Introduction

So I'm really going to look at live demo only. And so that gives me an opportunity to adapt everything.

So I have two options. And I'd like to get the feedback from each of you on what you would like me to demo more.

I can go take the example that Steve went through and show you what is possible using AI today and really kind of push it really far and kind of show you the limits of what's possible within like 20 minutes as I show you here.

Restrain it to what you might want to reproduce and how you can build on it yourselves at home and kind of keep it in the realm of what you might actually use on a daily basis. Which one would like me to kind of push the technology as far as it can to kind of show you what's possible within 20 minutes? And how many would like me to prefer look at what you would use tomorrow?

Okay, so I'll do the first one.

Okay, all of this is live, so please, I want this to be as interactive as possible, and I'm literally thinking through it as I go through, okay?

Demo Overview

So I'm gonna start with Replit. How many of you have heard of Replit before?

No, it's a good start. Replit is, you've heard of Replit before.

No, I didn't do it in the session, interesting. So Replit is an app or a platform that helps you, that can actually build apps for you. So you give it a direction.

Building a Recipe App

What do you want to build? And it will build the app for you. And we're gonna do that live now.

So we were talking about a recipe app. or build me a recipe app that takes recipes from users non authenticated or takes ingredients from users.

Actually, wait, let me think about I'm going to change it slightly differently. So build me a recipe collection app.

Users can suggest what ingredient collection users can suggest ingredients and they will populate in a list.

I'm going to start with this. Users don't need to be authenticated. Very, very simple.

So the whole purpose here is really to show you just what we can actually do with this technology at the moment. What's gonna happen here, like extremely simple here, right?

I'm just saying I want an app and I want all of you to be able to give me ideas for ingredients, and then we're gonna go and take it a little bit further.

Implementing Features

Okay, so it's now going to make a plan, help you create an ingredient collection application using React and Express. App will allow users to suggest ingredients that will be displayed in list format.

We'll keep it simple without authentication, nothing there. It is telling me if I wanted more, add ingredient categories, implement ingredient voting.

Ah, I'm going to do that. Normally, I never check another box because it's a little bit like, do I push it or not? It is live demo after all. And I'm just going to hit go.

I'm a software engineer by background, by the way, and this still blows my mind every time that I try it. Because this is what it's doing now. It's actually writing the code for me, doing both the front end and the back end, the stuff that I used to do for a living.

And I can guarantee you I was not that fast. And I made many, many more mistakes.

Reggie remembers. Life-threatening. Exactly.

Life-threatening, indeed. This is going to take another minute or so. I don't think it's going to take much more than that.

Is it better than bots? Sorry? Better than Bolt?

It is easier to deploy. Replit used to be what's called a DevOps platform. Basically, it has all the infrastructure to deploy an app and make it run in the easiest way. And that makes it very easy to go from something that is built to something you can actually use with everyone here in the room.

But it's approximately the same. Bolt, Lovable, Replit, those are kind of the three.

Deployment and Usage

You can see here it's actually created the database for me as well, for anyone that has any remote technical ability. It's literally doing everything. Oh, and there we go.

Ingredient Collection App. And it's asking me, how does the Ingredient Collection App look and function? Can you add ingredients and see them displayed in a grid?

Um, well, let's try test ingredient. Oh, wow. It's actually, um, test description and ingredient.

Oh, wow. Okay. It's done that. Yes, that works.

I mean, it doesn't always work on step one, to be clear. This was actually interesting that it did the full thing on step one. Basically, it's working.

We can proceed with implementing the ingredient voting feature that was requested. Okay. Would you like me to proceed? Yes. Proceed.

So it's now thinking again. Actually, let me see. I wonder if it. If it got stuck for a second. No, it's it's going okay.

It's obviously looking at my it's really looking at the app that it has. It's thinking it's not just doing what we were talking about before with prompt engineering. We're just predicting the next token or technically it is, but it's doing that in very complex processes and it's coming up with a new plan in order to now do the voting that we talked about that it talked about before.

