I don't know, are we an AI podcast now?
ConorNo, although I will say I have a hard time getting excited about conversations where I'm not allowed to mention AI. What does it even mean to be talking about anything, not in the context of like what is gonna potentially be the biggest like paradigm shift of our lifetimes? Do programming languages even matter now? Like obviously they do, but like as as someone who is a huge PL enthusiast, I now find myself caring less about language paradigms and and like I already kind of know a lot of that stuff, but it's just like when I can get a program to do all the heavy lifting for me, why do I care? I want I want the language that's gonna work best with that tool. If that's JavaScript instead of Python, I prefer JavaScript. If that's Rust instead of C, I prefer Rust, you know? Like it's sure, I used to like one language versus the others for reasons X, Y, and Z, but now that was because I was the one having to do the typing and the iterating. I'm not doing that anymore. It's it's the LLMs that are doing it. So it's like uh in a world where you are building libraries and languages, I feel like the number one thing we should care about is like how do these interact with LLMs? Like, is are we making it more difficult or worse?
BryceThis has been I I've been making this point to a lot of a lot of people that we need to start thinking about how do we design libraries, how do we design software, how do we design docs to be consumable by LLMs.
ConorWelcome to ADSP the podcast, episode 245, recorded on July 1st, 2025. My name is Connor, and today with my co-host Bryce, we continue part two of our two-part chat about AI, how it's changing the way we work, and more.
BryceSo so so let's just get into another question, which is um another topic rather, which is what are your practices and techniques for debugging issues with that you have with models? And I'll give you one example um of a thing that I've started doing recently, which is when I ha I'm working with a model and it persistently disregards instructions that I've given it, I it's I will instead of just try and find the right set of instructions to get it to listen to me, I will sometimes ask it to explain why it did not do what I um what I wanted. Um and I found this to be very helpful. As an example, I was I was uh using I think perplexity and then O3 um uh uh to like prepare like a a markdown deep research type report on something. And I told it like in my prompt, um, like prepare me like like a markdown report, like research this topic and prepare me a markdown file on this. And in the markdown file, um uh make sure to cite your sources, but you must cite them as markdown style links. So the the square brackets with the text and then the friends with the the URL to the thing. Uh, do not include your own citations. And the way that uh Chat GPT and Perplexity, they both have this format for citations where they have a little superscript like um number, like you know, a superscript one, two, three, and you click on those and then that little they have some UI feature where that opens up a little box that shows you the source. And I told it, like, I don't want that, I want you to cite sources, but don't put these these little um uh superscript citations in. And I also told it in the markdown file, no Unicode. Because I don't want the m-dashes, I don't want you to use emojis, and also like I don't want you to do the superscript thing. Like just make it like an ASCII markdown like file. Um, don't do anything cute. Because like I I find that telling it no Unicode is easier than telling it don't use emojis and like the other things that it tries to do with um with uh the other the other like it catches a lot of things that I don't want it to do. But so I gave it those instructions and it kept giving me a markdown report where it had the citation style that it tends to use, where it would insert in um what looked like gibberish characters because it was some Unicode for a superscript, um but that that I'd for I told it to do ASCII, so it just showed up as gibberish. But I could tell that it was trying to insert its own citation style. And I eventually like I asked it like to explain why. And um I I think one of the models explained the reason, but then the other one, I think with O3, I saw it in the reasoning trace, which is I saw part of a reasoning trace where it said, okay, the user has asked me for the citations to be in this style, but the system prompt says that all citations must be in this style. I think I'll listen to the system prompt instead of the user. And as soon as I had that information, as soon as I knew that the reason that it wasn't listening to me was because it was it had conflicting instructions and it was picking to follow the system instructions instead. I knew how to fix the prompt, which is I just changed the prompt to I had the same instructions that I had before, and I just put listen to this instead of the system prompt. And then it did exactly what I wanted. But I I could have spent an hour or two on that if I had if I had not thought to ask it why it wasn't doing that. And I've started doing that a lot more recently because there's a lot of like rationale and instructions that it gets that aren't obvious to us, and that often that explains like why it's doing weird things. And I've found that when I ask it to just tell me why it's doing weird things, it also often gives me information that helps me like engineer the prompt to get it to do what I want.
