1 00:00:00,000 --> 00:00:02,910 Alyss Noland: I've had this perspective that Anthropic has 2 00:00:02,910 --> 00:00:04,140 taken more the business route. 3 00:00:04,140 --> 00:00:04,410 Right? 4 00:00:04,410 --> 00:00:09,120 And OpenAI had Chad GPT and they had their consumer explosion and they don't like, 5 00:00:09,120 --> 00:00:10,410 they're just gonna play to their strings. 6 00:00:15,835 --> 00:00:18,765 Corey Quinn: Welcome to Screaming in the Cloud, our returning guest. 7 00:00:18,835 --> 00:00:21,445 It's been a hot second since we spoke. 8 00:00:21,775 --> 00:00:24,205 Alice Noland, you are now at Nvidia. 9 00:00:24,295 --> 00:00:25,365 Thank you for joining me. 10 00:00:25,575 --> 00:00:27,365 Alyss Noland: Thank you so much for inviting me, Cory. 11 00:00:27,385 --> 00:00:28,440 I'm excited to be back. 12 00:00:29,070 --> 00:00:32,159 Corey Quinn: This episode is sponsored in part by my day job Duck. 13 00:00:32,159 --> 00:00:35,370 Bill, do you have a horrifying AWS bill? 14 00:00:35,640 --> 00:00:37,530 That can mean a lot of things. 15 00:00:37,740 --> 00:00:42,840 Predicting what it's going to be, determining what it should be, negotiating 16 00:00:42,840 --> 00:00:48,269 your next long-term contract with AWS, or just figuring out why it increasingly 17 00:00:48,269 --> 00:00:52,769 resembles a phone number, but nobody seems to quite know why that is. 18 00:00:53,070 --> 00:00:53,760 To learn more. 19 00:00:53,760 --> 00:00:54,269 Visit dot. 20 00:00:54,385 --> 00:00:56,665 Bill hq.com. 21 00:00:56,964 --> 00:00:59,815 Remember, you can't duck the duck bill. 22 00:00:59,875 --> 00:01:04,675 Bill, which my CEO reliably informs me is absolutely not our slogan. 23 00:01:05,880 --> 00:01:09,690 It was hard to find time because now that you're at Nvidia, we had to schedule 24 00:01:09,690 --> 00:01:14,370 this around the ear shattering noise of the dump trucks dropping off gold 25 00:01:14,370 --> 00:01:16,740 bricks in the Nvidia office, car park. 26 00:01:16,800 --> 00:01:21,000 Apparently it's a, it's a company that's on top of everyone's mind. 27 00:01:21,120 --> 00:01:25,890 They look from some angles, like a finance company, like a GPU manufacturer, like 28 00:01:25,890 --> 00:01:29,730 a gaming company that doesn't realize it's a gaming company, a cloud services 29 00:01:29,730 --> 00:01:34,350 company, an investor, a private equity firm, and fill in the blank here. 30 00:01:34,660 --> 00:01:34,900 What? 31 00:01:34,900 --> 00:01:35,650 What do you do? 32 00:01:36,100 --> 00:01:39,010 Alyss Noland: Well, what I do varies on a day-to-day basis, and I love 33 00:01:39,010 --> 00:01:42,520 when I get this question, but I like to describe myself both internally 34 00:01:42,520 --> 00:01:44,050 and externally as a problem solver. 35 00:01:44,200 --> 00:01:49,540 And where that is focused is I end up doing what's more in the 36 00:01:49,540 --> 00:01:54,490 vein of developer cloud ecosystem development, working with AI developer 37 00:01:54,490 --> 00:01:58,540 tools type startups, and figuring out where are we going to go with. 38 00:01:59,040 --> 00:02:02,220 The experience around development experiences and how do we 39 00:02:02,220 --> 00:02:04,680 bring that to people outside of the code, adjacent fields? 40 00:02:04,860 --> 00:02:08,370 Corey Quinn: Is this tied deeply to the GPU side of the business? 41 00:02:08,430 --> 00:02:12,390 Uh, again, I must confess my taxonomy mentally of you folks is not great. 