, because the software hasn't been updated in 20 years. So it'd be cool to find an agent that I could deploy across Or I mean, we would save like thousands of hours if I could do something like that. And they're going to pay me for doing that.
And also, one comment around that isI'm super interested in how far we can go there, right? So can we, I think we can do a mini mobile fast with this agent, but can you deploy it at what level is this scalable? You know, like going from just using it, like these videos do on Instagram, like I just created my own app, yeah, but how can I actually do a scale that into code that is working in production environment or something?
And for me, that's one of those most important. I think that's the gap right now in two. Do it your own thing and making that-a product that you can then offer to someone early in the year. Out there, right.
Yeah, in my case, I, Similar, I've been hearing about it, reading about it, but I don't have actually engaged with it. So I don't know how to create an agent, basically. I want to know. And the purpose of it is, at least for me right now, of course, there probably will be a lot of other things when I engage that probably we'll think about. But for now, it's basically like an extension of me. It's like an agent that can help me reply emails, an agent that can help me in my work, what I'm already doing.
Yeah. What we're thinking of having the sessions to be, even if you miss one, you can still go to a new one, because it should be independent enough. That's I think one of the part of having the brain that I told you, the documentation of all of it, is that anyone can access a meeting, the recording of a meeting, or the transcript and understand what was done. Maybe for us, have this, cards or flash cards about what tools we're presenting, how do they use them, what is this useful for, and like to persist that knowledge into our memory that file in our database for this session.
Thank you for doing it. How are you going to be communicating via WhatsApp, I guess?
Yeah. Also, we can create a group, maybe WhatsApp, to do the one time.
I personally have programming well right now, TA, until 4:20. From 3:00 to 4:20.
Because we were thinking parents because most of us we don't have classes but we can find more space in the U. Sorry, Thursday.
Another thing I was just saying, because I once read aOh, I had a claw with that one. We were saying. The research was basically that this person had access, the club had access to all of their machines, right? And part of the job of the club was revisiting their emails and so on. So at some point they send like a Like emails, like trying to signal an affair or something like that. And then Cloud started to-yeah.
And then when Cloud's call was not being accepted or this person threatened Cloud to just kill it, Then it started to like, "Hey, it would be such a shame that your wife knew about this." Like blackmailing the computer science. Yeah.
So I guess one thing we can talk about is how you even establish guardrails with Agents like because you want to give it control to the extent that it can do tasks for you but also don't want to give enough power to Black Million about your affair, I guess.
I have no conception of how this stuff costs. If I want to make a personal assistant and have it help me apply for jobs, is that like a dollar a day? Is it like 10? Yeah.
Really like, include like three or four agents with different like skills and sensations about Caribbean decor letter, my CV, updating the scope of the requirements in Azure Posting. And For me, I think something Anyone to share withLike my way to do it. the text or the latex file that is exactly like the same as the Harris that Doe does in the world. but now we can program the way to use it for. Because people are already doing this We don't want to-so we want to know that behind those for accessing the new opportunities.
And it's almost one of the strengths of having these groups So for example, myself saying, or any of you, I spent like three hours on this and I shouldn't do that because this way was way faster. share those experiences for those who don't have to.
Yeah, one question I had was, Do you guys like a format in which there's a live coding component in the sense that we would walk through a tutorial together? Or would you prefer something where we find an interesting tutorial, we say, let's all explore this at home, and then come back and talk about it Do you guys have a sense of like what? You'd prefer-Or we can even come together and do mini projects in terms of forming groups.
If there's a longer term thing you want to work on that can't get done in an hour.
Many times. Yeah. That's where we get better. I'll say if I have to take something home and do it, I will fall off that wagon so fast.
So I think as we have access to these tools, The chance of doing that isAnd not to hold him to it, but James got really excited about this idea and he, for MPP folks, James Turk is a big CS professor in the CAP program, but he got excited about the idea and talked about possibly being able to give us the resources for some accounts. So that would enable us to also... You don't put the tokens by yourselves, but yeah.
Was there? Anything else that we haven't talked about that you guys are interested in? When we're thinking about the direction of the group,I think some of you are also taking this design beyond you.
And maybe we can also get some credits for us We're trying to investigate how that works in the University. But we can also add some-Great to you, it's totally amazing. That class was so cool. We'll have to. It's good. This is super cool. But yeah, I think maybe I'm quick. Are you pregnant? Any cloud, JVP, Courser, what are you paying for? Courser? Courser.
I prefer cursor and cloud. I was paper to him and just$40. I was using I think it's fine.
Now I'm experimenting with. One good thing of colleaguesIt doesn't have limits. I don't, like you don't have any limits. Mm. Right now. Right now. Like cloud, it has like, it's,Okay, like at some point, like I'm just doing some benchmarking with some models. for OCR. And every two, 20 minutes, it's like, OK, wait for five hours. That's so stupid. And you're like, OK, no, I'm not paying more. But with Codex, I never--How's the weight?
I've been experiencing that Coursera, for example, keeps you on limits on buying no more than you're And for doing that, it kind of-optimizes the power of the model that it has available.
So for complex taxes, it's just like, More advanced models, but for executing a plan. So once you create a plan for creating something, then execution is using the cheapest models to execute because once you plan and do the heavy thinking, it should be cheaper. So it will be interesting to, for me at least, to test.
All the, like with the same prompts. Test different AISee you live. Like, and, like, see... like do some benchmarking or AI Have you tried open call?
Also, like a quick CLI tutorial, what was super nice. I was super scared that we can CLI before the class. And right now, I think I'm not going to use-Yeah, anything else? See you later. Same, yeah.
But yeah, we can share some of those that are of that free or--Cool. Yeah, so that's-The beginning of this, I think, and our hope is to you know, like to start building these things together. Do it. I want also to get more of the first years involved, so if you want to maybe get some of your friends here, yeah, that would be fun. Yeah, yeah, yeah.
Yeah, we can do a few things. I think we'll send a survey for timing, and then we can also-we'll summarize the notes from today. I think we can look at the notes and then maybe write down maybe three or four ideas for the next session, and we can all take a vote on what we want to do.
No, I think that's a good point. Yeah, the idea is doing weekly because if we wait more than that, we then forget and don't do it, right?
So yeah, one hour weekly and-Gotcha. Took me now. Yeah, I think that's the way to go.
Uh, And then maybe also just if you didn't ask me, can you just message us so we know to add you to the group? Yeah.
So yeah, talk to your friends if you have some of them.
Which one? The one that like pseudo came out. My posts, have you guys heard of them? No. Is it better than others?
This is like... An accidental data leak reveals the existence of a new and graphic model like this. And--They changed the chat GPT to the original chat GPT. Like they were sitting on a model, but thought it was like not the right time. Really? Yeah. Yeah.
And we also wanted to give us a name for this. We were thinking maybe Agents for Good or something. That's too good for a good thing. But you know, to have the-The group started. Do you have ideas?
We should do a bit of our chemical policy related to the AI. I'm just wondering, I see the benchmarking, and I just say we should do a benchmarking for those, for policy specific tasks, but I don't know what we could see. Thank you.
Something I want to have an agent do is replicate a paper fully and also critique a paper that I've written or anticipate what it would say in a review for a journal.
Also, you could write a policy proposal.