Quarter-Life Capital

Matt (00:01.073)
So yeah, so we're just talking about like the hype cycles with AI, right? And I think for young people and everyone alike, everyone's thinking about, is this actually the new revolutionary technology that's gonna change the world forever? Or is it just another hype cycle similar to dot com bubble? I actually read an article recently that talks about like somebody who has lived through the dot com bubble actually saw lot of similarities between the dot com bubble and the AI bubble. But Matt was just talking about like, there's AI has real utilities, whereas in dot com, maybe there's

George Pu (00:01.081)
So yeah, so we're just talking about like the hype cycles with AI, right? And I think for young people and everyone alike, everyone's thinking about, is this actually the new revolutionary technology that's going to change the world forever? Or is it just another hype cycle similar to dot com bubble? I actually read an article recently that talks about like somebody who has lived through the dot com bubble actually saw a lot of similarities between a dot com bubble and the AI bubble. But Matt was just talking about like, there's AI has real utilities, whereas in dot com, maybe there's

Soham Mehta (00:01.102)
So yeah, so we're just talking about the hype cycles with AI, right? I think for young people and everyone in life, everyone's thinking, okay, is this actually the new revolutionary technology that's gonna change the world forever? Or is it just another hype cycle similar to dot-com bubble? I actually read an article recently that talks about like somebody who has lived through the dot-com bubble actually saw similarities between a dot-com bubble and an AI bubble. But now it's just talking about like, but there's AI has some real utilities, whereas the dot-com

Maybe there's not that much utilities. I wanted to hear your thoughts. No, I, um, firstly, I do think, yeah, AI has so much utility. I think, uh, when we were first starting with artificial intelligence, like Chad GP three was out and then like subsequently 3.5 at that time, I was kind of like in this bubble phase because it was just like toy. It was just a toy essentially. I could just like say something back to you. There wasn't really logic behind it, things like that.

Matt (00:31.017)
not that much utilities. Just one of your thoughts. no, firstly, I do think AI has so much utility. I think when we were first starting with artificial intelligence when like, ChadGT 3 was out and like subsequently 3.5. At that time, I was kind of like in this bubble phase because it was just like toy. It was just a toy essentially. could just like say something back to you. There wasn't really logic behind it, things like that. Now we're seeing like, for example, even at my firm, we started implementing like a...

George Pu (00:31.035)
not that much utilities. I just want to hear your thoughts. Yeah, no, firstly, I do think AI has so much utility. I think when we were first starting with artificial intelligence, the Tragi P3 was out, subsequently 3.5, at that time, I was kind of like in this bubble phase, because it was just like toy, it was just a toy essentially, right? It could just like say something back to you. wasn't really logic behind it, things like that. Now I'm

Soham Mehta (00:55.924)
Now we're seeing, like, for example, even at my firm, we started implementing like our own version of a GPT. And just the level of AI adoption from our white collar companies and so much other has like really evolved. And that, guess, in terms of like apparel to a dot com bubble, I still think the dot com was like a very useful time because of the fact that if that much investment didn't go into the dot com, I don't know how many websites could have been created. think creating websites is one of them.

George Pu (00:56.793)
saying, for example, even at my friend, we started like our own version of a GPT. And just the level of AI adoption from white collar companies and so much other that's like really involved. And that I guess, like a parallel to the dot com bubble. I still think the dot com was like a very useful time because of the fact that that much investment then you go into the dot com.

Matt (01:01.963)
GPT and just the level of AI adoption from our white collar companies and so much other that's like really evolved and that I guess in terms of like apparel to the dot-com bubble I I still think the dot-com was like a very useful time because of the fact that that much investment in going to the dot-com I don't know how many websites could have been created I think creating websites is one of the most revolutionary uses of the internet

George Pu (01:22.393)
websites could have been created. think creating websites is one of the most revolutionary uses of the internet. And that kind of changed a lot of how we did things.

Soham Mehta (01:25.496)
most revolutionary uses of the internet. And that kind of changed a lot of how we did things. And one of the things I do think there will be a parallel of is the same way when websites first started, there was so many and people didn't really know how to go through and navigate between a lot. And then I just got like, centralized to like maybe like five or six websites. And now today we use more than anything else. I think there will be a point AI to where like it'll be two or three LLMs that become so much more ingrained.

Matt (01:29.103)
changed a lot of how we did things and one of the things I do think there will be a parallel love is the same way when websites first started there was so many and people

George Pu (01:32.313)
One of the things I do think there will be a parallel is the same way when websites first started, there was so many and people were really on it, both were navigating a lot. And then they started like, centralized to like maybe like five or six websites. And now today we use more than anything else. I think there are only one AI tool or like two or three algorithms that become so much more in-brain. I remember Simon Olken was talking about like for OpenAI, he hopes that children never have to like, they think of it as like actual software.

Matt (01:41.864)
They just got like, centralized to like maybe like five or six websites. And now today we use more than anything else. I think there will be a point AI to where like it'll be two or three LLMs that become so much more ingrained. I remember Sam Altman was talking about like for open AI, he hopes that children never have to like, they think of as like the actual like soft, not software operating system, right? And this idea that like it remembers everything about you since you were a little kid.

Soham Mehta (01:53.954)
I remember Sam Altman was talking about like for OpenAI, he hopes that children never have to like they think of as like an actual like software operating system, right? And this idea that like it remembers everything about you since you were a little kid. I think at that point, there actually starts to become like network effects as well for an LLM. And so I think that would be really interesting to see.

George Pu (02:02.267)
operating system, right? And you're not protecting a lot of things since you are a little bit.

Matt (02:08.581)
I think at that point there actually starts to like network effects as well for an LLM. So I think that would be really interesting. I personally do see the value of like AI obviously I think we talk about them on the pod quite often. Right. But then and it's in a lot of my daily workflows and also my company's workflows. But I also see there's a sign of bubble that you can not just ignore. Windsor just got bought for three billion dollars by open AI.

George Pu (02:08.599)
I think at that point, it actually becomes a network of likes as well for NLM. So I think that would be really interesting. I personally do see the value of AI, obviously. I think we talk about it the pod quite often. And it's in a lot of my daily workflows and also my company's workflows. But I also see there's a sign of bubble that you cannot just ignore. Windsurf just got bought for $3 billion. Let's buy OpenAI.

Soham Mehta (02:20.27)
Right. then, and it's in a lot of my daily workflows and also my company's workflows. But I also see there's a slight of a bubble that you cannot just ignore. Windsor is about $53 billion, it's bio-inhabited. And Johnny Ive, the former chief designer of Apple, which basically is like an iPhone. He had his, he had his start at a new thing, a new device that he partnered up with OpenAI a year ago. So the company didn't exist a year ago.

Matt (02:32.619)
And Johnny Ive, the former chief designer of Apple, which basically designed the iPhone, he had his startup had a new thing, a new device that he partnered up with OpenAI a year ago.

George Pu (02:32.659)
And Johnny Ive, the former chief designer of Apple, which basically designed the iPhone, he had his startup, he and his startup had a new thing, a new device that he partnered up with OpenAI a year ago. So the company didn't exist a year ago and now it just got bought for $6.8 billion by OpenAI. And none of us know what the hell is it. Like, is it going to be a flop or is it going to be like a real thing? And who's going to buy it for that much? Is it like a sign of bubble mat, think?

Matt (02:45.095)
So the company didn't exist a year ago and now it just got bought for $6.8 billion by OpenAI. And none of us know what the hell is it. Is it going to be a flop or is it going to be a real thing? And who's going to buy it for that much? Is it a sign of bubble map you think? But I feel like one of the things here is that when you look at the dot com companies, a lot of them, it was great for the internet and propagating the internet, but most internet websites weren't that useful. Like, okay, yeah, pets.com.

Soham Mehta (02:47.694)
And now it just got bought for $6.8 billion by OpenAI. And none of us know what the hell it is. Is it to be a blog or is it going to be a real thing? Who's going to buy it for that much? Is it like a kind of bubble mapping thing? But I feel like one of the things here is that when you look at the dot.com companies, like a lot of them, like, it was great for the internet and propagating the internet, but most internet, you know, websites weren't that useful. Like, okay, yeah, that's dot.com. Like, yeah, that's pretty cool.

George Pu (03:02.603)
is that when you look at dot-com companies, a lot of them, it was great for the internet and propagating the internet, but most internet websites weren't that useful. Like, okay, yeah, that's pretty cool. can pay online for the first time. Amazon started before the dot-com bubble and propagated through it. But even it was as useful way back then as it is now.

Matt (03:15.049)
yeah, that was pretty cool. You could pay online for the first time. Amazon, you know, started before the dot-com bubble and propagated through it. But even it wasn't like as useful way back then as it is now. Whereas like the ease of like AI becoming very useful is more about like how can you wrap it in a way that makes it useful, right? Like, I don't know, like myself, like having...

Soham Mehta (03:16.814)
pay online for the first time, Amazon, know, started before the dot-com bubble and propagated through it. But even if it wasn't as useful way back then as it is now, where it was like the ease of like AI is very useful is more about like how can you wrap it in a way that makes it useful, right? I don't know, myself like having done like a lot of programming lately, I've come to realize that...

George Pu (03:29.081)
Where it was like the ease of like AI becoming very useful is more about how could you wrap it in a way that makes it useful, right? Like, I don't know, like myself, like having...

Matt (03:40.91)
done like a lot of programming lately, I've come to realize that the majority of the world moving forward is going to be figuring out how to get the real world into the context of an LLM. Right? There's going to be two types of people in the future. Those that make LLM smarter and those that teach them the world. Right? Like, and, and so, those that are creating wrappers, it's like, the question is, is this a bubble or not? And it depends on whether, how far this propagates, right? Because if it continues to happen, like we might be at just at the beginning of wrappers

George Pu (03:40.983)
done a lot of programming lately, I've come to realize that the majority of the world moving forward is going to be figuring out how to get the real world into the context of an LLM. There's going to be two types of people in the future, those that make LLMs smarter and those that teach them the world. And so those that are creating wrappers, the question is, is this a bubble or not? And it depends on how far this propagates. Because if it continues to happen, we might be just at the beginning.

Soham Mehta (03:44.238)
The majority of the world moving forward is going to be figuring out how to get the real world into the context of an LLM. There's going to be two types of people in the future, those that make LLM smarter and those that teach them the world. And so those that are creating LLM, the question is, is this a bubble or not? And it depends on how hard this propagates. Because if it continues to happen, we might be at the beginning of wrappers.

