#107 Unraveling the Future of AI: The Developer's Perspective with Steven Schrembeck

Machine transcription; unreviewed and may contain errors.

#107 Unraveling the Future of AI: The Developer's Perspective with Steven Schrembeck
Source: interviews/107-unraveling-the-future-of-ai-the-developer-s-perspective-with-steven-schrembeck.mp3
SHA-256: dafaf5befbdae16c8396a4be98ce485f00e8bfd41d00c935bcc745908cd673c0
Model: scribe_v2 | Transcribed: 2026-10-10T18:00:15.990234+00:00
Machine transcript — uncorrected. Speaker labels are local to this recording and do not identify people.

[00:00:00.060] Speaker 1: [upbeat music] Hello, and welcome to a new episode of "The CTO Show with Mehmet." My name is Mehmet, and as you know, in each episode I discuss different topics about emerging technologies from AI, digital transformation, cybersecurity. And also I sometimes have guests with me who are subject matter experts in one of the domains I usually talk about. And today I'm very pleased to have with me Steven Schrembeck, who's joining me from the United

[00:00:34.960] Speaker 1: States, from Georgia. Steven, thank you very much for being on the show today. I will keep it to you to introduce yourself, what you do, and what you are up to.

[00:00:45.420] Speaker 2: Hey, Mehmet. So thanks for having me on. So I do three things that are of relevance. One, so I'm the founder of a startup called Impossible Labs, so it builds AI stuff. It sounds very fancy, but I'm basically a solopreneur plus AI, [chuckles] plus contractors, and I, I really like it that way. Uh, beyond that, I make videos on a channel called Seeking Minima, so that's what brought us here today. That channel has one purpose, and that is to help people use technology instead of being used by it, and that's sort of my goal.

[00:01:20.420] Speaker 2: And then the, the last part is I write stories, science fiction and fantasy.

[00:01:24.680] Speaker 1: Oh, wow. I love this combination, really.

[00:01:28.160] Speaker 2: Mm-hmm.

[00:01:28.380] Speaker 1: Now, uh, I know Steven also writing like you, you do software development as well, right?

[00:01:35.680] Speaker 2: That's right. Yeah. I didn't mention the one thing that like pays the bills, right? Um, so I guess AI does pay the bills. Um, but yeah, so I've been a software developer for 12 years, something like that. Um, big corporate software developer. Uh, I made the rounds. I worked at Amazon for a while, uh-

[00:01:53.460] Speaker 1: Okay

[00:01:53.460] Speaker 2: ... worked at smaller companies. So it's been good to me, uh, but it's not, it's not as exciting as, um, cutting edge tech, which is always where my heart has been.

[00:02:05.820] Speaker 1: Uh, that's great. So Steven, just wh- out of curiosity, like you have a mix of, you know, multiple things at the same time. So what I would say drove you to choose this path, uh, for your career?

[00:02:25.180] Speaker 2: So I've moved through different kinds of software development and first of all, I can't help myself. That's the easy answer, right? [chuckles] It's, I can't help but be interested in things. Writing is something I haven't been able to put down for a long time, so that's just like something that won't leave me alone. Um, and same with cutting edge tech, right? Uh, I moved into software development because I liked building things. I'm not one of those people that loves code for code's sake. I, I like it as a tool, and while I am fascinated with building things and understanding how they work,

[00:03:00.120] Speaker 2: I'm more interested in the opportunity to solve cool problems and help people. And software was just the easiest way to do that 'cause as you know, it's easy to set up, tear down, like there's, there's no sandbox like software. Um, so I think that's really what drew me in. And then AI is just... Honestly, it felt like modern day magic. I started building models in 2016, uh, when deep learning was having a first or second renaissance, depending on how you look at it. And I knew that this was computer magic [chuckles] and I wanted to know how to do it. Uh, so that's what drew me into AI.

[00:03:36.500] Speaker 1: That's very, very cool I would say. Now to- l- I know this is the hot topic. It's on everyone's mind, right? So it's AI. Um, how do you perceive it from not only a developer perspective, a solopreneur perspective, how do you look to AI? Do you see, do you see it as, you know, the technology that will, you know, make really people's life easy or do you see it more as a, just a cool technology that we can do cool stuff with

[00:04:11.340] Speaker 1: it? How, how do you perceive really the technology? Because there are a lot-- And the reason I'm asking you, Steven, this question, there are a lot of debates recently with all the hype that happened after ChatGPT and all these technologies. What's next? What's going to happen in the world? So from a developer perspective, a solopreneur perspective, what you can tell us about this?

