In this episode, Conor and Bryce chat about AI, how it's changing the way we work and more. Link to Episode 244 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast: Twitter Conor Hoekstra: Twitter | BlueSky | Mastodon Bryce Adelstein Lelbach: Twitter Show Notes Date Generated: 2025-07-01 Date Released: 2025-07-25 AI Poll Results All of Conor's Vibe Coded Projects Cursor Claude 4 Vittorio's Camomilla GPU Mode ADSP Episode 238: Recommended Podcast Discussions on AI & LLMs ADSP Episode 239: Claude-Poisoned Dev Sipping Rocket Fuel CoRecursive Episode 113: When AI Codes, What’s Left for me? My AI Skeptic Friends Are All Nuts - Thomas Ptacek ThePrimeTime - How WE Use AI In Software Development Intro Song Info Miss You by Sarah Jansen https://soundcloud.com/sarahjansenmusic Creative Commons - Attribution 3.0 Unported - CC BY 3.0 Free Download / Stream: http://bit.ly/l-miss-you Music promoted by Audio Library https://youtu.be/iYYxnasvfx8
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
ConorHow would you describe LLM coding tools like Cursor, Windsurf, GitHub Copilot, Cloud Code, Codex, and others having changed your dev workflow? A, it's rocket fuel. B, it's very helpful. C, it's mostly unhelpful. And D, don't use them slash avoid them. And this is much closer to the poll that I actually cared about. Because for me, it's it's rocket fuel. It's completely changed, you know, how I work. I asked on Twitter, uh, Mastodon, LinkedIn, and YouTube. Welcome to ADSP the podcast, episode 244, recorded on July 1st, 2025.
ConorMy name is Connor, and today with my co-host Bryce, we chat about AI, specifically AI-assisted coding tools like Cursor, Windsurf, and more. Now this is doubly a good transition because I mentioned the language app that either it's on my list of things to build, but I'm also happy for someone else to build it. And then also uh within that episode talking about the cool CUDA toolbox talk, and you've seen the demo for this. We're gonna share our screen, folks, and we will make sure to walk you through this, and there will be links in the show note uh for all of this stuff.
ConorWe're gonna talk about for the next uh hour plus. Might cut it up into two episodes. We might make it all one episode, folks. We don't know. I now have a website called uh it's just a subdomain off of my GitHub.io, and it is uh vibing projects. It's matrix themed, folks, and everything here is vibe coded. We I assume including the website. Including the website. It it looks beautiful. And we're so the first one here is uh uh a summary of the polls.
ConorI did two polls. We're gonna talk about that in a sec. But I mean, Bryce, you've seen this and it's just it makes me so happy. It's I hope the future of like the landing page for like the KUDA library ecosystem. Basically, I I might even, if I have the energy, which I never do, I always say that. I say, if I have the time and effort, I'll go back and I'll I'll find a clip. Um, but I'm pretty sure in that episode when we were talking about this effectively a slide from your presentation, I was like, this is amazing. We need to do something more with this. Like I pointed out that like the Rapids logo wasn't purple and uh it, you know, it was hiding the QDF and QML and friends.
ConorAnd so I I at that time I had this vision. I'm not even sure how or specifically I articulated it, but there's a chance that I basically described this and now it's here, and definitely you like because it wasn't in your presentation, you were the one that mentioned the categories, and uh and so now it's all very interactive, it looks fantastic. Uh it took me, I think it's beautiful. A few hours. I mean, initially the one shot was pretty good, but it didn't look anything really close to this other than the fact it was a four by six grid. If we leave this and we come back here, we've also got a summary of the two polls.
ConorUh I know you saw at least one of them. I did. Um the first one that I asked, which I kind of regretted, not regretted asking, but it wasn't really the question I wanted to ask, was what category best describes you about AI? Is it A maximalist, AI will transform everything rapidly and profoundly? B pragmatist, AI is useful, but needs careful deployment. C, skeptic, AI is overhyped, limited, or socially harmful, or D doomer, AI poses catastrophic or existential risks. I kind of feel bad about the wording, because I I got AI to generate this for me.
ConorAnd in hindsight, I you know, I think B talks about like uh being careful and C talks about social harms, and anyways, it wasn't the best poll. But here are the results uh across Twitter, Mastodon, and LinkedIn. Twitter got 156 votes, Mastodon got 57, and LinkedIn got 147. Overall, if you total up the votes across the websites, uh Maximalist for A got 7.7%, so that was the uh second least common. The most common was B, Pragmatist, AI is useful but needs careful deployment.
ConorUh that had 48.7%, so almost half of folks. Skeptic was C at 38.3%, and uh Doomer was D at 5.2%. I got one comment, I can't remember on which platform it was, but they were like, oh, you've you've biased the polls by using the word Doomer. It's charged language. Listen, okay, I didn't come up with the term Doomer. That's the term of art. P Doom is a thing, the probability that you think, you know, the singularity is gonna happen, we're all gonna get wiped out. Um it's the term of art. Uh so don't don't uh hate the player, hate the game.
BryceNow now now the the getting to the what you said earlier about the poll, the the one problem with the poll is that it sort of conflates two things, which is one, how widely like I think what you're trying to ask about was like, how widely and rapidly will AI be deployed and useful versus in this there's a second thing which the poll is sort of conflating into that question, which is the risk of it.
ConorYes, yes. Which I so I wasn't a huge having read it in hindsight, I kind of which is why I ended up sending a second poll, because on top of it not being the best wording, um I honestly wish I just had a like left mac maximalist pragmatist skeptic doomer and then like let people infer what they will about it. Like uh that probably would have been a better poll. How did you vote? I voted maximalist. I am so high on AI. It is it's absurd. Like um, like this this website we're looking at, I basically uh I just gave it the data from the polls, and I was like, I want a a you know, static page, and like it unnecessarily gave me these pie charts and donut charts, uh, but it kind of looks cool.
ConorAnd I was like, give me some nice little animations and stuff. And like if you click on these, they hover, and it's it's all like not really necessary, but like the ability to build static pages that like summarize information. Like, where do I have it open right now? I built this yesterday in like 10 minutes. Uh, we have two different wedding lists, one from this Wishjoy site that uh people RSVP'd to, and then another one from our our uh what do you call it? Like travel agent that did all the bookings for people that wanted to book through uh her. However, like those lists are not the same. Some people use their formal or like official legal names with the the travel agent, whereas they go by something else colloquially, which is what they used for uh the RSVPing.
ConorAnd so uh my fiance, Shima, she asked me to like consolidate these, and I was like, well, you know, let's have some fun. So I just went and asked, I was like, can you build me a wedding guest like consolidator? I have two different lists. Some of them have uh exact name matches, so I want like you to auto-match them, which is this auto-match button. But then after that, I want to be able to like manually just like click on one from the left, one from the right, and then have those paired, and then like remove them from both of those lists. Uh like and so it the one shot gave me basically what this looked like with different styling. The second question I had to ask was like, oh, like please remove them, which I think it actually was already doing, but I just uh I didn't notice.