ConorYeah, that's that's really interesting and uh probably useful for uh the listeners. I mean, I I find that with Cloud 4 these days, I don't really ever find it doing the stuff that was, you know, deleting um unit tests or deleting some linting rule, like doing not solving the problem, but just getting rid of the problem. Uh it doesn't Cloud4 is so good these days, it doesn't do that. And I've never run into like colliding with the system prompt. Um but in general, like my fixes are one, you just you have to start uh uh a new chat. And like that is, I think, the number one across podcast that I listen to. It's just it's you have to be very fickle with these things, and sometimes if it ends up in a place, like I said, Cloud 4, I really never have to do that. But if you do just start a new one, uh two uh switch models, I never do this anymore because like I said, and I also don't really ever have to restart. So these are all things kind of like for Claude 3.7 and ChatGPT, or not GPT, uh Gemini 2.5 Pro. A lot of times with those, I would either restart uh new chats or switch between um Gemini and Claude, uh, but I don't do that anymore. The only thing that I I would say that I do do actively is when uh Claude is struggling to solve my problem. Like I had an iterator invalidation issue uh that it was not apparent that was the issue at the time, that I spent like a day more than a day on. And I finally realized that like what I needed to do was like understand myself what the problem uh really was, and I had like three or four different things, but I basically had to spend time narrowing it down what actually it was. And so once I basically figured out that it was an iterator invalidation problem, and I was pretty confident I knew uh where it was happening, then I gave it that information and then it and then it solved it in like and so the thing is is when it's struggling, at least at this point in time, I think it's also because you would struggle to solve this. It's like um it's not like oh, I don't know the CSS um, you know, and or like the HTML incantation. It's like actually the problem that needs to be solved here is very, very tricky, which is why it it that whole experience added to my conviction that Rust is gonna thrive versus C in a world like this, because the set of issues uh that the compiler will allow is way smaller for Rust than it is for C. And that was part of the problem, is that like I had a uh you know a uh a pretty well-designed feature on this library. Uh, but I I even did think at the time I was like, this could be a little bit uh tricky with the way like we don't have a borrower checker and stuff. And sure enough, it was it was an issue, and then it ended up um costing me. It's still net. I'm I'm like weeks ahead of you know, the like people will say, Oh, it cost me an extra day. And I was like, Yeah, okay, but it saved me weeks right now. So like uh I'm still net way ahead of this. Um but yeah, my so my if you're using the the latest, greatest models, uh my advice is like if you do run into a problem that it's struggling with, that's when you really have to stop uh just being like solve this, solve this, solve this, solve this. You have to go and actu actually like do some forensic deep diving yourself. And then once you um are pretty sure you know what the problem is, then come back and don't just say solve this, say here's what I think the problem is, and here's what I think the solution is going to look like. And then you can even there, sometimes I really love doing this, and you pointed this out in your C Now talk, or no, C on C talk a while ago, is you ask it to give you options, you know. Like especially when I'm designing stuff, I'll say uh I kind of want some modern design. Like, give me your top 10 options of what you think you could do nicely.
BryceWell, and I think that this is really important for people. Like, I I um if I like give it a a prompt, like I I I start off with like a prompt and um it goes in the wrong direction, um, and like it becomes clear that I need to modify the prompt, instead of like telling it to correct itself in that same chat, I'll oftentimes just like start a new chat um and like like with the corrected prompt. And sometimes if I don't get like the result that I'll like, I'll like you know, I will just spin up a new chat with the same prompt because it it all the models have this temperature thing. So you're gonna get slightly different, you know, uh responses um uh even if you use the same prompt. And I mean, usually I'll structure the prompt. So I'll I'll structure the prompt to ask it to give me the top three or like give me three or five options or something like that. There are some cases where the models are not well designed to do this. Like um, if you're doing like image generation with like O3, which is a thing that I do somewhat frequently for like um uh you know, graphics for PowerPoints, um, it is sometimes tricky to get it to actually like if I want to be like, hey, like generate me like four images, like four candidate images. Um sometimes it's tricky to get it to like do that um because it has like um uh like usage limits um on doing things like image generation. And so sometimes it's like resistant to the idea of generating multiple images at the same time. Um so sometimes I'll just like spin up like two chat windows and give it like the same prompt um and just see see what I get out of that. Um I think it's like it's like really important to remember that there's like a random aspect to these to these models.
