Algorithms + Data Structures = Programs - Episode 287: AI Takes & AI Taxes

Episode Date: May 22, 2026

In this episode, Conor and Bryce chat with Marco Franzreb Salgado about the state of AI, whether AI should do our taxes and more.Link to Episode 287 on WebsiteDiscuss this episode, leave a comment, or ask a question (on GitHub)SocialsADSP: The Podcast: TwitterConor Hoekstra: LinkTree / BioBryce Adelstein Lelbach: TwitterAbout the Guest:Marco is a software engineer at NVIDIA, where he works on improving the nvCOMP library, which offers fast GPU implementations of multiple data compression formats. For the past couple of months he has been working on a GPU implementation of the rotate algorithm.Show NotesDate Recorded: 2026-05-05Date Released: 2026-05-22ADSP Episode 237: Thrust with Jared HoberockADSP Episode 284: GPU RotateADSP Episode 285: GPU Rotate (Part 2)NVIDIA nvCOMPGOSIM PARISLife update: Zig, AI, unemployment, and moreIntro Song InfoMiss You by Sarah Jansen https://soundcloud.com/sarahjansenmusicCreative Commons — Attribution 3.0 Unported — CC BY 3.0Free Download / Stream: http://bit.ly/l-miss-youMusic promoted by Audio Library https://youtu.be/iYYxnasvfx8

Anyways, we're down, we're down, nobody cares. Nobody cares. Did we lose listeners over the AI hot take? Certainly, 100%. I don't know whether I can find somebody senior to bring in here who's going to be a good culture fit for the team and like have the right energy for like this like new like shifting, changing evolving landscape. And my hot take is, and like him and I talked about this a little bit, my hot take is not

everyone's going to make it. Welcome to ADSP, the podcast, episode 287 recorded on May 5th, 2006. My name is Connor, and today with my co-host, Bryce, we continue and conclude our chat with Marco. In this episode, we chat about all things AI. We get an AI hot take from Bryce, and we discuss whether we should let AI do your taxes. All right, everybody's recording now, and we're recording in the meeting. Bryce was already recording when he joined, and so he started talking.

and we'll let him continue to talk. But before we do that, we have to, because when I was editing last time, at one point we were going to let Marco tell us a little bit about MV comp, but I interrupted to get more details on his path to Nvidia and not having been programming like a few years ago. And then we never came back to MV comp. So one, just put it on the stack. It doesn't need to be at the beginning, can be at the end.

We'll put MV comp on. We also did get one, or technically, I think two questions from a single listener that we will put on the stack as well. So there's a couple topics on the stack, actually, and while we're at it, we'll put the second implementation of the GPU rotate, which everybody is waiting for. That's the third item on the stack.

All right, now over to Bryce. Feel free to tell us what you were telling us. Hang on. Hang on. I got to finish typing the thing to my agent. All right, now he's busy. Now he's busy with AI. That's the thing I was going to tell you

is like, I just got back from dinner, like sat down, got on this call to start, recording and at the same time I was like I got a I got to check how things are going I think that's the problem kind of that before if you wanted to get some sort of work done there was a hurdle because it's like okay if I want to do something I have to spend some time getting back into whatever I want to do and getting acquainted with what I was doing and so it takes a couple of minutes half an hour so you can't just hop in do work for 10 minutes or you couldn't and now you can because it's like okay what was my agent doing what do I have to

it now so that it keeps working. So you can fit everything in in five minutes. And I think that's what makes it so intense now. That's actually one question that I wanted to ask you guys is how you're dealing with context switching because I find it that now it's so exhausting, so much more exhausting than six months ago. Yeah, I mean, I would say it's less than five minutes sometimes. Like last night. So today is May 5th, tail end of the day. Well, I mean, it's very much the tail end of the day for Marco because you're over in Europe. Actually, yeah, what conference are at you at, Bryce? I'm at Go Sim in Paris. Go Sim? Yeah. It's some sort of agentic conference.

Go Sim. All right. Well, you're in Paris. All right. So both of you, it's a tail end on East Coast still. Let me do a better. It's an agenic AI conference. I don't know what Go Sim stands for. I'm sure I could be a better job. It's pretty good. So last night was May 4th, may the 4th be with you. And for the hardcore Star Wars fans out there, which includes me now, maybe not a couple years ago, but now I'm a massive fan. It was the two episodes season finale of Mall Shadow Lord available on Disney Plus, and I think

that's the only place you can get it. And my wife, beautiful wife and I were watching last night, and she is also a huge Star Wars fan now. And I can't remember if it definitely wasn't in. in between episode nine and 10. I sat on the couch for that, but at some point, I think we were watching John Stewart right before that,

and then I got up to get like, you know, a bubbly or something, and then I was like, I went up the stairs, and she's like, where are you going? And I was like, I just want to go check on Codex real quick, just real quick, just like 30 seconds to see if it's still iterating, because now I've gotten into the habit of, like, coaxing it into basically, it'll only launch six subagents at a time, but if you tell it like, you know,