George Pu (04:09.977)
of wrappers rapping different things for the world. Because let's be honest, if you look at any of the models that exist today, they could be way more big. Which means that we have not pulled out too much. AI is actually more capable than what it's actually being used for. Which makes me feel like, okay, is this the end of the bubble or the start of the bubble? Because I feel like there's a lot more growth to go from there.

Matt (04:11.033)
wrapping different things for the world because let's be honest like if you look at any of the models that exist today they could be used for way more things which means that we have not caught up to like AI is actually more capable than what it's actually being used for which makes me feel like okay you know is this the end of a bubble or the start of a bubble because I feel like there's a lot more growth to go from there I think personally like okay for internet there was there well you need to have an adoption curve

Soham Mehta (04:11.15)
wrapping different things for the world. Because let's be honest, if you look at any of the models that exist today, they could be used for way more things. Which means that we have not come out to where AI is actually more capable than what it's actually being used for. Which means, like, okay, you know, is this the end of the world or the start of the world? Because I feel like there's a lot more growth to go from there. I think for something like, okay, for internet, there was, well, you need to have an adoption curve.

George Pu (04:33.835)
I think personally, like, for internet, was, there, you need to have an adoption curve when you came from knowing nothing about the internet and getting a PC and actually using the internet, right? So there's that, there's such a learning curve. But when it comes to AI, I guess there's not really a learning curve. You just go in chatgpt.com and you can type in responses and you can get something back from it, right? We're not even touching the sort of like infrastructure.

Matt (04:40.593)
when you came from knowing nothing about the internet and getting a PC and actually using the internet, right? So there's such a learning curve. But when it comes to AI, I guess there's not really a learning curve. You just go in chat.com and you can type in responses and you can get something back from it, right? We're not even touching the sort of like infrastructure.

Soham Mehta (04:40.622)
when you came from knowing nothing about the internet and getting used to the internet, right? So there's that little question of learning code. But when it comes to AI, I there's not really learning code. You just go to chat.com and you can type in responses and you can get something back from it, right? We're not even touching this sort of infrastructure just yet, even like two and a half years after. So sort of agree with like, at pets.com, there's like no value compared to where we are today with the same chat you've been given.

Matt (04:58.567)
just yet, even like two and a half years after. So so do you agree like pets.com is like, there's like no value compared to where we are today with the say, chat GpT and cloud back in internet days? Yeah, like I think the way I would also look at it, I do think there was still value in pets.com, like any websites. It just like, it revolutionized the idea that like you can do something with like this computer or like laptops and things also weren't gonna like catch as much popularity.

George Pu (04:58.667)
just yet even like two and a half years after so like so do you agree like pets.com is like there's there's like no value compared to where we are today with let's say chat gpt and claude back in internet days yeah like i think the way i would also look at it i do think it's valid as like any websites it just like and revolutionized the idea that if you can do something with like this computer or like laptops and things also like catches much popularity

Soham Mehta (05:09.4)
back in the internet days? Yeah, like I think the way I would also look at it, I do think there was still value in fest.com or just like any websites. It just like, revolutionized the idea that like you can do something with like this computer or like laptops and things also won't kind of like catch as much popularity. It propagated like a lot, it's like the entire industry that became technology, right? So I think there always was value for like any companies that started there. And you always go through AP testing, right? Like,

Matt (05:26.247)
It propagated a lot of the entire industry that became technology. So I think there always was value for any companies that started there. And you always go through AP testing. With AI being created right now, again, what we were just saying, it was really weak before when it first came out, was popular, people were playing around with it. And yeah, I do agree.

George Pu (05:26.265)
It probably is a lot of the entire industry that became technology. So I think there always was value for any companies that started there. And you always look at the testing, right? With AI we created right now, again, what were you just saying? It was really weak before when it first came out as popular and people were playing around with it. And yeah, I do agree right now. I think it's very strong in the sense that you can actually use it for genuine tasks. And we haven't found...

Soham Mehta (05:37.23)
with AI being created right now, again, what we were just saying, it was really weak before when it first was like came out as popular people was like playing around with it. And yeah, I do agree right now. I think it's like very strong in the sense that you can actually use it for like genuine tasks. And we haven't found the right I guess, like form factor for it to become the most useful. It sounds like opening eyes really trying to do that. And it makes sense. I think right now because of the fact that the iPhone

Matt (05:48.551)
it's like very strong in the sense that you can actually use it for like genuine tasks and we haven't found the right I guess like form factor for it to become the most useful it sounds like opening eyes really trying to do that and it makes sense I think right now because of the fact that the iPhone or just phones in general this handheld device has been so long

George Pu (05:54.361)
the right form factor for it be found to be most useful. It sounds like OpenAI is really trying to do that. it makes sense, I think, right now because of the fact that the iPhone or just phones in general, this handle of the voice has been so long used by everyone, has become a daily part of our life. It makes sense that there should be that next evolution.

Soham Mehta (06:06.722)
general, this handheld device has been so long used by everyone, has become a daily part of our life. It makes sense that there should be that next evolution. And whatever that becomes, I would assume that AI has to be the integral part in that. There are, are like, there are there there are are are there there like, there like, like, there are like,

Matt (06:09.639)
used by everyone has become a daily part of our life. It makes sense that there should be that next evolution. And whatever that becomes, I would assume that AI has to be the integral part in that. are many wearable startups that have failed spectacularly, like the HumanePing and the others. Why do you think they're not successful compared to whatever Johnny?

George Pu (06:16.537)
and whatever that becomes, would assume that AI has to be the integral part of that. There are many wearable startups that have failed spectacularly, like the HumanePing and the others. Why do you think they're not successful compared to whatever Johnny Ive and Sam Olmans building? I have one I've building, so in that sense I'm not sure. But I think the point is, with all these other like...

Matt (06:31.375)
Sam Holman's building. I have no idea what they're building so in that sense I'm not sure but I think the point is with all these other like gags that come on they always feel gimmicky and I think even like the first phones were always gimmicky when they first started it takes time for like up until the App Store got created right like the iPhone didn't have as much like utility right I remember like I was listening to like a Steve Jobs like

Soham Mehta (06:32.872)
I want to have an audio with their building. in that sense, I'm not sure. I think the point is with all these other like gadgets that come up, they always feel gimmicky. And I think even like the first phones were always a gimmicky when they first started. It takes time for like up until the App Store got created, right? Like the iPhone didn't have as much like utility, right? I remember like I was listening to like a Steve Jobs like presentation on it. He was talking about the App Store being the radio. Like the way he like visualized this was incredible.

George Pu (06:40.727)
gadgets that come up, they always feel gimmicky. And I think even like the first phones were always gimmicky when they first started. It takes time for like, up until the app store got created, right? Like the iPhone didn't have as much like, utility, right? I remember like I was listening to like a Steve Jobs like presentation on there. He was talking about the apps for being the radio. Like the way he like visualized this was incredible. But if that app store didn't get created, all of that value into the iPhone wasn't gonna happen.

Matt (06:56.267)
presentation on that he was talking about the app store being the radio like the way he like visualizes was incredible But if that app store didn't get created all of that value into the iPhone wasn't gonna happen Well, I feel like with anything that's going to be used instead of a phone It has to have a particular function that's gonna be better than a an app on a phone, right? And we've just never had an AI tool that's reached that we've had other types of devices that are

Soham Mehta (07:02.414)
But if that app started and get created, all of that value into the iPhone wasn't going to happen. Well, feel like with anything that's going to be used instead of a phone, it has to have a particular function that's going to be better than an app on a phone. We've just never had an AI tool that's reached that. We've had other types of devices that are, you know, relevant to that. an example of this is in crypto is hardware. Right. The reason that it does not just in now because it's more secure.

George Pu (07:07.225)
Well, I feel like with anything that's going to be used instead of a phone, has to have a particular function that's going to be better than an app on a phone, right? And we've just never had an AI tool that's reached that. We've had other types of devices that are, you know, relevant to that. Like an example is crypto is hardware wallets, right? There's a reason that it's not just an app because it's more secure.

Matt (07:21.083)
you know, relevant to that. Like an example is in crypto is hardware wallets, right? There's a reason that it does, it's not just an app because it's more secure. And now we have hardware wallets now that are starting to have their own app stores. So then the question is like, what are they going to do with AI? That's going to be that much more useful that you can't just get, you know, with, with a phone app. And I, don't, I don't really know what that could possibly be. I think the most popular answer has always been like the glasses. Like everyone's like, where's these like the medical.

George Pu (07:29.431)
And now we have hardware wallets now that are starting to have their own app stores. So then the question is like, what are they going to do with AI that's going to be that much more useful that you can't just get, you know, with a phone app. I don't know what that could possibly be. think the most popular answer is always with your glasses. Like everyone is like, where are these like, the pedo glasses? Google glasses. Yeah. The fault with the Google glasses is that people walking down the street kept walking in a defense post. You know, because...

Soham Mehta (07:29.452)
And now we have hardware wallets now that are starting to have their own app stores. So then the question is like, what are they going to do with AI that's going to that that's more useful that you can't just get with a phone app? And I don't know what that could possibly be. I think the most popular answer has always been like the glasses. Like everyone's like, where's these like the meta glasses, anything like that. Google glasses. Yeah, the problem with the Google glasses is that people walking down the street, walking in the...

Matt (07:48.584)
Yeah, but the problem with the Google glasses is that people walking down the street kept walking into and defense posts and you know because the problem was is it was a an Interface on top of your vision instead of integrating into your vision, know, like instead of it being like right here on the table There is you know something that I'm looking at it's like it's like a screen in my face So I and I feel like that's so hard to do because our eyes react so quickly like AR is such a difficult thing to

Soham Mehta (07:54.798)
defense post. Because the problem was it was an interface on top of your vision instead of integrating into your vision. Instead of it being like, right here on the table there is something that I'm looking at. It's like, oh, there's a sprint in my face. So I feel like that's so hard to do because our eyes react so quickly. Like, AR is such a difficult thing to do. So the new glasses definitely aren't that different from Google Glass.

