Interview #9: Joey Houser

Noah Weisblat: I'm here with Joey Houser. Joey, thanks for coming on. Talk to me about, a little bit about your background. What do you do? What brings you here?

Joey Houser: So I am Noah's friend, first and foremost. I studied mathematics at Fordham University. I then accidentally got a job working at Citibank. A lot of these old banks are lack innovation, and I hated it. So I wanted to look for something new, and so I decided to go to grad school and study for a master's in computer science. I got a minor in computer science and computer science is just the extension of math into the real world a little bit. So I thought that would be a good fit. And pretty early into my, in my studies I was approached to start a computer vision startup identifying medical instruments with AI. We worked on that for about a year. We just had trouble, on the direction we should go and funding and just got burnt out after a while. And now I was approached by someone else to work on a machine learning project. I guess identifying defects in industrial crystal growth. These are used for lasers and a lot of lithography machines and stuff to make chips. Super super complicated in industry. Super complicated data set I'm working with right now. But I'm working in that. And then recently I was approached to start at ReadySet, which I just started this week. ReadySet Surgical, and they do vendor tray management for for hospitals for surgeries.¹ Very cool.

Noah: And I wanna talk to you a little bit more about some of the machine learning stuff you did and some more in depth robotics, all that kind of mathematics based AI tools. But we'll start just going over the AI space in general, so just. General take on the AI space, where do you think it's going? What do you feel about it so far?

Joey: So I'm a big fan of LLMs, but honestly I don't know if I see too much growth in the LLM market. Effectively they can just get more accurate, right? You can set up better AI agents. And they're good at thinking, but I see the growth in AI and you have to think of AI as a more broad, encompassing group of software tools. So you have these agents, you have LLMs, which incorporate agents a lot of times.² You have computer vision and then you have machine learning and probably something else that I'm missing. But. Machine learning is more math heavy. It will rise on data feature extraction, blah, blah, blah. That you'll probably see some improvement in some algorithms people will make, but nothing crazy. I really see the connection of AI into the physical world with like robotics and computer vision, like what I was working on. That will be a huge segment of growth going forward, I think. And then I also would say in the world of genomics, which is what I'm studying right now at Case. That will also be really big in terms of creating new drugs, identifying certain cancer related genes stuff, a lot of rare genetic diseases and creating new vaccines and stuff like that. That will be, that, that will be really powerful. And especially once we can start testing drugs in silico, that'll probably be like one of the greatest achievements of mankind, I think.

Noah: How far off that is? I don't really know. When you say testing drugs in silica, what do you mean by that?

Joey: So right now right now when you test drugs, they go through at least clinical trials, and you'll still have to do that. But if you can basically get the idea of how the drug will work and react within the human body before you even bring it into human trials you can effectively scale your drug creation or at least your drug research. It it gets a lot cheaper. So right now you have a lot of biomed companies that will, they'll, IPO in the stock market, and they'll just be penny stocks and effectively they'll just burn cap until they get an FDA review. So the FDA basically says yes or no. And so the stock either skyrockets or it goes to zero, right? And so theoretically you can have a better idea on how that clinical trial will look beforehand. So that, that will be super powerful. Yeah, that's the in silico stuff. And like there's gonna be, it's gonna be really cool once we get, start getting quantum computing and stuff like that. But that's not really AI. I guess that's, they'll integrate that I guess AI into quantum computing. But yeah that, that'll be really cool once we start getting more of that. But that's probably another like 15 years off.

Noah: I've seen a lot of research and people talking about. The use of AI in personalized drug treatment. So really being able to understand somebody's specific anatomy Yeah. And all of that. And then making drugs based off of very specialized or specific ways of their body. Is that, how realistic do you see that being?

Joey: That's super realistic. So once you can sequence somebody's, I, we can sequence people's genomes pretty accurately now. And so if you know someone's genomes, you're gonna, you're gonna know. Probably, which cancers are caused by which genes not all of 'em are known yet, but you're gonna know like what person is super likely to get this genetic disease or this cancer, right? And lifestyle factors obviously play a play play a part, but you're gonna know that and you'll be able to know how drugs will be able to. Due to their genome, some people are allergic to peanut butter. I was just using peanut butter. Some people are allergic to peanut butter.³ That's not a drug, obviously but that's stored in their genome. And so that, that type of genetic that, that use of genetic code to create personalized drugs yeah, will be super, super powerful especially towards genetic diseases.