queue up 100 sub-agents, and as soon as like one is done, just like launch another one and keep launching them, and you just tell it, like, I'm going to step away for a couple, I'm going to go have dinner, I'm going to be away for a couple hours,

but I don't want you to stop working while I'm gone. A lot of the times, if you tell it the right incantation, it'll just keep working while you're gone. And if you get that incantation correct, you can just pop upstairs, check to see if it's stopped, look at the work for like 10 seconds,

commit it, and then like hit the up arrow key, launch the same magic command, and then poof, it's back off to the races for another couple, hours. And so the point being is, do you even need five minutes? And my wife was like, what? No, no, no, we're done work. And I was like, I'm not going to do work. Do you not know about stop hook, by the way, Connor? Do I not know about what? Most of the harnesses have hooks that have on certain events, like if it tries to read a file or

if it tries to stop. So you can just install a stop hook that tells it to keep going. If you want to keep going hook. It's not, it's not stopping on anything. It's just concluding that it's done, it's finished doing the task that I assigned it. I warn you to be very careful with this because it can end up in a loop where it just sends a short response of I'm done. And then the stop hook triggers, no, you're not. And then it just loops on that forever. And that's very token wasteful. But aside from that problem, it's very useful.

I'm working on figuring out how to get past that. Yeah. Anyways, I didn't answer your question, Marco, but it's definitely a thing. Like, I was having dinner a couple weeks ago with a buddy of mine who works in Vancouver. And he was saying it's almost like exhausting now because, a lot of the mundane tasks that sometimes, you know, it would be a Thursday, kind of a long week, and you're just mentally tired, and you'd find that mundane task that's going to take an hour, and then, like, you'd just do that.

Now, like, those mundane tasks, they're gone. They've evaporated. So all that you have to do now is, like, hard thinking work. And it's hard to do that, like, constantly, where you're going to sit down and you know that, like, the work that you're going to do is, like, I'm going to have to sit here and really think about the structure or, you know. you know, it's not easy work. And so you find yourself mentally exhausted, like a lot quicker because there's no, like,

updating the docs or changing some markdown file, like all that stuff. I feel amped. I feel amped all the time, Connor. I'm not saying I feel exhausted. But when he was saying that, there are definitely times where, like, Friday rolls around. And I do feel like more, like, mentally zapped than, you know, the before times. And so how do we deal with this? I don't know.

You go for more runs. Yeah. I don't know, Bryce. I guess Bryce is amps, so he doesn't have any issues with this. I'm building things, Connor. Yeah. I mean, you showed me what you're working on. Yeah, we can't talk about it yet until it's done on the podcast.

But like, I don't know, I'm pretty close to having the orchestration layer set up. And then we can start doing the real fun stuff. But I think like, I want to say like two weeks, two weeks maybe. Yeah. Have you guys heard of, I do think that I'm like 75, 25, 25 in the, 75 just so amped, like, it's hard to fall asleep because you're so excited, which is interesting because I've listened to some podcasts that say the opposite, like, that they're the exact opposite.

They've never been less motivated before in their careers because, like, AI is taking a lot of this work. I find myself the opposite, and it's just like, and I was listening to a couple podcasts that we're talking about what Arnda Carpathie was talking about. I don't know, I'll link it in the show notes. But he was talking about, like, this idea of, like, neural computing, and it's just, it's to be like just in time apps and like just in time websites. And like I completely see that being the future. Like just the other day I added some feature where I could sort my race times. Like I have a

website that lists all of my 5K, 10K, half and marathon times over the last like decade plus. And I had a half marathon over the weekend. It didn't go well. And I wanted to know like what was the, what was the ranking of this in the 27 like half marathons I've run in my lifetime? And so, like, I'm not going to go and, like, manually eyeball it. Like, that's impossible. So I just went to cursor and was like, hey, add like a sort mechanism so that when I'm filtering, if I hit, if I just click on the header of this column, just immediately sort it. And then I said also, too, once it's in sorting mode, inject a basically column that shows

you one, two, three, four, five, because it didn't have the numbering. Sure enough, one shot at that. And, like, why do I have to sit down at a computer, launch a cursor instance, and, like, tell it, that's what I want. Like, in the future. you're going to be able to talk to this website and it's going to just in time generate you like a version

or like an edited thing where it's just say like and you can imagine like being an Android auto like in your car you're driving and like all the time what I want is like show me all of the restaurants that are open like past you know 930 p.m. that are within like a five clock.

It is terrible at that right now. Like it cannot tell you for the life of it like what's open currently especially like if I want a specific type of muffin you know that should be a thing in the future. It should be able to identify every single restaurant Why do you think it's bad at this? Oh, it's just terrible.