[00:04:37.940] Speaker 2: So a lot of the time the hype isn't real. So I also, you know, I'm very interested in Web3 mostly again for like coordination technologies. I'm terrible at, uh, [chuckles] I'm terrible at riding the hype wave. I'm always interested in the tech and the values. Um, so for AI, this time the hype is kind of real, and it was last time too. Uh, if you'll recall, I don't know what, four or five years ago, the last time deep learning really took off it was for vision networks, and this is really where the killer use case was. And the hype was real, it just wasn't evenly distributed. Now, as, as you know, diffusion networks

[00:05:12.940] Speaker 2: for generative AI and large language models, uh, are the latest hotness and the hype really is real again. Um, I think the reason that we might not see another AI winter between here and AI really doing very impactful things to society is because we have a lot of Lego blocks that we've built up. If you followed AI development for the past, you know, decade or so, it keeps increasing. But what has happened is that everyone was increasing in their own silos. So, you know, nat- natural language processing was

[00:05:47.920] Speaker 2: getting better. Um, time series analysis was getting better. Just transformer networks, all, all the vision networks, all of these things were improving. Diffusion, autoencoders, all these things were improving separately. And what happened is people from one domain connected a piece to another domain. So in this case, it was reinforcement learning plus natural language processing. Um, and plus-- And that was basically you connect two pieces from two verticals that have been advancing for a long time, and the result was exponential increase. It could do things. The difference was

[00:06:23.000] Speaker 2: we have generative networks. We've had GANs for a long time, but they're sort of obtuse and really hard to train. The difference is we needed, we needed an interface that understood what we meant. That was the big, like, unlock here, is that we could always sort of get networks to do really good things narrowly, but we couldn't interface with them in a, in a very good way. If they know something like, um, like a large language model does, how do you get the information out of it? And I think that natural language processing was an interface moment. So the combination of

[00:06:57.960] Speaker 2: all these individual advancements is really why you're seeing huge changes now, and the amount of individual advancement has not stopped. So I would be very surprised if there were not more verticals to combine and again, create another ridiculous pivotal moment every six months or so. But you might see lulls in the meantime. So I'd say the hype is real, sort of, right? [chuckles] People always tend to overhype no matter what you do.

[00:07:27.740] Speaker 1: Um, actually, Steven, you said something important which I discussed couple of, uh, episodes back. I, I go solo sometime and I said like I'm, I'm an old, not very old guy, but I'm an old guy enough to remember all the hypes in history and I don't remember a moment, you know, similar to what we are seeing today. Because I was like maybe 10, 11 years old when the first time I heard about the internet. You know, I was 14 years old when I tried the internet and then, you know, later on the smartphone, the

[00:08:02.240] Speaker 1: Web2. Um, yeah, like there was sometimes exaggeration, sometimes, you know, things going, but you know, the, the cycle used to take very long until we see the next thing. But with AI, I mean generative AI and the NLPs and, you know, these all models that we are seeing today, I think we, we, you know, as you said, I like the word you said, an AI we get-- we will not see an AI winter. I love this expression and I think we are going very fast. Now from a, I would say developer perspective, right? What do you think

[00:08:37.040] Speaker 1: would be the best applications that let me make it very general that everyone can benefit on other than, you know, the, I would say the sometimes very simple things we see, write me an email, write me this. So I believe like from your perspective, Steven, you see a bigger picture. Can you share that with us from your perspective?

[00:09:02.300] Speaker 2: Yeah. I think another reason why this time it's different is we have a very general model, a shockingly general model, meaning it can, it can reason about things, right? Whether it doesn't really matter what's happening under the hood, all that matters is the input output results in something that looks like commodified intelligence. It's a brain in a box. You can spin it up. Now it's not super smart. It's like maybe average adult smart, and then in a couple domains it's exceptionally smart, like law or some solving SAT problems. But what matters is that you have

[00:09:37.140] Speaker 2: common sense reasoning in a box. That's the kind of thing that honestly I expected to take a lot longer and I think a lot of people did. I did not see this one coming next. There was a number of hurdles to artificial general intelligence and I didn't think this one was gonna get knocked over first or next anyway. So really when you think about how can I use this, what could you ask somebody that you could pay minimum wage or just somebody with no training of average intelligence to do for you on a computer? Now you can do that for instead of $15 an hour, depending on where you live, you

[00:10:12.080] Speaker 2: can do that for 20 cents and you can do it a thousand times faster and you can have 100,000 of them working. So what can you do with that? Well, any information processing is if your job is information processing, that's AI job now. Um, and you're gonna do something else. So it's kind of hollowing out the middle. It's hollowing out people who work on-- If your job is to turn one piece of information into another piece of information, if there is a broad skill set and data set for what you do, you're gonna be

[00:10:47.160] Speaker 2: replaced pretty quickly, um, depending on the economic output of what it's worth to replace you. Except if you're a specialist. There is, is very hard to replace specialists. So I know a person whose job is to audit, uh, oil and gas refinery, uh, machinery for safety. And even though his job is basically to analyze data and spit out, okay, safe, not safe, do maintenance, don't do maintenance, replace, he's not gonna be replaced for a very long time because it's such a specialized knowledge set that he's fine.