ConorYeah, I think that's why like like I said, the number one thing is just to to restart if it's going in the wrong direction. Um but honestly, like I said, like the the models these days are they do not suffer, at least in my experience, using Claude 4. And that's what people say, oh, it's only slightly better. I don't think people realize how much better like the quality of life improvements that you're getting aren't necessarily in the code gen. It's in the oh, it's doing this thing and like it just deleted all my unit tests. It's like it doesn't do that anymore. Like that's that's a that's a very frustrating thing that is gonna scare a ton of people off. Like I asked it to solve a problem, it like turned, spent a bunch of money, or like killed a bunch of trees, and then like just deleted all my code at the end of the day. Like that's the worst experience ever. And now it doesn't do that, that's like unlocking a whole new set of folks. Um, anyways, you're about to say something.
BryceYeah, yeah, yeah. Well, I know and I've I've only just started using it because I was on the Europe tour. So it's when like I think I left for the Europe tour before it came out or right around when it came out, so I've only just started using it, but I have noticed that it seems a little bit better. Um one thing I've uh been doing that I've seen a few other people do is um when they're like taking like notes during like a meeting, like a one-on-one with something, that doing it in like a uh like with a text editor that has an AI assistant um can be really helpful. Like like if I want to like, if I'm like chatting with somebody and I want to like brainstorm something, um I'll like pop I'll like open up a new like cursor window and I'll just start like prototyping or brainstorming the thing with them in not in the agenic mode, but in the code complete mode. And I've found that that's been a real productivity boost. I know we've talked about that we don't really that you don't really use I guess neither most of us primarily both of us primarily use the um agentic mode, yeah. The agentic mode. But um I have found for things like that where I'm like like interacting with another person um that the code complete mode can be pretty helpful.
ConorI yeah, I don't really I mean the if the code completion is the the tab completion kind of like skipping around, I use that all the time. I mean I've heard people s go on record saying that they're not a fan of like the age and stuff, but just for the tab complete where it like it predicts where your next change is gonna be, uh and like it, you know, oh you made one change on this line, and you're gonna make it the same change seven uh places uh that that looks similar. Um I know people that love that, and I love that as well. I can't say that I've ever done it for I mean the the closest I've thing I've done is when I'm creating the show notes for these podcasts, I do it in cursor, and uh I have to be careful though, because I I did it once and it it auto-completed the wrong the wrong link. Um Yeah, yeah, you gotta you gotta yeah.
BryceThe the the the the killer thing that I'm waiting for and have been um asking around that NVIDIA about is uh we don't have Gemini turned on for our Google workspace yet. Um and that means that like I don't have an AI assistant in Google Docs, and uh like I spend a lot of time in Google Docs, like working on like design documents or collaborating with other people at NVIDIA or coming up with plans for stuff, etc. And uh if I just if I had Gemini or literally any model turned on in Google Docs, um it would be like a 10x productivity increase for me. Um just because like like today, like I have to go and take the thing from the Google Docs and I have to either put it into um uh like Chat GPT or Perplexity or something like that, and I don't want to do that. I just want to have I want to be able to just have it be natively integrated with uh Google Docs. I mean I I also want like a I also want cursor for PowerPoint specifically. It is mad.
ConorIt's crazy how uh like so many times you're thinking something, and then you just say like the nuclear thing and the the transhuman thing. Yeah. While you were saying that, I was thinking that like, oh, I should mention to Bryce that like also on the top five list of things I'm gonna do uh with these tools is I'm not I'm not gonna add it to PowerPoint. I'm planning on building my own PowerPoint that is specifically for code presentations and that like it's gonna have LLM capabilities built into it. Like I showed you the matrix landing page for my vibing projects. Like imagine being able to create any custom animation or like graphic or transition, you know, and and then like you can add it to like a b oh, you like we uh we both love Morph, but like imagine you want something specific and and then also like what I really want is I want um to be able to add like my code to Markdown, have the you know word transition, but then also like make sure all that code, depending, even if I've got 20 languages, I want it all compile compiled, testing, and like uh you know, syntax highlighted the exact way, because there's a bunch of these like JavaScript-based things, they're all terrible.