So it's updating the database schema because now it has to include the fact that you're all going to be voting on different ingredients and It is, this is the thing, and this has only been for the last month or so that it's been able to do this. It's actually applying an upgrade to the database to make sure that it's changing the various columns there without losing any existing data. It's not, it's actually able to deploy this, and this is something that it was not able to do three months ago. It was always getting stuck at these types of things.

So, one question between you. You say now it actually builds the database and everything. Is the database running on their server, or is it something you can apply?

Real-world Applications

So this has now gotten to a point Where in the last two weeks I built three apps specifically to streamline operations within the company that I run.

One of them at the moment is about running these events. And so I'm about to release it. I did two in the last two weeks. This is the third one.

But rather than having a very clunky Excel sheet to coordinate for us, it's 10 to 18 events every month. We have organizers, sponsors, pizza and beer, venues, like everything around it right now sits in a Notion database actually. And I'm building the app.

The way that I'm building this app is literally I have meetings In between meetings, I put a few requests into Replit. I wait if it works.

I have another meeting. So, okay, did it work? Not okay. Another few requests.

And it's literally just building it in between meetings as I go from one to another. Now, I know that I'm on the early adopter curve by a mile here, right? But still, that is definitely how all of this is working.

Testing and Debugging

Okay, here we go. Can you try voting on an ingredient? Ah, okay. Failed to vote.

I'm actually going to open the console here. Just going to say didn't work. Got this. Yeah.

All I'm saying, I'm literally just copy pasting what the error was here on the server. I'm not even not even trying to figure out what this is. And this is almost like this partially touches my pride.

I am a software engineer. I'm not even looking at what is going wrong. I'm just like, it's going to take me too long to figure out what might go wrong. So I'm just going to tell this is the error. Go and figure it out.

Here you go. I don't know what it got. OK, see the voting functionality for it. Let me fix the issue by checking the validation and error handling in our voting endpoint.

Good. OK, it's redeploying restarting. And it's going to ask me again. There we go. There we go. Now it seems to work.

You might think, OK, well, what can you do now? Here, I'm going to look at, I'm going to actually say yes, but I'm going to look at deployment. I can just shortcut it, set up deployment, approve and configure, deploy.

How is this running? Which model is running in the background there? Do they have an interface to check No.

So it's $20 a month to use this. I don't know how many software engineers you work with, but they tend to be more than $20 a month. Sorry, inside joke for those that work with software engineers.

Yeah, $20 a month to get access to this. They have a combination of models. Some of them they actually train themselves, but most of it is done by Claude Sonnet 3.5 at the moment.

They use some 01 in the background. But this is the other bit, is that these models are still evolving really, really quickly. So 03 is coming out today. I heard rumors it was 10 p.m. Eastern time.

So, I mean, I'm going to be asleep, but the 03 is supposed to be a dramatically better model than 01, which means literally these people will just be able to swap out the model and everything that this thing got stuck on before will suddenly start working. So that's the other thing, these platforms are getting built and literally they don't do anything and their own products will start to get dramatically better because the underlying models are still getting better.

Cost and Efficiency

So for 20 bucks a month you can deploy as many apps as you want? No, you will get, you will hit a limit. There is a usage. You get $50 a month of usage credit with the $20 a month.

And then if you use it like me, I end up using, well, it's really this month that I ended up using it much more. Now that I'm deploying production apps, I'm getting to like a hundred dollars a month, but that's, I'm using it quite intensively.

So here we go. We have ingredient library, Joshua replica app. I can open this here.

And are you locked into their infrastructure? Infrastructure, no, it's a Postgres database. So the only thing I would have to do is I would have to, all the source code is mine, I can literally port that. I would have to have my own Postgres database somewhere. But it's all defined in code, even the schema, if you saw, the schema itself was actually defined in code. So the, where do we have here? The schema is there. So I can literally just copy paste this into another server and it will go and do it for me wherever I want to go and take it.