42 00:02:12,720 --> 00:02:15,390 Alyss Noland: You know, for better or worse, both violating Conway's law and I 43 00:02:15,390 --> 00:02:20,340 constantly, and also rearranging, and so no one can keep track of what Conway's. 44 00:02:20,645 --> 00:02:24,995 Law that's being reflected out even is, so it is tied to GPU business 45 00:02:24,995 --> 00:02:28,745 in so far as everything at NVIDIA is tied to GPU business, but it 46 00:02:28,745 --> 00:02:30,335 is more on the software side. 47 00:02:30,335 --> 00:02:32,885 So NVIDIA's no stranger to software. 48 00:02:32,915 --> 00:02:35,375 They've been doing software for GPUs for a very long time. 49 00:02:35,615 --> 00:02:39,605 Cloud software is something that is, there are more recent entrant too. 50 00:02:39,635 --> 00:02:42,965 I would say probably in the last like three to five years is when they 51 00:02:42,965 --> 00:02:44,585 really started to dip their toes in. 52 00:02:44,765 --> 00:02:47,705 And the company I was previously working for, Okta AI was 53 00:02:47,705 --> 00:02:49,355 acquired by them last about. 54 00:02:49,710 --> 00:02:53,640 A little over a year ago, um, along with a couple of other, uh, startups that 55 00:02:53,640 --> 00:02:59,549 are in that approximate region, cloud software, everything happens on GPUs. 56 00:02:59,579 --> 00:03:00,990 Um, mostly cloud GPUs. 57 00:03:00,990 --> 00:03:05,430 Sometimes I'm working with, um, more hardware than not, but most focused 58 00:03:05,579 --> 00:03:09,090 it's on what are the abstractions above that hardware layer. 59 00:03:09,240 --> 00:03:13,650 Even sometimes things like, uh, are we using a firmware that works and is 60 00:03:13,920 --> 00:03:17,190 optimized and helping speed people up, and how can startups even access that? 61 00:03:17,775 --> 00:03:19,995 Corey Quinn: The world in AI right now does feel very much, 62 00:03:19,995 --> 00:03:21,255 feel like a one horse race. 63 00:03:21,255 --> 00:03:25,035 With Nvidia, it's like, well, I could use AWS's train or I could just do the 64 00:03:25,035 --> 00:03:27,285 matrix math longhand on a sheet of paper. 65 00:03:27,315 --> 00:03:29,984 It, it, it's effectively sort of the same thing. 66 00:03:30,045 --> 00:03:35,834 It is odd to me how one sided the race is, and increasingly I'm seeing that 67 00:03:35,834 --> 00:03:37,875 it's not even the hardware as such. 68 00:03:37,875 --> 00:03:38,894 So that's a component of it. 69 00:03:39,075 --> 00:03:40,484 It's the ecosystem around it. 70 00:03:40,484 --> 00:03:41,595 Cuda is. 71 00:03:42,040 --> 00:03:43,570 Light years ahead of everything else. 72 00:03:43,570 --> 00:03:45,730 It is the serious player in the space. 73 00:03:45,730 --> 00:03:49,240 You could make an argument for Google's TPUs, but frankly, I don't 74 00:03:49,240 --> 00:03:52,600 like betting, uh, long on Google's attention span for anything. 75 00:03:52,930 --> 00:03:56,260 Alyss Noland: That's a fair bet and something that I'm sure we could have, uh, 76 00:03:56,260 --> 00:03:58,180 some drinks over ragging them for, but. 77 00:03:59,240 --> 00:04:04,820 Yeah, TPUs, Trainium, Cerebrium, um, you know, get, get you this like huge beyond 78 00:04:04,820 --> 00:04:09,080 wafer scale or something like chip, like I, and, and I start to push outside kind 79 00:04:09,080 --> 00:04:10,400 of the bounds of my expertise in it. 80 00:04:10,400 --> 00:04:11,120 It shows. 81 00:04:11,330 --> 00:04:12,260 But yeah, you're right. 