Noah: Yeah. One more question on the general AI space, then we can jump into some ML robotics stuff. Which LLMs are you using right now, or which AI tools are you using right now?

Joey: So I actually got turned onto Claude from you. And so I've been using Claude Code and Opus, and I really like 'em. I thought G PT five was awful. I don't know. It was way worse than four. Somehow. I at least thought I, it was close to unusable. It was way slower and just, ah, yeah I couldn't stand it. I use Gemini just 'cause I get it free with through case. I don't really like it very much. I think it has a hard time picking up on context earlier in the thread. Yeah, I've just noticed like in long threads Geminis is annoying sometimes.

Noah: Makes sense. And do you have an opinion on which of those companies you think you feel best about long term?

Joey: Google is here to stay. I can tell you that OpenAI just received a hundred billion dollar investment from from Nvidia. That was crazy. Yeah, that was wild. So I think they're here to stay. That would be crazy if you I don't know how you fumble that. So they're probably here to say Microsoft also backs them. So I, I would say they're here to stay as well. Let me, I, grok X AI I'm skeptical of. I haven't used it much, but to me it just seems like Elon has his fingers in too many pies and he's. He, he is outcast from open AI. All that drama that he seems to always get in. And so it seems like he's just grasping a Oh, I have to have something like that too. So I'm skeptical on that. That would be if you have the couple, that would be the one. I would say is not gonna do very well. You do have the Chinese ones like deep seek is the only one I've seen. And that, that seems pretty good. The Chinese are definitely funding a ton of AI projects, though mostly right now they're trying to, they're trying to outdo ASML in the lithography space for making chips.⁴ So that's like they, they're trying to go after the manufacturing arm, I think to be like domestically capable, and then they're going to probably end up from there trying to compete with Nvidia. So that, that will be interesting. And that's also why I'm confused why Nvidia is investing in OpenAI if they're. They, there's an entire country that basically has it out for 'em. The second largest economy in the world basically has it out just for Nvidia. And instead of reinvesting in chip design or some other architecture like quantum computing or something like that they're, they are a little bit, but they're investing a hundred billion in an open AI, which is just gonna be circular. 'cause open AI is just gonna buy their chips. So I it's almost a stock buyback at that point.

Noah: Yeah, definitely. Yeah. And what you said about Brock too, I agree with, I think Brock needs to focus a lot more on the physics and the Tesla aspect of it. All the humanoid robotics versus trying to be a Twitter. Spot for everybody. I don't,

Joey: Yeah, exactly. And Tesla, you'll see a lot of stuff in the financial world complaining about Tesla's valuation, but Tesla through its cars driving everywhere and through a lot of the stuff Elon's been funding for years now has just probably the mo the largest visual data set in the world right now. And so in terms of robotics and self-driving cars. Tesla's behind right now, like with who's the company at Waymo. And there's probably some other guys who are competing. I think I've seen I don't know their names, but Tesla's behind them in terms of you're not seeing Tesla type, Tesla branded Waymo's driving and picking up places. But once they really start releasing releasing their models into the wild and just start letting cars drive by themselves. They have just so much visual data that you'll see tons of 'em flying, right? It'll just be exponential, I think. So that, that's way Waymo is very slow to pick up specific cities. And the the hardware's really expensive 'cause it has the big lidar thing and they have to connect to the actual car. Tesla just already has. All of that built into every single car they built for the past 10 years or so. And they have all this visual data and there, there's the problems with, oh, when it's dark outside, when it's rainy, it's okay, you just don't have the Teslas drive, hands free in the rain. And when it's dark outside, I that's not that crazy. And also there's lights for the most part outside in cities. So you just have Uber and cities, I don't know.

Noah: Yeah, no, that makes sense. That's a perfect segue too. So you've done some work training data sets, and that's what you just mentioned with the cars. They have to understand a billion or however many different images and videos and what they're seeing right in front of 'em. So just talk to me and talk to me. Explain I'm five, how do you train a data set for something that you're building?

