Introduction

So I'm Josh, by the way, I'm a co-founder and CEO at MindStone. Normally I would have introduced this meetup tonight, and now I'll do the last talk. And I'll try and revamp how I normally introduce it at the end.

How many of you have heard of Replit? Okay, not even half the room.

How many of you have tried Replit Agents before? much less.

This is gonna be fun. Hopefully this works.

Can you hear me when I, I'll try and push it a little bit closer.

Concept of Agents

What I'm gonna try and do, I'm gonna put my notes here. For the last year, year and a half or so, this whole concept of agents has been thrown around quite a bit.

The idea is that large language models, you can often give them one question, they'll come back with an answer, you can iterate on it. What if the agent, what if a program could iterate on itself and interpret its own results in order to figure out when it gets to a result?

Live Demo Overview

And that's what we're gonna try and do now. So the whole thing here is a live demo. So I'm gonna create an agent.

We're going to live build a feedback app. Which, in this case, I want your feedback on how tonight went.

So I'm going to say, build me an app that allows people to rate three talks with one to five stars. Hmm. and give free text feedback as comments.

And then I'm gonna say, let people ask a question and vote questions from others up and down. I'm gonna hit build.

Building the Feedback App

And this is honestly one of the most interesting things that I have seen AI be able to do recently. So what is happening now is it's actually looking at what I want to build. Now, this could have been anything.

I was literally just text inputting what app I want to build. It's going to build a plan.

It's going to actually suggest if I want to include additional features. So here, do I want to include user authentication or an admin panel or implement real-time updates or add data visualisations? I'm not going to do any of that. I'm just going to build the thing that I asked it to.

But it's interesting to see that it was suggesting if I wanted to go further than what I had actually said. What you're seeing now is it's live coding the app.

File by file. It is going through. And on the left-hand side, you can see the files being populated.

It is doing multi-file process where it's building out the database, the models, the user interface. And it is going to live build the app that we just asked, which is going to be about three talks.

People will be able to rate one to five stars. You'll be able to input a free text question. And people will be able to vote these up and down as we go through.

What is interesting is that it's not only able to build it, it's able to execute it.

Execution and Iteration

1Look at the logs that come out when it executes, figure out if there is a mistake, rerun itself, correct the mistake, and then go on. And that's the loop that will go through here.

So it created the front end, it installed Flask. So it's actually installing various modules on the server as we speak as well.

Now, this does take sometimes 30 seconds. I'm sorry that building an entire app sometimes takes more than a minute.

Sorry?

$20 per month. Not unlimited, but I have yet to hit the limit. And I've been doing live demos quite a lot.

So it's now created the index. It's created the database, well, I mean the link, and now we're gonna wait for another few seconds.

If it hits a point where there is a problem, it actually comes back to me and it should be asking me for questions. So here you can see it's restarting it. Here, you can see it's accessing its own pages. Not get through the UI, but just through the server to figure out if there are any logs that come out. So it's figuring out, is there anything that is going wrong here that I need to improve on?

Challenges and Troubleshooting

I'm just telling it, you know, this is what I want to happen. Okay.

It's asking me for an open API key. Let's go and get that. Okay.

Interesting. I don't know what was going wrong. But I am just going to go into creating a new API key, demo, copy, going back in here, and I am saving the API key.

Wait, you can't have done it already, no? No, okay, so it's thinking.

It renamed, wait, renamed the talks one, two, and three, editing main, let's figure out what it's gonna do next.

So the other interesting bit that's happening here, so it's actually taken my API key and it's now going to implement the integration with OpenAI to answer the questions that you will ask as we go through. Any questions as we wait for another minute? The tool's called Replit. R-E-P-L-I-T.

Okay, it's done talk one, two, and three. At least that's good. Let's figure out if it's doing the rest. Any other questions?

Right now, it would probably have a problem with that. It doesn't actually handle images inside. I say that, I've not tried it. I would be interested to figure out what happens if you have a UI, yeah, if you have like a sketch, a napkin type of thing in mind, yeah. Very good point.

Okay. Let's just see what happens here. Submit question. Yes, okay.

Talks have been renamed, but the questions aren't yet answered by OpenAI's GPT-4. Okay, so it somehow didn't get the, the integration properly yet, but it's already gotten the API key, so I wonder... This is as advanced as it's gotten.

I don't even know if it's written some of the code for OpenAI already or not. I don't even bother to figure out where it would be. I literally just say, it's not working. Make it work.