So QR code generator. Actually, sorry. Here. Come on. Just want to make this all easier for you guys. OK, if you can scan that QR code. Scan that QR code with your phone. You should all see the app pop up that you could live use. And I want you all to be putting in the ingredients. Please go and put in some ingredients in this. Everyone scanned it? We're all good. I'm going to go back here. Wait for some of these to start populating. Oh, there we go.

You can all vote on ingredients, by the way. So this is, I mean, again, I'm not doing anything, right? This is not the apple. That's not the apple. Okay, great. Put in some more and then vote on the ingredients you want me to use. Minus one, no apples. I want some tortilla. Add orange. What is beamy? I gave the explanation. Yeah, asparagus with broccoli head. It has a head like broccoli. Okay. So now please vote. Now please vote on the ingredients you want me to use because it's not done yet. We're definitely going to go two steps further. Tortilla is the one that is going to start winning here for the moment. Yes! I totally agree.

Okay, so you know what? I'm actually... Can you order ingredients in the list by best votes or most votes up? I'm literally just lazy. And so instead of trying to figure out which of the ingredients are getting most votes, I'm just telling the app, can you please sort this for me so I can have a look through? I'll flip this back here. Well, you can see tiny change it lit. I mean, it was so tiny that it literally just flicked through. Now it is relaunching. I go and have another look here. Okay. Oh, wow. We actually had people. Some people clicked on voting up quite a lot, I imagine. So I'm going to go and take just these two rows. I'm going to take these two rows.

Advanced Interactions

Now I'm going to flip to ChanchiPT. And I'm going to take some of what Steve was talking about before. So in ChattyPT, I'm going to do, imagine you're an aspiring restaurant owner. You want to launch a restaurant based on only a set of ingredients. The set of ingredients is the following. I think White Rabbit used that. White Rabbit? I don't know White Rabbit. Eggs and avocado. Well, there you go. I like that restaurant. Now, now, wait, what did I say here?

Okay. Give me five recipes or the five most appealing recipes you can create based on the ingredients. Walk me through your thinking before giving me your answer. How many people know why I would have finished with that last sentence? Walk me through your thinking before giving me your answer.

Because you are not yet using all three. That's one reason. Why would I not be, well, would I use it not on O3? Because they do that automatically now. O1 and O3 do it. Weirdly enough, you still get better answers when you do that, by the way. Really? Yeah. They tell you not to do that. Why would you, yeah, so good point.

Why would you ask the model to outline its thinking before giving you an answer? Anyone know? So the way that these models work is that they're constantly predicting what's called the next token. A token is somewhere in between a few letters to a word. And that means that the first word is part of what determines the second word. It becomes part of the input that determines the second word. The first paragraph becomes part of the input that generates the second paragraph. So it is literally taking its own output, putting it back into input and determining what comes out next.

Now, that means that if you ask it to give you an answer and then tell you how it got to that answer, it will just justify whatever answer it gave you. But if you tell it to outline its thinking before giving you an answer, the reasoning, the thinking becomes part of the input that generates the output, the actual answer at the end. Your answers will be dramatically better. The hallucinations will go down. It is the biggest differentiating factor like on any of your prompts that you can use.

Chain of Thought Reasoning

It's called chain of thought reasoning for those of you that are interested.

So I have top five recipes here And now I'm going to say now imagine you're an experienced storyteller.

Built me a presentation outline. That sells why we should open this restaurant. Include some of the recipes. And actually, I'm going to go one more thing.

Again, same thing here. Walk me through your thinking before giving me the outline.

Okay. So here it's now giving me a presentation because we're opening a restaurant. We have these 10 ingredients and they're going to be the determining factor of what's gonna make our restaurant.

So start with a hook, emotional connection and curiosity. We need to capture attention immediately, correct, define the problem, why this restaurant and why now, present the unique selling proposition. Correct. Why will it succeed?