Joey: So in computer vision, it's a little bit different than than just classical machine learning on a just what I'm doing with the crystal stuff, right? So normally what it is you have this vector of information in a, basically an Excel file, right? And it just has a bunch of rows, and that's all the computer sees is just numbers. And then it says, okay, based on kind of the connection between these numbers, this is what we think will be this value, right? So you throw in, I don't know. You say somebody had somebody trains as a pitcher for 10 hours a day for a month, this is what and this is their height. This is what I think their, your, their ball speed's gonna be, stuff like that. And so you have a bunch of pitchers training at different times and their height and maybe their weight or something like that. Throw that in a computer. The computer then predicts, okay, this is what I think. This is how I think like how fast they're gonna throw, right? So that's just like the basic machine learning and that, that was getting big. When I was in college, in the late 2010s people were at least that there was still a ton of stuff about object oriented programming. Everybody wanted to do software development, but machine learning was a big interest there. At least at the school I went to, with computer vision. The, there, there's different ways of doing it. So there's pure classification. This is gonna sound super boring, right? But you have like pure classification. This is a mug, right? Okay. Then you have detection, which is this is a mug and I'm gonna draw a bounding box around it. It's gonna track it, right? And so in, in a lot of robotics, you have to have that detection aspect. Okay? So it's like a whole other layer of complexity, right? It's pretty easy right now to detect whether or not this is a mug. But if you have a visual data set and it's trying to detect where the mug is and how big it is. You need somebody to sit down. And this is what Elon did that I thought was really genius. You need somebody to sit down and just draw bounding boxes around that mug in different environments. And that's super expensive and super hard to do. And Elon was taking his giant visual data set and he was paying people, I wanna say in Arizona somewhere. Like 20 bucks an hour, literally just to draw boxes around trees and cars and all this different stuff just to build this big labeled dataset that he could then feed into an algorithm that's similar, as I said, with the, baseball analogy. And it's a lot more complicated. But so something like that feed that into algorithm that then can be deployed on. Its either through its robots, but really in its cars to detect okay, this is a person crossing the road. You need to stop. This is another car getting into that lane, you need to turn over. Something like that. So that's really they, they have a, they have, they definitely have the largest data set. I would be blown away. They do not have the largest labeled data set. And that's where Tesla's valuation comes in.

Noah: I think that's super interesting. That's a very good explanation.

Joey: You'll see. I was gonna say, Elon built one of the largest data centers, at least the largest one I've seen. People talk about just to train these visual AI models in I think he built it in Memphis and then he couldn't get the power connection. It had something like a hundred thousand H 100 Nvidia GPUs couldn't get the power connection. So then he, and he built that like right across the border in Mississippi. He just built like a huge natural gas power plant just to power this data center that's like all to power and train on its visual AI data set. So it's. Pretty cool.

Noah: Explain how you see this, NVIDIA's investing a hundred billion dollars in open AI.⁵ There's 10 gigawatts of energy or compute that's going into that deal. For open AI to use the Hoover Dam produces two gigawatts. Like all this energy and power. What's using it all up? Is it the training? Is it, what is it?

Joey: Yeah. Yeah so it's the training, what exactly people are training on. Obviously I know the visual AI stuff needs a ton. So for reference in my startup, I had a pretty, I had a small data set that I labeled myself. To get the most features out of an image, you want the highest resolution, obvious. That means that, if you have a 1920 by 10 80 box, that's, 19, 20 data points on the row. And then 10 80 or I guess inverse 10 80 data points in the columns. And then, yeah, 1920 rows. So multiply that, that a computer has to look at all of that, right? So you start throwing in more images and you rotate 'em, blah, blah, blah. You add some more complexity and the what's called VRAM visual RAM which is what is used for all this, it just skyrockets instantly. I, yeah I was easily able to train stuff on 80 gigs and I did not have a very big data set 80 gigabytes of VRAM. And the, to give you some reference of the power, when when I was living at home, I would have my desktop with two GPUs. What I'm recording this on now two, two GPUs running two different mod training, two different models. And I was working out of the sunroom and the sunrooms not heated very well in the winter. It was probably 10 degrees outside. But it was getting hot in there because these GPUs were running at full speed, just cranking out heat. So it, it's a ton of power that's needed. What exactly people, why Microsoft is building a new data center every day or every three days. I don't really exactly know outside of just cloud infrastructure I guess but what people, cloud infrastructure is definitely a big part of it. What people are training, I still don't really know. Honestly. I like visual AI. I definitely know what Elon is working on. Obviously I know. Anybody can know what the LLM guys are working on, but there's a billion different other companies and I really do not know exactly what they're working on. Right now, my, my machine learning stuff, it can run on my CPU just fine. Yeah it's a small data set and we're doing more conventional machine learning with model stacking. But it's not we do not need these crazy GPUs yet if we get a larger data set, probably, I dunno.

