Bryce: 00:00
And so your your estimation of how rapidly this will be put into production across the industry, your your your meter may be a little miscalibrated. And two, I think because we work at NVIDIA, we're in a bubble and we don't realize how early things are. But like like generally, I think you are more optimistic about what the current model's capabilities are. I am constantly frustrated by what the current model's capabilities are, but can still get a lot done with them. And this is the reason why the C evolution process is just not competitive. C like a three-year cycle does not reflect the rate at which our industry is changing. Think about how much our industry will change in three years. So the next C standard will be C29. Given everything we know about the revolution undergoing our industry right now, how can we possibly expect to get anything meaningful out of C29?

Conor: 01:10
Welcome to ADSP the podcast, episode 301, recorded on August 12th, 2026. My name is Connor, and today with my co-host Bryce, we chat about my existential crisis due to AI, AI fluency, the potential death of conferences, auto-auto research, and more. Alright, I mean, should we pick up where we left off? You believe humans still needed. I am doubtful. Tell me your topic. Alright, my topic, I mean, my topic is dovetails off of, you know, if we assume, you know, humans aren't really needed for coding, you know, what happens? I've been spending a lot of time th thinking about this like post-AGI, you know, kind of world we live in, and it's the title of the test episode that I had in my head, which may or may not change by the end of this recording with you, is The Illusion of Necessity. And in a nutshell, or a larger nut, with or without a shell, the idea is that I've come to realize that like technical conferences, to tech technical committee meetings, a lot of stuff is just, you know, it's like that Matthew McConaughey scene from The Wolf of Was Wall Street. Where he beats his chest, it's a Fugazi, it's a Figazi. It's all a Fugazi. You know what a Fugazi is? Fugazi, it's uh fake. Yeah, Fugazi, Fugazi, it's a wazi, it's a woozy, it's a f fairy dust. It doesn't exist, it's never landed. It is no matter, it's not on the elemental chart. It's all just kind of we just made it all up, you know? Do you know what I mean? It's like Microsoft, Bloomberg, all these companies that use C. They are companies that created value for some definition of the word value for the world, you know, they're they're profit-generating entities, and they used C as a tool to generate that value. And so, you know, they started saying we should get together in person because if this is a tool that you're using to generate value, you can justify sending people to these places to meet, even if it isn't necessary. But it's like, you know, these people are creating us value, we gotta keep them happy. Let's send them to Hawaii, let's send them to whatever, all over the world. And and then like these changes in the language, which some definitely are good, you know, move semantics, you know, helps speed up programs, etc. Not all of it's bad, but you could argue that 50% of what the C committee does in terms of library development, which I know, you know, you were library chair for a while. I don't I don't mean to shoot on your past, but you know, arguably we could have just had a package manager and called it a day, you know? So like it's like these meetings lead to language changes and library standard library changes, which then lead to an ecosystem of conferences where people are meeting to do conference stuff, and it's all just like, you know, other communities evolve their languages without, you know, meeting in person. And it's just what what's the whole what's my whole point? Is just that like I think the real reason that all of this stuff ends up being created is just to keep people happy. You know, there's two people that I will not name that I have an immense amount of respect for that said the C committee is just a social club. At the end of the day, that's what it is. It's people that like getting together to talk about technical stuff, and yeah, I I I do not I do not disagree with that.