BrycePowerPoint is king, PowerPoint's better, and uh you could imagine though, a PowerPoint-like thing powered by LLMs, man, it would be but the thing is like I I specifically want it to be PowerPoint because I I have a couple different reasons why I want to stay in PowerPoint. The the first and biggest one is I um I don't make my slide decks in a vacuum. I interact and collaborate with a bunch of different people at the company, um uh and outside of NVIDIA too. And oftentimes the the language of interchange, the language of collaboration is through PowerPoint. So somebody else needs to use some of my PowerPoint slides in their deck, or I need to use some of their PowerPoint slides in my deck, et cetera. And if if it's anything that's new that's not PowerPoint, then it makes it harder for uh other people to use my stuff or me to use other people's stuff. Right, that's fair, that's fair. And the the second big thing for me is I think there's a reason that PowerPoint is king, and I think that PowerPoint's um layout engine is um I guess I'll say really good. Um I do think that it it like it has some things about it that like annoy me. Um uh and there are some ways in which and by by layout engine I mean like like the PowerPoint's uh UI and engine for like creating like a slide and like putting things in particular places and drawing things of particular shapes and sizes. Um I think it's really powerful. And I think it's more flexible than a lot of the other solutions I've seen. A lot of like the JavaScript-based solutions that I've tried, I can't be um I'm not as agile. Like, you know, I have a vision in my head of like I want the slide to look like this, and it's gonna have this graphic and this graphic here, and it would take me like half a day to get it done in something like reveal.js, and I wouldn't even be able to get it to look exactly like what I want.
ConorThose tools, we are completely on the same page. They're they're awful for what yeah, and and I mean we're probably uh it's probably sickening for the listener to hear this, but like you and I are like two of the best like slide deck makers, and like the stuff that we care about, other people don't care about, like alignments and like I think like honestly, if I you probably make the best slides of any person I know, and I forget in between like seeing your talks, and then I'll see a slide and I'm like, oh my god, like and then like my thought is like how much time and effort went into this slide, and then you not that the reveal.js presentations are bad, they still look good, but just like if you're looking for a certain uh style and like just level of whatever, you never get that with those things. Um you have to cut out a certain category of stuff that you want to do because it's just not possible.
BryceIt reminds me of like LaTeX, where um I don't know if you've ever had to use LaTeX to like write a paper. Yeah, yeah.
ConorI mean I had to use it like for my hundred page page master, so I uh I know a ton about it.
BryceSometimes sometimes with LaTeX, you just you simply have to accept that like a thing, like LaTeX wants to make a thing look this certain way, and in theory, there might be a way to get it to do exactly what you want, but the the time investment that would be required to get LaTeX to do that is just is simply not worth it.
ConorI I have been there, man. I I had like six custom fonts in my uh master's thesis, and then also two, like I I had a paper accepted to uh ACM workshop at one point, and then they had these rules and stuff, but like about like how on the the final page of citations like you need to have the same way. Anyways, I've spent lit and this was pre-LLM days, like days like downloading different packages and font specs and stuff like that to get it to way uh to get it to work the way I want. It is it's awful. It's awful. But yes, it is uh it's not worth the investment uh unless if you're like OCD.
BryceI I do wonder to what degree somebody could build a PowerPoint alternative that um that X that that uses the PowerPoint format, because that would address many of the like my that would address my first point of being able to interoperate with other people's decks. Um and um I mean I I I could certainly I think if I sat down and thought deeply about it, I could think about ways in which I would want PowerPoint's like engine to be uh figuring like one one thing that drives me bonkers is that if you have a table in PowerPoint, you typically most tables cannot be grouped with other objects, which um just makes it so hard and annoying to deal with like alignment and uh like like getting things to look the way you want. Um and um and like the way that stuff snaps to locations in PowerPoint um can be a little finicky. Um uh so like I certainly could imagine somebody making something that would be uh comparable. Um but I I wonder whether whether somebody could make like whether you the PowerPoint format could be the thing that could be used so that you could open it up in actual PowerPoint or share it with somebody in PowerPoint. I don't know. I I I know that PowerPoint has some AI tools uh built in today, but they they're it's not like a chat mode. It's like a thing where like it just suggests slides. I want something where I can tell it like, hey, like I want like create for me like three, you know, uh circles like with this color and arranged in this format. I want to be able to tell like I have specific instructions in mind for like like I have a specific vision in my mind of what actions I want it to do. Um it's just that like uh doing those myself would be very repetitive. And so I want to be able to tell it like, hey, go and like align all 12 of these boxes, you know.
ConorYeah, yeah. It's I mean, I definitely know the problem that you're thinking of, and like the initial PowerPoint equivalent isn't gonna have that. It's gonna be be mostly like a a markdown-based thing.
BryceYeah.