When you are, why? Why? Why? Because it has not figured out to like hook. And also too, Google Maps in general. Like how do you query Google Maps? Say I'm driving from, you know,

I will say, I will say, listen, I don't, I, I, I, I, I, my Twitter takes have gotten a little bit spicier recently. And I hate to use this platform to shit on Gemini, but I just have to. I tweet, so I had this thing where like I needed to ask, I needed to update some calendar invites. I needed to add a person, like invite a person to some Google calendar invites, and then I wanted to change the color of a bunch of Google calendar invites. It was like a pretty simple ask. I sent the prompt to Gemini from like, you know, my Google calendar page. Because there's a little Gemini button on my Google calendar page. And it was immediately, it was just like,

I can't do that. And then I went to chat GPT and I did the same prompt and chat TPT, which, what company makes chat GPT again, Connor? Open AI. Right, right. And who makes Google Calendar? Google? That's right. That's right.

Right. And so ChatGPT's Google Calendar integration worked perfectly in the first time. And so now, like, it's actually really good. Like, I've been doing all my calendar management from the Google Calendar-like plugin thing in ChatGPT. But every time I try to use Google's AI to do anything with Google's products, it sucks. It's terrible. It's just absolutely awful. It can't do anything in Excel.

It can't do anything in email. It can't do anything with my calendar. First of all, it's Sheets, not Excel. Excel's a Microsoft product. I don't want to be the pedantic guy on the podcast, but Excel is from Microsoft. Google Sheets is from Google. And you're thinking that's a distinction without a difference,

big difference, folks. And I haven't really tried to do much with integrating stuff. And in general, I don't use ChachyPT anymore because it's just terrible. That being said, Gemini has been offline all day. What do you mean? GPD 5.5 is great. 5.5 is amazing, but like if you're just using it as like a free user going to like the web app to ask some question, I gave up on Chad GPT like months ago because. You're not,

you're still not paying for a Jad GPD subscription? I don't want to pay for like if it's just simple questions. Like if I want to know how to convert some Fahrenheit to Celsius or I want some random thing. I'm not going to tell you about my cloud bill for all the various stuff I've been building this. Also, too, I don't want to be using my corporate account for like personal questions, you know? Like if I'm, if I'm looking for muffins, I don't need a video knowing that I'm... I don't use my... Like, I have personal subscription to chat GPT and actually to Gemini.

Although, you know what? I'm really... I really should make a note to go cancel that Gemini subscription. It's so funny because everyone's like anecdotal experience. They have one bad experience. They go look at Gemini and Gemini happens to be better that day. And now I've been using Gemini for the past. But like Bryce is saying that, you know, chat, GPT,

is way better. I have no commentary on the quality of the model itself, but my default has been to use chat GPT for all my chat like tasks. And the times when I've tried to use Google's AI has been when I've needed to do something that involved tool use with Google's product. And it sucks. It's so bad. Yeah.

I don't know. What's your experience, Marco? What are you, what are your, I mean, both you can answer your own question and then, do you prefer Gemini, chat GPT, none of the above? You're just sitting up outside touching grass and enjoying life. I've been using perplexity. For my web-based stuff, I use perplexity because you can choose the model according to what's the best at the moment.

And for me, I mean, whenever I need to look something up on the internet, it's more than good enough. And then for coding, I'm still with ClodCode. I've been seeing that apparently Codex and GPD5.5 is supposedly better, but I don't know, cloud code works well enough for me, and I'd rather just stick with that because who knows, maybe in two months, another model comes from Anthropic, and it's now better than 5.5, and I don't want to be switching between Codex and ClodCode all the time, because I have also my conversations. You're not thinking big enough. The reason why you need to be multi-harness and multi-model is because

when you start using this stuff enough, you're going to run out of your corporate token rate. And sure, I could go and ask for more. But my whole, like, I just want to move fast. I don't want to be down for, like, a day or two or three days because I need to explain to people why I need a hundred million more, you know, tokens or whatever. You know, so, like, I have, like, I now have, like a system, like a cascading system of various different ways that I can get inference from NVIDIA, where, like, the worst case is, like, I go spin up some B200 nodes on Brev and just, like,

serve myself Deepseek V4 Pro or something. Like one of the quantized versions that'll fit on like a B200. Yeah. So yeah, you got to have like a strategy, like, you know, a fallback scheme in case your preferred model is not available. I'm not that big of a token abuser. You really destroy the cloud code tokens instead of like a month? I think I'll, I think I will run out of, I don't think I should say what like the

Nvidia cap is on like dollar spend like by default. But I'm already in the power user group and yes, I think I will expend it because like not not right now because right now I'm just putting together this orchestration system. But once I get back to actually doing like runs of stuff, yeah, I'm going to burn through it very quickly. I think that's a great nickname that you've come up with Marco for Bryce is Bryce Adelstein Lelbeck, aka the token abuser because boy oh boy, is he abusing tokens. That's for sure.