[00:11:22.240] Speaker 2: So if your job is pretty generic junior coder with no specialization or pretty generic junior designer who just does websites, you should consider using other things. Now I think your question was to ask how can people use this? That was sort of like the negative side-

[00:11:41.540] Speaker 1: Yeah

[00:11:41.760] Speaker 2: ... which gets a lot of attention. But the positive side is true too. This is commodified intelligence in a box. If something, if there's enough data, you can generate it. But that's the very general thing. I can tell you that what I'm using it for is to remove, is to really take control of my sort of- To fight back against the attention economy, I guess is the right way of putting it. So just for instance, yesterday-- Actually this morning and last night, I wrote a tool that uses OpenAI's Whisper to transcribe

[00:12:17.012] Speaker 2: podcasts. So it takes from audio to text to give you a transcript, including speaker attribution, so who's talking. And then it takes that, pumps it into ChatGPT-4 through the API, [chuckles] and then I run it through a prompt where I tell it basically, "All right. Here's the questions I'm gonna ask. Here are the things I need to know about this podcast. Here are the viewpoints I wanna see." And then it just curates this report. So there are certain podcasts, let's say like macro investing, right? I don't really care about this too much. I wanna know roughly, you know, where the markets are. I don't wanna pay attention to it. News is-- makes you feel bad anyway. But now,

[00:12:52.132] Speaker 2: with the push of a button, here's everything I need to know distilled down into a s- tiny amount, and I can curate that for many podcast feeds. And this is just stuff I'm making. So I am using it to sort of get more information and learn more than I could before, and to underst- to learn quickly and to synthesize information. I'm using it like intelligence in a box. "Hey, read this for me."

[00:13:18.152] Speaker 1: Mm-hmm.

[00:13:18.532] Speaker 2: "Hey, distill this down and..." But only in the places where my goal is just information. So if you wanna be matrix mainlined, just plug me in. Like, I just wanna know how to do this. This is a, a killer use case right now for anybody. You don't have to know how to program. Just interfacing with the, the chat clients is enough. Uh, you could be learning really, really quickly right now. And so this is the sort of-- This is the trade.

[00:13:45.592] Speaker 1: Yeah.

[00:13:45.652] Speaker 2: So you have an opportunity where AI is really, really useful at helping you learn just about anything that makes you very marketable right now. Or you can ignore it [chuckles] and sit around and wait for it to-- your job to be slowly eaten away. Um, so it's like you can ha- you can really stand out right now if you're willing to do the work and learn and try these things out, or you can, you know, wait on the sidelines.

[00:14:11.012] Speaker 1: Wow. You know, like what you are trying to build is, uh, something similar. Uh, I mean, not exactly the same idea, but I wanted to... I rely on preparing this podcast on some news feeds. So, and one of the use cases I thought is let ChatGPT or whatever API read these feeds and just select which ones are important so I can choose a topic, for example. Now, you mentioned the word prompt several times, and there's a hype, if we can call it, about

[00:14:46.212] Speaker 1: prompt engineers, right? So prompt engineers and, you know, from your perspective, and let's discuss it both, you know, from a real software development perspective and plus, you know, a realistic use case perspective. How important is the prompt when interacting with models like ChatGPT?

[00:15:08.732] Speaker 2: Ideally, over time, the prompt becomes less important. If you have a good natural language model, the goal is for it to be a better interface. So I'll, I'll just take a step back and do a little technical explanation that I-

[00:15:21.012] Speaker 1: Yes, please

[00:15:21.352] Speaker 2: ... hopefully doesn't put people to sleep, but I think is very useful for understanding how these things work. So at the center of these large language models, and even stuff like diffusion models, um, is something that researchers call latent space, which in my opinion, they just use a technical sounding term to describe something that's very complicated. I like to think of it as just this big information soup. It's encoded in a sort of machine language. It's just a representation that the model has learned about the wor- the world in the most dense format, right? This is the center of

[00:15:56.392] Speaker 2: everything it knows. This is where it's recorded all the information it knows, all the patterns it needs to find. You can think of it, it's like its repository of knowledge, but it's not in words. It's in, technically in embeddings. Uh, it's just hyper-compressed information. So it knows a lot of stuff about a lot of stuff. You know, the, the OpenAI models have read a bulk of the public internet, so they know a lot of things. In fact, they know a lot more than we can pull out of them. And so a lot of the

[00:16:31.112] Speaker 2: trick is how do we get out-- I know you can do this. How do I get you to know what I mean and so that I can synthesize, either do these operations, write this code, perform these tasks, help me-

[00:16:42.752] Speaker 1: Mm-hmm

[00:16:42.952] Speaker 2: ... read this thing, understand what I'm trying to get? So that interface gap between I know you have this information in your latent space, in this ocean of information, I know you can do this. I'm trying to get you to focus only on what I want you to do, and there's some misalignment there. So there's a gap in the interface, which is do you know what I mean, and can you-

[00:17:05.692] Speaker 1: Mm-hmm

[00:17:05.712] Speaker 2: ... sort of pull it out of what you actually know how to do? So that's what prompt engineering right now solves for, is can you shrink that gap to be better at synthesizing what you want from these large language models? So right now it is useful. I-- It will probably be useful for a while, but in theory, as things get better, that interface should get better, and the models will get better and better at understanding what you want so that you don't have to use all these tricks of the trade. So I wouldn't say there's a super long shelf life on this skill, but

[00:17:40.972] Speaker 2: I could be wrong.