George Pu (07:57.721)
was it was an interface on top of your vision instead of integrating into your vision you know like instead of it being like oh right here on the table there is you know something that looking at it's like oh it's like a screen in my face

So I feel like that's so hard to do because our eyes react so quickly. Like AR is such a difficult thing to do. So the new glasses though, are they like, are they that different from Google Glass? Or are they just using a new approach? Like the Matic glasses, the Ray button ones are pretty successful I heard. I would imagine they're like, I don't know. I haven't looked at them. I mean, my big problem is like always like for me, like I have a very strong prescription. So if I'm going to get those and it needs to work with my existing vision and most of the time it doesn't.

Matt (08:18.537)
So the new classes that are they like are they that different from Google class or they're just using a new approach like the meta glasses the Ray button ones are pretty successful

Soham Mehta (08:22.478)
Are they just using a different approach, like a matter glasses, the three buttons or three successful and creative I would imagine, I don't know. I mean, my big problem is like always that for me, like I have a very strong prescription. So if I'm going to get those, to get a boost of five dozen. And I feel like that's also a big problem for all these, you know, glass creations. Like 50 people wear glasses, you know, and have something like a prescription that's not going to...

Matt (08:28.199)
I would imagine though, I don't know to be honest. I haven't looked at them. I mean, my big problem is like always like for me, like I have a very strong prescription. So if I'm going to get those, then it needs to work with my existing vision. And most of the time it doesn't. And I feel like that's also a big problem for all these, you know, glass creation, like 50 % of people wear glasses, you know, and have some type of prescription that's not going to be quite proper. Whereas a phone is so easy, like, cause everyone can look at it at a distance, right?

George Pu (08:42.683)
feel like that's also a big problem for all, you know, glass creation. Like 50 % of people wear glasses, you know, and have some type of prescription that's not going to be quite proper, whereas a phone is so easy, like, is everywhere. Right? Yeah. And I think like for AI...

Soham Mehta (08:51.202)
be

Matt (08:58.203)
AI I think the adoption I think we've been debating this like the society's been doing for the past few years But I really think what it comes down to is like coding and encoding has really had a product market fit for the market I'm not too sure if like writing your emails for you and all that stuff. It's like that much of a product differentiation Whereas coding I think has really taken off Well, I feel like even even for like writing emails like what I've noticed myself doing is like building up context inside of chat-tubety or Gemini and like

George Pu (08:59.019)
I think the adoption, I think we've been debating this, like the society has been debating for the past few years. But I really think what it comes down to is like coding. I think coding has really hit a market fit for the market. I'm not too sure if like rendering emails for you and all that stuff, it's like that much of a product differentiation. Whereas coding, think has really taken off. Well, feel like even like writing emails, like what I've noticed myself doing is like open up context inside of catch me tea or Gemini and like...

Soham Mehta (09:20.974)
What I noticed was like building up context inside of Chechiti or Japanite. like now it knows what kind of context it is. And to answer this, I generally want to talk about these things. I'm not sure how important is to use national apps to a context around these things. I feel like that's the same thing.

Matt (09:27.383)
Now it knows about my contacts and knows about the people that I talk to and like, Hey, like bring up this context from here and answer this. and here's this email. And I generally want to talk about these things, but makes, but, but I don't, not sure how to word it, but word it in my tone and don't use dash dash. it doesn't seem like AI, right? So people are building like context around these things. I feel like that's the same thing people are doing with coding. And what I've actually realized is that, having things that are readable by an LLM has become so much more important.

George Pu (09:27.467)
Now it knows about my contacts and knows about the people that I talk to and like, hey, like bring up this contacts from here. this, and here's the I barely want to talk about these things, but I'm not sure how to word it.

But word it in my tone and don't use dash dash so it doesn't seem like it. So people are building that context around these things. I feel like that's the same thing for coding. what I've actually realized is that having things that are readable by an LLM has become so much more important. The other day we were working on user flows in an app on Figma. And I literally said to my business partner, was like, hey, we need to get in perfect, which is like a format that...

Soham Mehta (09:51.854)
Realize is that having things that are readable by the way, it becomes so metaphorical. The other day we were working on our computer for us in an app, Sigma. And I literally said to my first partner, was like, hey, my reading is getting working. Or that, you know, is just like a text file. Yeah, so it's like a free software, right? So that's it's just like, like, flows by having food for, so that it's easier for me to do that.

Matt (09:57.21)
The other day we were working on like user flows in an app on Figma and I literally said to my business partner I was like, like we need to get this in mermaid which is like a format that You know is it's just like a text file Yeah, I think that creates the flow chart, right? So have have what exists in Figma in the flows but have it in code form

George Pu (10:08.185)
Um, you know, it's just like a text file Yeah. So have, have one existing thing about in the flows, but have it in code form, um, so that it's easier for AI to read that because it can go in instead of you having a screenshot, it can read that. then you convert that into flows and then can look at your existing code and then it can merge the two. Right. So having everything that can do that format that is also available in code format is now become more important than ever.

Matt (10:17.927)
So that it's easier for AI to read that because it can go and instead of you having to screenshot it it can read that and then it can convert that into flows and then it can look at your existing code and then it can merge the two right so having everything that you do in a format that is also available in code format is now become more important than ever

Soham Mehta (10:35.206)
So you're saying that like the AI is kind of like the nucleus in the situation that's like the center over like everything when your terms like however you're coding whatever project you're doing

George Pu (10:42.073)
how really important the logic is. I think Figma has done something that's really, I think it's a little bit like a glimpse of what's going to happen in the future. So Figma definitely has this like dev mode, right? But dev mode is for humans, not for AI agents. So it basically built this like new tool called Tunnel or something. But basically you open up a local host, which is like a local server on your computer that will show exactly the design details and code to your

Matt (10:44.263)
I think Figma has done, so I was actually doing the same thing with Figma. I think Figma has done something that's really, I think it's a little bit like a glimpse of what's going to happen in future. So Figma definitely has this like dev mode, right? But dev mode is for humans, it's not for AI agents. So it basically built this like new tool called Tunnel or something. But basically you open up a local host, which is like a local server on your computer that will show exactly the design details in code to your AI agent. So you actually launch that.

Soham Mehta (10:50.946)
So Vigma definitely has this death mode, right? But death mode is for humans, not for AI agents. So I basically built this new tool called Tunnel or something, but basically you open up a local host, which is like a local server on your computer that will show exactly the design details and code to your AI agent. So you actually launch that tunnel and then you can give this Vigma design to your AI

George Pu (11:12.027)
AI agent. So you actually launch that tunnel and then you can give feed this Figma design to your AI agent. Oh my god, that's amazing. I'm going have use that George. So that's a new thing I see the theme.

Matt (11:14.217)
tunnel.

And then you can give feed this big one design to your AI agent. Oh my God. That's amazing. I'm going to have to use that George. So that's a new thing. I can see the theme like it's slowly transitioning. So the future may be more about, okay, how can you build your app that's like readable by AI agents? Right. And I think before I was like, okay, it's not quantifiable. I can't see myself like what is this? How does it mean? And now I really saw it's like, okay, AI agents can actually do a lot of shit. Yeah. It's like basically like everything that you experience on your computer should have the ability to basically right click and inspect elements no matter what the

Soham Mehta (11:19.054)
Oh my god, that's amazing. I'm gonna have to use that for a week. That's a new event. I can see that thing like it's slowly transitioning to the future. be more about, okay, how can you build your app that's readable by AX, right? And I think before I was like, okay, it's not fun. I can't see myself. What does this account mean? And now I read this twice. Okay, AX is a manual. Oh, shit. Yeah. It's like basically like how anyone with experience on a computer should have the ability to basically right click and inspect elements no matter what the application is.

George Pu (11:24.313)
gets slowly transitioning. So the future may be more about, okay, how can you build your app that's readable by AI agents? Right? And I think before I was like, okay, it's not quantifiable. I can't see myself. What does this help? does it mean? And now I really saw it. like, okay, AI agents can actually do a lot of shit. Yeah. It's like basically like everything that you experience on your computer should have the ability to write with and inspect elements no matter what the application is. that you can navigate through everything. That's the other thing I've noticed as well is like for certain interfaces, like say if I'm using Google Cloud,

Matt (11:45.174)
application is so that you can navigate through everything and that's the other thing I've noticed as well is like for certain interfaces like say if I'm using like Google Cloud or AWS right and me having to go and navigate through okay I want to set up this you know this virtual machine or I want to set up you know this this other thing I have to go through all the different permissions it's way faster for me to just ask the AI hey I want you to do this thing it runs writes the CLI commands and then just runs them right so me doing any

Soham Mehta (11:46.958)
navigate through everything. that's the other thing I noticed as well is like for certain interfaces, like say if I'm using my Google Cloud or AWS, right, and we having to go and navigate through, okay, I want to set up this, you know, this virtual machine, or I want set up, you know, this other thing, I have to go through all the different permissions to do that. for me, it just asks the AI, yeah, I want you to do this thing. It runs the CLI commands and then just runs them. Right? So we do it.

George Pu (11:54.219)
cloud or AWS, right. And we have to go and navigate through, um, okay, I want to set up this, you know, this virtual machine or want to set up, you know, this, this other thing. I have to go through all the different permissions. It's way faster for me to just ask the AI, Hey, I want you to do this thing. It runs right to the CLI command and then just runs them. Right. So we can do it in the past instead of me having to get familiar with an interface. I'm just talking to a bot that's going to write the commands in the background that's going to do it for me.

Matt (12:14.857)
instead of me having to get familiar with an interface, I'm just talking to a bot that's going to run the commands in the background that's going to do it for me. So everything needs to be LLM accessible. And if it is, you're like 10 times faster. Yeah. So I think, I think I just saw like the question is like, okay, are

Soham Mehta (12:14.99)
Instead of me having to get familiar with an interface, I'm just talking to a bot that's going to run the commands in the background that's going to do it for me. So everything needs to be LLM accessible. And if it is, you're like 10 times faster. Yeah. So I think, I think I have like the question is like, okay, are we in type cycle? Because like, I'll be in type cycle even though like Matt and I were saying, okay, they're in those use cases. Then we as, I think specific for a code perspective, we find it really useful.

George Pu (12:23.213)
So everything needs to be LLM accessible. And if it is, you're like 10 times faster.