Noah: Yeah, sure. Question robotics. A lot of people talk about robotics and then there's also humanoid robots, which are essentially robotics that look human or can be interpreted as human, can do human tasks.⁶ Where do you see robotics going in the next few years with AI in mind? And how do you think AI is impacting the growth of robotics?

Joey: So I'm skeptical on humanoid robotics. I think just the batteries alone have gotta be hard. I don't know, like how long do they need between charging and stuff. But whenever you add a new joint or something like that you add a huge degree of complexity back in the sixties. Like where the whole robotics craze started was, I wanna say GM introduced something in their factory where they just had a robotic arm that arc welded something in their car factories. And it was such they showed it on some big TV show and it was such a big deal that Kennedy had some whole commission on. How they were going to stop robots from basically like ending the careers of every single automaker. Obviously there are still there's still the auto workers union. There's not as many people employed in it. But they, that craze is still died. I think though, once you start integrating it with, with really cloud computing if you if you can distribute the the computing stuff offsite. That's what I, I don't exactly know how they're doing it now, but I'm skeptical. You can put a computer and all of these different joints and powertrains to move some guy around somebody's house and make coffee and stuff. With localized computing, it just seems it and not have it run outta battery in 10 minutes. That's my view. I think more specialized robots not this universal replace people type robot. I think that's just a marketing gimmick. More specialized robots will be very powerful in the future that, that is self-driving cars. Those are specialized just to drive safely on the road. They are robots. They work they have their own powertrains. They work on their own. They, I would assume, actually I think Tesla does do all the computing localized, I think. But you could, they, they still have internet connection. They could definitely compute up in the cloud. Maybe they do that. But yeah if you can really distribute the the computing power. You get some robot like a Roomba, which doesn't really need too much computing power right now, but say you could find some use case for a Roomba to pick up specific crumbs or something like that. It's gonna be super difficult to have a Roomba identify what an oat is versus, some piece of bread, right? Or at least if it does right now, you have to have some big computer on it and it's gonna be super energy. Ineffective, inefficient. If you can distribute the the power, basically you just have some receiver and you have some basic computer that says, Hey, this is what I just saw. Send the pictures back to the cloud. You have, you have ba, you have open AI that can just work out of the and this is what AI agents are, but you have one of the most powerful softwares ever created in huge data centers that can just feed to a single robot on the ground. It's super powerful. That's where I see it going. Not the humanoid stuff, but like stuff that's much more specific.

Noah: Yeah, I see it a little bit more in between those two examples with things like construction or Amazon where they're specifically designed to build X, Y, or Z or move X, Y or Z.