ConorAnd then also like the the Wishjoy uh RSVP list, it included everyone, including the people that uh did um declined uh the invitation. So I was like, I need like a filter button. And then sure enough, like it built in like exporting to a CSV with it. Like 10 minutes of work to do this, and uh um Ashima is uh she's doing hospital work, so she's out of town right now. And I like Google, uh uh what do you call it? Google meeted. I was like, sweetie, you gotta look at this. It's it's so cool. And it's like your ability to build these tools, like probably it would have been only five or ten minutes more work if I had just like skipped this step, gone to an Excel spreadsheet, because I'm I'm good enough with Excel.
ConorI could have done this. But this was like it was it saved me probably like five or ten minutes, and it was just so much like the power that you have.
BryceWell, it would have been it would have been five to ten minutes to do the data analysis, it wouldn't have been five to ten minutes to make a UI for it.
ConorOh yeah, yeah. Just to like get the all we need is like a uh unique list of names of the people coming at the end of the day. Anyway, so just your ability to build these kind of like single page things it is just phenomenal.
BryceCan you make sure we we are SVP'd, right?
ConorYes, yes. You were on the list of folks that like so at the end of it too, like there's a bunch of people that are SVP'd yes, but didn't book through the travel agent. So I think there was a list of 24 folks, and so then I have a button that says add people that self-booked, and then in the CSV it mentions in parentheses when it you export it that they they were self-booked. Oh nice. And so then we're gonna just send that back to our travel agent because and also the wedding list people or the the wedding resort folks because they need this full list. Um but just like you know, like this kind of tool, how often is this kind of tool would it be useful uh in when you're just doing stuff?
ConorAnd like uh after when I showed uh uh Shima this, I was like, I was I was just so excited, I was like, choose something, like well, what do you want to do? And then we ended up um if I take this out of full screen mode, uh she has to do like billing codes depending on the type of work she does, and she practices in a bunch of different specialties. So she she's a licensed family uh medicine doctor, but her specialty is public health and preventive medicine, uh, but she also specializes in addiction. Uh so she was like, it's a very complicated world. Can you just like uh get it to go search up all the OHIP, which is the Ontario, I actually don't know what the hip stands for, uh, but it's like the Ontario health card.
ConorI I guess Americans aren't aren't familiar, but like different countries have free health care and the doctors they they bill stuff. And and so uh at first it it came up with all the specialties, but then she was like, uh, there's so many codes, like just give it the four that I need. And then it has some fuzzy fuzzy searching stuff. It's not perfect. And at one point she also was like, Oh, it's like it's making up codes and whatever. And I was like, Well, yeah, if you if you're asking it to go and search online, like it'd be better if we just had like a PDF that it could scrape. And she's like, Oh, like let me send you the PDF. Then we dropped the PDF down.
ConorShe said it's still not perfect and she needs to review it, but like already she when uh she was looking for like a smoking code and she was like, Oh, it's interesting because in the PDF it's listed under community medicine. It technically, I think she said was a hospital surfamily thing, but they labeled it as public health. So it wasn't perfect, but she's like, but this is the code and that is the price. And uh she's like, you know, if I spent an hour with this polishing it, like this will make my life so much simpler going forward. And she and like literally like her excitement for the first time ever, like truly, she's like, I now understand why sometimes I have a hard time like getting you to come to sleep.
ConorBecause like she's like, I can now think of like 20 different applications that like I could build something like this that would make my job and my life like so much easier.
BryceI I it's interesting because I also have infected Ramona with the uh with the AI uh bug. And we only we only have one like family Chat GPT subscription. So we're using like one account. So like if I go to Chat GPT, I can see like all of her queries too. And like every day there's a new thing that she's been researching. Like yesterday, she just got a new tennis racket, and like it hits a little bit different than her old tennis racket. And so she's been using AI to like come up with like a training plan uh that she could use.
BryceAnd she was like, like, I want something, it's like a PDF that I can print out. Um, it's got like a lesson plan, and I was I was just like, you know, I didn't I didn't even have to like encourage her that much. Like I showed it to her, I showed her like how to start using AI tools like a couple months ago, and now she uses it for uh like a surprising number of things.
ConorUm is she using just the the ChatGPT interface? Because I didn't know it could generate PDFs. Or you're saying that you hooked her up with no.
BryceYeah, chat chat chat GPT can generate PDFs, can generate um almost any file type. Like if you ask it for like an Excel file, for PDF, for a whole bunch of things, yeah, it can do a lot of that. I think I think it does I I I mean I would guess that at least in some cases the way it's gonna do the PDF generation is because it has it has its Python scripting environment. So it may just um it may just like run Python code to generate the PDF. But I think it has a native tool for PDF generation. But for like some other file formats and for some scripting tasks, you'll see if you're using like the reasoning mode and you look at the reasoning trace, you'll see that it's calling out to a tool to do the um to do the work.
ConorRight, right. Yeah, it's I mean Shima is so excited, she basically was like, I need access to because I'm using cursor in Cloud 4. Yeah and she's like, I need I need this now. Because she when we were building this, like I showed her that all I do is talk in English. And I like this is HTML, CSS. I'm not a I'm not a web dev. I you know I know enough if I need to fix something because it the the model can't figure it out. But like these days, it it never if if it messes up some alignment, you just say, Oh, hey, sorry, the the pie chart when you expand it is slightly outside the border. Could you just like fix that?
ConorYou don't have to even say, like, I know that what you want to say is increase the padding so that it doesn't do that, but like you can just use English. And and I really think I was telling her that like there's this layer of stuff that you need to know. Like, you need to know command line, you need to know git if you want to host something on some website, and then you need to know like at one point she wanted a certain color, and then she's like, Oh, it's this color. And I was like, Well, can you give me the hex code for that? Like, you know, it'll it'll it'll we're not at the point where you can just drop a screenshot in a folder and then it can like you know uh figure out the color from that uh PNG. No, I mean I I tried it, I tried it.
ConorSo like uh it we're at some point we will get there. Um but like currently at this point in time, like there's there's such a thin layer of stuff that you need to know um in terms of like you know, if you're trying to host some website, you just like in put an index.html at the root of some GitHub repo. I could teach her that stuff in like an hour, and like all she needs to know is like when a model tells her that it can't do something, ask her, ask them uh to write a Python script for you. Like if it says, oh, I can't go and like scrape this stuff, it's like okay, well just give me a Python script that can, because there's no limitation on what you can do with Python, and Python can do anything.
ConorAnd so it it's just like uh, you know, anyways, to get back to the poll. Uh so it's good that our partners are.
BryceI have a question about the diagnostic code lookup thing. Can you go back to that one?
ConorThe diagnostic code?
BryceYeah, yeah.
ConorOh, sorry, I just billing one. Do you mean the the billing side?
BryceYeah, yeah, yeah. So, okay, so so you your what uh my question was gonna be what is your data source here? The the data source is that there's there was some PDF that you fed it in, and then it just embeds the information about the PDF into this. This is just like a static page.