[00:17:43.472] Speaker 1: And that could be [chuckles]-- You know, like, because yesterday, when actually in the, during the weekend, one of the things I was trying to do is to let the model tells me how best it prefers to be interacted with. And-

[00:18:01.052] Speaker 2: Got it

[00:18:01.392] Speaker 1: ... uh, I spent some time until, and actually you, people might laugh, but I told ChatGPT after I get the answer I wanted, I said, "Okay." Can we write an e-book about the... And put some examples, and it did. So as per ChatGPT, it prefers to have the following: a context, specificity, instruction, purpose or goal, and limitation. So this is what I was told by ChatGPT, and it gives me an example, by the way. If you ask me this way, exactly as you said,

[00:18:36.268] Speaker 1: Steven, I know that I should search exactly in this. You know, I don't have to go search the whole language now model, because I know exactly what I have to go. So this is maybe where the prompt engineering, but I believe, you know, I have a little bit of technical background and I know a little bit how these models work. I think the more they are trained, the better they get, so even you will not need even, you know, an exact prompt to get the information out of it, right? So that's, that's very, very cool, I would say. Now, here I want to ask you, now we have this moment, I would say,

[00:19:11.768] Speaker 1: what do you think other technologies will get combined or let's say, you know, used together with the, you know, all this AI, um, technologies? Like for example, personally, I think automation with AI will play a huge role. I started to see, you know, someone talked about autonomous bots that they can interact, you know, with each other. What's your view on this?

[00:19:37.408] Speaker 2: All right. So the use... Just that's like three good questions.

[00:19:40.308] Speaker 1: [chuckles]

[00:19:41.088] Speaker 2: Um, so I think that there's a lot of-- It will be combined with everything. It's intelligence, uh, so it, it gets combined with everything. But I think the ones that are sort of juicy and next, I actually don't think it's automation. Now, I could be wrong. There is a lot of value there. So I think it's more like narrowly there are new kinds of automation that were just unlocked and those, yeah, that's greenfield. That'll be anything that required like common sense reasoning. I think that those things... There is-- Okay, automation will move forward. The reason I don't

[00:20:16.108] Speaker 2: think that this is the automation moment, I mean physical movement and manufacturing, stuff in the real world, is because physical stuff is hard. [chuckles] It's hard and expensive and physics is finicky, and moving stuff around in the real world requires precision and rules and money and a lot of testing that is a lot slower than doing stuff in the digital world. And this is something that surprised me, but I, as soon as I saw diffusion, I don't know, what is that? A year ago, something like that. When stable diffusion first came out, I knew immediately

[00:20:51.148] Speaker 2: that, oh, I was wrong. It's not automation. It's, it's coming for knowledge workers first. It's coming for everything. So that sounds very sinister, but like, because that's where the data is. Not only is that where the da- the data is, but the people building these things, this is what they understand, right? They understand how to use computers, how to do information jobs, how to program, how to do design. So not only do they have the data sets, but this is a dom- set of domains they understand. So knowledge work will be automated actually before robots. So

[00:21:25.588] Speaker 2: hilariously, it's the plumbers and the mechanics and even the factory workers of the world that will probably, uh, be gainfully employed, you know, longer than the sort of Silicon Valley types, which is im- highly amusing. Um, but again, it, that sounds sort of dark. I think the other ways I would combine this, I, I can just tell you what I'm doing. One of the products I'm most excited about is called Intent. Um, so I'm trying to build a system that literally watches your screen. So this is a combination with other data streams.

[00:22:00.988] Speaker 2: So it's, it's using vision nets to watch your screen and it knows what you're doing, uh, all the time. It's private, so it runs on just your computer.

[00:22:10.208] Speaker 1: Mm-hmm.

[00:22:10.428] Speaker 2: And it just creates a stream of-- It answers the question, what did I do all day, right? It literally just keeps a log because it can understand what you're up to. Uh, right now you're doing an interview and it can log moment by moment. So at face value it's just analytics, but the real goal is to do executive control software. So sort of you tell it what you wanna do and it sort of keeps you in line. It's like-

[00:22:34.128] Speaker 1: Mm

[00:22:34.488] Speaker 2: ... "You shouldn't be doing that. Why are you up at 2:00 A.M. watching anime? You shouldn't be doing that. You said you didn't want to." So it's sort of like outsourcing your higher mind to keep your lower minds in check.

[00:22:45.808] Speaker 1: Mm-hmm.