These are all things that should be taken into account. Again, by using chain of thought reasoning in this way, if I don't agree with any of those points, rather than talking about why the result is wrong, I can just say, actually, you know what? It's not about the unique value proposition. I want you to think about something else. I can push back on the reasoning rather than on the result that came out.

but it's now creating a presentation. So spice and sweet, a culinary adventure, captivating, opening Hokukei. What if your next meal could surprise you in the best way possible? People want more from their dining experience. Introducing the concept, sweet, savory, and spiced.

Yeah, I get that. Actually, you know what? I might actually want to go here.

Bold dishes that redefine expectations. So what did we actually get to?

Tropically spiced pizza. Spanish tortilla with a pineapple twist. Spiced strawberry banana dessert wrap. Savory sweet pineapple salsa and banana strawberry spiced smoothie. I mean, that's... I would be interested enough to at least give it a go.

I'm going to take this. And I'm going to put it in.

Using Gamma for Presentations

Has anyone heard of gamma? Have you heard of gamma?

You might wait one minute. Okay, I'm going to hit create paste in text.

Continue the only thing I do here as I'm literally taking that presentation that was created by chance to be right. I can have a few configurations. I want to take it inspiration for that.

You can change the image model and stuff like that. I'm just going to hit continue. And, okay, I'll take the theme that it suggested.

Yeah, gamma. And I hit generate. Yeah.

A full slide deck about this restaurant that we are opening. Captivating flavors, a culinary adventure. The dining revolution begins.

It does the imagery. Look at this. Signature flavors, unforgettable combinations, oregano and spice, fruit infusion, tomato and tortilla, bold dishes beyond expectations for sure.

Look at that. It's actually doing the images to our dishes here, by the way, which is not bad. Those who wanted the, I mean, this is different to the Hawaii, but a tropically spiced pizza with pineapple, there we go. And tortilla crust.

A sensory experience beyond the plate, market-ready fusion cuisine, So it's actually hallucinated some numbers, to be clear. I mean, 30% estimated market expansion for bold flavor dining. They're high satisfaction with innovative menu items.

All of those are invented numbers. They don't actually exist. But the time it took me to get to an actual deck that starts pitching this? Pretty good.

Now, I'm going to go one step further.

Creating a Song with Suno

Now, imagine you are an experienced songwriter. Write me an indie rock song that pitches why we should open... I'm sorry. Open this restaurant.

How many of you have heard of Suno before? Only four. It's going to be a lot of fun as well.

So it's now writing me Sweet and Spiced, which is a song. I'm going to take... The other thing is when you use ChatGPT later in the day, it actually gets slower because the US has woken up.

The best time to use ChatGPT is in the morning because, well, we have the advantage that they're not actually awake. Do you use Claw to talk for this sort of thing? I have.

Is your preference ChatGPT or...? It depends on the thing. Cloud tends to be better at writing.

ChatGPT has a few features around it that just make it a little bit easier to use within fully defined practical workflow. So when it's about reasoning, ChatGPT tends to be better.

Also, I have the pro account, which is the $200 a month thing. And so you can switch between these different accounts and the pro model like honestly blows everything else out of the water. So whenever I need to go through, that kind of switches it for me.

The issue I have with this tool is like when I want to create a nice presentation, let's say I have to write this whole prompt on GPT, for example, or would it be a better idea to even build a custom chatbot or just a custom GPT as it is essentially made for gamma app? for the building step, the slides and the explanation? Absolutely.

Would you do that? So I do all of my presentation brainstorming in ChangeBT or Cloud. And then once I know what I want to be in the presentation, then I get one of those tools to outline, okay, give me a full outline of everything I want to present.

And only from that point do I go into Gamma. At the end, can you automate the whole process at the end? Well, that's more than what I just did in terms of the creation.

So what was the thing that you would want to automate more? Maybe if I can connect the gamma API key with make. Okay, I see what you mean.

Yeah, so I guess you could have the GPT write to like a Google Sheet and then you could pull the Google Sheet in. No, actually gamma is not part of make or Zapier, so you can't do that. There is no.