ConorYeah, so uh it all of these are just static HTML pages. Initially, we just said uh can you go get the OHIP codes for all the different specialties? And then we gave it one example of like a, I don't know, 957 or something like that. Uh but then very quickly, Shima was like, well, it's missing a bunch, that one doesn't exist, uh, so like I don't know what's going on here. And then I then I mentioned the PDF thing, and she's like, oh, like, and then she dropped it into a WhatsApp and uh I downloaded it, and then she was like, Okay, this is way better. So yeah, it was a PDF. I actually thought it was gonna build a Python script to scrape it, but apparently it it was uh it knew about some tool called PDF to text on my Ubuntu machine that it just called directly and it was like a um I don't even think it was optical character recognition.
ConorI think it was literally just like it just took the raw text out of the PDF file. And so then from there it did something with that stuff.
BryceI wonder whether um like I wonder how much more challenging this would be if if if you wanted to like integrate it with like a live data source. I mean I don't know, I guess it would not be.
ConorI mean these codes don't change um Okay, they don't change that frequently, yeah. Yeah. I mean oh I'm I'm I'm assuming like uh But but if these this is it would fix the accuracy problem.
BryceIf you if there was some online place that lists all these codes but like not in like a searchable way, then you could just tell it like just get the data from here. You know, just like have like a like a have like some JavaScript thing that goes and fetches this and like imports it.
ConorI imagine in the antiquated world of medical record systems, they do not have some API at endpoint that you can go ahead and get a JSON object with all this stuff. Um I'm guessing that and I think that's why this was the first thing that came to mind for her. And she's like, well, one of the trickier things is like most people they only operate in one specialty. Like they specialize as a family doctor and that's all they practice. But uh because she um her specialty was public health and preventive medicine, the first two years of your five-year residency is in uh family medicine.
ConorSo right off the bat, you basically become a family doctor. Um, then she did public health, but then she did a specialization, so she's operating, and then she also practiced as a hospitalist doing like remote hospital work. That's what she's doing right now. So she operates across four specialties, and she's like, sometimes, you know, uh I don't work in uh hospital work as frequently as I work in addiction work, so like I'm not as familiar with the codes, and like you really need to know your codes in order to bill properly. It's like one of the most important parts. And so, like, it she does all the billing correctly, but it's just like a pain, and I'm I'm assuming the fact that she handed me the PDF is like that's what she's she's looking up in like this large PDF that lists everything else, every single specialty and all these different codes when really like she knows that there's some smoking billing code or some like uh what was one of the other ones, like counseling, you know.
ConorSo you just type in C-O-U-N and all the different individual um counseling uh or things that mention counseling in the description, they just immediately show up. She's like, you know, this is what I need. I just need something that can like filter out across all the different codes and specialties that I care about what the billing codes are. And it's like, you know, and that's you you hear, which is kind of what we're we're uh we're burying the lead of this, uh, you know, both of you and I are um uh very high on this technology. But when we go to the AI poll, so the second poll that I answer uh asked, which was actually closer to what I cared about, uh a poll for devs.
ConorHow would you describe LLM coding tools like cursor, windsurf, get up copilot, cloud code, codecs, and others having changed your dev workflow? A, it's rocket fuel, b, it's very helpful, C, it's mostly unhelpful, and D, don't use them slash avoid them. And this is much closer to the poll that I actually cared about. Because I, for me, it's it's rocket fuel. It's completely changed, you know, how I work. Uh and if you look at the overall, and this one I actually, so I I I asked on Twitter, uh, Mastodon, LinkedIn, and YouTube, this only lasted 24 hours, so it wasn't as long as the other one. 55 votes on Twitter, nine on Mastodon, uh, 78 on LinkedIn, and 120 on um YouTube.
ConorUh the number one answer averaged was don't use them, avoid them at 35.7%. It's pretty even with B and C, which is unhelpful and helpful at um 24.4% and 32.5%, but uh in last place by a long shot is 7.4%. Kind of interesting. I know Mastodon is a smaller thing.
BryceWith the other poll too, where the Mastodon results were very different than the other results.
ConorVery, very different. It's 78%. Don't use avoid them.
BryceUh um like like LinkedIn and YouTube are also different, but the thing I noticed about the Mastodon results both here and the other poll is that the Mastodon results are very have a very extreme skew to one perspective.
ConorYes. Uh so they mastodon seems to be anti-uh AI the most for whatever you know your definition of that is. But it's anyway, so this is the backdrop of like I and and I, you know, I released two shorter episodes pointing people to other AI content, podcasts, articles. One of the articles was All My AI Skeptic Friends Are Nuts, which I thought was fantastic. It's where I've gotten the terms uh sipping rocket fuel, and uh he also refers to Claude Poison devs. I uh I don't necessarily like the word poison because it's got a negative connotation, but I uh it sounds better than like Claude Supercharged or Claude Enlightened.
ConorUm, but I definitely am like if it's if the word is Claude Poisoned, I am super clod poisoned. And I I yeah, I want to talk to you and I want to talk to other folks. Um like I just saw on my podcast uh feed what happens when all the dev jobs are gone or AI takes all our jobs, and it was from the co-recursive podcast. And uh there's a ton of discourse around like whether these things are good, they're bad, they're useful, not useful. And anyways, that's what I just want to chat. Like, how and there was another podcast by the Primogen and Friends guys, there's like four of them TJ Dax and Casey something.
ConorAnd Casey who said he was using them 0%, uh, and then the other ones were anywhere from like 40% to 80%. And anyways, I'm just like I'm I am so flabbergasted on how people can find no utility. If your argument is that it's bad for the planet and I disagree with like the legal implications of copyright and whatnot, okay, fair enough. But it like the utility, it's like this, I think it's more life-changing than like the internet. It's completely changed the way I work.
ConorIt's gonna be changing the way that like all these professions are working, in my opinion. Anyways, I'm curious to get your thoughts. I think probably the listeners uh have heard enough from me. What are your thoughts? Uh how what have you seen amongst like the people that you're um working with on a day-to-day basis? Because you interact with more folks than I do, like on research. So, yeah. Over to you.
BryceUm, yeah, so I mean, uh let's first talk about the um is it is AI coming for everyone's jobs? Um you know, I'm not sure if we've used this analogy before in the podcast. I don't think so. But uh in the 1980s, the first spreadsheet um software came out, um, like uh uh This account and Lotus 123. Um I think those were the first, or at least they were the first to like, you know, uh be commercially viable.
BryceAnd uh the accounting industry was like in a panic. They were like, oh, there's gonna be no more accountants you know, 20, 30 years from now, there will be no accountants because the spreadsheet will kill accounting. Um and I mean that's obviously not true. All that's happened is that accounting has become less tedious. And you know, it used to be like before before you had a spreadsheet software, and before you had like accounting software, accountants would just have books. They would have paper, everything would be on paper.
BryceYou go back and you look at these pictures where accountants would create these huge, huge, like physical sheets. You'd have a huge, a huge table where you'd have a physical, a physical sheet where you'd do all the accounting, and you'd have cells, you know, uh rows and columns just like you would in a spreadsheet. And that's how you would that's how you would you would do all the accounting work. Um and it was like awful. It was incredibly tedious and like a lot more mistake prone. And and then um and then you know the spreadsheet software came out and it didn't it didn't kill the accounting industry, it just changed it.