[00:22:46.108] Speaker 2: And, but ultimately I wanna combine information from other things. So wearables is a big one. Uh, anything that... AI just wants data. So anytime-- So I have an Oura Ring, right? It has all my sleep data as well as continuous like biometrics. Wow, that would be great to incorporate into the thing that knows everything about my goals and what I'm doing all the time, and these things just stack on top of each other. So I would say anything that has a consistent high quality amount of data or data in large volumes is going to be very

[00:23:21.368] Speaker 2: useful. If I had to bet on a dark horse interaction with AI that most people don't see coming, I'd say it's probably brain computer interfaces. Um, most likely in terms of consumer EEG.

[00:23:35.568] Speaker 1: Mm-hmm.

[00:23:35.828] Speaker 2: Um, so I think Meta and Apple both have EEG coming out incorporated into wearables. Um, it can do more than you would expect. Um, so that's probably... Plus informa- brain computer interface information in, you can do that through sensory augmentation. Won't go into it, but turns out the brain is pretty good at getting information into it if you train it. Um, that one is also pretty interesting. So I think people are gonna start to chafe at how slow information goes out and into their brains and they're going to be interested in more or

[00:24:10.728] Speaker 2: less hooking up to these systems all the time, and these are the tools that enable that. You may have asked another question, but those, those are the answers I got for you.

[00:24:20.676] Speaker 1: No, no, that's, that's completely fine. And, uh, it, it made sense because, you know, anything which is, as you said, generating data, actually these language models lives on data, right? So they, they, they need to suck data to give us back what we want. And I think maybe something related to, to, uh, if we think consumer perspective sensors, you know, anything that comes from cameras and this side. In industrial, uh, vertical, I can think also about sensors and this kind of thing. So yeah. Now a, a question

[00:24:55.716] Speaker 1: that came to my mind just out of curiosity, honestly, like of course I'm too much into the, uh, you know, anything similar to OpenAI and like, and there are a couple of others, but when it comes to, you know, something like Stable Diffusion, like, uh, Mindjourney, like, uh, Da- uh, DALL-E 2, I want to understand personally. So, so consider me a, a guy that, that he's not from a technical background now. So for me, what I see is that these are algorithms or whatever you want to call them to

[00:25:30.616] Speaker 1: generate for you medias, right? So whether it's say photos or, you know, kind of paintings, but what are really the real uses or real world use cases that we can get out of, of these like, uh, Stable Diffusion or, uh, uh, Mindjourney?

[00:25:52.316] Speaker 2: I think people would probably be shocked about what is currently in development. Um, it takes a long time to build production systems for really-- You get this Cambrian explosion of like people eat up the easy use cases first. Just slap a user interface on ChatGPT or, um, Stable Diffusion. Cool, there you go. Now you can edit photos. That's great. [chuckles] Cool. You can make people disappear in photos. All right. Now what? You can generate art. Who cares? Um, but you should really consider this more like... By the way, there are other really good generative

[00:26:27.115] Speaker 2: networks. Transformer networks are competing against, uh, Diffusion now, and I think it's only a matter of time until they end up working together. They have different strengths. So generative information is really, it's, it's not slowing down either. I think you should look at this not as generating media, although it does do that, and that will have a profound impact, uh, particularly in terms of needing identity solutions, um, and probably the death of the public anonymous internet in a way. Um, but that's okay. Uh, it will bring other good things too, but it's more like

[00:27:01.816] Speaker 2: generative information. So here's an example that I, I like to give. You should think of these things as being really good at generating things to about 80% or 90% fidelity. Uh, they're not so good at generating it to 100%. So let's say you're, I don't know, you're an aircraft designer at Boeing, or you design cars. Your, your job is gonna look more like sitting back, crossing your arms and saying, "Okay, I want something that is sleek, powerful, uses this last year's model. Show me 40 different combinations."

[00:27:36.696] Speaker 2: Whoop, here they are. That one. [chuckles] Okay. Now given this one, make four different variations of this. Run it through the, you know, the stream or the, um, the tests for wind resistance and all this stuff. Great. I'll come back after lunch. Okay, here it is. Here's your model. That would've taken you how long with clay models or even just like going in with computer models?

[00:27:58.456] Speaker 1: Wow.

[00:27:58.556] Speaker 2: It's un- unbelievable. The amount of testing an idea, I- the i- the cost of testing out ideas is going to zero, and that's gonna have a very big impact on innovation, which is not really well captured right now or understood. When it costs you $100 or $1,000 or even $10 to do something and it takes you 10, 10 hours, 20 hours to test it, you don't test that many things. You have to go with, by necessity, the things that look like they're gonna have ROI. But when the cost of testing

[00:28:33.555] Speaker 2: a thousand things is basically zero and you can test a thousand things in a minute, things change dramatically. You, you can explore horizontally a lot more than you could before. You can test out really dumb ideas, and sometimes those dumb ideas turn out to be really interesting, useful ideas, and that's gonna happen everywhere. So it's not just generating media, it's generating information of any kind. Uh, but it will be-- The reason we see media is because we have a public open internet and everyone dumps all their data online, so guess what they trained on, is the stuff that there was a lot of

[00:29:08.376] Speaker 2: data for. But there is data for other stuff. It's a little harder to get and it's a little less lucrative upfront and a little harder to build, but it's coming. So, uh, any information fundamentally, whether it's science experiments, you know, generating novel drug compounds, all of this stuff is just information. Genetics. So if you want to test something out to just 80% or 90% fidelity really, really, really fast, and then that last 10%, 20% is still done with humans and maybe regular computers,

[00:29:44.156] Speaker 2: that's a big change. It's a couple orders of magnitude, I think, in terms of like speed of innovation. So that's, that's a huge use case for the generative networks.