Bryce: 04:44
The pandemic was a great opportunity. I said ever since I joined the C committee, I found the process of language evolution to be archaic and just burdened by process and acronyms and unwritten rules and guidelines and procedure. And I tried my best to modernize and improve it, and we had a real opportunity with the pandemic to modernize our process, to move to asynchronous work. And a large part of the committee was like, nope, we should just shut down and not do any work during the pandemic. We can only possibly evolve this language by meeting in person. And I mean, uh the the the argument on the other hand is that asynchronous work, not everybody can make time during their regular job to do asynchronous work, and that there is there is an in there is a value to the intensity of and the focus of everybody being in the same place at the same time, and that when you're there, you are 100% focused on the work that you're doing for the week. And I do think that that that that is true, that you do end up 100% focused for that time. However, I've never found the argument compelling that people can't make time asynchronously to work on the C ⁇ standard throughout the year. And I think that, you know, there's a large there's a large part of the C ⁇ committee that that for whom it is a hobby and not part of their day job, and those people would be disenfranchised by moving to a different process. But I actually think that was beneficial. We had too many stakeholders in C. The people who should be making the decisions about C should be the people for whom working on C is their primary job. So the implementers and the other key stakeholders in the language. Like that would have been better. I think that would have been a better model. But during the pandemic, there were many opportunities for the committee to modernize, and the committee rejected almost all of them. And in fact, many, many things, innovations and improvements that were made during the pandemic, my impression is that they have been rolled back, and that now we have moved back to more or less the same process that we had beforehand, which was entirely in-person meetings. And this is the reason why the C evolution process it's just not competitive. C like a three-year cycle does not reflect the rate at which our industry is changing. Think about how much our industry will change in three years, right? Like the next C standard will be C29. Given everything we know about the revolution undergoing our industry right now, how can we possibly expect to get anything meaningful out of C29? Like a a three-year cycle, right? Like how is the process that we've previously used for evolving C ⁇ , how does that make sense when everything is changing? So no, I I I don't I don't think that C ⁇ will continue to be relevant for that reason. The the the length of the cycle isn't it the length of the cycle is an artifact of the archaic process that is used to evolve C ⁇ . And there has been no it's very hard to get people to vote to disenfranchise themselves, but that was the only way that that could have happened. But the industry's moved on. I don't think we C ⁇ does not have the same level of priority that it used to have. And the the the barrier to entry for language and compiler evolution is a lot lower now. And at least for my industry, for the the compute industry, most of the interesting work doesn't happen in C ⁇ . It's a legacy language there. Most of the interesting work happens in in dialects or or compilers, and Python dialects or compilers these days. Something we've talked about a lot on this podcast before. But like, was it at all just a social club? I think there was a time in the 90s in which the the the mo the collaboration model of the C committee was the only model that could have worked. So it it literally started like at the beginning of the internet era when we did not have the tools that we had today. So, no, it it was not always just a social club, but the fundamental s flaw was that the committee modernized C but not never modernized itself. And that is why it has failed.

Conor: 09:36
So I want to separate two things here because there's the C committee as a is it successful, is it not successful? We both agree it's not going well, not going the right direction. That being said, you know, a part of my whole existential crisis thing and realizing that a lot of the ecosystems that exist in the world, technical ecosystems, whatever, other ecosystems, you know, we've created them, and a huge aspect of them is the social component. Whether or not you agree that like the C plus plus committee is like primarily a social club or not, you know, you can agree, disagree. But the thing that I guess it leads to my existential crisis is that I I think social clubs are good, you know? Whether or not they should be, you know. Oh, there's a fire truck in the background. Whether or not, you know, they should be whether the C plus plus committee that has the task of evolving the language and the library, and a lot of people are paying a lot of money to fly these places, you know, is a side point. It's more that like there's a part of me that's like sad because, you know, the I I don't go to the C plus plus committee and haven't for years for other reasons. And I still go to conferences, but are conferences gonna disappear too?

Bryce: 10:57
I think language No, no, I actually I I said this at PlatziConf in Mexico. Gave uh it was actually a very interesting talk about mentorship and sort of my career journey and what and and how to volunteer and get involved in tech communities. One of the points I made there was that in the current age where every internet social media platform and forum is filled with AI-generated content, I actually think that there is a higher value placed on in-person events because it's more genuine. It is hard to filter it the signal-to-noise ratio on social media platforms, on online communities, is a lot higher than it used to be. And so if you're getting started in this industry, how are you going to make connections on the internet? I think it's harder than it used to be, and I think that there is still a great value in in local events and in conferences. And one of the points I made is like, yeah, if you want to get started in tech, it is important to have a network. And the best events are the events that you can get to. So you don't necessarily need to go to a big conference if there you can find a conference or a meetup in your community, and if there isn't one, you can start one. But I do think that these have a great value, maybe even more of a value than they had in the post-pandemic pre-AI era.