BryceUm I mean there's there are still plenty of accountants. I don't know whether there's maybe fewer accountants now than there used to be, but um it's not like accountants are going away. It's not like it's an industry that's dying. Um the same for like actuaries, right? Like, you know, the the the invention of uh the invention of spreadsheets and of other like computer software that actuaries now all use today um uh did not make the field go away. It just changed it. Um it made workers more efficient. And I honestly I think it made quality of life a lot better because I think if you were an accountant in like the 70s and 80s and you were leaning over this table, like filling in stuff on this paper, I think you'd be a lot happier to like be doing it in software.
BryceUm I don't think that software engineering's gonna go away. Um I think that the field's gonna change. Um you know, there's every time I think in in the history of economics that we've seen great increases in efficiency, um, we've seen like industries change, but not like there's a couple exceptions, like you know, uh agriculture. You know, there it used to be like a large percentage of the workforce was uh you know uh worked on a farm.
BryceOr I think in like I read an article the other day where like in the 1880s, um, I think it was that like 30% of the United States GDP was uh was the production and sale of firewood. Um like that's you know that's obviously no longer the case. Um but I think generally every time we've seen great increases in uh in productivity and efficiency, we've seen quality of life uh improve for people. Um and I don't think like we're at a point where uh like where we're gonna live in some utopian society where we don't need to have humans involved in like work at all.
BryceUm so yeah, I I I don't think that I don't think that all of our jobs are going away anytime soon. They may change, though, in their nature.
ConorI do wonder if there's an analogy though in your accountant um story or a recap of history, in that I do think that there are a lot of industries, including software, where there are kind of levels of work. You know, there's the senior engineer and there's the junior engineer. And specifically, if you're a junior engineer that's building these, you know, s spas single-page applications, and like it's just HTML and CSS. And in the other ones, I think about the first one that comes to mind is like there's lawyers and then there's paralegals. And the paralegals are doing this kind of and so I wonder in the accounting industry back in the day, was there like you know, the CPA accountants and then like some other job that would like do some kind of manual whatever?
ConorAnd I I do think that like a lot of those jobs are gonna vanish because you know you don't need to hire uh a web dev person if you're just built for the fancy, you know, JavaScript D3 animated, beautiful, whatever. Uh I'm I mean I'm sure Chat GPT can help you, but like it's not the stuff that is the bread and butter uh that that they're just like they can knock out one of these in like 10 seconds. Um so I do think that there is some impact, and I wonder uh if similar things have happened in the past, but it's just like kind of it kind of gets forgotten in the annals of history.
ConorAnd because uh I don't know, I I guarantee you that there's a bunch of listeners being like it's not exactly just that like a bunch of accountants were employed and they changed the way it worked they worked. Because now it's like there's a bunch of software engineers that are employed that are now like 10 times, 100 times more productive with these tools, and they don't need to hire junior devs. And so now like a small team can get done a lot more with a lot less like human resources. Um I don't know what your response to that is.
BryceI do think that um, you know, change means that some number of jobs that exist today will not continue to exist. But I think that the software industry will continue to exist and thrive. Um and I think maybe there will be fewer jobs, but I don't think it's gonna like be a decrease in order of magnitude. And so there's the there's a principle in economics, I I can't remember the name, but basically the the key of the principle is that when when the cost of something uh drops significantly, uh you tend to see um uh great increase of uh in usage.
BryceUm and this this came up uh uh when there was all the scare about the deep seq model a couple uh months ago, where they're like, oh, like deep seek, you know, they they trained a new cheaper model, and you know, does that mean that the whole AI industry is gonna go bust? And the answer is like no. Like if the if the cost of um training and deploying and using AA models goes down by 100x, that's gonna mean that there's actually gonna be more usage of AI because it'll be more affordable.
BryceAnd in the case of like a web dev, um I think if you want to have a website built today, if you're like a company you want to have a website built today, um like a simple website, I think it's like five to 10K. Um and there's a lot of businesses out there that still don't have great websites. And in particular, if within the US, it's pretty common for companies to have a website. Um but you go and you know, travel overseas, a lot of countries do not have websites for things. Um like from this is always bothers Ramona.
BryceShe loves playing tennis. Every time we travel, she wants to you know go find a local tennis club. They never have websites. Like we're in we're in the middle of nowhere in Austria or we're in uh Turkey or we're in Bulgaria, and she can like she can find on Google Maps like a tennis club, and maybe they have a phone number, but they don't have a a website that has you know the details about how to contact them, et cetera, et cetera. Um and so if the if if the cost of getting a website developed by a web dev drops from like five to ten K down to like you know a hundred to five hundred dollars, because now the web dev, instead of them having to spend like a week or two doing this thing, they can just they can just feed it into an LLM, they can pop out like 10 times or 100 times more websites a year, and the maintenance burden also decreases for them, then they can drop their prices.
BryceAnd then I think that means that a lot more companies, a lot more small businesses, a lot more you know, organizations will have websites, you know, uh a lot more things will become digitized. And I think if you look at the world and usage of technology in the world, there's still like a very long tail of the world that's not digitized. Um and you know, processes that aren't automated, that aren't, you know, uh uh handled in software, tracked and software, uh, companies that don't have websites, etc., like there's still like a really long tail of that.
BryceUm like the total the total uh market for um tech stuff uh is much larger than than the currently realized market. So I think I think that if you're a web dev, you probably should start using AI, and probably the change that's gonna happen is that your costs are gonna go down. Um, and that's gonna mean that you're just need you're gonna need to find you're gonna have to figure out how to bring in more customers.
BryceUm and your business is gonna be more of a volume business. Like you need to find a way to automate more of your of your business um uh to be able to survive in this new world.
ConorAaron Powell Do you think this is the the key reason behind the pulls and kind of the don't avoid or use or you don't find it that helpful is that there's a bunch of folks that are worried about the impact it's gonna have on like our industry?
BryceYeah, um I think so, but I I also I mean uh one one of one of the like you know Bryce's laws is that like inertia is king in uh in the software industry. Um you know, the the almost always the reason that we make decisions in software is around inertia, you know. Um usually like when we're making like poor decisions. Um there's a lot of people talking about in the AI space that like distribution is king, right?
BryceThe like the distribution is the key thing. Like if you have distribution, like that's your moat. Um and uh and I think that that sort of gets to the point here that that if you're able, if you've got a captive audience, if you've got users already, even if you're just an AI, even if you're just wrapping some frontier models, the hard part is not developing or the developing the models is hard, but um uh there's always gonna be a new, you know, great model, you know, coming out, you know, six months from now, right?
BryceThat's gonna disrupt the the field. Um and uh uh the thing that really gives you long-term survivability is if you've got a customer base that's already bought into using your thing, because uh they're like the for them to switch away from using your thing, the other thing would have to be like 10x better. Um like like you know, the open AI doesn't have to have the the best models because like w we use we use chat GPT as a verb, right?
BryceLike, oh if you're gonna use a uh like a large language model for something, somebody says, like, oh, I'm gonna go chat GPT this, because like it was the first on the scene, like that's gotta be by far the most popular place where people um go to interact with um with a model, um with a large language model. And so like they've got a huge moat because of that. Like they don't have to have the the best model. Um uh and uh I I think that you know for many, many years I just used like Vim and VI.