Now the other thing with it, so this is not foolproof, like live demos, live demos, we'll see where we get to. One of the things that does happen is when it's not able to, okay, now it's installing OpenAI, so probably it wasn't doing everything before.

One of the things that I do run into sometimes is if you're building something and you get into a loop that it doesn't seem to be able to figure out, it's actually often better to just start from scratch because it's like, well, I only spent 10 minutes building this app. Might as well start from scratch. And suddenly you're out of the loop because you realize it got stuck somewhere. And so 1if you start your initial description of the app that you want in a slightly more defined way, it actually is able to get to the end result much faster.

Technical Details and Comparisons

Yes, sorry, you have a question? Right now, Replit is restricted to the tech stack they have, which is Python, Flask, and Postgres and stuff, but there is, I was hesitating between doing Replit and a Cursor demo, so Cursor obviously gives you much more flexibility.

If you are an engineer by background, probably Cursor will give you a little bit more. If you don't know how to code, Replit is probably where you want to start.

Okay, can you add new questions? Let's figure out what happens now.

Who am I? Still not answer, whoop. Question.

I'm gonna, I don't think that's it, but we'll see. Test. Nope, still not answering when I create one or when I add one.

So, again, two minute delay here. There's one more thing that I wanted to add to this as we finish.

So, any more questions? Yes, Alex.

Impact on Roles

It might be like going forward, but if you really fast forward, how does this affect the relationship between... That is a very good question. Where does the relationship between product managers and engineers get to when product managers can just get the AI to build the stuff rather than anyone else?

It is a really interesting question. I don't have a definitive answer. The only thing that I do know is that it is only going to become dramatically better.

What is clear is that the models are still getting better and our implementation of it is still getting better. This is still based on GPT-4 type technology. It's not even working on O1 yet, which came out last week, which has advanced planning and kind of goes a little bit further.

It is not based on Opus 3.5 that is going to come out or GPT-5 that's going to come out, which is all going to go further. And then we're going to look at tech stacks that are going to expand because right now one of the reasons these things work is that they try to define the tech stack in order to prevent errors from occurring.

So you kind of say you can only build with these play blocks because we know they don't have errors in them when you combine them. So there are a few hacks that are coming through. But the ability to go and create these is going to go much, much further.

Okay.

Advancements and Future Potential

Let's see. Don't know what's going to happen now.

AI answer could not be generated. Okay. That's a shame.

I'm going to pass this because live demo because I wanted to do one more thing. Okay. Answers are not being generated.

Let's not finish. I've never tried this before, by the way. I'm just telling it, like, let's not finish this feature.

Final Feature and Conclusion

This feature. But instead, I want a button that takes all questions and creates a song out of them using OpenAI.

Let's figure out if it's gonna do this one, because that was gonna be how I wanted to finish this, but we'll see if it, I've never tried actually telling it not to do an executed feature, so we'll see.

Any other questions as I go through? Yes?

Absolutely. Everything. Everything here.

All the code is here. This is all generated. You have all access to it.

You can hook this up to your GitHub, basically execute this locally, however you want to do it. It is totally there.

So this is building live apps with everything that you can imagine that you've been using. Like literally, how to put it? Zapier will hook up multiple apps and it gives you a defined list of tasks that you can go through.

This is actual software that you can tell it to do whatever you want it to do. Like you can build an email app using this, you could go and, anything really, anything that you've been using on a machine.

So let's just figure out where it's at. Okay, it's not, see now it's getting stuck, which is a shame, but I wanna show the last bit then.

Okay, which is here I'm going to hit deploy just for the sake of it so that you can all actually use this as we go through, because the whole idea was you go and use the QNA app whilst I was doing this. But that didn't quite work.

Any more questions as it deploys? Yes. Yes. Sorry, yes.

It's a very good question. I would not have the answer to that at the moment.

I actually did this demo a few times before and it It seems to... I don't know where it went wrong in this case. Somehow it got lost into answering the question.

Yes?

Yeah, absolutely. It's entirely your intellectual property. I say that.

Okay, so here we are. You can go here. Actually, let's go QR code generator.

If you all... Try this out. Go and try this QR code.

Okay, so I'm gonna flip it back if you're still not there. I'm trying to figure out how many phones are still up. One, two. There we go.

So now you can rate and you can ask questions. Put in any question that you have.

Interesting. Oh, wow. Lots of people have actually left comments on the first talk already.

And then here you've got tests. And thank you, everyone, for putting in a whole bunch of test questions.

You can actually vote these up and down. And you can see them populate to the top.

So the app actually does work. It's just not answering the questions that I really want. If I had another 10 minutes, I would do that.