Conor: 12:38
To a certain extent, uh, I agree.

Bryce: 12:40
I guess the part I'm worried about is like Hang on, you you're every time you're talking, you uh you've got to say that again, you broke up.

Conor: 12:47
I've got my recursive self-improvement agent running though.

Bryce: 12:51
But we can start it up again once I What is it improv what is it recursively self-improving?

Conor: 12:57
Uh well I've I've I figured I've got my little uh kernel agent program, and you know, I'm trying to get it to improve, and I just figured I was like, you know what, why don't I just put my loop inside a loop and tell it to improve itself?

Bryce: 13:11
Yeah, that's I've I've been I've been looking at doing that, at doing auto research. Yeah, I've been looking at that recently.

Conor: 13:17
This is like auto-auto research though.

Bryce: 13:19
Yeah, yes, that's that's sp that's specifically what I've been calling it is auto auto research. Yeah, it's very uh compute intensive though.

Conor: 13:26
Yeah, I haven't cracked cracked it the code perfectly yet, but it is very colorful to look at. Anyways, I mean the question is is like, you know, there are conferences that I want to go to because it's fun. But I I you know I I I feel like I should be careful with the words that I say, you know, because it's not like I don't want to get like I don't want to be uninvited from conferences and stuff, but I just I don't see the value in going to a C conference or going to Oh yeah, I I haven't I haven't been to a C conference in a while.

Bryce: 13:56
I I'm going to a GenTIC conference.

Conor: 13:58
Yeah, but that's the thing. It's I was gonna say like insert any programming language conference really is like I just yeah, I I don't I don't really see the value from a like technical perspective. Like I I could make that argument two, three years ago and it's also enjoyable to go to the conferences, but now I'm just like if it's not AI related, but then it's kind of sad because then like you're you're kind of like not abandoning a a community you were a part of, and this is like the whole like I I feel like I kind of bouncing back and forth and not coherently make my point, but like I think that community is like super important, is a it's one of the outcomes of my existential crisis, is that like being a part of a community, whether it's technical, whether it's like friends and family, it's very important. It's like it's like part of the essence of being human, in my opinion. If you don't have like a community, it's like that's what leads to depression. Or I'm sure like one of the things that contributes to depression. And so, like to be in this moment where for a while I was going to either functional conferences or C conferences or Python conferences, and then realizing that like there's a huge shift in the industry that I work in, which is now kind of like rendering those communities, like I it sounds harsh to say obsolete, but like that's kind of how it it feels a little bit like. And anyways, it's just coming to turns. That's why, like I said, like the illusion of necessity, like were was it even necessary before? Not really, but like you can justify meeting up because it is good for continuous learning, and it's very it's very healthy to like uh especially if you're a remote.

Bryce: 15:40
Nah, I I I still I still get a ton of I still get a ton of value out of interacting directly with our with like users and with tech communities at conferences.

Conor: 15:50
But like you said, you said you started going to agentic conferences now.

Bryce: 15:53
Not C ⁇ conferences. Yeah, but like I still go to Python conferences, I'll probably still go to some Rust conferences. Like the events that I've gone to changed a bit, but and like the focus has changed. Like I used to go to conferences to go for the talks, now I go almost exclusively to to like talk to people. But like I I don't know, I may I may I may go to less events in the future just because I feel like it's cutting like I feel like I have more to do as an IC contributor now than I have had in a long time, and that feels more impactful, and I want to spend more time doing that, and I don't know that and and the amount of travel that I do cuts into that. So I may like I've turned down speaking opportunities uh through part of the rest of the year. I may I may do less travel in the future.

Conor: 16:41
So like rattle off to me, because I know you were at a couple of python events, like what are the agentic conferences you've gone to, and like how have you been invited to speak at these? Have you applied to speak at these?