BryceLike I wouldn't use an IDE, like I I learned my ways of doing things, I didn't want to change, right? Like I was happy with how I was doing things. It change would require work, would require me to stop what I'm doing and spend some time to set up a new thing. And so I was resistant to the change just because like I was doing something that was working, I didn't want to go and you know change, change how I was working. And I think that's the way for a lot of engineers. Um so I think it's I think it's I'm sure some of it is people being afraid of it. Um uh and some part of it I think is people not trusting it, people thinking, oh, it's not gonna be good enough.
BryceUm that's that's a piece of it too. It would be interesting to know from those people who said that they don't use AI how many people have tried it, how many people have tried it and then gotten frustrated and decided to not continue. Because it is certainly a frustrating experience. I've had frustrating experiences using um cursor, uh uh, both on the like the model doing dumb things. I've had like times where like cursor's been buggy, um, you know, all sorts of things like that. Um and you know, it it it got it took me a couple days to like learn how to use it effectively.
BryceAnd there were certainly times where I sort of got into like a rabbit, like went down a rabbit hole of like, oh, I'm asking it to do this thing, it's just not doing it correctly. And instead of me just going to do it manually, I'm just like, I'm gonna keep hammering on this thing and yelling at this model until it does what I want. Uh, because I'm very frustrated and like it should be able to do what I want. Um and so I wonder if some part of it is people have tried and run into some of the limitations and then decided like, oh, well, it has limitations, and and so because I was already skeptical, I'm just gonna assume it's not good enough.
BryceUm and uh I that that may be part of it. Um uh but I I think a large part of it is just inertia, it's just people haven't tried it because um uh because they they're it's hard to get people to try new things.
ConorYeah, I I will say, I mean, we've all I think we've all been there. If you're using these tools, is it definitely you can end up in these rabbit holes? And I get I guess I mean one of the thoughts that when you were saying that is that I wonder if a part of my maximalism on these tools is that I am an expert in C. Right and it's already such a painful experience, like half the time. And like I'm an expert, but just sometimes you run into some like if you're doing bleeding at cutting edge stuff with like template metaprogramming or const expert metaprogramming, it's just like the errors are awful, and like it is literally like there are talks by like Vittorio Romeo and friends of like tools that help you diagnose, and like you know, he's I know Vittorio's written like a tool, I can't remember what it's called, link in the show notes, but like it will it'll like collapse the nestedness of your like template metaprogramming compile time errors so that like you can parse them better.
ConorAnd like every C developer that like gets to a certain level knows that you you develop a skill in like parsing these, that like, oh, and when you have like the nested templates, you go to all the way down, and that's where the error will be, and you just skip all that type information. And it's like it's a painful experience, and I have to become an expert in stuff that I don't want to be an expert in. It's just like I want to write code and I want to ship software or like you know, ship libraries, and so much of the stuff, like I think I mentioned in a different conversation that like the cognitive load that I have to carry in my head is a C dev.
ConorAnd so when a tool comes along that like is you know like Moses parting the Red Sea and being like you don't need to do any of that stuff, you can just like you can all of that boilerplate and not just like documentation, unit tests, you know, and you're they say don't get it to write your unit test, but like do you know what like writing a Google test, a G test, or like a doc test? Like, there's so much scaffolding, and then you can just go and like modify the test for the test case that you want. Like it might you you want to definitely verify, you're not just vibe coding your unit test, but like, do I really want to like, oh, what's what's the macro for the top level thing?
ConorThen there's the subcase thing, and you know, if you're switching between unit test frameworks, they've all got different macros for stuff. Oh, that's right, you're not allowed to use a lambda inside a macro because it doesn't work that way, so you have to use function objects if you're doing anyway, it's just like the AI knows all that, and you just you can hold your hand, it's a very lovely experience. And then every once in a while you do end up having to smash it over the head with a bat, but like that's no more painful than what I was doing before. And so I wonder if people that are coming from like Ruby or Rust that like have much better tool chains and it's less painful, are they less willing to tolerate when the models don't do well?
ConorThat was kind of a thought I had in my head.
BryceI think that that may be true. Um I think that Yeah, that may be true, but then on the other hand, I think that for people writing lower-level systems code, large language models tend to do a much poorer job. There's this uh there's this uh community called GPU mode, which is like a uh GPU kernel hacking community. Um and one of the projects they've been working on recently is developing a uh a better training set and model for writing CUDA kernels. And the reason for this is, and this was like as of like three months ago, these evaluations, but they did evaluations of like how do the state-of-the-art models do it writing CUDA kernels.
BryceAnd I don't remember um the exact numbers, but I think it was something like uh for like 70% of the test cases, the the state-of-the-art models like could not even produce code that would compile. And of the 30% that did compile, only like 5% would run. Um and of the 5% that would run, like only like 1% would run fast. Um so like that's like a pretty bad failure rate.
BryceUm that like only like 1% uh uh of the tests, uh uh the state of the art models could only like get the correct code for like 1% of the tests. And by correct code I mean like, you know, you have the reason you're using a GPU is because you want it to go fast. So if the code's slow, you know, you don't care. And so like one thing that they started doing is they um they run like a coding contest on their Discord um where they have coding challenges uh to like write an optimal kernel for something, and the coding challenges are all just an exercise to collect data, to create a curated data set of what good kernels look like, so that they can train or do fine-tuning on a model to teach it how to write good kernels.
BryceUm and and uh what I'm getting at here is that if you're writing some lower level systems code, if you're writing like concurrent or parallel code or performance-sensitive code, which is often what you're doing if you're in C, the models are not necessarily as good at that as they are as like um, oh, I want a website for doing X, Y, or Z. Um and I I think you actually have to do a little bit more hand holding to get them to write like correct concurrent code or correct like performance uh oriented code.
BryceUm so uh it's interesting because it it may mean that, well, yeah, the large language models mean that a lot of the painful parts of dealing with C, like error messages, etc., um, can be automated away. It may require a little bit more human intervention um and somebody who's more senior and who understands what the correct code would look like to be able to get them to generate uh good results.
ConorYeah, that's fair enough. I mean, uh I I definitely have had that experience where when I'm coding CUDA C libraries, uh you really need to it the directions it wants to take you without guidance are typically not great. Yeah. And even when you spell out what you like, a lot of times they don't do it. I mean, I just gave that presentation the other day uh with Thrust, where I was like, all right, there's a s there might be a single algorithm with the first word and second word of this algorithm name, and it still did not reach for that algorithm.
ConorIt was like, okay, yeah, I'll call these two separate things. So a lot of times, yeah, you you need the expertise to know what you want. Um, but if you have that expertise, I still find it more pleasant. And I I wonder like if it's uh a a lot of people, they they love the craft of writing code, and so they just have no interest in a tool that is gonna do the writing for them. And I used to think that like I loved writing code. Like I I I I know I like it, but like I don't know. I I these tools to me, I I feel like maybe that's the thing is I do I do love, and that and that's maybe it explains why all of my YouTube videos are on these simple little leak code problems where they give you the function definition and the return type, and like all you have to do is implement the algorithm.