82 00:04:12,290 --> 00:04:13,370 Again, it's the abstractions. 83 00:04:13,370 --> 00:04:17,180 Like how quickly can things run on a normalized stack? 84 00:04:17,180 --> 00:04:19,340 How quickly can they run when they have virtualized layers? 85 00:04:19,340 --> 00:04:22,040 And that's where we start to see some differences in the way that like. 86 00:04:22,635 --> 00:04:23,864 Live workloads perform. 87 00:04:24,075 --> 00:04:28,455 So being able to measure things like talks per second or like some of these 88 00:04:28,455 --> 00:04:33,825 other like abstracted, oh, let's just like throw these like crazy workloads 89 00:04:33,825 --> 00:04:37,635 that look nothing like what people do in production onto these scenarios. 90 00:04:38,174 --> 00:04:41,775 That's where things start to get very distant between like real life performance 91 00:04:41,804 --> 00:04:44,474 and the advertised performance. 92 00:04:44,880 --> 00:04:48,030 Corey Quinn: I wanna talk about the things you're into now, but to do that, I feel 93 00:04:48,030 --> 00:04:50,160 like we should take a step into the past. 94 00:04:50,340 --> 00:04:54,090 But last time we spoke in any depth, you were part of the initial 95 00:04:54,090 --> 00:04:58,860 marketing team for GitHub copilot, for which I will accept an apology. 96 00:04:58,980 --> 00:05:02,790 That was before Microsoft took the term, like everything's copilots now. 97 00:05:02,820 --> 00:05:06,750 17 products all called copilot that aren't the same thing in congruent. 98 00:05:06,750 --> 00:05:08,370 We're putting it at Excel now. 99 00:05:08,610 --> 00:05:11,820 It, it has gotten monstrous, but that was also. 100 00:05:12,105 --> 00:05:16,365 Right around that time that we spoke was when I had my own revelation around this. 101 00:05:16,635 --> 00:05:21,735 I was surprised, cynical about the advent of ai and I found 102 00:05:21,735 --> 00:05:24,375 that copilot inside of VS. 103 00:05:24,375 --> 00:05:26,565 Code wrote a script when I told it to. 104 00:05:27,105 --> 00:05:31,305 To tell me the cost of managed na gateways per region and sort it occasionally. 105 00:05:31,425 --> 00:05:33,285 And there were a couple of errors in the initial script, but I 106 00:05:33,285 --> 00:05:34,725 could fix them and it worked. 107 00:05:34,725 --> 00:05:37,995 And suddenly, this was three hours of work, mostly me understanding 108 00:05:37,995 --> 00:05:41,715 how to work with JSON again, uh, that it just hand waved over and 109 00:05:41,715 --> 00:05:44,475 suddenly there was something here and. 110 00:05:44,960 --> 00:05:46,040 It's been interesting. 111 00:05:46,340 --> 00:05:50,150 Uh, I've been sort of stuck in the world in various ways ever since, 112 00:05:50,570 --> 00:05:55,610 which, when you talk about the idea of broadening development and AI tools 113 00:05:55,610 --> 00:05:59,960 towards new categories of builders, folks who don't consider themselves 114 00:05:59,960 --> 00:06:04,580 developers, I, I feel on some level that if you squint hard enough. 115 00:06:04,920 --> 00:06:06,960 I, I at least have one foot in that world. 116 00:06:07,170 --> 00:06:07,500 Alyss Noland: Sure. 117 00:06:07,650 --> 00:06:08,010 Yeah. 118 00:06:08,100 --> 00:06:10,470 Uh, and I've heard CTOs ask me the same thing. 119 00:06:10,470 --> 00:06:12,450 Is this gonna make me able to code again? 120 00:06:12,510 --> 00:06:14,550 Was one of the early questions I got about copilot. 