Bryce: 16:53
I went to Gosim. I went to Gosim Paris uh Paris. There's Agent Conf that's in Poland that I heard about. There's another Gosim that I I might go to. There's a there's a bunch of these conferences out there. There's a bunch popping up. I don't know which ones are like necessarily the most popular. There's like the PyTorch conference, the uh ML Sys, like ICML, like all the big AI conferences. Like, I think those ones are still like you know super important and relevant to go to. And like you do also have to keep in mind that like the we're still very early on in the AI era, and uh it's gonna be like this is an industry that moves very slowly and is a lot of inertia. And like, yeah, like things are moving fast now, but if you look at like adoption, like adoption is still relatively slow, and like I I continue having cut conversations with people, like one of the biggest takeaways I have from going to conferences is talking to people and realizing the disconnect between where I am on AI and where other people are on AI. I did this thing that I may try to turn into so I I have a vision. So I I I I let me let me take a step back. I don't know how to teach CUDA anymore because I I I for the last 18 months have like been the person maintaining and developing a lot of the CUDA educational content that's on NVIDIA's accelerated computing hub. Like I rebuilt all of our CUDA Python stuff from the ground up. I sort of took on the maintenance of the CUDA C stuff that our colleague Georgie maintained, and I I've sort of been the steward and shepherd for this content, and I've been doing a lot of teaching and training sort of prior to the long horizon agentic revolution at the start of this year. And then since that start of that revolution, my interest in and focus on the educational content has decreased a lot because it just feels so outdated, like popping up a Jupyter notebook and having people like type code into it. And so I I did this thing at EuroPython and EurosciPy Sprints that was kind of impromptu. And I have to give a big thank you to Modal. Modal is a serverless GPU provider, so they provide GPU compute for inference, for training, and also for like sandboxes and notebooks. And so instead of you like renting a GPU node for the where you're charged per hour or whatnot, they let you like spin up a GPU environment where you can like run one function on a GPU. And they have this checkpoint and restore thing. So like you can run one function and then like go do other stuff and then go run another function later in the same environment. And what they'll do is they'll cache your environment somewhere and then then they'll restore it later, and they can restore it very, very quickly in a matter of milliseconds or something like that. Anyways, modal people were at EuroPython, and I basically propositioned them on like Friday morning for a pile of credits to go do this sprint, and uh they were able to get me set up very quickly, which I greatly appreciate. And uh what I did for the sprint is a prototype of what I'm I'm gonna call Auto Research Academy. And our listeners may or may not know about GPU mode. GPU mode is this kernel coding compet this kernel GPU kernel programming community, and they regularly run these GPU kernel coding competitions. And one of the reasons to run these GPU kernel coding competitions is that they are a great way to generate high-quality data on what good GPU kernels look like and what optimizing U kernels look like. And that data is very useful for people that want to go and train models to be better at writing CUDA code. And the kernels themselves can be very useful. Some of the kernels generated by this coding competition. So the this training data or this data from GPU mode can be very valuable. One, the kernels themselves are very valuable. Some of the kernels developed for these coding competitions have influenced things that have gone into open source libraries or NVIDIA products. The data itself is published openly, and people can use it when they're doing research on developing new open models or using it in their model training to make their models better at writing GPU kernels. But so we did a bunch of contests, the GPU mode did a bunch of contests over the summer, and those contests were all focused on these linear algebra kernels, so QR solvers, Chalesky solvers, and eigenvalue solvers. And one of the contests was ongoing during the EuroPython and EuroSciPi sprints. So the premise of the sprint was you're gonna come and learn how to compete in these contests, and you don't need to know anything about GPU programming because what you're gonna learn is how to set up an agentic loop to compete in these contests on how to do auto research. So no GPU knowledge is required. And I just gave people, you know, an inference key for NVIDIA's inference hub. So I just like set up an inference key for seven days. And I just, you know, told people just go wild. Do do whatever. And we had, I think, eight or nine people show up in total. And it with the sprints are over a course of two days. And it was absolutely amazing. Some people had never done any GPU programming before at all. Most people had never used AI in this way. They had maybe used cursor, some of them had used Claude Code or some form of agent, but nobody had done long horizon tasks. Part of the reason for that is because, you know, it's expensive. A lot of these people were using their own personal subscriptions, or some of them were telling me, yeah, we're not allowed to use AI at work at all. Like, you know, I've used it for some personal projects, but we're not allowed to use this in work. So they've not really had a chance to use it in anger. And this is one of the reasons why, like, I specifically wanted to give them access to like unrestricted access to a variety of frontier models because I wanted people to have the experience of what you can do in that setup. And I gave everybody a sandbox and the inference keys, and then we had these modal credits, and I just let people go wild. And I think I tweeted about it that like every like we had four or five people who sort of got working loops up and running and submitted, like placed in the the between top 10 and top 20. I I ironically, some of the people who did best were some of the people who had the least knowledge about CUDA. And I I attribute that in part to unguided versus guided. That the people who knew about CUDA were more likely to guide the model to say do a specific thing to keep the human in the loop versus the people who had no knowledge about CUDA had no option but to do something unguided. But everybody did really well. Everybody had submissions, like everybody who stayed for both days had submissions on the leaderboard, and a lot of I think it was very rewarding for everybody involved. And I came away with that, with this idea that, like, okay, this is how you would teach CUDA programming going forward. And that's where the idea for Auto Research Academy comes from. And the way I would do auto research academy is as follows. So it would be two hours a day in the morning at like a week-long conference like GTC. And it instead of being a tutorial, it would be more like a study group. So you come and meet for two hours in the morning or so, and part of it is set up in teaching you, okay, we're gonna, here's how we do our agentic loop, here's the problem that we're trying to solve. It would be some sort of GPU mode-like problem that everybody would be competing in. So some part of it would be set up, getting everybody onboarded. You would get access to inference, you would get access to compute, you'd get access to sandboxes so that you could play around with this. And then the other part of this two hours every day would be a lecture. But the lecture wouldn't be on coding. The lecture would be on theory and system fundamentals. Like have somebody like Stephen Jones give a lecture on like the basics of CUDA, of like how a GPU works. Or like on the next day, have somebody give a lecture on the math behind the particular problem that we're writing a kernel for. But none of the lectures would have anything to do with code. They would all be about the theory. And as the week goes on, in the mornings, you know, you would you would basically you'd come in and sort of it would be a check-in on like, how are things going? What problems have you had? Because in between each two-hour study group session, your agents have been working on and doing stuff. And so when you'd come to check in, it would sort of be your time to like report in on your progress and like chat with people who might be able to help you if if you've run into some weird problem or something. And yeah, that this is this is I think the future of how I would teach CUDA. Yeah, I would if if I I will pitch the idea for GTC 2026. I already know, I already know like the five speakers that I would have speak, and that would be a pretty great lineup. It was it was really something else. Like, you know, we had some students who come and, you know, the thing is like you think like, oh, but if the agent's doing all the work, like what do they learn? Well, just people couldn't be more wrong about that. Because I guarantee you, if I had pitched a sprint of like come right CUDA kernels, half of those people wouldn't have shown up because it would have been too intimidating. But instead, I specifically said, you don't need to know this because we're just gonna learn how to automate it. And yeah, like they didn't write the code themselves, but after the two days, you know, on the second day, when people started having results, like we started talking about like, okay, how do we analyze these results? And like, what's the actual problem that we're solving? Like, what's the math behind the problem that we're solving? And like, what does a good solution look like? And what like what is your agent doing versus your agent? And we started like looking at like, okay, well, you know, to understand why this worked, then you need to understand this piece of how GPUs work. And then I would show them this is how this aspect of GPU programming would work. And sort of like because they it was their agent, they felt like an ownership of it, you know, and they felt motivated to try to understand what it was doing. And so, like, instead of learning by doing something yourself, you have the agent do something, and then you, the human, had to try to figure out what did it do? And did it do something valuable? And like, why did this work and why did that not work? And through that process, I think those people learned more about CUDA than they would have in other settings. And they also learned a more valuable skill than writing CUDA kernels. They learned how to automate writing CUDA kernels and then how to evaluate the results of that automation. And that is the skill I would want to teach people going forward. So, yeah, like I think I think I would delete all of the current way that we teach, and that this is what I would teach going forward.