ConorI love doing that stuff. You know why? Because there's no cruft, there's no ceremony. You're just it's purely writing the algorithm. And I think that's the thing is I love that kind of coding. So much of programming and software engineering is not that. It's it's scaffolding, it's ceremony, it's you know, dealing with errors, it's documentation, it's uh you know, unis unit test framework like setup. Yeah. It's becoming a second uh expert in another, you know, language called CMake in order to get your build system to work.
ConorAnd and uh and you know, I I remember before I had like my white belt or whatever the second level belt in CMake was, it was so painful trying to get like a project to build and to pull in something. Now it's not that bad because I got my blue belt or whatever belt, and so and and also CMake and has gotten a lot better in terms of the utilities they provide you. If they've got whatever the CPM fetch thing and you just point it at a GitHub, it's it's a lot better than it used to be. But just like I guess my my my my thing here is that like the part that I love about coding, I still get to do. And I can just ignore the other 90%.
ConorI don't I don't know. I'm just I'm trying to figure out why people are so averse to this this um what I think is like the best thing to happen in our careers.
BryceWell, and I mean one uh like to that point, one uh uh you know, on this question of is AI gonna eat software jobs, um it's not like we as an industry, you know, it's not like sh like software famously, you know, doesn't like ship on time. It's not like we as an industry have a great track record of delivering, you know, uh bug-free code like on the deadline that we set sent out set out to uh to achieve, right?
BryceLike uh, you know, we we things software projects tend to be over budget, late, um uh and have uh a lot fewer features than than we'd intended. Um so you know, we could we could have a 10x increase in software engineering efficiency. If we had a 10x increase in software engineering efficiency, I don't know that it would necessarily mean that we would have fewer software engineers. It might just like, or even fewer software projects, it might just mean that like, you know, we can deliver higher quality stuff.
ConorYeah. That's fair enough. I mean, uh, what is the when you when you're saying that it came to mind the uh what's his name? Dunky Videos or whatever, is a YouTube comedian that kind of makes funny videos about uh uh video games. And he had one on GTA 6, which I I think that's the I don't I don't play GTA, but whatever. It's one of the rock star games that's very hyped, and it's it's been delayed like years in the beginning. That was like the the the the punchline of the video is like that the trailer came out, and then he was like, Whoa, look at look at the graphics, whoa, look at the graphics on these graphics, and it's just a very beautiful trailer.
ConorLook at the graphics on this graphics, and then he's like, Unfortunately, we'll never be able to play it because it's like delayed until 20 2036 or whatever. Yeah, anyway, so yeah, it's uh yeah, famously we uh our industry is not. Great for um shipping things on time. I mean, like, where is uh Apple's intelligence? Like, are they still don't really have AI? Like they're working on what's going on. They're working on it.
BryceUm well, and it's it's um I mean, I think we have to keep in mind that the the um the models and the technologies that we have today are still pretty rudimentary and limited. I um I often get frustrated when dealing with uh with Claude in uh in cursor or any model that I'm using in cursor. Um because it's like I'm working with a junior engineer, but um a junior engineer who never learns.
ConorUm and um cursor rules, cursor rules, buddy.
BryceYeah, yeah, but but I I think that I think that things like cursor rules and all the other equivalents of cursor rules that that other you know frameworks have are a a crutch for um uh for like a bigger problem. And the the bigger problem is that we need to have we need to have like uh uh AI systems that have a first class long-term memory that like things like where like when I'm working with something, like it learns over time.
BryceUm and where it it doesn't like decohere over time. Because the problem is like if you um uh you know you can't have a a single session with a model that just goes on forever because eventu eventually it gets too large and eventually there's too much context or stuff falls out of the context window and things start to decohere. And and uh, you know, uh open AI has some memory system, uh, but as far as I can tell, it's basically where it's just like a tool where um when you're in chatting, if you have memory turned on with open AI and you're chatting with um with one of their models, and you tell the model something like, Oh, I you know, um we we like restaurants that don't have food that's deep fried or too salty.
BryceUm uh like you know, it it has a tool where it'll be like store a fact that the user has this preference, or store a fact that the user's name is this, et cetera. And then it's got a database that has all those facts and it can go and pull those out. Um maybe that's maybe that's what the you know, what the optimal memory s looks like, and maybe maybe that just needs to be matured a little bit. But I suspect that probably the next wave of big innovations that we'll see in models will be um some big revolution in memory and in long-term memory.
BryceUm and uh and that'll give us agents that um uh have greater persistence and that uh learn from us over time. Um and that sort of grow with us.
ConorYeah, it's I mean there's definitely like a uh you learn the patterns of like you said, the first couple days you were using it, and then you kind of got you figure out it's like if you don't use cursor rules, you've probably found out that if you have a repo that has a building.md, you shouldn't just ask it to build. You should, to be safe, say build using at building.md. And then it never fails with when you tell it where. But if you don't tell it, uh depending on how it's feeling that day, uh, you know, you might have 75% of the time it does the right thing, but usually what it ends up doing when it doesn't do it is it does something that doesn't work, and then it greps the you know the the repo, then it finds a building.md, then it goes, but it takes an extra 30 seconds for it to do the grepping, and depending on what model you're using, it could be faster or slower.
ConorBut like you learn the golden path of like one, just add a cursor rule, or if you're not doing that, you just always point it at the building file. And it's true, like why does a model, after the one time in one session, does that not just like uh uh what's the word, uh propagate to like every other session that when the first thing it should do behind the scenes when you ask it to build if you're in uh repository, is to go and check if there's a building.md, because obviously like there are the instructions.
BryceI'm sure you could write a cursor rule telling the model to write cursor rules. Um uh you know, like you could write a cursor rule that says like when I tell you something, you know, that's a best practice that you know you should you should do in the future, like you know, like write a cursor rule for that thing. I'm sure you could do something.
ConorOr just have some feature built in that whenever you get interrupted by the user, uh because I I imagine like I don't have it set up right now. I still am typing all my commands, but like I'm pretty sure I'm gonna set up so I can just start talking to this thing and like and speech to text. The other thing I was thinking is like so often my workflow now is because I'm using Cloud Force on it, which it usually takes you know 30 seconds to a couple minutes to maybe even a few minutes, depending on the task and how complex it is. Um it takes some time. And so a lot of the times I'm like juggling. I'm juggling multiple things.
ConorUm and I tend not to do two uh like AI coding sessions. I'll be working on docking. No, no, no. Well, because the thing is like I don't you're definitely not doing it in the same repo, or at least I'm not going to.
BryceSo it's like what you do is you create multiple repos, you create multiple checks outside the repos. There's a couple different techniques for doing this.
ConorYeah, I uh I I don't like I I like my workflow. My maybe we should talk about that at some point with the workflow. But currently mine is I've got my single repository, I'm using cursor, uh Cloud Force Sonnet, uh, I have the you know the agentic window, and um I I basically I have everything that auto turns on, but NVIDIA doesn't let us do the total Yodolo mode, but it'll let you go and search on the web and stuff. And so, you know, I've I've got it, I've got multiple monitors, you know, so I've got it on my main monitor. I ask it, give it quite a bit of detail, it sends off to work, and then I'm gonna go start like you know working on some other task or like um task, you know, um putting them on some stack of like things that I I want to do.