[00:29:53.196] Speaker 1: That's, uh, you know, like it, it's, it's like very enlightening, I would say, thinking about doing simulations using these technologies because they can generate different, let's say, prototype for you. But there are some people that they are arguing that, um, we might be too much depending on the existing knowledge that already these- ... models they have, so it will be lazy to generate new content for the language models to use. What do you think about that?

[00:30:29.580] Speaker 2: I think how much does it cost to get the best aircraft modeler or the best car modeler to do something? A lot of money. How much does it cost to get 80 to 90% [chuckles] of the best? Still probably quite a bit. So the difference is, okay, yes, if even if we-- let's say I'll, I'll buy that argument, and we can only get to even 80% of the average of any skill, given it has enough training data. Who cares, right? All it's doing is repeating

[00:31:04.540] Speaker 2: existing patterns. Well, what matters is that I don't have to learn a skill. I can have 0% car modeling knowledge, and now I can model a car. That's a very big deal. Uh, just in the practical perspective, like for my stories, right? I can do the cover art. I can't do cover art. Before this, I spent $250 or $500 for like incredible concept art, and it took weeks of back and forth, and it was a huge pain, and half the time it doesn't end up looking right. Now I can just, just, I don't know, I'll just throw stuff at the wall. I'll just do it for four hours, and I've generated a couple hundred

[00:31:39.440] Speaker 2: pictures, and it costs me basically nothing. You can imagine that for any field. Like, you don't have to learn a skill to do it at 80% capacity of average-

[00:31:51.520] Speaker 1: Mm-hmm

[00:31:51.520] Speaker 2: ... like the average professional. That's a very big deal. And so you don't even have to advance state-of-the-art in order for it to matter significantly. It's about democratization of skill set.

[00:32:05.560] Speaker 1: Well, that, that's another point of view, I would say, which is also valid. Uh, Stephen, can you tell me, because I know you, you write sci-fi also as well, right? So, uh, can you tell us a little bit more about this side? Uh, I, I'm sure that I have a lot of my audience who would be interested to, to know more about it.

[00:32:24.600] Speaker 2: Sure. Yeah, it's just for fun. Um, but I do occasionally try to write-- Sometimes there's a combination of these. So it's sci-fi and fantasy, so the fantasy is mostly for fun. I always ha- try to have a lesson. Um, so I've written a story recently called "Reverie." That is a story about I think that humanity doesn't have a lot of great things to run towards, like collective goals. I think we have a lot of things we're trying to avoid, but that's not super motivating. Who gets up every day saying, "Oh, I hope I don't-- Hope I'm not in pain today"? That's not super motivating.

[00:33:00.080] Speaker 2: Um, so I, I tried to come up with here's where I think we should try to go as a species. Um, I'm writing a story now called "Hellmakers." This is about... This is why I smiled when you mentioned, um, s- I think you mentioned something about, um, bots, automated bots.

[00:33:18.920] Speaker 1: Yeah.

[00:33:18.980] Speaker 2: So already I've seen people talk about somebody has already had the bright idea. You know what would be great? We could-- We need to open source these models because, you know, people probably don't want us to have access to this. It sh- everything should be free and open. You know what else we should do? We should put this on decentralized computing platforms. This is a good idea. And I'm just like, this is the worst possible idea.

[00:33:41.720] Speaker 1: Wow.

[00:33:41.760] Speaker 2: Let me write a story to tell you why. So it's a, it's basically a story about the founder of a decentralized digital autonomous worker protocol who they just start like a decentralized computing, uh, blockchain protocol, right? Um, the goal is that it uses a consensus mechanism, so you can't turn it off easily unless you can turn off all the nodes on the network.

[00:34:05.000] Speaker 1: Oh.

[00:34:05.600] Speaker 2: You can't turn it off. And it's permissionless, so anyone can insert money, and you insert a prompt, pull a model off the shelf, and you go tell it to do something, and it will go do whatever until it runs out of money. Uh, and you can't turn it off. So I show this is probably not a good idea. So the punchline of the story is that the, the, uh, creator of the protocol ends up using this to get back at their co-founders who cuts them out of a deal and, as the name implies, creates a bot that literally just publicly tries to ruin this other

[00:34:40.560] Speaker 2: person in perpetuity and cannot be turned off. So this is not good. We should think very carefully about not having an off switch for our machines. It just seemed like a very dumb idea to put them on a box that you can't turn off frankly.