Yeah, so I think I guess the question is like, okay, are we in this hype cycle? Because like, are we in a hype cycle? Even though like, Mad and I were saying, okay, there are those use cases that we as and I think, I think specific for a code perspective, we find it really useful, right? For like email summaries and stuff, I personally don't feel like it's game changing for me. Document summaries, not game changing. Coding, yes. That's the one thing that's like game changing. So do you think like we're in a hype cycle per se? I don't think it's a hype cycle, sure. I have heard people all day.

Matt (12:33.095)
in this hype cycle because like I read hype cycle even though like Matt and I were saying, okay, there are those use cases that we as, and I think specific for a code perspective, we find it really useful, right? For like email summary and stuff, I personally don't feel like it's game changing for me. Document summaries, not game changing. Coding, yes, that's the one thing that's like game changing. So do you think like we're in a hype cycle per se? I don't think it's a hype cycle for sure. I have heard like a couple of people like talk about like

Soham Mehta (12:44.234)
For like email summaries, I personally don't feel like it's game-changing for me. Docular summaries, not game-changing. Coding, yes. That's the one thing that's like game-changing. So do you think we're in a hype cycle, not per se? I don't think it's a hype cycle for sure. I have heard like a people like talk about like a doom cycle in the sense of like existing coders that think that's Armageddon for them in a sense, which I also don't agree with. I think just the amount that AI can help grow like an entire economy is going to be revolutionary in that sense.

George Pu (13:00.505)
I do think on the sense of like existing coders, I think that's on my getting for them in a sense, which I also don't agree with. think the amount that AI can help incorporate into our economy is going to be revolutionary in that sense. And I just don't think there can be a hype cycle in this sense because I remember reading this idea that like.

Matt (13:00.825)
cycle in the sense of like existing coders something that's Armageddon for them in a sense which I also don't agree with I think just the amount that AI can help grow like the entire economy is gonna be revolutionary in that sense and I just don't think there can be a hype cycle in this sense because and I remember reading this it's this idea that like

Soham Mehta (13:14.762)
And I just don't think there can be a hype cycle in this sense because I remember reading this it's this idea that like Yeah, I will be able to notice things that we miss for so much right for example if somebody has like a PhD in like a very very very specific like physics it physics topic Yeah, I can take the context and put into biology and I actually make that bridge right and if you can build enough of these bridges throughout the world Which is people that are curious?

Matt (13:21.735)
Yeah, I will be able to notice things that we miss for so much right for example if somebody has like a patient like a very very very specific like Physicist physics topic yeah, I can take that context and put into biology and I actually make that bridge right and if you can build enough of these bridges throughout the world, which is people that are curious The amount that we can build in like 1050

George Pu (13:22.393)
Yeah, will be able to know things that we miss for so much, right? For example, if somebody has like a very, very, very specific like physics topic, yeah, I can take the context and put it into biology and I can actually make that bridge, right? And if you can build enough of these bridges throughout the world, which is people that are curious, the amount that we can build in like 10, 15 years, don't think we should ever wonder in that sense. True. But I think where AI gets us, AI gets us 90 % of the way there. Right? If you sell a...

Soham Mehta (13:41.102)
The amount that we can build in like 10, 15 years, I don't think we should ever underestimate humanity in that sense. But I think we're AI. This is AI that's 90 % of the way there. But if you tell AI, hey, build me an app that does XYZ, it will get you most of the way there. But the other 10%, like when it gets wrong, like, oh, this is slightly out of place, or this will make it weird, or people are going to not like what this flow looks like. Like humans are able to get the other 10

Matt (13:43.559)
I don't think we should ever underestimate humanity in that sense. But I think where AI gets us is AI gets us 90 % of the way there. If you tell AI, hey, build me an app that does XYZ, it will get you most of the way there. But the other 10 % that it gets wrong, like, oh, this is slightly out of place, or this wording is weird, or people are going to not like what this flow looks like. Humans are able to get the other 10 % there, which I think is perfect because basically every human is able to get to that 90 %

George Pu (13:53.618)
It will get you most of the way there, but the other 10 % like that it gets wrong like oh, this is out of place, wording is weird, or we're to like not like, you know, what this flow looks like, right? Humans are able to get the other 10 % there.

Soham Mehta (14:08.064)
Yeah

George Pu (14:08.761)
I think it's perfect because basically every human is able to get to that 90 % and then it's 10 % that's going to be important. So any task that's being done in the world can be done by 10x faster, you still need humans to do that last 10%. And I just feel like easiest, people are saying it's Armageddon for software engineers because at this very moment, people aren't hiring as much because there hasn't been any new companies that have come to fruition that are building up any more things that have adapted to...

Matt (14:13.513)
then it's the 10 % that's going to be important. So any task that's being done in the world can be done like 10x faster, but you still need humans to do that last 10%. And I just feel like the reason it's like people are saying it's Armageddon for software engineers is because at this very moment, like people aren't hiring as much because there haven't been the new companies that have come to fruition that are building that many more things than have adapted to like, you only need to do that last 10%.

George Pu (14:38.073)
Like, oh, you only need to do that last 10%. I think a thing I fear is a lot of human, including myself, become overly reliant on AI. Right? And I think that's big issue moving forward. If you're relying on AI to do your tasks, then are you really any more good at doing what you're doing? Because I see something we talked about before, Matt. It's like the harder before is the easier.

Soham Mehta (14:38.114)
Like, you only need to that last step. I think the thing I fear is like a lot of like, I've human, including myself, become overly reliant on the action. Right? And I think that's what that's a big point forward, if you're relying on the action to your tasks, then like, are you really any more good at doing what you're doing? Because I see something that we talked about before, man, is like, the harder before is the easier. Right? And I'm just wondering, like, okay, if I'm overly reliant on AI.

Matt (14:43.079)
I I fear is like a lot of like human, including myself, become overly reliant on AI. Right? And I think that's a big issue moving forward. If you're relying on AI to do your tasks, then like, are you really any more good at of doing what you're doing? Because I see, I see something like we talked about before, man. It's like the harder before is the new easy. Right? And I'm just wondering like, okay, if I'm overly reliant on AI and us as a society overly reliant on AI, we're going to let code that are bad.

George Pu (15:01.889)
Right. And I'm just wondering, like, okay, if I'm overly rely on AI and also society over the right, we're going to let code that are bad ship through and they're going to be new vulnerabilities. you know, but I feel like it depends on how you're doing it. Right. Like you, like before this, you showed me, was it Claude code AI Claude code, Claude code, which runs in your terminal. You spin it up.

Soham Mehta (15:06.19)
analysis society over here on AI. We're going to let the code that are bad ship through and there are going new vulnerabilities. But I feel like it depends on how you're doing it. Before this you just showed me a Claude code AI. Claude code. Claude code which runs in your terminal, you spin it up, you're like hey I want you to do x, y, tasks and it just goes through and it doesn't. Sometimes it shows the different

Matt (15:10.339)
ship through and there are going be new vulnerabilities. But I feel like it depends on how you're doing it. Like before this you just showed me...

Was it Claude code AI? that what it's called? Claude code Claude Claude code which like runs in your terminal. You spin it up and you're like, Hey, I want you to do X, Y, Z task. And it just goes through and it does it. But, and it, and and it shows some like sometimes it shows the diffs of the different files, but you're not able to, um, as easily, you know, go through and see where those files are to see where the diffs are. Right. Whereas it, say if you're using cursor and you're using Claude or you're using chat GPT or using Gemini. Well, it's going to show you in exactly.

George Pu (15:27.319)
And you're like, Hey, want you to do X, Y, tasks. And it just goes through and it does it. But it shows some like sometimes it the files, but you're not able to as easily, you know, go through and see where those files are, to see where the diffs are. Right. Whereas you're using quad or you're

Soham Mehta (15:35.918)
you're not able to as easily, you know, go through and see where those files are to see where the dips are. Whereas, say if you're using Cursor, you're using Claude, or you're using FragiKati, or you're using Gemini, well, it's going to show you in a task with what file something was happening in, and you're going to be able to click in and see how it was thinking for the last 10 seconds. And so what I found as a person, I could even learn, you my open-source base, you know, five or 10 times as fast.

George Pu (15:46.713)
ChatGPT are using Gemini. Well, it's going to show you in exactly what file something was happening in, and you're going to be able to click in and see how it was thinking for the last 10 seconds. And so what I found as a cook for me, I've been able to learn even on my own code base, you know, five or 10 times as fast and actually get context. But if you're someone who's just coming in and saying like, Oh, you're let the AI do all of it and I'm not going to learn, then you're going to fall behind.

Matt (15:49.817)
what file something was happening in and you're going to be able to click in and see how it was thinking for the last 10 seconds. And so what I've found has occurred for me is I've been able to learn even about my own code base, you know, five or 10 times as fast and actually get context. But if you're someone who's just coming in and saying like, oh, I'm just going to let the AI do all of it and I'm not going to learn, then you're going to fall behind. I feel like the key thing is to become extremely proficient at the last 10%.

Soham Mehta (16:04.878)
and actually get context. But if you're someone who's just coming in and saying like, I'm just gonna work with the AI, I do all of it, and I'm not gonna learn, then you're gonna fall behind. I feel like the thing is to become extremely proficient at the last 10%. If you're extremely proficient at the last 10%, and while the other 90 % is deep within, and you're learning faster on that, then you're gonna become like a tech developer in no time. And what I've noticed as well is that the best developers,

George Pu (16:14.801)
I feel like the key thing is become extremely proficient at the last 10%. If you're extremely proficient at the last 10 % and while the other 90 % is being written and you're learning faster on that, then you're going to become a like a 10x developer in no time. And what I've noticed as well is that the best of, you know, the best developers that were around five years ago, also the ones that are using AI the most proficiently now, which, you know, it's just kind of ironic. Yeah. think we also have to like understand

Matt (16:19.72)
you're extremely proficient at the last 10 % and while the other 90 % is being written and you're learning faster on that, then you're going to become like a 10x developer in no time. And what I've noticed as well is like the best of, you know, the best developers that were around five years ago are also the ones that are using AI the most proficiently now, which, you know, it's just kind of ironic. Yeah. And I think we also have to like understand that like humans are all these

Soham Mehta (16:33.646)
that were around five years ago, or also the ones that are using AI the most proficiently now, which, you know, it's just kind of ironic. Yeah. And I think we also have to like understand that humans are always going to be like the lazy people and like the more ambitious people. I think that you're just changing your definition to become a delegator. You're delegating like a large portion of this to an incredibly smart assistant. That's all this is, right? If you hire a new employee that's incredibly talented, you still should be reviewing their work. should.