ConorBut like uh at times, you know, you you are waiting, it's like you would hope that it was a bit quicker so that I don't need sometimes it's like five minutes, and I'm like, I don't need five minutes. Like I'm kind of going back and forth. What I'm thinking is like I spend a lot of time running, listening to podcasts. I would love to just like ambiently, and maybe this is like an unhealthy thing, but like I enjoy doing this stuff so much, I would like to ambiently be like just always talking and like working. Um so like if I'm going on an hour run or a two-hour run, that's like two hours of lost productivity with Claw.
ConorWhere like if it's and if I tell it like take a little bit more time, because I don't want to every you know 30 seconds it be interrupting my podcast and my run. But if I say, you know, check back in with me every five minutes, I would be so happy to like not even stop what I'm doing. You give me a pair of like the futuristic meta uh bands or whatever they're calling, meta lenses. Um like they're not here yet, but give me a pair of like you know, Oakley sunglasses where I I need a little bit of visual. I can't just do it based off of whatever, but it gives me like, you know, in my little futuristic holographic thing, I stop running because you know I don't want to uh hit a bike or whatever, but you can imagine, like walking around, even while you're you're traveling, you're touring, and just constantly making, checking in and being like, okay, yeah, that's a great direction, or no, the next step is to do this.
ConorAnd like, how much time does it take to like vocalize and speak that? It's like 10 seconds, 20 seconds, and uh I just feel like that would be such a cool thing to do, is like, you know, you're you could you could then technically work and like just go for a walk. Maybe you're not even exercising on a daily basis, but now these tools like make it's like, whoa, I can still continue to absolutely take a meeting or or or or go. It's like uh I we we already have a ton of flexibility because we both work remotely. Um but like I just I think it's gonna completely change um you know the way that we work and and yeah, it I don't know.
BryceIt's uh Yeah, I I I um I agree. And you know, for me that's one of the really appealing things about it because um uh I I've had all these ergonomic issues for years. Like I can't spend eight or ten hours a day at the keyboard. I can maybe spend like three or four hours a day at the keyboard. Um and like sometimes not even that. Like I've had I've had issues for years, and I, you know, I figured eventually technology would solve this problem, but I thought it was gonna be like the the brain chip. Uh fun fact, one of my one of my childhood friends was like employee like number 10 at Neuralink.
BryceUm so I probably had an in. I probably could have got got him like to, you know, to get one of the chips installed into my brain. But I figured that was gonna be the thing, was that I was gonna get the brain chip and then be able to talk to interact directly with the computer. But um, but I think the it it seems like the technological solution is gonna be is gonna be AI. It's gonna be that eventually I just, you know, I um maybe I'm at a keyboard, but I'm not, you know, typing everything out. Um maybe I'm typing out one prompter, you know, every five minutes or so and letting it work.
BryceUm, or you know, I'll I've been thinking about doing the speech to text thing so that I can go like completely keyboard free. Um I do I do find that like with the current like cursor UI, like I do, you know, when I'm going and reviewing a diff or something, um, or uh like I I still have to, you know, go and like touch things, um or like you know, selecting context is a little bit of a of a hassle. So but I'm I'm sure that there's ways, like I'm sure that you could go a lot more compute like uh uh keyboard and mouse free if you wanted to, um, which is like a really exciting prospect to me.
BryceSo yeah, I've been thinking about trying out the speech to text. Although it may drive Ramona crazy to hear me yelling at the uh the robot all day.
ConorYeah, I mean, well that's part of the reason I don't do it right now, but you know, ideally uh at some point we we can have a bigger place and uh separate separate rooms, ideally on separate floors. Uh because Shima also she works um remotely uh a lot of the times. She has like virtual clinics and whatnot. Um yeah, I mean when I hear you say that, I worry because Shima gives me a hard time for being uh what's called a transhumanist, where you know we're there's a percentage of the tech bros that are super happy about the future of you know integrating and augmenting our intelligence with like hardware.
ConorAnd I posted a tweet uh a week ago that I asked uh ChatGPT to summarize like the the timeline of AI like levels, because I finished watching a movie called Mountainhead, which is a kind of dystopic look at billionaires and how they're destroying the world with tech. And uh anyways, it was like we're currently at uh ANI, artificial narrow intelligence. Next is AGI, which is general. After that is uh uh SI, superintelligence, and then the two after that are transhuman, which is like integrating with uh hardware, and then posthuman, which was like uploading your cloud to the brain, uh or uploading your brain to the cloud.
ConorAnd anyways, when you're talking about like getting a chip in your head so you can talk to the robots, I'm there like being like, oh man, that just sounds so awesome. Like I don't want to be, I don't want to be patient zero. But I and I know most of our audience is software developers, but I do wonder, like, is the because I I there was a uh New York Times daily episode where they were commenting on on some stuff, and they were their their view with like people like us is that we are like broken like morally and that like we're we like we don't care about society. And it's like I definitely care about society, but I also just like why should we put a little box around us of of like what we are or like should or should not, you know, like like do?
ConorUm I don't know, maybe I'm not articulating that correctly, but it's just like it's not that I don't care about society, it's just like um we live in a world that's defined by like our history, and it's just hard for people to fathom what the future could look like, you know? Um Yeah.
BryceYeah, I mean I I I think it's I think generally when we've had great scientific advancements or great advancements in productivity, yeah, like the Industrial Revolution, there were lots of downsides, you know, like children losing arms in factories, like stuff like that. Like, you know, it's not all good, but I think on the whole, um, great advances in technology have improved quality of life for humanity. Um I mean, I think you look at quality of life today versus 200 years ago, I think it's hard to argue that things aren't better now.
BryceUm you know, that doesn't mean that doesn't mean that all the technological advancements have um have only been used for good or have only had good outcomes. Um and I I don't mean to I don't mean to s to suggest that the that the positive outcomes justify bad stuff that's happened along the way. Um and certainly there's lots of lots of ways in which AI is being used or will be used that may um uh harm people. Um, you know, things like surveillance, you know, i i it certainly could be something that could become a tool of um of nefarious actors, nefarious uh or countries that uh do not enjoy the the rights that our two countries have.
BryceUm but um I mean I think I think on the whole i it's hard to imagine that this won't be good for society uh unless you're you're of the the perspective that you believe that uh pea doom is likely. Um and that that you know i it's just hard for me to imagine how that scenario plays out um because technology is at the end of the day very, very brittle.
BryceUm and it would have to be not so brittle to really be a threat to us. Um but it's I also I also feel like um it's almost impossible to control the development of AI. Like if if we if we as a society like decided that like we weren't gonna build AI because it was too dangerous, I don't know how we could possibly control that. Because it's not like like look at like nuclear weapons. Nuclear weapons are something like that.
ConorI was literally just thinking the same thing in my head.
BryceYou you like for nuclear weapons, there's a specific set of physical materials that you have to control access to to create them from being cre uh created. And like that's something that like you can you can monitor and you can observe um uh and you know detect through through satellites to some degree. Um and and even with all that, we've only had like an okay track record of you know of of keeping new countries from developing nuclear weapons.