[00:34:56.180] Speaker 1: [clears throat] Um, a question that I ask now, you-- it, it became a, um, traditional question for me. Like, um, do you think that AI is allowing all of us, humanity, to reach, to reach singularity?

[00:35:20.500] Speaker 2: Maybe. Um-

[00:35:22.660] Speaker 1: That's a, that's a philosophical question, I know, so [chuckles]

[00:35:26.240] Speaker 2: The key is in what you said at the end, which is all of us and humanity to reach the singularity. Yes. I actually think that positive outcomes are the most likely scenario in the long term, and there's a lot we can do about it, and a lot of things that we could be doing about it that are actually quite feasible that are probably not well known, so it's something I'm focused on. But yes, I think technically speaking, yes, this is the way you would need scaled intelligence. And I think the most important part of AI as moving us towards singularity is that

[00:36:02.060] Speaker 2: maybe is also underappreciated, is that it doesn't have cognitive biases. Uh, unlike our own brains, we can reprogram it, and we can program away the blind spots that we know we have. It doesn't matter if you know about anchoring bias or recency bias, um, if you still d- you experience them whether you want them or not. You still make dumb choices all the-- If you're tired- And hungry, you're gonna make bad choices. Like, this is just how the human brain works. You can't program around it. But that's not the same for our machines, which means that we should be able to program them to our ideals. And

[00:36:37.032] Speaker 2: as long as we can keep advancing them, I don't see why, you know, you can't have wonderful, incredible outcomes from that. So in short, yes, I-- but I think it's neutral. It depends on what we do.

[00:36:49.072] Speaker 1: That's, that's fair enough, I would say. Um, be- before I will, you know, keep it for you to do a final conclusion, not related to AI, related because you are a solopreneur, right? So how, how, how do you, do you see it? Like, what's your experience being a solopreneur? Because again, this is the, this is the second, uh, common question I'm asking to my guests if they are solopreneurs, like why you opted to be a solopreneur, like why you're not part of-- I, I know that you, you worked for some companies before, and I did by the way as well. I'm a

[00:37:24.112] Speaker 1: solopreneur kind of now. So tell me your experience as a solopreneur and why we are seeing more solopreneurs, uh, in, in the field.

[00:37:36.952] Speaker 2: So I'll answer the last one first 'cause it's, it's fairly simple. I think we're seeing more-- people are more isolated in general, but I think we're also more empowered by technology in general. Like I mentioned, like I, I can do a lot of things I don't know how to do thanks to advances in, in AI and other software, right?

[00:37:54.192] Speaker 1: Yeah.

[00:37:54.292] Speaker 2: Um, I don't know how to compute compound interest. I mean, I could do it by hand if you gave me the formula and a lot of time, but a spreadsheet can do it. Uh, so I'm enabled to do a lot by myself, and because I'm working in software and AI, which is like software for software, it's-- I can-- I have a lot of leverage. Um, so I am enabled to do it. Um, but also at the same time, it's often easier just to pay contractors, and I find that there is a sort of cultural zeitgeist movement towards, uh, people who want

[00:38:29.412] Speaker 2: their own autonomy, and they wanna interface with other people as people, and they don't want to be an employee. So I found it very easy to just, "Hey, what's your rate? Like, let's figure out a deal, let's figure out a partnership." And that has been the easier way to go. So that's the answer to the second part. So the first part, my experience as a solopreneur, I think that the challenge, especially with AI, right? So selling software development is fairly straightforward because people understand kind of how to, how to value it. The biggest challenge of getting

[00:39:04.312] Speaker 2: paid to build stuff with AI is how to help people understand the opportunities that they have and how to help them value it. It's-- There's a lot of work that has to be done. You don't have to do a lot of work to tell people about regular software systems. They already know how to reason about it. There's a lot of competitors they know what to compare it to. It's very simple. But for, for AI, there's a long conversation of education. Okay, here's what is possible. Okay, now I need to fully understand your business so that I can analyze,

[00:39:39.232] Speaker 2: okay, where's the opportunity? And then from there, I have to help you understand what the value is, that we can justify the cost, which is higher than regular software. There's a lot of steps. So-

[00:39:49.832] Speaker 1: Yeah

[00:39:49.972] Speaker 2: ... it's a very high trust thing. Also, I noticed that most businesses-- So again, this is just money from consulting, not building my own products. Most businesses would benefit... They're not even fully utilizing software. They're n- definitely not even fully using non-software solutions. And so AI is often like, why would you bring this really complex solution in? 'Cause it's sexy, it's got buzzwords. But I've found so far, for the most part, price optimization, retention prediction, uh,

[00:40:24.792] Speaker 2: collaborative filtering, just-- even just regression, like linear s- um, logistic regression, just statistical models. These are often like 80/20, so most people don't need the big guns. That is the cool stuff for sure. I'd say generative AI or the large language models changes things a little bit. But by and large, you don't need AI for most solutions. So a lot of the challenge is finding the people who really do have a slam dunk use case and building the trust, telling the story. So

[00:41:00.672] Speaker 2: yeah, that's, that's what it's like, is that there's a lot of legwork. That's sort of the AI side.