Joey: I remember Boston Dynamics, I've not heard from in a while. They were the forefront of all this, like a couple years ago. And one of the things they were trying to do with the humanoid robots, which I think makes a ton of sense, it's boring. It's not, like somebody, you wake up and some robot has just made you coffee or something like that. It, it's boring but it was just a robot that was going to unload boxes from a truck, right? Yes. And so that sounds very boring, but. If you're picking up boxes all day out of a truck you are going to get injured. It's super expensive to ensure that person is going to get hurt probably for life at some point 'cause he is just bending down and picking heavy stuff up all day. You can have a robot do that. That is so powerful. And then, like technically somebody's job got replaced, but then you just have that person go, you get a more cost effective business and that person can be trained to actually drive the truck or something like that. You can move more boxes that way. You don't have to have somebody just literally break their back moving boxes. So that's that type of stuff. Yeah. I definitely see that when it gets into construction, getting into somebody's house and planking boards will be that. That's, I think, far out. I think that's super hard to do. But moving, honestly, I would say some of the demolition work would just like moving drywall from inside of the job site to the trash bin that I could definitely see a robot doing. That's, once again, demo's. Another thing just like boxes, picking heavy stuff up off the ground, identifying it. Bringing it to another location. You don't even necessarily need the robot to identify where the location is. You could probably just geolocate it and say, Hey, just throw it here. All I need to do is just. I know what scrap junk is. You could, honestly, you might not even need to do that. You could probably just have a laser pointer, say, Hey, pick that up. And it just knows like exactly how to reach and grab stuff. That's, that you get into I only took a little bit of robotics sadly got hired. Not sadly. I yeah. No, I was very excited to get hired. But I, I had, I dropped this class and I was going to take all this stuff and basically it's all it's difficult because. You're training, you are having an arm go out and reach something and you don't really know the pressure that it should put on this object when it picks it up. So that's one of the really difficult things and that's why you see these robots that are super slow to move and pick stuff up. It's a difficult problem.

Noah: Yeah, that makes sense. I'm curious to see how it goes because it's, robotics I think is one of the most fascinating use cases of AI, so I think we'll see.

Joey: Yeah, I think so. Where that turns out to,

Noah: okay. I got one or two more questions for you. Just obviously you went to, so you went to Fordham you're back in Cleveland, you have jobs left and Right. You're you're working on a couple of different projects. What would your advice be to somebody who's either just graduating high school or just graduating college and trying to get into the computer science field? Or trying to just get started with anything AI related, or not AI related. Just if you could just talk to somebody in that space, what would you tell them?

Joey: I would say read a lot. That has been super valuable to me. I think. All this nonsense I just threw at you over the past, like however many minutes we've been on here. I'm telling you, when I graduated college, when I graduated high school, I definitely didn't know it. But when I graduated college in 2019, I knew none of it. Reading helps a ton. Networking is also super important and not necessarily networking where you're going to these events and you're trying to grift with people. I'm talking like, go find, your friend's dads help a lot. But random people you meet in coffee shops, random guys, old guys, you meet in bars, that's another one. I know all the young guys are going out trying to get the girls. You go out and you talk to some old guys and that is very beneficial for your career. What else? I would say on the technology front. Just start trying to play around with something you are like mildly interesting and just like get, get familiar with basic machine learning models. So I was very interested in in energy and so I was trying to create an oil price predictor algorithm. Now I think through, I did not know about Fred Data, but if you want any sort of economic data, you can just get an API with the Federal Reserve and you can download whatever you want and you can make some, basically just at chat, ask chatGPT whatever LLM you want, how do I make some sort of predictor model? And then you can start optimizing it. You can start playing around with that type of stuff. You start doing deep learning stuff that's really fun. That, and then I really. I enjoyed going to grad school. I really liked it. I think Case is a, is an incredible school. Fordham was good, but cases, and maybe this is just because I'm taking grad programs instead of a, undergrad program, but the professors really know what they're talking about. It is really innovative. There's just so much going on. Yeah that's what I would say. Read as much as you can. Network, play around with something, and then, grad school's really not a bad idea. I as annoying and as expensive as it is I, I have, I don't, I would not have gotten into all this stuff without having gone to case. I would just be a bum in finance, I think doing basically janitorial work. So yeah.

Noah: Cool. Anything that we did not cover today that you wanted to talk about?

Joey: I, let me think. No I think that's probably I would say also there, there is the bioinformatics, computational biology stuff. I think a lot of biologists now that Alpha three came out. And won the Nobel Prize in Biology two years ago, one year ago. There, there are a lot of biologists who are getting really interested in the computational biology space, and I think on the biology end, they're going to start doing more experiments with that in mind, which should be really interesting. Yeah. Yeah, a lot of the protein folding stuff can get really cool.

Noah: Cool. Joey, I appreciate you coming on. Excellent conversation.

Joey: No problem.