BryceI mean, well look at look at current events um uh and look at events from you know past 10 to 20 years with uh North Korea. Um AI is not like that. Uh to develop AI, you just need like some dude in their basement who has computers, and okay, maybe you need access to large quantities of compute to to really do it. Um and so maybe there's an argument that that there's a similar like material constraint. But um I don't think we're going to like you know burn all the data centers down entirely because we need them for other things.
BryceAnd if you don't get rid of all the data centers or control tightly control all the data centers, I don't see how you can how you can keep people from developing AI. Um and so it it it seems like it's a futile effort to try to put regulatory guardrails on it. Um and in terms of like technical guardrails of of things like aligning models and stuff like that, um, it seems like we get really mixed results right now. So I I um part of my perspective is like if the AI is gonna kill us, then it's inevitable that we'll develop the AI that will kill us.
BryceThere's not really much that we as society can do to stop that. So if it's gonna if it's gonna turn us all into paper clips, um then we're sort of doomed anyways. Um and if it's not going to, then uh then that's great. Um and so there's not much for us to do because it's either inevitable doom or we're gonna be fine.
ConorOh I mean, I would never word it that way, because I'm sure there's some listeners, but I I do completely agree. I've always said uh that we were born arguably at the perfect time, because most uh AI folks they say that if the singularity is gonna happen and the paperclip, you know, some s version of that scenario happens, it's gonna be roughly 2050. Some people are saying sooner now because they didn't realize things would accelerate this quickly. But the point being is that, like, uh, you know, we're in our 30s, 2050, another 20 years, hit our 50s.
ConorUh, you know, we were alive, we got to see it all, and then it could end. But it's like w we are not gonna stop that. Like, whether uh, you know, it's like there's all these different agreements globally. There's always one country that's like, no, no, no, you guys do that, but I'm gonna go ahead and like you know, make money and uh we're gonna we're gonna ignore this. And so that's like there's no there's no stopping this. Um and uh Yeah, so I I kind of agree. Like if if it's gonna happen, it's gonna happen. But I don't think it's gonna happen. And I just think it's gonna be like I I I used to say when I was uh a bit younger that I was I actually think I thought it was suboptimal that we the time we got born.
ConorLike imagine if you knew that planes were like like traveling by plane was a thing and you were born like in the 1860s, and then like you only at the tail end did you start to see it, and then like it became like you weren't able to travel unless if you hopped on a boat, it's a whole thing. It's like you don't get to go uh and and so now the planes, I have the same view of like technologies that we don't exist. At some point, you're gonna have like shoes that give you the ability to fly, and it's like I can't believe I live in the time where I don't get to fly with my shoes, you know? It's like there's some there's some equivalent of that.
ConorThat like, is it is it shoes that make you fly? Is it your ability to like teleport from one place to another? It's some crazy thing that we can't fathom that like at the tail end of my life, I'm gonna start to see the like the beginnings of, and I'm gonna be like, oh damn, like it would have been good if I had been born like 50 years later so that I could do that. And I feel that like AI potentially is that thing that's gonna unlock all this stuff, and uh, you know, potentially people are gonna be able to live like way longer. And anyway, I don't know.
BryceI I I guess part of part of my part of my thinking here is that I think that um I do think that like almost any technology, AI and mistakes with AI um will likely cause damage to society and communities. I think it's more likely that this happens not through ill intent, but just through bugs and accidents. Um in the same way that like, you know, there have been various disruptions of internet of the internet and of the stock market by, you know, uh bugs in in in tech.
BryceUm uh, you know, like there have been like things like the flash crashes from like uh uh automated like algorithmic trading uh uh companies that you know had a bug in their algorithm, something like that. Um I I think that for AI to present a real existential threat to us, you'd need a few things. One, you would need a an actual artificial general intelligence, which we're still pretty far from, I think. I still think that that's something that we won't see for five to ten years.
BryceUm uh you'd need that, and it would need to be a really a persistent thing um that uh is like a uh an ongoing and long-running process, right? And it would need to be incredibly resilient. Um if it's a thing that lives solely on the internet, um I don't know that that is uh sufficient to present a threat to us because uh I actually might disagree with that.
ConorYou know, one time in Toronto uh payment systems went offline for like four hours, and it caused like pandemonium. Like, you know, you you couldn't go to a bank to withdraw money, like you couldn't everyone was freaking out, and it it was it was like four hours of not having credit card things working, and like the payment systems were just all down. It was nuts.
BryceBut but there's a big difference between pandemonium and uh the robot turning this all into play paperclips, right?
ConorUh yes, but like uh you could imagine a world where uh with like deep fake and like misinformation and like shutting down the internet, it would just lead to absolute like anarchy. And it wouldn't be the end of the world, but the amount of damage that uh it could cause would be um like a word bigger than massive.
BryceI don't need like you know, completely like I can imagine I can imagine substantial damage, but not like existential end of human life damage.
ConorUm yeah, yeah, not existential, but like, yeah, there'd be uh millions and millions of people in the existential damage.
BryceI think you would need to have an artif a persistent artificial general intelligence that has a um uh a physical presence, which means that it would need to have robot robots or drones or like control of physical resources. And it would probably have to have control of those resources in like multiple uh you know resilient communication mechanisms that uh can't be disrupted by the various militaries and governments of the world.
BryceUm and that actually that that actually seems like that's a pretty high bar to clear to me, at least.
ConorI just realized like this conversation started off like an hour ago being like, yeah, like I don't get why people aren't excited about Clock 4 and uh and like you know, I think it's really changing the way the workflow is. Is it because they're worried about jobs, and now we're talking about like what are the actual parameters that would need to exist in order for the the uh AI and the artificial general or superintelligence to wipe us out? Yeah, yeah.
BryceI mean, like the the um I I I do think that like if you had like a scary like rogue model, some sort of scary persistent rogue model that that escaped containment and was like out there on the internet, um I I my guess would be that world governments have greater a greater ability to uh uh to turn off the uh the pipes of the internet and control things than we might imagine.
ConorYeah, yeah. I mean that's that's the case. A lot of countries they can just turn on and off the internet whenever they want, right? Um it would be pandemonium, but you know, uh and and then Yeah, if if you were talking about existential, like end of life, like humans don't exist, like yeah, that's definitely not what would happen. There would be a bunch of folks that survived. It wouldn't be a good situation, but like we would be able to rebuild for some definition of the word rebuild. Uh 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.
ConorThanks for listening. We hope you enjoyed and have a great day.
BryceLow quality, high quality, that is the tagline of our podcast.
ConorThat's not the tagline. Our tagline is chaos with sprinkles of information.
In this episode, Conor and Bryce chat with Zach Laine about C++26 Reflection vs CRTP, AI, kids & more. Link to Episode 313 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP:...
In this episode, Conor and Bryce chat with Zach Laine about the C++26 contracts fiasco. Link to Episode 312 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast:...
In this episode, Conor and Bryce talk with Zach Laine about his career and the career changes he has made. Link to Episode 311 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials...
In this episode, Conor and Bryce talk about Bryce's next chapter and where he is leaving NVIDIA for. Link to Episode 310 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The...
In this episode, Conor and Ben chat with Jason Turner about default construction in C++. Link to Episode 309 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast:...