George Pu (16:44.635)
humans are all these are all these people, more vicious people, I think they're just changing their definition of a delegator. You're delegating a large portion of this to an incredibly smart assistant. That's all this is, right? If you hire an employee that's incredibly talented, you still should be reviewing their work pretty sure. But if you have this specific task that they can do, like no problems, then yeah, that would be kind of like a population. Yeah, for a non-technical work, like so how, where do you see this is going? Because Matt obviously talked about the technical side of things.

Matt (16:48.272)
But then you're just changing your definition to become a delegator. You're delegating like a large portion of this to an incredibly smart assistant. That's all this is, right? If you hire a new employee that's incredibly talented, you still should be reviewing their work, you should. But once you find out this specific task that they can do, like no problems, then yeah, you can kind of let them off the leash, things like For non-technical work, like so where do you see this going? Because Matt obviously talked about the technical side of things and like how...

Soham Mehta (17:03.234)
But once you find out that this specific task that they can do, like no problems, then yeah, that you can kind of like let them off the leash, things like that.

George Pu (17:14.555)
and how a certain number of people can do this a lot better. For the non-technical stuff, guess sometimes people do argue, for example, accounting, or email writing, or contract drafting, or lawyers work. Those work the competitive, the moat of each individual, maybe less than coding. I don't know if that's fair argument. Where do you see people who are just relying on AI in other industries other than tech, or other professions? There's a lot of I view this. One is by coding.

Matt (17:15.682)
certain number of people can do this a lot better. For non-technical stuff, guess, like sometimes people do argue like for example, accounting, know, or

let's say email writing or a contract drafting or lawyers work. Those work the competitive, like the moat of each individual is maybe less than coding. I don't know if that's a fair argument, but where do you see people like who are just relying on AI in other industries other than tech or other professions? There's a lot of ways I would view it as in like two different lenses. One is the fact that think coding was like the easiest like MVP, like to be able to prove out that like AI can do a lot of things because coders were the ones that were making the AI, right?

Soham Mehta (17:32.91)
I don't know if that's a fair argument, but where do you see people like, or just rely on AI in other industries other than tech? Or other professions?

George Pu (17:52.372)
So like, for example, whatever regulations, right? So like right now you need to bridge an accounting with like, some natural standards and technicals of like, an LM score well. And so I think that is like,

Matt (17:52.615)
So like they kind of have contacts from both sides of things Like for example in accounting there's like very specific rules, right? If you're under IFRS, US GAAP or ASPI, Regulations you are. It's like right now you need to bridge an accountant with like so now she understands the technicals of like an LLM so well and so I think that is like I guess Still a little bit of a bridge away in terms like the most valuable stuff I've seen so far is again, we have internal AI

Soham Mehta (18:03.054)
regulations you are it's like right now you need to bridge an accountant with like some of that understands the technicals of like an LLM so well. And so I think that is like I guess, still a little bit of a bridge away. It was like the most valuable stuff I've seen so far is again, we have internal AI now at our company that you can feed it like these like 300 page documents, non scanned OCR. So I'm just talking about like, ones that are like with handwriting things like that.

George Pu (18:14.073)
still a bit of a bridge away in terms of like the most valuable stuff I've seen so far is we have internal AI now out of the company that you can feed it like it's like 300 page documents, non-scanable CR so I'm just talking about like ones like Android and things like that.

Matt (18:22.566)
you can feed it like these like 300 page documents, non-scanned OCR. So I'm just talking about like ones that are like with handwriting.

like that and can read through all of it and it can actually give you context tell you understand where it is and so for example if like I have to read a lot of contracts for my job right and the way you you try to find specific pages is by you control that for specific words that you think might hit now I can control that for context right and I want you to think about how much that changes everything and that same thing Matt when you're talking about like oh I can understand my code so much better I can understand my contracts so much better because I'm gonna search for context

Soham Mehta (18:32.238)
and can read through all of it and it can actually give you context until you understand where it is. so for example, if like I have to read a lot of contracts in my job, right? And the way you try to find specific pages is by you control F for specific words that you think might hit. Now I can control F for context, right? And I want you to think about how much that changes everything. And that same thing, Matt, when you're talking about like, oh, I can understand my code so much better. I can understand my contracts so much better.

George Pu (18:32.301)
You can read the wall and you can actually give you context to understand where it is. so for example, if like I have to read a lot of contracts for my job, right? The way you try to find specific pages is by you control F for specific words that you think might be. Now I can control F for context. Right. And I want you to think about how much that changes everything. That same thing, when you're telling me like, I can understand my code so much better. I can understand my contracts so much better. Cause I can just search for context.

Soham Mehta (19:00.302)
I can just search for context like hey is there a clause for a right of first refusal for example or whatever else there is like and this becomes even more valuable because when you're going into like these multinational companies that they maybe have a contract in Paraguay they have one in the United States completely different languages and so then the way the clause is framed is also going to be different the words that they use and things like that right and again like the only hard part about this is language right

Matt (19:01.49)
Hey, is there a clause for a right of first refusal for example or whatever else there is like and This becomes even more valuable because when you're going to like these multinational companies That they maybe have a contract in Paraguay. They have one in the United States completely different languages. Yeah, and so then The way the clauses framed is also gonna be different the words that they use and things like that, right? And again, like the only hard part about this is language, right and the L

George Pu (19:02.235)
a clause for a right of first refusal for example or whatever else there is like and this becomes even more valuable because when you're going into like these multinational companies that they need that are comparable you have one in the united states completely different languages and so then

The way the clauses range is also going be different. The words that they use and things like that, right? And again, like the only part for about this language, right? And the LLM can kind of the same way you're talking about, like if it can just access it, it's just data for it. All of this is the same to it. Wait, which company you guys using for like to be able to get 300 documents in? I don't think you can say, So, we're using all the elements we're saying, G4, and.

Soham Mehta (19:29.336)
And the LLM can kind of the same way you're talking about. Like if it can just access it, it's just data for it. All of us is same to it. Wait, which company you guys like to be able to get 300 documents in? I don't think you can say it. So we can switch between all the LLM sites like Llama, GPT-4, So was this an internal tool that was built within your company? okay. And so like the LLM that's being used is like, yeah. Fascinating.

Matt (19:30.503)
the same way you're talking about. If it can just access it, it's just data for it. All of it's the same to it. Wait, which company are you guys using to be able to get 300 documents into I don't think you can say. So we can switch between all the LLMs. There's Lama, GPT-4. So was this an internal tool that was built within your company? Okay, interesting. And so the LLM that's being used is yeah. Oh, fascinating. It's probably Seminite. No, because my friend...

George Pu (19:48.249)
So was this an internal tool that was built within your company? Oh, okay. Interesting. I love how it's being used as like, how fascinating. It's probably Gemini. No, because my friend, one of our friends from Waterloo is actually working on an AI startup that's basically doing the same thing, but for more, we can literally pass in like 300 documents at a time.

Soham Mehta (19:58.254)
No, because my friend, one of our friends we've worked with, he's actually working on an AI server, basically doing the same thing, but for other years, we can literally pass in like 300 documents at a time. He was talking to me, he put all the engineering challenges and all that, and he had to have like, this is where there's like nobody using our servers, and then suddenly we get 300 documents at once, but we're suddenly getting a thousand documents, and then they're doing it multiple times. I'm like, oh my God, yeah, that's like a really hard like, scaling problem.

Matt (19:59.495)
One of our friends from Waterloo is actually working on an AI startup that's basically doing the same thing. But for lawyers, where you can literally pass in like 300 documents at a time. And he was talking to me about all the engineering challenges involved with that. Hey, yeah, we'll have like instances where there's like nobody using our servers. And then suddenly we get 300 documents at once, or suddenly we get a thousand documents and then they're doing it multiple times. And I'm like, my God, like, yeah, that's like a really hard, like scaling problem. But meanwhile, you have to stay scaled up in

George Pu (20:09.529)
And he was talking to me to put all the engineering challenges involved with that. Like, Hey, yeah, we'll have instances where there's like nobody using our servers. then suddenly we get 300 documents at once or suddenly we get a thousand documents and then they're doing it multiple times. And I'm like, Oh my God, like, that's like a really hard scaling problem. But meanwhile, we have to stay scaled up in the meantime and be like, you know, paying that AWS or G cloud bill. It's like a, it's like a hard like engineering problem to even build.

Soham Mehta (20:27.726)
But meanwhile, you have to stick to the same art the entire time. you take that in your own US, it's a lot of bill. It's like a hard, like, engineering problem to even build the wrappers. Well, that's an interesting idea. I never thought of it from that side. Yeah. Speaking of wrappers, wrappers usually get like a bad rep. know, wrappers usually get a bad rep. OK, why don't you open a attitude thing that you do? Or tell them your business, like, yeah, I can't open it. I can explain your logic in less than a minute. Things like that.

Matt (20:29.498)
meantime and be like you know paying that AWS or G cloud bill it's like a it's like a hard like engineering problem to even build the wrappers you know that's interesting yeah speaking of wrappers I guess like wrappers you should get like a bad rap

George Pu (20:40.068)
Yeah, of wrappers, guess like wrappers you should get like a bad rep, know, man, like wrappers you should get a rep like, okay, I'm gonna ask him open the eye to the same thing you do or like tell me your business idea I can code it I can explain your logic and like less than a minute things like that. Like where do you see this wrappers phenomenon? Do you like actually believe in some of them? Well, I heard an interesting quote from he was either the CEO of Windsurf or CEO of Cursor who said

Matt (20:44.678)
You know, man, like wrappers, you should get a rep, like, okay, why doesn't open AI to the same thing you do? like, tell me your business idea. can code it. I can explain your logic and like less than a minute, things like that. Like, where do you see this wrappers phenomenon? Do you like actually believe in some of them? Well, I heard an interesting quote from, it was either the CEO of Windsurf or it was the CEO of Cursor who said the biggest challenge with AI, um,

Soham Mehta (20:57.422)
Well, I heard an interesting quote from the CEO of Windsurf or the CEO of Cursory who said the biggest challenge with AI is not feeding it to data, but getting it to focus on the right thing or focus on the right information. And I think that comes a lot from my book, and how you're doing prompt engineering.