Yeah, no API for Gamma. Yeah, I looked at that. So now I'm going to go into Suno and I'm just copy pasting the lyrics that we get from Suno.

What was the name again from this thing? Sweet and Spiced. And I'm going to do Indie Rock.

Is it the free version? No, this is like $20 a month. Most of these are like $20 a month.

I was going to say, if you had to pay a musician to go and play this, I can guarantee you it'd be more than $20. That's the V3. That's the version 3.

This is version 4. We've been stuck in the same old scene. Burgers from

You know what? I am going to use my own audio because it's much stronger, actually. So I'll just use that.

Between the word and flavor broke the rules. Took a risk, rewrote the tools. Savory meets sweet.

A little heat, a little sweet. Close your eyes, take a bite. This is not bad!

So let's try the other one for a second. We'll see, cause this one's not too bad actually. But we'll see.

Totally different style. rules. Took a risk, rewrote the tools.

Savory meets the sweet. A little heat, a little treat. Close your eyes, take a bite, tell me this will feel just right.

Dirt, spice it up, sweeten down. Take a risk, take a ride. Let's serve up something wild inside.

Pineapple pizza, strong like weed. Switching the dreams, banana beads. Open the doors, light up the flame. This restaurant's gonna change the game, game.

Conclusion

So on that note, I will finish there, but hopefully it gives you an idea of just how much, I know we kind of, I tried to jump off of the theme from Steve as we kind of looked at recipes and stuff, but the, this has absolute practical application to basically, I think what most of you probably do on a fairly regular basis. Like I use gamma, Every second day for presentations, I use an AI assistant to brainstorm specific strategies, vision, preparation for a client.

I don't use Suno that much. Creating songs is not really part of my day-to-day workflow.

But how many of you have heard of Notebook LM? Notebook? Notebook LM. Notebook LM is another tool.

If I had more time, I would have demoed that as well, but basically that's more of an audio tool that you can use every day. It's an audio way of exploring any type of data that you want to throw at it. So you can throw thousands of pages at it, and then you can have an audio conversation with the data. It's literally, it creates a podcast out of whatever you throw at it, and then you can have a conversation with the podcast hosts, and it's the same.

None of these tools existed. So Gamma existed a year ago, Suno existed, but it was in version two, nowhere near where it's now. And the others didn't exist. So in the last two years. And it is just changing fundamentally how we think through things.

The agent bit. Actually, I do. You know what? I'm going to finish with one last thing.

Future of AI in Work Environments

Because I showed you the agent from Replit, which was just building the app. But what is important to understand is how that's going to go way beyond just coding. Generative AI ally. Let's see. Where is this? Here you go.

At Krafton, we are revolutionizing the battle royale genre with generative AI. Introducing PUBG Ally, the world's first co-playable character designed to play like a human teammate.

Hey Ally, I'm looking for a level 3 vest and 5.56 ammo. Keep your eyes peeled. Found a level 3 vest and ammo nearby. Pinged it for you. Enemy spotted. I'm covering. Is there a vehicle nearby? Found one. I will come to you. I'm ready to get into the action now. Where can you place them? On it, on the left. Wait, I'm knocked out. Crawling to you now. Just hang on. I couldn't have done it without you. I know. I carry for a living.

So I'll finish there. But that's just to show in terms of this was a game kind of thinking about, but this concept of a co-pilot, basically, if you understood what was going on, you had one player and an AI, and you're just giving instructions to the AI, and the AI goes and executes whatever that is. In this case, it was like, I need a vehicle. The AI goes out, figures out what the environment is, gets a vehicle, brings it to the individual.

Okay, now this is the next step. In a work environment, this is exactly what is happening at the moment and what we are all going to go through literally this year. Agents are the big thing this year. We will all get into this situation where you're just giving these various commands to AIs that are all executing in parallel on various things, and your job ends up being orchestrating whatever comes back and then trying to create the best result out of that based on whatever problem you have in front of you. And I really don't think people are ready for just how fast this is coming. So yeah, hopefully that was useful.