Conor: 28:27
Yeah, it's it's funny that you've had this whole thought exercise because I've not in the exact same articulation, but once again to come back to the existential crisis, I've been thinking like if programming languages don't matter, like all of the insights and valuable things I feel like I have to talk about are like in that space. And then I got I was even talking to my wife, Shima. I was like, you know, like so, like, sure, okay, you find a new AI community or whatever, but like, what am I gonna talk about? And and then she made me realize that like I don't think that like what I now call like AI fluency is what do you call it? Like, like I don't think I have anything to say there, but the truth of the matter is is like most people aren't even close to being like AI fluent, and like both you and I are, and that itself is like, and that's what you're describing, right? Like putting people in front of these agents and like having a workshop over a week, and then like you don't realize like how much you've internalized how to use these things well, and um you know that I'm not gonna go start like a company called like AI Fluent, you know, Academy, but you know, if I were gonna go do something, like it would be an idea, right? Because it's like on one hand, you're correct that the adoption rate of the larger industry outside of tech, but just like the white-collar spreadsheet intellectual work has adopted this stuff very slowly. On the other hand, though, I think that it's gonna like it's gonna be like a tidal wave of disruption way faster than people think. Because if you're a company that doesn't adopt this stuff, you're gonna be displaced by a new competitor that is operating this way because it's just impossible to compete. And yes, the the large Microsofts and Salesforces and whatever, they're still gonna have the way that they operate, but just like I don't think in five years, like the landscape is gonna look anything remotely close to like what it looks like right now.