George Pu (21:08.697)
The biggest challenge with AI is not feeding it the data, but getting it to focus on the right data or focus on the right information. And I think that comes a lot from the prompts and how you're doing prompt engineering and actually from a lot of trial and error of figuring out the best prompts. it's like, what is the way to a cursor at the end of the day? It's a wrapper that they continue to evolve and improve over time.

Matt (21:12.226)
is not feeding it to data, is not feeding it the data, but getting it to focus on the right data or focus on the right information. And, and I think that comes a lot from like the prompts and how you're doing prompt engineering. And that comes from a lot of trial and error of figuring out the best prompts. So it's like, what is windsurf or cursor at the end of the day? It's a wrapper that they've continually evolved and improved over time where the prompts come at just the right moment based on the models that they're using. And so it's almost like.

Soham Mehta (21:27.118)
And that comes from a lot of trial and error of figuring out the best prompts. it's like, you're user for a person at the end of the day. It's a wrapper that they continue to draw and look for for time where the prompts come at just the right moment based on the models that they're using. And so it's almost like a finely tuned wrapper that exists around it that like, okay, yeah, you could start from scratch and try to do that. And so like everyone says, oh, well, these wrappers are a problem.

George Pu (21:37.497)
the prompts come at just the right moment, take from the models that are using, and so it's almost like a finely tuned wrapper that exists around it that if you had to start to scratch and try to do that, it's going to take all that front on error that they've already done. And so everyone says, all these wrappers are a bubble, but there is work that, was prompt engineering work that was done to get to that point, and as new models come out, that's going have to be done again and again and again.

Matt (21:42.073)
like a finely tuned, you know, wrapper that exists around it that like, okay, yeah, you could start from scratch and try to do that. But it's going to be, it's going to take all that trial and error that they've already done. And so like everyone says, all these wrappers are a bubble, but you know, there is like work that there was prompt engineering work that was done to get to that point. And as new models come out, that's going to have to be done again and again and again.

Soham Mehta (21:56.974)
There is work that there's a lot of engineering work that was done to get to that point and as new models come out that's going have to be done again and again and again. And I feel like that's basically the hardest work that needs to be done from now on, which is like, okay, for lawyers, what are the optimal wrappers? For accountants, what is the optimal wrappers? I have the optimal wrappers, like there's the coders, right? We've got Claude Code, we've got Kurser, we've got Rinser. Now everyone else needs to catch up.

Matt (22:07.047)
And I feel like that's, that's basically the hardest work that needs to be done from now on, which is like, okay, for lawyers, what are the optimal wrappers for accountants? What is the optimal wrappers? We already kind of have the optimal wrap wrappers for coders, right? We've got Claude code, we've got cursor, we've got windsurf. Now everyone else needs to catch up. Um, what do you, what do you see this like? Because, uh, what Matt just mentioned, I think it makes sense too, but also like an argument could be made and said, okay, whatever wrapper you're building, isn't it easily replicable?

George Pu (22:07.097)
And I feel like that's basically the hardest work that needs to be done from now on, which is like, for lawyers, what are the optimal wrappers? For accountants, what is the optimal wrappers? We already kind of had the optimal wrappers for coders, right? We've got Claude Coden, we've got Kerser, we've got Winster. Now what else needs to catch up? Yeah. So what do you see this like? Because what Matt just mentioned, I think it makes sense too, but also like an outcome could be made. It's like, okay, whatever wrapper you're building, isn't it easily wrapped?

Soham Mehta (22:27.285)
Yeah, so what are your reasons? Because what Matt just mentioned I think makes sense too, but also like an opportunity could be based on... Isn't it easily applicable for us to just do something similar and just solve those problems and maintain it and just do something the same? I think people say that about almost anything and everything. I remember a couple years ago when my friends even said that about Bitcoin. It's like, oh, this is just like one simple protocol, right? Just work it. Yeah, exactly. Can I not just create my own version of it? I think the problem becomes...

George Pu (22:36.013)
couldn't like us just like do something similar and just solve those problems and maintain it and just do something the same?

Matt (22:37.141)
just like do something similar and just solve those problems and maintain it and just do something the same. I think people say that almost anything and everything. I remember a couple years ago when my friends even said that about Bitcoin. It's like this is like one simple protocol right? Just fork it. Yeah, exactly. Cannot create my own version of it. I think the problem becomes there's also a value in the story. There's a value

Soham Mehta (22:56.354)
There's also a value in the story. There's a value in just like how it looks, how much you get used to the look of it. For example, even with like chat GPT, like I, the was a Sora, the voice or whatever their default voice. got so used to that when I switched. Yeah. When I switched to Gemini, it actually kind of bugged me. I had to go back to charging just for the voice, right? Like, it was just a cleaner, like talking experience. so I think that's the point.

Matt (22:58.655)
like how it looks, how much you get used to the look of it. For example, even with like chat GPT, like the was a Sora the voice or whatever, their default voice. I got so used to that. When I switched to Gemini, it actually kind of bugged me. I had to go back to chat GPT just for the voice, right? Like it's just a cleaner like talking experience.

George Pu (23:05.049)
I to be like, the was a sort of whatever. I got so used to that. I think that's the point with wrappers is as humans you get like one certain way of doing things. Nobody really likes change. So even in the case where

Matt (23:23.971)
So I think that's the point with wrappers is as humans you get like used to like one certain way of doing things. Nobody really likes change. So even in the case where two products are completely substitutable, we still see like that one product that like people just like made the product market fit with being the one that wins. Like for example, think like Uber versus Lyft, like Uber is so much more prominent.

Soham Mehta (23:25.672)
with wrappers is as humans you get like used to like one certain way of doing things nobody really likes change so even in the case where two products are completely substitutable we still see like that one product that like people just like made the product market fit with being the one that wins like for example i think like uber versus lift like uber's so much more prominent than like lift i'm sure a big thing is that it's like and i also just don't like uber me personally

George Pu (23:34.627)
two products are completely substitutable. We still see that one product that people just made the art more fit with being the one that wins. For example, think Uber versus Lyft. Uber's so much more prominent than Lyft. I'm sure a big thing of that is like, I also just don't like Uber, me personally. I still hate the look of the Lyft app. It's so disgusting. Do you think we're gonna have loyalties in the future? I can imagine, you know.

Matt (23:49.849)
sure a big thing is that it's like I also just don't like personally I still hate the look of the lift app it's so disgusting. you think we're gonna have like loyalties in the future like I can imagine you know you know dad in the future disciplining his son like no son we're a chachabi tea family you know. I think there definitely has to be.

Soham Mehta (23:54.446)
I still hate the look of the Lyft app. It's so disgusting. Do you think we're going to have like royalties in the future? can imagine, you know, know, dad in the future discipline, like, no son will be charging me teeth at me, you know? I think there definitely has to be a I don't know in terms like you or your kid, but for example, the whole point is, if all of LLMs are trying to create a context that are decades long, whichever one you start with as a kid is going to be the one you have to kind of continue with.

George Pu (24:02.233)
you know, dad is usually disappointing his son. Like, no son, we're a Chachibut family, you know. I think there definitely has to be. I don't know if there's like a new or a different. But for example, the whole point is if all of these elements are trying to create a context that are decades long, whichever one you start as a kid is going to be the one you have to kind of think Totally true, yeah. Because like we were just saying, you took a lot for you to build context, right, from Chachibut. Like, oh, they're writing that voice, this and that. You just.

Matt (24:10.631)
in terms of you with your kid, but the whole point is if all of these elements are trying to create context that are decades long, whichever one you start with as a kid is gonna be the one you have to kind of continue.

Soham Mehta (24:23.526)
cause like what you were just saying is like, took a while for you to build context, right? For chat GP for like, like right in my voice, this and that you just can't, you just can't keep doing that. and starting with a brand new LLM. You're going to have that on a complete basis right there. Like where you have like, like this company that primarily uses Gemini and then like, we'll use it for certain things. then like, what I've noticed is like for specific things, like in our company, lot of the like documents, like legal documents or.

Matt (24:23.983)
Like what you were just saying is like you took a while for you to build context right for chat GP for like like right in my voice this and that You just can't you just can't keep doing that and starting with a brand You're gonna have that on a company basis I think like where you have like like this company like primarily uses Gemini Yeah, and then like only uses chat GPT for certain things and then like but what I've noticed is it's like for specific things like in our company a lot of the like documents like legal documents or context documents are all

George Pu (24:31.673)
You just can't keep doing that and starting a brand new online. You're going to have that on a company basis. Like where you have like, this company, like primarily you Gemini. Yeah. then like, only for certain things. then like, but what I've noticed is it's like for specific things, like in our company, lot of them, like documents, like legal documents or context documents are all with like, with a Gemini, but then everyone like personally use the chat GPT but then all, you know, for coding all with.

Soham Mehta (24:52.738)
context documents are all with Gemini, but then everyone personally uses ChatGpT, but then for coding it's all with Cloud. So it's almost like very specific elements of a user-specific use case and stuff. I mean, just use LinkedIn for more jobs, use Facebook more for my speed, right? I feel like different little niches will always get created with the companies.

Matt (24:53.917)
with like.

With Gemini, but then everyone like personally uses chat gpt but then all, you know, for coding it's all with Claude. So it's like almost like, like very like specific, like specific LLMs are used for specific use cases almost. I mean, just like you use LinkedIn for more jobs, use Facebook more. Right. Like I feel like, yeah, different little niches will always get graded with like the companies. I do think habits is like a huge thing. A few years ago I was talking to like a VC and they were talking about like, okay, we're just, we just have an investment in FinTech company that focuses on teams.

George Pu (25:01.059)
So it's like almost like very specific, like specific algorithms are used for specific use cases almost. I mean, just like you use LinkedIn for more jobs, use case for more investment. Different little each of those get created in the companies. I do think habits is like a huge thing. A few years ago I was talking to like a VC and they were talking about like, we're just, we just have any investment in FinTech company. The focus is on teens, bank accounts.