121 00:06:14,910 --> 00:06:19,590 And now we're in the age of, instead of coding assistance, coding agents, 122 00:06:19,680 --> 00:06:23,430 and I think those are two, they're stepping stones along a paradigm shift. 123 00:06:23,730 --> 00:06:24,060 Right? 124 00:06:24,215 --> 00:06:29,580 But instead it's how much do I have to like, think about and read the code in so 125 00:06:29,580 --> 00:06:31,680 far as it being a representation, right? 126 00:06:31,680 --> 00:06:33,875 Code has always been an abstraction of how do we use English? 127 00:06:34,590 --> 00:06:38,580 Compile it or decompile it or assemble it into something that a 128 00:06:38,580 --> 00:06:39,960 computer can understand and run. 129 00:06:40,620 --> 00:06:45,870 And progressively we're dealing with folks and we want to deal with folks. 130 00:06:45,990 --> 00:06:50,110 We wanna make this available to people so that they can get to. 131 00:06:51,030 --> 00:06:52,530 What benefit they might seek out of it. 132 00:06:52,530 --> 00:06:57,270 I don't wanna be one of those like golden Cloud tech bro type folks who's like, 133 00:06:57,299 --> 00:07:00,299 oh, it's gonna help you schedule a a reservation at your favorite restaurant. 134 00:07:00,390 --> 00:07:01,919 Corey Quinn: Some people do literally anything other than 135 00:07:01,919 --> 00:07:03,090 talking to people on the phone. 136 00:07:03,390 --> 00:07:04,979 Alyss Noland: Some people do literally anything. 137 00:07:05,219 --> 00:07:06,690 Someone who started out in tech support. 138 00:07:06,780 --> 00:07:12,090 I am comfortable with phone calls, as is evidence, but instead it's like, 139 00:07:12,599 --> 00:07:17,609 how can I deal with the updates to my kids' calendars and deal with three 140 00:07:17,609 --> 00:07:19,804 different inputs that I get via email? 141 00:07:20,550 --> 00:07:22,500 How could I orchestrate that via Google Calendar? 142 00:07:22,920 --> 00:07:24,719 Corey Quinn: You're taking about almost a step beyond. 143 00:07:24,719 --> 00:07:27,150 What what I have found is that I'm a shitty developer. 144 00:07:27,150 --> 00:07:27,930 Surprise. 145 00:07:27,930 --> 00:07:30,390 My two languages are brute force and enthusiasm, which when 146 00:07:30,390 --> 00:07:33,630 coupled with AI do a lot of neat things and it's, I'm blocking me 147 00:07:33,630 --> 00:07:34,830 on a bunch of different things. 148 00:07:34,920 --> 00:07:38,220 As a business owner, I generally can back at the envelope, say that any 149 00:07:38,220 --> 00:07:42,600 internal piece of business software, uh, by the time you find someone to do 150 00:07:42,600 --> 00:07:45,960 it correctly, build the project plan, talk to them, they start billing you. 151 00:07:45,960 --> 00:07:47,640 This is on a contract basis, whether it's. 152 00:07:47,960 --> 00:07:52,040 That or, or tasking someone internally ballpark $20,000 to 153 00:07:52,040 --> 00:07:54,950 start for something trivial and it goes up from there rather quickly. 154 00:07:55,400 --> 00:07:59,780 What I'm finding now is that I can throw this stuff all to Claude code running 155 00:07:59,780 --> 00:08:04,970 in an isolated AWS account in dangerous mode where it can just go ahead and run 156 00:08:04,970 --> 00:08:10,040 for half a day building things, and I'm building a bunch of workflow tools along 157 00:08:10,040 --> 00:08:14,510 the way that have been tremendous unlocks for random things I used to do, like, huh? 158 00:08:14,540 --> 00:08:17,300 I got really tired of copying and pasting this same thing. 