Bryce: 30:31
I think it's gonna be I think it's gonna be a I think the long tail is gonna be a slower journey than you might think. I think that it's gonna be a long time before it's saturated. I mean, it was honestly the the most it was the most valuable part of EuroPython and EurosciPy for me was this experience and just the degree to which folks had not tried this before. And like I I I'm constantly running into people who've not tried or have not had the opportunity to try before, and who I I push to give it a shot, and then I check back in on them three or six months later, and they've their workflows have changed. And like the like in part because I I convinced them to to to try. And I I just think it's gonna take a longer to drive adoption than than you might think. I think that this this the frontier will continue advancing, but the frontier is a small group of people. And in in right now, it is very expensive to try stuff out. Like I do think, I do think to get good at this, you've got to make a couple $10,000 mistakes. And not everybody has the budget to make $10,000 mistakes. And that's not gonna be true in two to three years, because the frontier intelligence of today will get more commoditized, and the like Opus 4.5 level intelligence, maybe like OPIS 4.8 level intelligence, is really good enough for a lot of these long horizon tasks. Like those are the models, the models starting like January, February of this year are the models that really unlocked long horizon tasks, which have been the big game changer. And those capabilities will get commoditized and token prices will go down as new hardware comes out, as new open models come out, token prices will go down, and it will it will not always be the case where you have to make a few ten thousand dollar mistakes to really learn how to use this stuff. But at least right now, you gotta make a couple ten thousand dollar mistakes to learn how to use this stuff. I I do think there's a huge value to giving people an environment in which they can make but in which they can make those mistakes in not $10,000, but maybe a thousand dollar mistakes. In an environment where they can they can feel like they can experiment without having fear about the repercussions of failure. And I think that's one of the problems that the people have right now is that either they don't have access to tokens and frontier models, or they do have access, but they're afraid to to make those big mistakes. And uh even you, even you were afraid. I I don't I don't worry about my inference bill that much because Oh, I th I thought I was gonna say I was afraid of making mistakes.

Conor: 33:20
I was like, what are you talking about? But the yes.

Bryce: 33:25
You were afraid of you were afraid of your inference bill, right?