Soham Mehta (25:19.214)
And they were talking about like, okay, we're just the main investment in a company that focuses on teams, bank accounts, and doing like something out of the blue teams, because in the US there's no team-fabric bank accounts as standalone product. Their thing, their thesis, which I don't know today if it's true, proven to be true or not, was like, okay, if we can get a bunch of teams to use the product, they'll be using it as an adult also, and they'll grow up with this bank account, and they will be using this account more than their Chase or other bank

Matt (25:24.777)
accounts and doing like something valuable for teens because in the US there's no teens backing bank accounts and a standalone product their thing their thesis which I don't know today if it's proven to be true or not was like okay if we can get a bunch of teens to be using the product they'll be using it as adult also and they'll grow up with this bank account and they will be using this account more than their Chase or other bank accounts. Wasn't that the whole idea with like credit unions to get you in like I remember when I was like

George Pu (25:25.817)
and doing like something valuable for teens because in the US there's no teens backing bank accounts as a standalone product. Their thesis, which I don't know today if it's true proven to be true or not, was like, okay, if we can get a bunch of teens to be using the product, they'll be using it as an adult also and they will grow up with this bank account and they will be using this account more than their Chase or other bank accounts. Wasn't that the whole idea of like credit unions get you in? Like I remember when I was like...

Soham Mehta (25:48.174)
wasn't that the whole idea of credit unions to get you in? I remember when I was like 13 or 14, I got my first credit union account. I remember sitting there, the lady literally explained to me, hey, you're putting in your initial $5, basically, towards the credit union, and now you own a small portion of the credit union, I think. Oh, really? There's a whole spiel involved. I remember that, yeah.

Matt (25:52.808)
13 or 14, I got my first credit union account. I remember sitting there, the lady literally explained to me, hey, you're putting in your initial $5 basically towards the credit union, and now you own a small portion of the credit union type thing, right? And there was a whole spiel involved. I remember that, yeah. Back in the day. I don't know how prominent credit unions are in GTA or Toronto or big cities and stuff.

George Pu (25:52.857)
13 or 14, I got my first credit union account. I remember sitting there, the lady literally explained to me, hey, you're putting in your initial $5 basically towards the credit union, and now you own a small portion of the credit union type thing. they were like, there's a whole spiel in all that. I remember that, yeah. Back in the day. I don't know how prominent credit unions are in the

Soham Mehta (26:16.526)
Back in the day, I don't know how common credit unions are in GTA or Toronto, big cities and stuff, but I'm from a small town, people use credit unions. But the problem is, couldn't use that forever because, I mean, grew up with people using it, but I couldn't use it forever because that credit union doesn't exist in Toronto, for example. But that would be a problem for Coffee, because wouldn't want access it anywhere. True. If they were using the things for stablecoins, then...

George Pu (26:20.889)
GTA in Toronto, big cities and stuff. I'm from a small town, so people use credit unions. But the problem is, obviously I couldn't use that forever because, I mean grew up and continued using it, but I couldn't use it forever because that credit union doesn't exist in Toronto for example. Right, but that wouldn't be a problem for software because you can access it anywhere. True. If they used, if they supported stable coins then... Yeah, so I guess, yeah man, what are your thoughts about...

Matt (26:22.761)
But I'm from a small town so like you people use credit unions But the problem is like obviously I couldn't use that forever because I mean I grew up and continued using it But I couldn't use it forever because you know that credit union doesn't exist in Toronto for example, so True if they were used if they supported stable coins then

Soham Mehta (26:45.614)
So I guess, yeah, man, what are your thoughts about the future trajectory of AI? Because we've come two and a half years, I guess. And I always hear people say, I don't know what's going to happen next in this space. As someone who uses code quite a lot, I know your main area is not AI, but you use all the tools. Where do you see it go for the rest of the year? Well, I feel like the biggest, it's just going to keep improving for coders.

George Pu (26:49.561)
the future trajectory of AI because we've come two and a half years, I guess. And I always hear people say, like, I don't know what's going to happen next in the space, right? As someone who like uses it to code quite a lot, I know like your main area is not AI, but you use a lot of tools. Where do you see it go for like the rest of the year? Well, I feel like the biggest, like it's just going to keep improving for coders, right? But I see it expanding to more more like white collar jobs, like

Matt (26:52.691)
and half years, I guess. And I always hear people say, like, I don't know what's going to happen next in the space. As someone who uses it to code quite a lot, I know your main energy.

it's not AI, but you use all the tools. Where do you see it go for the rest of the year? Well, I feel like the biggest, like it's just gonna keep improving for coders, right? But I see it expanding to more and more white collar jobs, like it being introduced into law firms, it is being introduced for accounts, it being introduced for other white collar jobs where...

Soham Mehta (27:13.838)
But I see it expanding to more and more white collar jobs. Like it being introduced to the law firms, it being introduced for accountants, it being introduced for other white collar jobs where it just all it takes is somebody to create a wrapper that's really effective for a game to a of users and then suddenly it's good to go. And so I feel like that's the biggest thing that it's going to expand to. And so are we on the edge of like a bubble of the pot? Probably not.

George Pu (27:18.465)
it being introduced into law firms and it being introduced for accountants and being introduced for other white collar jobs where it just all it takes for someone to create a wrapper that's really effective for it, gained a bunch of users and then suddenly it's good to go. And so I feel like that's the biggest thing that it's going to expand to. And so are we on the edge of like a bubble about the problem? Probably not. Is there going to be a bubble at some point in the...

Matt (27:26.331)
You know, it's just all it takes is someone to create a wrapper that's really effective for it, gain a bunch of users, and then suddenly it's good to go. And so I feel like that's the biggest thing that it's going to expand to. And so are we on the edge of like a bubble about to pop? Like probably not. Is there going to be a bubble at some point in the, you know, AI trajectory? Probably. And maybe that happens in like five years, maybe in, you know, three to five years, we have some type of bubble that pops, but then

Soham Mehta (27:42.016)
Is there going to be a bubble at some point in the AI trajectory? Probably, and maybe that happens in like five years. in three to five years we have some type of bubble that pops. But then when it comes to looking 25 years in the future, it ends up not being a bubble at all because the bubble pops and then it just goes back to where it was, just like we had with the dot com. So I just read something that's called the ads revenue reaching like 10 billion.

George Pu (27:45.207)
you know, AI trajectory, probably, maybe that happens in like five years, maybe in, you know, three to five years, we have some type of bubble that pops, but then, you know, when it comes to looking 25 years in the future, it ends up not being a bubble at all because it, the bubble pops and then it just goes back to where it was just like we had with the dot com. So, So what do you see like, so I just read some news about opening as revenue reaching like 10 billion for the past year, which I think is 10B.

Matt (27:53.672)
you know, when it comes to looking 25 years in the future, it ends up not being a bubble at all because it, the bubble pops and then it just goes back to where it was just like we had with the.com. yeah. So what do you see? Like, so I just read some news about as revenue reaching like 10 billion for the past year, 10 billion, 10 B, which is super amazing. And I think before that was 2.7 billion or something like that. Holy moly. So it grew significantly, but the point is like, they're still not profitable. You know, the 10 billion.

Soham Mehta (28:10.35)

George Pu (28:12.761)
Wow. Holy moly. How are not profitable? What? are they doing? They're just melting the GPUs.

Matt (28:22.247)
How are they not profitable? Yes, they're not. They're profitable. Yes, that's my puzzle too. I don't know where they spend their money. They're just melting the GPUs. they're probably melting the GPUs. guess, Soham, where do you see this go? Do you think that's a forming of a bubble? Because it's such a popular tool. It's the best tool maybe in the history of the internet. And it's still not profitable even after all those years. Yeah. I think it's to the point where the same way way back in like...

Soham Mehta (28:40.302)
of internet and it's still not profitable even after all these years. Yeah, I think it's up to the point where the same way way back in like the mid 1990s it was like these computers where I remember Bill Gates said what do I need more than 10 kilobytes for on a computer right like I think we're at that kind of state with GPUs it's just like as advanced as they are I don't think they're advanced enough for where we're going in like 10-15 years I just think

George Pu (28:42.715)
still non-profit even after all those years.

Matt (28:52.113)
these computers where I remember Bill Gates said, what do I need more than 10 kilobytes for on a computer, right? Like I think we're at that kind of state of GPUs. It's just like as advanced as they are. I don't think they're advanced enough for where we're going in like 10, 15 years. I just think a breakthrough always happens where things like this, the cost just almost go to zero. Like when you have this much value that can, again, a company has worth 300 billion right now on the secondary markets, right?

George Pu (29:04.761)
where we're holding like 10, 15 years. I just think a break will always happens where things like this, the cost of the phone just go to zero. When you have this much of value, again, OpenAI is worth 300 billion right now on secondary markets, right? You have this much value and people, Nvidia is second most or third most valuable in the world. It cost us two years to complete GPUs.

Soham Mehta (29:08.04)
A breakthrough always happens where things like this cost is almost go to zero. Like when you have this much value that can again like open AI is worth 300 billion right now on the secondary markets, right? Like you have this much value and people are Nvidia is the second most or third most valuable company in the world in the past just two years because of the GPUs. When there's this much money in the market, there's just going to be like some kind of a breakthrough at a point. I think that's just like what the market's waiting for.

Matt (29:21.575)
You have this much value and people are Nvidia's second most or third most valuable in the world in the past two years because of the GPUs. When there's this much money in the market, there's just gonna be like some kind of breakthrough at a point. I think that's just like what the market's waiting for. And all these companies will become profitable. Like all of these, I think AI have to just come through.

George Pu (29:31.097)
When there's just no money in the market, there's just going to be some kind of a breakthrough at the point. And I think that's just what the market's waiting for. And all these companies will become profitable. All of these, I think from AI, will have to just come true. feels, a point when it's very odd, it smells like a duck, walks like a duck, it talks like a duck, it probably is a duck. I feel like we're in the timeline of AI, we're in the summer of 2012, in Bitcoin timeline.

Soham Mehta (29:37.87)
And all these companies will become profitable. Like all the thesis that I think in AI have to just come true. feels at a point when it's very obvious if it smells like a duck, if it walks like a duck, if it talks like a duck, it probably is a duck. I feel like we're in the timeline of AI, or like in summer 2012 in Bitcoin timeline, where in like a year or two years, they're going to figure out how to create specialized mining.

Matt (29:44.432)
feels at a point when it's very obvious if it smells like a duck it walks like a duck if it talks like a duck it probably is a duck. I feel like we're in in the timeline of AI we're like in summer of 2012 in the Bitcoin timeline where in like a year or two years they're gonna figure out how to create specialized mining

George Pu (30:00.409)
where in like a year or two years they're going to figure out how to create specialized mining hardware, ASICs. Right? why is it open? Like is there, is there an optimization that can be done on the hardware level for LLMs that are way more efficient than the GPUs? I feel like probably, right? the whole thing to make sure I'm not being optimized around the shock 256 hashing algorithm. That's why ASICs, if I came to fruition.