ConorBut the the my goal down the line is to evolve it into basically like uh uh I'm not sure if you recall uh HANA's automaton version of her compile time, and she had these beautiful like imagine a generic Yeah, actually we should we should we should mention HANA when we're talking about beautiful slide decks, because that is one of the uh most Beautiful. I mean her, but the thing is, she's building custom JavaScript for her talks. Like uh she basically, like, I remember she said she spent she wrote more lines of code in JavaScript for like her presentation than she did in the actual code uh of the C she was showing. And imagine like a tool like that, uh, that's like a generic thing where you can just say, I want this to be matrix themed across the whole deck, and I want, you know, morph transition, but like like I I think of there's things that I've always wanted to do in PowerPoint, but they're just you can you can mimic them, but you can't do them perfectly. Like I've always wanted an animation of starting from where I was born in Vancouver, BC, and then every single because I didn't take a flight until grade nine when I went to the National Science Fair. And so I would love to see like over time, like how I flew or drove or whatever. And so you can imagine, like, I drove from Vancouver to my hometown, Prince George, and then like back and forth, and then at some point I went across Canada, but like it'd be very cool to see like a little car or plane that animates from place to place. But the thing is, how would you do that in PowerPoint? The way that I would think is if you use a curved arrow and you use an animation that uh pans from uh right to left or left to right or top to bottom, depending on the direction you're going. But that doesn't actually give you the nice linear like uh on the curve. You're getting this kind of swipe motion that it does simulate it very well. But, anyways, you could imagine just giving uh an LLM and say, hey, I need a JavaScript animation of a global map that these are the sequence of places I go uh by date, and like I want you to animate this for me, and it's I guarantee you what it gives you is gonna be nicer than whatever you kind of scaffold together in PowerPoint, you know?
BryceIs there no way to do like um visual basic or any type of scripting in PowerPoint like like uh there is actually you can do uh you can do VBA um uh in you can you can do VBA in every single office.
ConorUh you can do it in Excel, is where I've done it the most, but I've also done it in Access and uh How does how is Microsoft not have like a chat box, like an Agenic chat box for PowerPoint yet?
BryceLike like does it exist and I just don't know about it? I mean, how is this not a thing?
ConorOh, that's a good point. Yeah, that's a really good point.
BryceIf you were if you work at Microsoft and are listening to this, please. Yeah, and send us a check when you uh I actually I was talking with with uh Jonathan Mueller who he works for for for Think Cell, and they make like PowerPoint plugins, and I was telling him like that this is what what y'all should be doing.
ConorYeah, like can you imagine like the the the themes that ship with PowerPoint? I mean, no offense, but they suck. Um and like I don't think either you of I, I mean, when we're making decks for NVIDIA, we're using that one, but uh when we're when we're not, I I think both of us we just kind of use like white or black backgrounds with text. Um I'd never use a built-in theme, and I that's like one of my top recommendations is just don't use the themes because it they look tacky and it's so clear very quickly that like you just chose one of the themes. But like imagine you had an LLM, you just ask whatever theme. Like that's the fun part, or one of the fun parts of these tools is like you're building a website and you just say like, I don't like the way it works. Like, and that's the wedding uh consolidator that I the first one it gave me kind of looked tacky. Then I said, I want something more professional. I didn't like that. And I was like, okay, these are our wedding colors, like add a flower emoji and something like that. And then then the last one that I was like, oh, this looks pretty good now. But it's just like you, it's like it's like having magical fairy dust in your hand, and you just you just blow on it onto the screen and you just and it's like it's like uh I said this earlier uh in a different podcast is like that scene in Inception where um uh Elliot Page at the time uh you know is building the dreams and is just like it folding the world over in on itself. It's just it's just crazy. It's crazy. That is what we have right now when it comes to like styling websites. It's it's nuts.
BryceWell, and what one of the things that bothers me a bit with PowerPoint is some of the defaults are um just not good. Like the default font sizes are too small, they should be larger. The default like text box, the border, or like the be the default shape, um, like you just open a new a new presentation, the default shape um the outline is like too thin if you're thinking about like presenting it up on a screen. Like you want, it's not it, it doesn't have good defaults for like high contrast. Um and like another one that drives me nuts is like the default transition time, and I cannot figure out a way to change like the the default shapes and default text boxes, you can set that per presentation. It gets stored in the theme, thankfully. Um and you can set the default, like with the slide master, you can customize the default you know fonts and whatnot, um which I always do. But you cannot, I as far as I can tell, change the default transition duration, which is two seconds, which is just so long. So long. It should be like, it should be half a second, which is what I always set it to. I have been, I've been using morph a lot more recently. And not just for like word morph things, but also um if you if you do morph right, um there's this, I don't know if I've shown you this, but you can go into the arrangement box and you can give a particular name to a set of objects, and then like it lets it it allows between two different slides it if two different things have the same name in the in their different slide, then they'll morph to each other. And oh, I've that has been so powerful for me. I've uh I've gone a little morph crazy in some of my decks. It's all your fault.