Conor: 33:28
Yes, and still am, but uh Yeah.

Bryce: 33:32
I'm at let's see, I'm checking right now. The last few days has maybe made it worse. Let's see. Why is location not Berlin? Why is this a filter on the AI spend uh dashboard that there's a filter that's on by default that is location is not Berlin? 1.3 billion cursor tokens, 50.1 billion Claude Code tokens, 63.9 billion codex tokens, and then 256 billion tokens on our internal inference hub. So that is 256 plus 63 plus 50 plus one. So let's say it's 370 billion tokens that I've spent year to date. Which is a lot. It probably puts me in the top, I don't know, the top hundred of Spend It NVIDIA, probably. But but like there's there's stuff that I've learned in that process that I don't think I could have learned without doing things at scale. Like, certainly I've made mistakes, but if I didn't make those mistakes, how would I have known that they were mistakes? You know? If I didn't go and try stuff, how would I have known? And there's certain there's certain behaviors that only emerge like at scale. And like there's certain stuff that only happens like when I'm a hundred hours into an auto research loop. And there's just there's no way to get a hundred hours into an auto research loop without spending a couple billion tokens.

Conor: 35:14
Anyways, we gotta cut everything we just talked about for the last 20 minutes. And uh how do we wrap up this episode? Because I probably need to go soon.

Bryce: 35:26
How are we wrapping up this episode?

Conor: 35:29
I don't know, man. It started off uh this wasn't the episode I thought it was gonna be. What did you think this episode was gonna be? I don't know. Like I said, uh it started off as uh existential crisis, conferences, committees, fighters, figazi, and then we started talking about being AI fluent and like auto research.

Bryce: 35:46
Uh I'm gonna diagnose you. I think you have two problems. What one, I think that I think that because you are you you're in the research side of things, you've not been exposed to like the difficult, like the challenges with like deploying this stuff in scale that other people have. And so your your estimation of how rapidly this will be put into production across the industry, your your your meter may be a little miscalibur, calibrated. And in two, I think because we work at NVIDIA, we're in a bubble and we don't realize how early things are. But like generally, I think you are more optimistic about what current models' capabilities are. I am constantly frustrated by what the current models' capabilities are, but can still get a lot done with them. So I maybe, I think that the current models are not are not as good as you think they are, and not as many people are using them as quickly as you think they are. So the the skills that you have are still valuable. I also think that understanding the theory and the systems behind software is a more valuable skill now than it has ever been in this industry. Like, this is maybe the only time that a CS degree is actually maybe really helpful because there are like it for strategic decisions, a human is still going to be involved. And so, like, understand like your knowledge of algorithms, your knowledge of of like programming language theory, like all those things are still going to be valuable in how you direct and use agents.

Conor: 37:31
I don't know. I th I think that like the fifth interview of the classic DSA data structures and algorithms interview where you did four algorithms and data structures, and then the fifth one was large-scale system design. I think the large scale system design is the thing that's really important, but like knowing the different flavors of like chunk by, chunk by key, you know, partition, stable partition. I don't know, man. I could I I hope I'm wrong, but I think you know, yeah, I I don't know. I I don't know. I don't think that stuff is is that useful anymore.

Bryce: 38:09
I think that knowledge of like minutiae of like language syntax and stuff like that is less useful, but like knowledge of like systems, knowledge of like performance, like all of that is valuable and field experience is valuable, and so I don't think that like I don't think that your experience is any less valuable because anything that you're experienced with, you're gonna be better at figuring out how to automate it well than other people are.

Conor: 38:37
Time will tell.

Bryce: 38:38
And also, like, like you also have to keep in mind, like, somebody has to figure out how to train these models and build these models and make them good at coding. And people who have skills like you and I do are probably uniquely well suited to work on those sorts. Just wait, eventually, eventually you're gonna get a little further down the road and start getting into like the RL and Evel's side of things.

Conor: 39:01
I mean, yeah, we got NemoTron. You know, uh, I'm not worried about my employability. Uh I'm more worried about just society. Nothing, nothing important, nothing big, just society as a whole. You know.