Matt (30:08.111)
Hardware, ASICs, right? why has no one, like is there an optimization that can be done on the hardware level for LLMs that are way more efficient than GPUs? I feel like probably, right? Like the whole thing with Bitcoin was they needed to optimize around the shock 256 hashing algorithm. And that's why ASICs, like came to fruition.

Soham Mehta (30:08.846)
Asics, right? what life is there? Is there an optimization that can be done on the hardware level? for LLM that are way more efficient than GPUs I feel like probably right like the whole thing with Bitcoin was they need to optimize around the CAC 256 patching algorithm and that's why Asics came to fruition and so what is the things that are going on in the background with LLM and can specialized hardware be developed?

Matt (30:30.327)
And so what are the, what is the things that are going on in the background with LLMs and can specialized hardware be developed that is way faster and way more efficient at that processing. And then that ends up making, you know, open AI incredibly profitable. And one thing I will say is that we don't talk about enough is, mean, China proved with DeepSeek that they can kind of do things under constraint and the amount of restrictions the USA has in terms of GPUs to China.

George Pu (30:30.329)
And so what is the things that are going on in the background with LLMs and can specialized hardware be developed that is way faster and way more efficient at that processing? And then that ends up making, you know, open AI incredibly profitable. And one thing I will say is that I feel like we don't talk about enough is, mean, China holds a deep seat that they can kind of do things under constraint and the amount of restrictions the USA has towards GPUs to China.

Soham Mehta (30:37.484)
that's way faster and way more efficient at processing. And then that ends up making, you know, opening eyes. And one thing I will say is that I feel like we don't talk about enough is, mean, China proved a deep seek that they can kind of do things under constraint and the amount of restrictions the USA has in terms of GPUs to China. I mean, like there's just like such an incentive even just for them to be able to create something brand new and like how they.

Matt (31:00.391)
I mean, there's just such an incentive even just for them to be able to create something brand new.

George Pu (31:00.523)
I mean, there's this smushing center even just for them to be able to create something brand new.

Matt (31:07.409)
take on this problem. Yeah, I think that's true. So I feel like we're just going to wait and see how the space goes. do feel more confident about coding. Like for someone that's like not always coding, think I've become more confident at my levels of coding. But I think the problem is like I sometimes over rely on AI to solve a few things that I'm not really that interested at. And I think that's potentially a problem. Like my back end developer always tells me that he is skeptical of using a lot of AI in back end development.

Soham Mehta (31:07.448)
take on this problem.

George Pu (31:07.459)
Take on small.

Yeah, I think that's true. So I feel like we're just going to wait and see how the space goes. do feel more confident about coding. Like for someone that's like not always coding, think I've become more confident at my levels of coding. But I think the problem with me is like I sometimes over-reliant on the AI to solve a few things that I'm not really that interested at. And I think that's potentially a problem. Like my back-end developer always tells me that he is skeptical of using a lot of AI in back-end development because of the error rate.

Soham Mehta (31:36.846)
because the error rate as a group produces is actually quite high and his point was like when you're designing the logic it's super important to make sure it's a really solid foundational logic when you're designing a backend system whereas sometimes if you rely too much on AI it's going have a lot of flaws and at first I didn't believe it but then I grew up being more accustomed to that idea actually I hope you get my point. Well I feel like the bigger problem is like just giving it enough context for it to go like

Matt (31:36.866)
because of the error rate that it's to produce is actually quite high.

George Pu (31:38.571)
is able to produce is actually quite high. And his point was like, when you're designing a logic, it's super important to...

Matt (31:40.44)
And his point was like when you're designing the logic, it's super important to make sure it's really solid foundational logic when designing a backend system. Whereas sometimes if you rely too much on AI, it's going to have a lot of flaws. And at first I didn't believe it, then I grew, became more accustomed to that idea. I'm like, okay, you're right. Yeah. Well, I feel like the bigger problem is like just giving it enough context from the get-go. Like I agree completely with your backend developer if you're not

George Pu (31:46.091)
make sure it's a really solid foundational logic when designing a backend system. Whereas sometimes if you rely too much on AI, it's going to have a lot of flaws. And at first I didn't believe it, but then I grew, became more accustomed to that idea. I'm okay. right. Yeah. Well, I feel like the bigger problem is like just giving it enough context from the get go. Like I agree completely with your backend developer. If you're not providing enough context on it, on the way you want it to be developed, right? Because as a developer, have certain preferences.

Soham Mehta (32:05.858)
I agree with your back end developer if you're not providing enough context on the way you want to develop, right? Because as a developer, you have certain preferences. So, you know, the AI is just going to go off on a tangent based on what it thinks would be optimal. But if you're able to design the system ahead of time and figure out like, hey, I want this, you know, to be engineered in this manner, with this architecture, because of this, then it's going to be much more effective at putting all those different parts together. And also,

Matt (32:10.313)
providing enough context on it on the way you want it to be developed, right? Because as a developer, you have certain preferences. And so, you know, the AI is just going to go off on a tangent based on what it thinks is the optimal. But if you're able to go and design the system ahead of time and figure out like, hey, I want this, you know, to be engineered in this manner with this architecture because of this, then it's going to be much more effective at putting all those different parts together. And also, if you keep a very keen eye on it, right, if you just run Claude code and it just runs in the background,

George Pu (32:16.045)
And so, you know, the AI is just going to go off on a tangent based on what it thinks is the optimal. But if you're able to go and design the system ahead of time and figure out like, hey, I want this, you know, to be engineered in this manner with this architecture because of this, then it's going to be much more effective at putting all those different parts together. And also if you keep a very keen eye on it, right? If you just run cloud code and it just runs in the background, you're going to end up with a product at the end of the day that's totally wrong.

Soham Mehta (32:35.566)
if you keep a very keen eye on it. If you just want to slide code and it just runs the background you're going end up with a product at the end of the day that's just totally wrong. But if you start it off and literally in the first message that it does you realize that it's doing the wrong thing and you stop it right there and you say hey no actually this is what I want and then you continue and you keep doing that right it might be every four prompts that you need to modify what it's doing then it's going to end up being on the right track but like that's the trade-off right because you can't just be you know

Matt (32:40.333)
You're going to end up with a product at the end of the day that's just totally wrong. But if you start it off and literally in the first message that it does, you realize that it's doing the wrong thing and you stop it right there and you say, Hey, no, actually this is what I want. And then you continue and you keep doing that. Right. There might be every four prompts that you need to modify what it's doing. Then it's going to end up, you know, being on the right track. But like, that's the trade-off, right? Is you can't just be, you know, a dormant coder. You need to be actively involved. It's kind of like having a junior developer that

George Pu (32:43.533)
But if you started off and literally in the first message that it does, you realize that it's doing the wrong thing. stop it right there and you say, Hey, no, actually this is what I want. And then you continue and you keep doing that. Right. There might be every full prompts that you need to modify what it's doing. Then it's going to end up, you know, being on the right track. But like, that's the trade off, right? As you can't just be, you know, a dormant coder, you need to be actively involved. It's kind of like having a junior developer that like is incredibly good at syntax, right? That's basically what AI.

Soham Mehta (33:05.85)
You need to be actively involved. It's kind of like having a junior developer that like is incredibly gonna sit back. Right? That's basically what I got. Yeah, I think, yeah, so. No, was just saying, I was just thinking there's like something to be said about like that it would be a mouse on, but like this idea of just like human like laziness in the sense that you just try to like one shot it or just like in three, four prompts. if you piecemeal almost anything, whether it's coding or something else to very, very small little sub-sex.

Matt (33:10.153)
like is incredibly good at syntax, right? That's basically what AI is. Yeah. I think, uh, yeah. I also think there's like something to be said, like that it would be, uh, well, mouse on, but like this idea was like human, like laziness in the sense that you just try to like one shot it or just like in three, four prompts, if you, if you piecemeal almost anything, whether it's coding or something else to very, very small little sub-sex, it's almost always like very accurate. I've seen some good examples of this as well.

George Pu (33:13.467)
Yeah, and I think, yeah, something. I also think there's something we said that it would just not sound right, because it's human laziness in the sense that you just try to one-shot it or just in three, four prompts. I like you piece me on almost anything, whether it's coding or something else, to very, very small little sub-sex. It's almost always very accurate. I've seen some good examples of this as well.

Soham Mehta (33:34.796)
It's almost always very accurate. I've seen some good examples of this as well.

Matt (33:38.981)
What's that meme? like the guy crawling in the desert and it's like, this direction is like water. And it's like, you edit this one file, manually go and edit this one file or 20 GPT prompts, you know? And then he goes for that one, you know? It's like, you could end up in a loop where it's just stupid. Like you just open the damn file and fix it, you know?

George Pu (33:39.031)
What's that meme? It's like the guy crawling in the desert and it's like, in this direction is like water. And it's like, edit this one file. Like manually go and edit this one file. Or try 20 GPT prompts, know? And it's the same that one.

You can get into the loop or it's difficult. You just open the damn file, fix it.

Soham Mehta (34:04.416)
Open the damn file. That's what I'm Yeah, I think we'll close it there. So for our listeners, there many other episodes that we discuss about AI, sovereignty, Bitcoin, and everything related to your 20s and 30s. So feel free to scroll through our other episodes and then follow us on Twitter, YouTube, and other platforms. If you have any ideas about what you want us to discuss, tweet to any one of us on Twitter or to our quater life cap Twitter account.

Matt (34:08.305)
we'll close it there, interesting discussions. So for our listeners, there are many other episodes that we discuss about AI, sovereignty, Bitcoin, and anything related to your 20s and 30s. So feel free to scroll through other episodes and then follow us on Twitter, YouTube, and other platforms. And if you have any ideas about what you want us to discuss, tweet to any one of us on Twitter or to our quarter life cap on Twitter account. We'll be monitoring that 24-7. So thank you.

George Pu (34:31.163)
Twitter account will be moderni24x7. So thank you.

Soham Mehta (34:33.705)
Thank you.

Matt (34:39.528)
All right.

Soham Mehta (34:39.854)
it.

George Pu (34:39.865)
It's there.