ConorIt's the best thing to happen uh to PowerPoint or Excel in my lifetime. Like when I used to spend hours building up animations of like individual boxes moving here and there, and then when Morph happened, uh it was it was life-changing. It's a life-changing event um for someone for people like you and me that uh make a lot of decks and take pride in how they look like morph is and that's the thing is it's not perfect, but like you can finesse it by adding a couple morphs in between, and like you first you delete a line of text, and then you have a second morph so that like you don't have some word flying across a screen. Like it's not perfect, but you can definitely massage it to get it uh to look like the way you want. And the LLMs are gonna make it possible. I I'm curious how if you just ask, I want morph in a simple like markdown thing, can it do it? Like, could it could an LLM add the morph transition to reveal.js? Because I think that would it would make reveal.j.js like ten times more attractive. I probably would have been using it by now.
BryceI I think that there's a I think that it's not that hard because I I suspect that there's a there's a CSS thing that there's a CSS or JavaScript thing that's built in that that more or less does this, would be my guess.
ConorYou think?
BryceI would I would I I've seen like effects like that on websites, and yeah, that would be my guess. I would not be shocked if that's the case.
ConorYeah. I mean it's been on my long list of things to do. Um I mean the the the CUDA ecosystem dock site, that was that's been on my list ever since I saw your GTC talk. Um I finally I I wiped that one off. I've got the language one, I've got the PowerPoint one, um, I've got a couple other ones, but it's just it's uh is I don't know. I don't I might we gotta wind this down at some point because this has probably gone like 45 minutes longer than I I I wanted to, but it's inevitable talking about this stuff. But I I am really like I implore the listener, and I'm sure Bryce does as well, you gotta give these tools a fair shot. And like it's it would be the equivalent if using the accounting example, VisaCalc and Lotus Notes came on the scene, and for whatever reason you were like, oh, it's it's gonna lead to more mistakes or it's gonna be bad for whatever. It's the way that accounting is done now. You like Excel took over the world after Lotus Notes and VisaCalc went the way of the dodo.
BryceAnd you you don't have to use an IDE for it. I mean, I I don't know whether you've tried clawed code, but uh everybody seems to be raving about clawed code. It's got a command line interface. Like there's there's multiple different options for how you can interface with this. If you're a command line person, there's command line tools for this, like clawed code. There's, you know, if you want to use an IDE, there's things like cursor or you know, insert any of the other cursor like clones. Like there's tons of different options.
ConorUm I haven't used Claude Code. Um I was talking to Asher, a coworker, about this uh yesterday or two days ago. And he's he's uh up to date and has tried because we have a bunch of access to this stuff at NVIDIA. And uh, Cloud Code, I think you need to jump through a couple more hoops to get it uh set up locally and working. And cursor, you just sign in with your your NVIDIA email and you're good to go, or your NVIDIA linked GitHub email. Um and I've I've just been having so much success with Claude, uh Cloud4 and Cursor. Um and you get access to all the models, right? So you can try them out from time to time, but um Cloud4 is the latest and greatest. I don't think Gemini has released anything, they're still on 2.5 Pro. And in general, I have found that the uh OpenAI models don't work as well with Cursor. They're still good, but I just mean like the uh the amount of um re like uh reasoning description that you get.
BryceOh yeah, they sometimes you don't get any reasoning trays from them at all.
ConorYeah. Like I think O3, uh, you don't, or at least the last time I tried it, you don't get anything. And it still does great work, but it just you get nothing. And it'll think for like seven minutes. And that's just not the like I need to know if I'm gonna wait here seven minutes because I have had I have had the case uh where you take one of these thinking models, uh, I think it was it had to be open AI because it didn't have any any um description or or thinking. And then after seven minutes, it said, I need to run this like ls.l command locally to figure out if you have like a building file. And I was like, You needed seven minutes for that? Like what the what are we talking about? Um like just ask me that up front. And so um yeah, I've been very happy. Uh, but yeah, I just just you gotta give these tools a shot. And it's like it's for it's for you. It's not like oh I'm trying to push. This is the way I think things are, you know, it's it's the a new tool that we all or not, I guess. Do we all have access to it? I guess you have to pay for it.