Bryce: 39:15
What what are your concerns about society? I thought you were concerned about our society of like the field of software. I can't help you with your concerns, your existential concerns about the impacts on society. I can't help you with the impacts of this on, you know, our industry.

Conor: 39:30
Oh, uh no. I mean, uh I'm gonna we're gonna be fine. We're we're we're good to go. I mean, I I mean more Then what are you what are you worried about? Like I like I said, I just think like a ton of communities I well I've realized, like I said, my my existential crisis, and it's just that like I think a lot of the motivation for certain meetings and whether it's a conference or a committee meeting was uh like it's actually social. It's not actually technical. And you can justify it with technical talks, but it's really a social thing. And I But why is that a problem? Why is that a problem? Because you're not gonna be able to justify it.

Bryce: 40:11
There's no technical justification anymore, and those those communities and conferences are gonna tell you're telling me that while our field undergoes a massive revolution that changes all aspects of software engineering, that continued education is not gonna be something that companies are gonna prioritize.

Conor: 40:34
Not at the types of conferences that we have historically gone to.

Bryce: 40:38
I think that all of those conferences are slowly gonna like all of those conferences are gonna either go away or they're gonna have more.

Conor: 40:44
That's what I mean. You just said it. You just said it. They're gonna go away. That is the point.

Bryce: 40:57
Because it will become a part of the whole industry. The the software development and AI assisted software development will become one and the same. It's like like you know, there's talks at there's talks about like package management at CP at CPPCon, like people talk about like version control and branching strategies, right? Like it's a C conference, but people still talk about like software engineering, and AI is gonna become inexplicably linked with software engineering. I guarantee you, if you go look at the agenda for any C conference, like a good chunk of the talks are gonna be talking about AI and vibe coding and code review and and the challenges of using AI for particular languages. It's it's everywhere, man. The conferences are gonna adapt.

Conor: 41:45
Maybe. And if they do, I don't think it's willingly.

Bryce: 41:47
I I I get the sense that like the world hates AI and uh you are correct that there is that that there is a large part of the world that yes, you're correct about that. There's a large part that is afraid, even within our industry. There's a large part that don't like it, and it does centralize power to some degree. And you know, the I I do think that big tech benefits more than smaller tech companies. And in particular, if you're a software freelancer, if you work at a smaller company, you just have a lot fewer resources, a lot less access to resources than other people do, and that's gonna be tough. But eventually this technology will commoditize and will be more widely available. I mean, just just imagine, Connor, if you will, what happens when the token prices go down a hundred X? When the stuff that would cost us, you know, a thousand dollars today costs, you know, ten dollars. It's gonna be game changing.

Conor: 42:49
I mean it already is game changing.

Bryce: 42:52
Yeah, but but but like once once doing like a billion token task costs like ten dollars, not a thousand dollars, then like everybody's gonna be able to do the stuff that we're doing. And that's really when the the like wait, we you're the thing that I think you you that you don't realize how early we are.

Conor: 43:12
I think AGI is right around the corner. We are, as Demis Hasabis said, former head of GDM, Google DeepMind, now I think he's head of Alphabet. They promoted him. We are in the foothills of the singularity. And you may think we're early, and probably for tons of corners of industries, you know, it is early, but I think coding is dead. And maybe this should just be episode 300. We were gonna we were planning on bringing on a special no?

Bryce: 43:46
No, no, no, no, because cut come on, because we we got we gotta do a Sean episode for episode 300.

Conor: 43:51
Alright, so this will be episode 301 then. So uh we've already heard the Sean episode. Be sure to check these show notes either in your podcast app or at adspthepodcast.com for links to anything we mentioned in today's episode, as well as a link to a GitHub discussion where you can leave thoughts, comments, and questions. Thanks for listening. We hope you enjoyed and have a great day.

Bryce: 44:08
Low quality, high quality. That is the tagline of our podcast.

Conor: 44:14
That's not the tagline. Our tagline is chaos with sprinkles of information.