Yeah. Yeah, yeah. But the thing is, you know, Nvidia has this great developer crowd Brev, and it's fairly easy to get access to GPUs in Brev, except just not Blackwell GPUs. And so I have been thinking lately, like, okay, like, you know, I have all this access to this Breb compute that I can use for the various projects I'm working on.

And so, like, in the worst case, I'll just figure out how to run on, you know, like there's a ton of H-100 availability here, so I can just run, you know, some models myself on all these H-100s, and then serve up my agents that are also running on this infrastructure. That's one of the reasons I had to pause to do this orchestration layer thing. And what do you do for Auto Run? Because you can't run Auto Run everywhere, anywhere you like. And I think that's where it really becomes powerful because having to be...

Right, but Brev is one of the approved sandboxed environments for Nvidia. It's not behind the VPN. All of the stuff I'm doing, none of it has access to any Nvidia proprietary information. All the stuff I'm doing is just experiments on a certain thing we can't talk about yet, but all it requires is just my agent to be able to look at code on GitHub. It doesn't need access to anything proprietary. It doesn't have access to any. My security models is that this is running in a cloud instance.

It has no credentials that can do any damage to anything. Yeah, that's definitely nice. Yes. So, I mean, I don't know if we fully answered with Marco's question. Bryce doesn't need to take a break because he's so excited. and... No, I got to take a break because I need to go

to see all this agent what to do. But yeah, I'm 75-25. And I am concerned like when I hear some of the stuff that I hear on podcasts

about just the motivation of certain developers being at like all-time lows. Yeah, that is concerning to me. Can I have a hot take? All right, go ahead. I mean, of course. Actually, wait,

this is the perfect, the perfect time. At one point, like 100 episodes ago, there was some Korean K-pop song, I was supposed to insert whenever Bryce had a hot take, boom. I just inserted it.

And I was talking to one of my more AI forward colleagues at our Shanghai office when I was there two weeks ago. And he's really like a very forward thinking guy. And I think he was really early in a lot of this hygienic stuff in ways that I didn't realize like a year ago. But like now I realize like, oh, he was totally ahead of the curve. And he said something about, you know, I have everybody in my team. is either an intern or a new college grad.

And like, I have slots for senior people in my team, but I don't know how to fill them. Because, like, we're doing all this agenic stuff, and all these young kids, like, they get it. And I don't know, I don't know whether I can find somebody senior to bring in here who's going to be a good culture fit for the team and, like, have the right energy for, like, this, like, new, like, shifting, changing, evolving landscape.

And my hot take is, and, like, him, him and I talked about this a little bit. My hot take is, not everyone's going to make it. I think that there's some people in some systems and parts of our ecosystem that are very resistant to AI. And if that becomes a persistent, stubborn attitude, I don't think they're going to make it. One of the themes of this conference that I've been at has been, so it's a conference about like open source and in the agenic era. Like, oh, and maybe not open source, but it's about open models, open weight. open source in the agenic era, like open AI, basically.

Not the company open AI, but, you know, open space AI. And one thing that keeps coming up is like, you know, our existing open source licenses make sense for, like, in this era. Like, what does open source licensing mean when all the code is generated by these models that were trained on all the code in the world that has varieties of different models? And then like the other thing that came up in a conversation today was this idea that like, open, and this has come a couple times, is open source dead?

You know, could open source be dead? There's a lot of these projects that are saying we don't want open source contributions. Like, no, no, we're not accepting any open source contributions or we're only accepting contributions from people within the team. The trend line that I'm seeing,

I think that saying, oh, we're not, we're using zero AI, we're not accepting any open source contributions. That's like saying, oh, we're not using version control. Like, no, we just, we don't use version control. That's not how we work. It's an unstoppable force. It's a, it's a change in the way that we work. And I think that if projects in systems and communities and people don't adapt, they may be left behind. And you and I, Connor, I think, had talked about this, about this idea that, like, it's hard, it's getting harder and harder to talk to colleagues who are AI skeptics. Like, to the degree that, like, if somebody is an AI, is AI skeptical, I almost can't have a meaningful conversation with them because the way in which I've worked has changed so much, the philosophy of how I've worked has changed so much that like

we're talking about two different jobs. We're talking on two different timescales of what's possible. We're talking using two completely different tool sets. And so it's like if they don't believe, if they just reject everything about what I'm doing, how can I have a meaningful conversation with them? And like, I'm not saying that everybody has to be as AI gung-ho as you or I are. But I think if you just believe that this is something that you can just, oh, it's a trend, it's got to blow over, or that you don't have to have, like, a real legitimate plan for how you're going to adapt to this world. If you just are denying it and think it's going to blow over, I think you're going to get left behind. That's my hot text.