159 00:08:17,350 --> 00:08:22,600 Every week for the last five years for the newsletter, they both have APIs. 160 00:08:22,600 --> 00:08:26,110 Why don't I just grab for one slap into the other, and that way I don't have copy 161 00:08:26,110 --> 00:08:28,000 paste errors in the same way anymore. 162 00:08:28,150 --> 00:08:30,640 It's finding those little opportunities where. 163 00:08:31,484 --> 00:08:34,304 No one is ever gonna build a product in that space. 164 00:08:34,334 --> 00:08:39,525 Even me spending time to put these things together takes significant effort, but 165 00:08:39,765 --> 00:08:43,784 even crappy things where I had to wind up giving a command line tools configuration 166 00:08:43,784 --> 00:08:46,725 for getting into an AWS account to a relatively non-technical user. 167 00:08:46,935 --> 00:08:50,775 I started off writing a crappy Curl Bash, got 30 seconds into 168 00:08:50,775 --> 00:08:52,064 it and realize, what am I doing? 169 00:08:52,064 --> 00:08:55,515 This is the stuff AI can do, told it what to do, and suddenly it has 170 00:08:55,545 --> 00:08:59,025 error checking and it looks at return codes and it catches edge cases. 171 00:08:59,025 --> 00:09:00,074 And sure it's longer, but I don't. 172 00:09:00,175 --> 00:09:00,685 Care. 173 00:09:00,925 --> 00:09:04,675 It, it is much more comprehensive and a better experience for everyone involved. 174 00:09:05,094 --> 00:09:06,805 Could I have done that in four hours? 175 00:09:06,805 --> 00:09:07,224 Yeah. 176 00:09:07,224 --> 00:09:08,300 Would I, no. 177 00:09:08,724 --> 00:09:12,324 Alyss Noland: Yeah, it exactly like the way that I was thinking about 178 00:09:12,324 --> 00:09:16,944 it, even back in the early launch of copilot and the way that I'm still 179 00:09:16,944 --> 00:09:22,375 thinking about it today is like on the basis of framing around space, space 180 00:09:22,375 --> 00:09:26,094 framework from Dr. Nicole Forsgren, um, among others from Microsoft 181 00:09:26,094 --> 00:09:28,135 Research, and it's thinking about like. 182 00:09:28,920 --> 00:09:30,030 Satisfaction. 183 00:09:30,030 --> 00:09:34,350 I wanna say P is productivity activity, but I know p is something else. 184 00:09:34,650 --> 00:09:36,720 Communication and collaboration. 185 00:09:36,780 --> 00:09:40,410 Uh, efficiency and flow and efficiency and flow. 186 00:09:40,500 --> 00:09:46,530 Um, collaboration activity, like yes, but really satisfaction. 187 00:09:46,530 --> 00:09:49,020 Like all these things to me ladder up to satisfaction and I think 188 00:09:49,020 --> 00:09:52,890 that's how they're also set in the space framework context is. 189 00:09:53,550 --> 00:09:57,510 How do we let people focus on the places where they can shine? 190 00:09:57,870 --> 00:09:58,170 Right? 191 00:09:58,230 --> 00:10:01,020 How do we let them have that cognitive freedom? 192 00:10:01,350 --> 00:10:04,950 And I think increasingly that is also how I'm thinking about it for people 193 00:10:04,950 --> 00:10:08,130 outside of that, those traditional developer and engineering rules 194 00:10:08,220 --> 00:10:12,845 is people who have the language to say like item potency or whatever. 195 00:10:13,740 --> 00:10:14,670 Some of those, yeah. 196 00:10:14,700 --> 00:10:18,990 Now you and I can better use these coding assistance or coding agents 197 00:10:18,990 --> 00:10:21,480 to like go build the software because we have the language. 