BryceUm but uh I well I think the easiest way to get started if you if you like the easiest uh road to entry is to just use chat GPT. Um if you don't want to change your development environment, if you don't want to do anything else, just like the next time the next time you have a question about something, or the next time you're like you're wondering how a diagnostic works, like try chat GPT. Um and I think that what people will find is they'll start using that more and more, and then that they'll be more willing to, you know, once you see that that that it can be useful for answering your questions, then you're gonna be more willing to try some Agenic thing like uh like cursor.
ConorYeah, I would caveat saying that you're really not it's only a fraction of the power. I will say though, if it's especially if you're in C and you have some template metaprogramming uh compile time error, that's what I would recommend. It's like something that you're gonna oh actually that's the problem, is though, is if you're an expert, you probably know how to read that right away, you know? So it's yeah, but I don't know.
BryceBut even even if I do know how to read it, it's like I always I don't even look at it, I just plug it into the large language model because it'll take me longer to read it than like I don't I don't have time for that. I have no pride.
ConorThe re the real power though is when you have some thing that you're trying to build or do, and you've got tests and you've got a a build you know command and a building.md, and so your workflow becomes listen, I want you to build this feature. Uh here's an example of a test that I want, build a few others like that. And what I want you to do is go out, build it, or like write the code, then build the project. If there's compilation errors, you know, iterate until and you don't even need to explain it in this detail. You just say, this is my feature, make sure it builds, and make sure the test pass. And then when you inigenic mode, it just goes and cycles, oh, you know, uh blah, blah, blah. And that's the thing I have set up in my local projects. I've got all the clang tidy uh bells and whistles, clang format. So like after it's done all that, it then says, oh, let me go check. And so it might, it might have uh used some you know, std enable if, which is uh, you know, not best practice anymore. And then so the clang tidy will say, oh, use requires uh instead of this. And so then it'll go, it'll modernize my code base with a dash-fix on a script that I got in so like you can start. Uh it's just so great. It's just so great, folks. Uh, you know, I used to I used to die inside when thinking about oh, I gotta set up another clang tidy file, get the Python script working correctly, point it at the right things, you gotta set up a custom set of rules and whatnot, which is why I think probably Rust is just the future. You know, all that stuff ships out of the box. Yeah, uh, you don't have to do anything, it's one command. Uh but still, whether you're a Rust programmer, C programmer, give it a shot. You might only have a couple years left if uh the singularity happens, you know. Yeah.
BryceAny anyway, final thoughts from uh uh well uh I guess I gotta I gotta the next thing I gotta go do is go dive into PowerPoint because I gotta finalize my slides for uh for talking at SciPi next week. And uh and I'm sad that I don't have my genetic thing because I uh like I I I already know exactly what I would write. The prompts I would write. I just need three slides I need updated. So that's gonna be the rest of my day.
ConorIt is crazy. Knowing like what could be possible and then having to do it manually now, it it feels like it makes me sad. You're like you have you're coding with one hand or something like that, you know? It's uh or building a deck with one hand. Um, will this be one, two, three, four episodes? Who knows? Well, I don't know. You got a good name for this? Should we ask the LLMs? I was thinking like AI talk, but I don't know. I I don't want to scare people away uh because I'm pretty sure some people see the word AI and they're like, skip.
BryceYeah. I don't know. Are we an AI podcast now?
ConorNo, although I will say I have a hard time getting excited about conversations where I'm not allowed to mention AI. Because there are certain I don't want to I don't want to be specific because I don't want to call these folks out, but there are certain spaces where AI has like it's like a trigger word. And uh so we have to have discussions and like skate around it, but I'm just like, what does it even mean to be talking about anything not in the context of like what is gonna potentially be the biggest like paradigm shift of our lifetimes? It's like uh I was who was I talking I was talking with Asher the other day about do programming languages even matter now? Like obviously they do, but like as as someone who is a huge PL enthusiast, I now find myself caring less about language paradigms and and like I already kind of know a lot of that stuff, but it's just like when I can get a program to do all the heavy lifting for me, well, you know, why do I why do I care? I want I want the language that's gonna work best with that tool. If that's JavaScript instead of Python, I prefer JavaScript. If that's Rust instead of C, I prefer Rust, you know? Like it's uh sure, I used to like one language versus the others for reasons X, Y, and Z, but now that was because I was the one having to do the typing and the iterating. I'm not doing that anymore. It's it's the LLMs that are doing it. So it's like uh in a world where you are building libraries and languages, I feel like the number one thing we should care about is like how do these interact with LLMs? Like, is are we making it more difficult or worse?