So, I mean, I'm interested, Marco, to get your thoughts on this as well. But, like, so I want to just reframe a little bit of, like, yes, we had that conversation. And let me put my biases if you are for somehow stumbling across his podcast for the first time. Bryce and I, and it sounds like Marco as well, but he can speak for himself. We're massive fans. I love AI. It's making, it's change in my life. Absolutely love it.

And it's, I don't think there are a ton of people that think it's like a fad. and it's going to blow over anymore. I think everybody is seeing they've tried, you know, since basically December of 2025, they've tried either GPT or the Anthropic models and they've seen that these things are useful. I think now, though, it's people coming to the realization, like, maybe we should get this guy on.

So I randomly saw this YouTube video. I think it had like either less than 100 views or less than like 1,000 views when I first saw it. I'm not subscribed to the guy. I have no idea why, YouTube recommended it to me, but I watched it and it was interesting. And it was basically a guy, he's a Zig developer. I won't know his name off the top of my head.

And he has a small YouTube channel with like a couple thousand subscribers. And he had been working on a, I want to say, source control, like an alternative to get plus or minus. You know, the details don't really matter. But he basically quit his job, worked on it for two years and hadn't really gotten the amount of interest in it that he would have liked. And so now he basically just posted. this video saying that he kind of feels a bit lost because he, you know, left his job in order

to work on this thing. And now AI's here and AI's basically replicating like half of the open source stuff out there. Is open source dead? That's an interesting question. Like, I don't think these models would be possible, or at least they wouldn't be as good as code, without the open source ecosystem. So it's kind of like a little bit of. That's certainly sure. Sure. Maybe is open source dead? Maybe. But like that's ironic because the fact that these models are so good is due to the fact that they had a corpus of open source code to train on. But that doesn't mean that open source isn't dead, right? Like, I think I said on a past podcast, like, anyone that's like working on some niche

language in hopes that it's going to like break into the top 20, that that language is not like a AI, you know, better for AI code gen. It's like a lost cause. And I think that's a really sad thing. If you were investing like a decade or half a decade in a language that now it's just like, well, we don't even really care that much about. programming language. Anyways, so I don't think it's necessarily like people that are anti-AI and don't believe it's a thing. I think it's more that like people are, people are anti-AI because they're

worried about like the disruption that it's going to cause in society. And this and like I think in that video, if it wasn't the Zig guy, I'm, I could be misattributing this. But basically it was either that internet podcast where someone said like, I didn't get into software engineering to like prompt LLMs to generate me code. Like I got into it because I, I liked the art of like crafting code. And like now that's not what being a software developer is anymore. Like what does that mean for my career and for like my meaningfulness of like, you know, like if I don't want to sit there prompting and like architecting systems and not actually

handcrafting the code myself, like what does that mean for me? And like I'm not, I don't have the answer. I just think it's an interesting like thing that he brought up that I hadn't really thought about it. I have had my own, like, mini existential crisis of being like, do I actually like programming? Like, I really thought I liked programming, but this new thing is like even better. I can move faster. I don't care about the code so much.

And anyway, so, like, I guess I fit into a bucket of people that manage to transition to this new style. And I'm still happy. You know, I might be a little bit more exhausted. I might be working harder, but I'm still happy. But I think that there's a huge fraction or percentage of developers that are not happy. They're not happy, like, prompting these. LLMs and it's just like what does that mean like should we care? I mean if it's effect if you go ahead.

Again, I don't know if they're going to make it. Well, it's not a question about whether they're going to make it or not. It's a question about like or I guess I guess maybe it's a moot point. You're just kind of saying like it doesn't matter. It's kind of like the nuclear thing or whatever like whether or not we do it or someone else is going to do it. So like you just have to get on the train or get left behind. I think that I think that there will continue to be roles for people writing code.

manually reviewing code, but I think there will be a lot less of that. Well, I guess, but that doesn't really answer the question. The question is, should we care about the people that are sad and that, like, they don't get to write their, like, four loops and algorithms anymore? Yeah. Should we care about our colleagues who have, you know, invested a substantial amount in their career and now maybe skills and things that they're passionate about?

because a lot of people in this industry are very, came to it because they're passionate. Should we care that that may be changing and changing very rapidly? Yeah, of course we should care. Of course we should care. And I think it's hard when we have so much excitement about this to have empathy for that audience. And I think you and I, Connor, probably don't have perspective on it because we are, we're happy to be in this, like, architecting mode. And for me, at least, it's been many years since I was really writing code.