[00:41:07.932] Speaker 1: Yeah. That's fair enough, I would say, and, uh, I have to agree with a lot of the points that you mentioned, Steven. Steven, any final thing you-- maybe something I didn't ask you, you want to, to, to say or share before we, we close?

[00:41:23.212] Speaker 2: Um, well, my producer and, um, the social media growth manager would probably hit me over the head if I didn't say go to the YouTube channel.

[00:41:33.732] Speaker 1: Yes, please.

[00:41:34.032] Speaker 2: Seeking Minima. [chuckles]

[00:41:36.052] Speaker 1: Yes.

[00:41:36.672] Speaker 2: Um, but-- So I mean, do that if you want to hear more words like this. But-

[00:41:43.092] Speaker 1: Actually, I wanted to ask where people can find more about you. [laughs]

[00:41:46.892] Speaker 2: Ah, yeah.

[00:41:47.032] Speaker 1: Like this one. Yeah.

[00:41:48.432] Speaker 2: I, I, I'll get better at this, guys. Um, but so what I wanted to say is, uh, I wanna-- the whole purpose of Seeking Minima in general is this is not a for-profit thing. Like, I can just build software. Like, there's no point recording videos or talking to people like you for, um, for a living. That's not what I wanna do. I am trying to help people use technology and to create better outcomes for everybody, and I think that we are at a fairly pivotable, pivotal moment, and I see a lot of, despite some of the

[00:42:23.392] Speaker 2: pessimistic takes so far, that's not really how I look at the world. I think that we have a lot of opportunity- To make a big difference right now. And I want to help people realize that they don't need to stand here and look scared. Um, there's a lot we can do in terms of, uh, AI alignment, um, wealth inequality. We have a lot of tools at our disposal, and we're at the really sweet spot where there's... Most of this stuff is greenfield. It's brand new. We have a lot of opportunity to use it well, to put common sense rules in place, and to really set it up to work for us

[00:42:58.400] Speaker 2: instead of against us or for a select few peopl- few people. And I think that, yeah, I'm gonna continue exploring those things. I have a lot of concrete ideas, but also I just like to synthesize other people's ideas. There's no need to be terrified of the future because, uh, we have a lot of control over it, and we're the ones who control it. I don't know. People seem to forget that. As if by the way, the future, what future happens is based on what we do. So if you don't like it, you should do something about it, and that's, that's what I'm doing. If you're interested in that too,

[00:43:33.400] Speaker 2: then yeah, I'd love to keep talking with you.

[00:43:36.720] Speaker 1: That's really great, and I think we are on a same mission, I would say, Steven. Because one of the reasons I'm doing all, you know, this podcast, I'm, you know, creating a lot of content is to raise awareness. This is first. And tell people, you know, like, hey, uh, you know, it's not like only the scary stuff you see in the media. You know, like people gonna lose their jobs. I don't know what. You know, all this negativity. So I'm trying to put some light on the positive side of the technology, and not just AI by the way, any technology.

[00:44:11.900] Speaker 1: Uh-

[00:44:12.220] Speaker 2: Sure

[00:44:12.300] Speaker 1: ... 'cause I, I, I love technology myself, like since I was a child. And I always believe that technology's main goal is to take us forward, make our lives easy. From business perspective, it's always, I'm repeating myself on multiple episodes, increase revenues, increase customer base, decrease churn, all these nice things that, uh, business people like to, to, to, to hear. And AI is no different than this. So AI will make, you know, um, your business thrive and, uh, better. And for

[00:44:46.940] Speaker 1: individuals, I'm saying exactly what you mentioned at the beginning also as well. Guys, like if you are scared of something, go study it. Like this is the best thing you can do. Because if you just keep watching it and do nothing, yeah, you're gonna be scared. I, sorry, I cannot help. This is why I'm trying to help you in generating this content as much as I can, of course. Um, so yeah, this is why we are aligned, and this is actually why I wanted to interview you, Steven, because when I read, uh, your bio, this what attracted me. And by the way, this is I do with all my guests. I will share, you know, the YouTube

[00:45:21.880] Speaker 1: channel, like anything else I can find if you want to share that with me also as well. So that would go into the description of the episode and desc- the description on the YouTube also as well. Well, we, we came to an end, Steven. Thank you very much for being my guest today. I really appreciate the time. And-

[00:45:39.420] Speaker 2: My pleasure

[00:45:39.900] Speaker 1: ... hope... Thank you very much. And for my audience, whether you are watching this on YouTube or you are listening using your favorite podcasting platform, don't forget to subscribe. And as usual, I'm telling you guys, I would love to hear your feedback about this episode or about the show in general. If you are interested to be a guest like Steven today, like don't be shy. Please come up with the idea that you want to discuss, and I will be more than happy to have you as guest with me. And as usual, until we meet next time, thank you very

[00:46:14.860] Speaker 1: much, and we will see you soon. [outro music]