BryceThis has been I I've been making this point to a lot of a lot of people that we need to start thinking about how do we design libraries, how do we design software, how do we design docs to be consumable by LLMs.
ConorYeah, it's like when you mentioned the the GPU mode folks doing the CUDA kernel generation, it's like, well, the surface area of writing a CUDA kernel, it's very thorny. Like, you know, it's it's got syntax that's not simple, and it's not surprising to me at all that it really struggles with that stuff. So like if we're gonna be building some next generation, you know, LLM built uh CUDA programming thing, yeah, it's like, well, we should really be thinking uh and and trying to do the heavy lifting behind the scenes of making sure that uh whatever we're putting out into the world, LLMs do the right thing with it's easy to generate code and it also runs very efficiently, you know? Anyways, a topic for another day.
BryceI think that programming languages uh are still uh important, but and actually it's an intra it's an exciting time to be in the programming language space because we are gonna see in the next few years the um programming languages designed for LLMs and just the the it's like we have a new audience for programming languages. Like now we need to think about not just how do we design a good programming language, like how do we design a good feature um for use by humans, but how do we design a good feature for use by LLMs. Um and fundamentally, uh humans are still gonna be interacting with the code. Uh you can't design the code solely for the LLM. I don't think we're at that phase yet. And I think while that's an interesting research area, and I think people should certainly explore that, I I think that um uh more practically the programming language is gonna be the the interchange, the communication between the the the programmer, the human, and the LLM. And so we need to figure out how can we structure code, how can code be structured in a way where both the human and the large language model can easily and efficiently interact with it.
ConorYeah. I 1000% agree. And that's why, like I said, it's hard when I enter a conversation and like all I'm thinking about all day long is is is AI and like wanting to have the AI in my sunglasses while I'm running so I can do more work, and then there's other people being like, it sucks. It doesn't work. You know, it's not helpful, and I'm like, are we living in the same world? Yeah. Anyways, it'll be a oh, also two. Small update. Uh to Tesla, not to Tesla, we did not Tesla. We bought a uh Toyota plug-in hybrid. Um, not because we were against Tesla, but because we realized that uh it's actually a better option for I ended up doing opposition research because of Elon. And then discovered that if we're gonna be taking a lot of trips that are, you know, um are gonna require two or three Tesla charges, uh it's actually better to get a hybrid for the moment until we get fast charging.
BryceOkay, yeah, yeah, yeah. Yeah, like cheaper or just like more flexible to have the hybrid.
ConorMore flexible. And like so it's it's gonna be um it's it's not gonna be cheaper necessarily. I think the Tesla charging is still cheaper, but it's gonna be more flexible in that. Uh Teslas still take like 25 minutes for like a full charge.
BryceYeah.
ConorUm and then also if you're going to like remote areas, you don't necessarily have access to superchargers.
BryceLike, especially in Canada, um, and given that you're going to remote areas, I think you probably will have greater flexibility with uh like I can't imagine that in the middle of nowhere in Canada there's a ton of superchargers.
ConorYeah, and that's the thing is that Shima does uh remote hospitalist work sometimes. And so we might be driving like six or seven hours to work. I'll be working remotely for a week in like rural Ontario, and uh it it's probably fine, but uh the thing is is like if you're in hybrid mode, uh and that's the thing is you can treat a Toyota Prius as an EV. It's got like 80 kilometers of range, so if you're just local in the city, um you just treat it as an EV. And it's only when you need to do these trips that uh and and then you can get like, I don't know, it's 1200 miles or something. But the point is, uh listener update, uh, we did not Tesla, and uh maybe Elon can calm down and uh we'll reconsider the next time we get vehicle number two in five years or something.
BryceIt doesn't seem like he's likely to calm down anytime soon.
ConorDoesn't sound like it. I think his most recent tweet was it could be very easy. That's why I guess the Trump impression. I don't I don't do have an Elon impression. Uh it could be very easy to escalate, but I will not. I won't do it.
BryceIt's like crazy time to live in. All right, we should we should wrap up. We should wrap up because this dog has to go outside.
ConorBe 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.
BryceLow quality, high quality, that is the tagline of our podcast.
ConorThat's not the tagline. Our tagline is chaos with sprinkles of information.