myself because I think we've talked about it before in the podcast. You know, I've had these ergonomic issues over the years that have sort of naturally pushed me into more of an architect role. I couldn't just sit at a keyboard and spend 10 hours a day, you know, building stuff. Now I can do that again. So yeah, I'm very excited about that. I can understand how for other people who are at the keyboard, you know, hacking away 10 hours a day and who really enjoy that, how they could be very devastated by having that taken away. Especially I understand it because that's something that happened to me. I used to really enjoy that, and then I physically could no longer do that. And so I had moments where I was really very much in the throes of existential crisis because I was like, well,

this was my whole career. What am I going to do now? Right? Like, I've been there. I get it. I had times when my ergonomic issues first started coming up where I really thought I was going to have to find another job or another career because I would not be able to be successful because the thing that I was good at, which is writing code, you know, I couldn't do it anymore. But, you know, those fears did, in my case, prove to be unfounded because I was able to, you know, adapt, go into a bigger role, you know, a role where I'm more architecting and leading than writing code day to day. And, I mean, I guess with AI, I ended up working out. But, yeah, like, it was really rough. That period of my life was a very rough period of my life.

and it was very hard on me mentally, in addition to being like the tough on me and painful physically. So yeah, I feel for people who are on the other side of this, I don't know, I don't know how to help folks. And like to some degree, I think that as we progress down this timeline, you know, as we progress down the shift towards this agentic era, I think people do need to reassess if you are not happy

working in the new ways in which we may work, you may need to find something that makes you happy. And I don't know what that is. I hope it continues to be like, you know, the same career path that you still have some stability, but that may not be the case for everybody. You know, if you're not happy working in this new way

and if the industry of manually writing code as a dying industry, then you may need to go find, and if you're, if you're finding it depressing and sad to be in that space, because the thing that you used to do is no longer is valued, then it might be time to go and figure out what else would you be passionate about. I don't know. I'm just rambling here. Well, I mean, I've got thoughts, but I'm curious to hear from Marco. What are your thoughts on all of this?

And also, too, I guess you've started, you've been, like, compressed, right? Like, you started programming for the first time, like, four years ago. And so, yeah, I'm curious. Like, what are your thoughts both, like, personally how you feel about this whole thing and also, like, more in general? Yeah, I think I would agree with you both in that, obviously, programming with AI is the future, and there's no way around that. People that use agents to program are orders of magnitude more efficient and more productive

than people who don't. I mean, just this week, I did something that took me two days of working on it, like, not here and there, not fully on it, something that six months ago, just six months ago, would have taken me a whole week at least. And so there's no way around that. Someone that uses AI is going to be more productive than someone that doesn't. But then I also have this feeling that right now it's very fun and very interesting because of the novelty, because you have so many options, you can do so many things. But I can see a future where maybe the joy will be taken out of it. Because right now they're not good enough that they do everything.

for you. I tell it what to do and then I still need to know how to code myself because I need to look at what the AI has done and then judge whether that's what I wanted it to do or not. And for that, I need to be good at programming because I need to understand what the code is doing and whether the idea that I had is actually what the code is doing or whether there's a better way of doing it. So it kind of takes the boring part away, which is actually writing it, but I still have to have the taste in what good code is and have the good ideas. But maybe, in a year from now or five years from now, it's already better at me than that. I'll just say, I think you said something really key there, and maybe didn't even notice

that you said it. You said it took the boring part away, which was writing it. And so I think maybe that is like, you know, we view it as boring, or I wouldn't have necessarily said that writing code was boring, but like I was mainly doing it in C++. And I wouldn't say it's boring. I always said it was painful. And that's one of my theories is that, like, depending on the language you were coming from, if it was C++ or C++, you love these tools. You never, you don't have to sift through, you know, template error message. You just put it in a loop. The AI figures it out.

Whereas if you were maybe a rust or a ruby or some language that was really pleasant to use with a really good ecosystem, that the fun part, you know, was writing the code. It wasn't necessarily, you know, and so you take that away. And anyway, so I feel like we're all in a group of folks that, you know, don't miss the typing of the, you know, the syntax. But I think some people actually really enjoyed that. Anyways, Bryce, you were going to respond to. But yeah, I think the question is that, will in five years the AI be better than us at basically everything?

And so where would our place will, where will our place be in that? Yes. So I'm recalling what I was going to say now. Yeah, significant amounts of babysitting are needed. You know, Connor, for a long time, there's been this trend where we've come on and talked about it. And I've, we've both been excited about AI. But my perspective has been like, oh, this this freaking thing. it keeps like I get so frustrated at it.

You know, I think we've talked about it a couple episodes about my like frustration and like I'll frequently I'll be, I'll tell it like, dude, come on. Like, do you think that was real? Like, I'm like passive aggressive with it. And I know this is not the best prompting style, but at this point, I just, I got to be me, you know. You know, like I'll tell like, you know, take this, take this crap out. Like, you know, what are you doing here? And like, when it, when it violates like agents.md or like some rules, that I've like clearly like written down.

I'll like do like what you do as small children. Like did you read the directions that I gave you? Oh, you did? Why didn't you do them? And that makes me feel a little bit better about it. But yeah, I don't know how it's interesting. There's like a duality here because on the one hand, there's what I was talking about earlier with.

My colleague was saying that like everybody on his team is a new college grad or an intern because they get it. They get how to use that. And on the one hand, like, yeah, I think, like, you need to have the right mindset, and young people are more likely to have that right mindset. But on the other hand, the crimes that I have seen committed that I was able to catch because I have, you know, 10 to 15 years of experience, like, you know, I'm building this little, like, orchestration layer thing. And, you know, like, security is a

concern. And, like, this has been, this was pretty clearly laid out in the design document. So I, I don't read a lot of the code, but I spent two days just iterating on the design document with the AI. I had it draft it and then I had it redraft it and redraft it until like I was every line of the design document I read. Like every bullet of the design document I was comfortable with. Like I reviewed that very, very carefully. And like, you know, in particular, like the security section. Because again, this thing doesn't have access to any, you know, proprietary information. It's just like it's running in a sandbox. Nothing really that terrible can go wrong, but still, you know, best practices. And, you know, there was a very clear plan for

what stuff it needed to expose to the outside world. And it just completely ignored one part of that. And, you know, sure, if I had read every line of code, I would have noticed that, maybe, but there's no way I could have read every line of code. And so the reason that I noticed it is that, like, eventually I caught on to like a subtle detail about like the architecture when I was talking to it or like I was asking about some like subtle thing. And I just had an intuition where I was like wait, hang on a second. If it's saying that this works this way, then that means that like X, Y and Z and that almost certainly has to be super insecure. And keep in my, I have no background in, you know, software security, web services, etc. I just like have a bunch of general industry experience. So I sort of have a sense of like, you know, what could be a problem, what would not be a problem. I know it. I know it's least what I don't know. But somebody who's less experienced and who doesn't have all of the like war stories and scars that, you know, a more senior dev has, I don't know. I feel like they would maybe miss a lot of these things. But maybe it's the sort of thing where, you know, in the old day, in the old days, you would make those mistakes yourself. And then the horrible thing would happen,

you know, the security bug, the corrupting your database, etc., would happen. to you early on in your career, and it would be a learning experience. And maybe now it's just that, like, it's a lot quicker for you to make those mistakes, and maybe also you learn faster from it. But, like, I think that the people keep asking about like, okay, but like, how are people, how are young people going to learn if they're not writing the code themselves? Well, I think that if you just go in, like, vibe code something and it's like a huge big mess and, like, has a bunch of subtle bugs in it, like, you're a very very very very. You're a very going to reel a lot. Like, that's going to be a learnable experience, too. I don't know about you,

but like early on in my career, I wrote some like messy, you know, piles of garbage too. And like, I learned from that. And I don't think that's any different. Like, I don't think that the learning process is that much different. Like, you're going to, you're going to make mistakes. You know, this is a thing that lets you make mistakes at a much faster rate. But you're, everybody's going to make mistakes as part of the learning process. Yeah. And I guess it depends on the kind of person you are. but like one of the things super early on with these tools that I started calling it or referring to them is it's the democratization of expertise. And I don't mean that with respect to code. I mean that with respect to a lot of other white-collar professionals that you might need at some point in your life that you have to pay money, you know, real estate agent, lawyers for real estate, taxes, etc.

If you are the kind of person that wants to dedicate a Saturday to becoming an expert and, you know, insert whatever, you can. Like this year, I basically filed my own taxes alongside my tax advisor. So, like, I have my tax advisor do it. But then at the same time, I was like, I'm pretty sure with these tools, you know, I can figure out how to file all the correct forms. Sure enough, like, I did a better job and, like, caught a bunch of things that, like, my tax advisor didn't even catch because I would just sit there with the model and be like, listen, this is my situation, like, enumerate the list of things that I could possibly be

filing as, like, taxable income reductions. And, like, your tax advisor that you're paying whatever X hundred dollars to, file for a single year is not going to, you know, I mean, technically I guess I could take a meeting with them and then I could pepper them with questions. But like in my experience from the people and the relationships that they have with their tax advisors, you don't get that kind of like service unless if you're going to some like bespoke place. Anyways. And so like I've done this like next year I plan to file my taxes on my own, you know, and there's like a ton of stuff where you can basically teach yourself very rapidly like the expertise that you would offload to some other professional. I mean,

there's still some things that you're going to need actual actual people for. Be sure to check these show notes, either in your podcast app or at ADSP thepodcast.com for links to anything we mentioned in today's episode, as well as a link to a get-up discussion where you can leave thoughts, comments, and questions. Thanks for listening. We hope you enjoyed and have a great day. Low quality, high quantity. That is the tagline of our podcast.

It's not the tagline. Our tagline is chaos with sprinkles of information. Let me give you some financial advice. I disagree. I disagree with your advice, but go ahead. There's this quote about like, God, let me see if I can find it. There's this quote about like a computer cannot take responsibility for something so it should not be allowed to make a decision that I think is kind of applicable to the idea of having to file your taxes.

The reason I pay someone to file my taxes is that if something goes wrong, I have a person I can blame. And, you know, some, some potential ways of getting restitution if there's a very serious error in my taxes, right? That, that I have some, I have some insurance and protection from a serious error. And if you file your taxes with AI, you don't have that, that protection. And so I'm willing to pay the money for that, not, like, regardless of whether or not, of how correct or not correct my taxes are. Yeah, sure, sure, maybe they'll miss things. maybe AI will help me identify things that they missed in the taxes. But, like, that's actually even more of a reason for me to continue paying them.

Because, like, if they do make mistakes, like, you know, I can point that out to them. But, like, you know, if something, if I made some very serious error filing the taxes on my own, I'm the only person responsible, and I have no course for restitution. I don't know. I don't really understand. And like I said, I do disagree. But you're saying, but, like, so what's the worst case that people are going to accuse you of committing tax fraud and you're going to end up in prison? No, like, like.

What are you worried about here? If a tax error occurred that cost me like, you know, tens of thousands of dollars or hundreds of thousands of dollars, like an accountant filed my taxes and led to that error, like I have various ways through the legal system that I could go about recouping to some degree my costs. So basically you're going to say if there was taxes that you owed, but because of your advisor, you didn't pay them. You're now going to force your advisor via the legal system to pay it for you. It gives me, it gives me somebody else who's a bad look, man. It's a bad look. It gives me somebody else who is culpable, somebody else who is on the line,

who is responsible for doing this thing. And that if they make a mistake. We might make this tax section bonus content after the outro. It might be a 20 minute after the outro. But, uh, please, please don't, Connor, just please don't rely on your, Please don't rely on AI solely to file your taxes. No, no, no, no.

That's the thing. I file my own taxes. Like, I do all the calculations. I'm just talking to the AI. Like, there's just like, honestly, I think the... I take it back. I would rather have you use AI to do all the taxes than have you spend time calculating the taxes yourself.

No, no, no. And that's the thing. I'm a former actuary. Like, I love spreadsheets. Like, I used to love doing my taxes before my tax situation got really complicated. And honestly, like, filing your tax. is incredibly simple if you have a simple situation. I do not have a simple situation ever since working for Nvidia because they're an American

corporation. And like the way I get compensated, it comes through like different various whatever compensation mechanisms. And so it got more complicated. I offloaded it. But now with AI, like, and the main thing is, is like there are certain specific forms that like if you do not fill out that form, it can cost you like ridiculous amounts in terms of like fines.

But like once you make sure you got all those forms, and there's no worry anymore. There's no worry anymore. So that's the thing that like, I guess in defense of you, although you're working in the States for an American company. So it shouldn't be that tricky. But it's that if there was like a fine that you ended up having to pay, not taxes that you owed in the past, but like a fine that they're levying on you, then I guess you could go after the tax advisor. But anyways, we got way off in the weeds because I mentioned taxes accidentally. It just happens to be April just passed. April 15th is for the Americans.

April 30th is for the Canadians? What are we? Are we done with AI? Are we talking about GPU rotate now? Are we answering questions from the people? There was some other third topic on the stack. The MVP comp. Yeah, what are we taking this conversation?

Yeah, we should rotate to GP rotate. Yeah, maybe what we'll do too is this episode that just happened, because that was at the what, 50, 46 minute mark. I'm sure we're going to lose listeners over this episode. That's all right. the tax stuff is now bonus content it's going to be after the outro and I'll put some little thing and we got a little get lost in

the you know a tax conversation if you care you know listen if you don't feel free to stop. That one's going to come I'll just say this the point I'm making with the tax situation is less about the specific example of the tax situation. It's more about the general principle of why I there are

certain things that I prefer to pay a professional money for even if I could do them myself just because paying the professional money for it gives me somebody else who's responsible where if there's a fuck up, I can go to that person and there can be restitution. I see. Have you. I mean, if that's if that's the way that you operate, I understand. My thing, though, is I think that there's the only person that will maximally care about like your financial well being, like maximally care about it is you. As

soon as you start outsourcing to someone else, like they are now looking out for themselves, first and foremost, not for you. Like, maybe you have an amazing financial advisor, tax advisor, whatever other advisor, but like, or maybe if they're in your family or they're a good friend. What? Trust but verify. Well, that's the thing. If I'm going to verify and do their whole job for them, why am I paying them in the first

place? Because it's less work for me to verify a work product by somebody else. This is the same thing with AI. Like, instead of me writing 10,000 lines of code, it's less effort for me to review an audit a 10,000 lines of code, to review a PR from my tax accountant or my lawyer or whatnot. I'd rather do that than have to go and do it all myself. I don't know if, I mean, in certain cases, I think that's true.

In the case of this specific example, my, like, verifying was basically just, like I said, replicating the whole process myself. Anyways, we're down, we're down. Nobody cares. Nobody cares. Did we lose listeners over the AI hot take? Certainly.

100%.