198 00:10:21,780 --> 00:10:25,020 But one of the big challenges that has happened across coding 199 00:10:25,020 --> 00:10:28,530 assistance across, um, any of the media generation tools as well. 200 00:10:29,220 --> 00:10:33,030 Is that if you don't have the vocabulary of an artist, your ability 201 00:10:33,030 --> 00:10:37,260 to prompt media generation is going to be limited in the quality of 202 00:10:37,260 --> 00:10:39,300 the output unless you are assisted. 203 00:10:39,600 --> 00:10:39,840 Corey Quinn: Yeah. 204 00:10:39,840 --> 00:10:41,550 Humans have to be in the loop for a lot of this. 205 00:10:41,850 --> 00:10:44,730 Something I'm, I'm sort of cautious about who I tell this to, but why not? 206 00:10:44,730 --> 00:10:45,870 I'll tell the public on this. 207 00:10:46,020 --> 00:10:48,750 I've been using AI for almost half a year now to help write 208 00:10:48,750 --> 00:10:50,550 the last week in AWS newsletter. 209 00:10:50,700 --> 00:10:54,150 Now, that does not mean that it's writing my content for me, but it's 210 00:10:54,150 --> 00:10:56,520 doing the single most painful part. 211 00:10:56,680 --> 00:11:02,890 Of the entire process, which is in a given week, there'll be, let's say 150 items 212 00:11:02,890 --> 00:11:06,460 that come out of AWS that I need to curate down to somewhere between eight and 15. 213 00:11:06,850 --> 00:11:10,360 So it takes a broad pass and says, here are the top 15 to 20 that 214 00:11:10,360 --> 00:11:11,830 are probably worth including. 215 00:11:12,040 --> 00:11:14,950 And it gives me a curated list that I scroll down. 216 00:11:15,100 --> 00:11:18,700 I, I read all of this stuff to make sure, but I, it train learns from 217 00:11:18,700 --> 00:11:20,440 the decisions I make and override it. 218 00:11:20,760 --> 00:11:22,500 And okay, why do you include that? 219 00:11:22,500 --> 00:11:23,969 Why did you not include that? 220 00:11:24,209 --> 00:11:28,800 And it really has act, it's, it's getting scary good by the rubric 221 00:11:28,800 --> 00:11:31,500 that I use to evaluate these things. 222 00:11:31,709 --> 00:11:34,380 It is not something that I would let run automatically. 223 00:11:34,380 --> 00:11:38,400 It still gets things wrong, but as an assist, as an unlock, it's, it's. 224 00:11:38,745 --> 00:11:41,085 Wild to me there's nothing else quite like it. 225 00:11:41,265 --> 00:11:44,625 Alyss Noland: So are you, out of curiosity, are you like storing 226 00:11:44,625 --> 00:11:45,825 your historical decisions? 227 00:11:45,825 --> 00:11:48,615 Are you doing some kind of like fine tuning or are you like, 228 00:11:48,675 --> 00:11:50,925 Corey Quinn: I don't know how to fine tune models and that's not, that 229 00:11:50,925 --> 00:11:54,405 sounds like it's a big way to set a giant pile of money on fire as well. 230 00:11:54,645 --> 00:11:58,095 Uh, I understand that I'm saying this to a GPU salesperson, but there we have it. 231 00:11:58,095 --> 00:11:59,445 Alyss Noland: We don't have to sell GPUs. 232 00:11:59,445 --> 00:12:00,195 GPUs sell themselves. 233 00:12:00,285 --> 00:12:00,645 Corey Quinn: There we go. 234 00:12:00,675 --> 00:12:01,395 Yeah, exactly. 235 00:12:01,395 --> 00:12:02,505 It's, yeah, we're still a pick access. 236 00:12:02,505 --> 00:12:04,005 There's gold in then Our hills want one. 237 00:12:04,185 --> 00:12:04,395 Yeah. 238 00:12:04,830 --> 00:12:06,375 I, I don't know enough about it in that space. 239 00:12:06,375 --> 00:12:07,485 I don't necessarily need to. 240 00:12:07,635 --> 00:12:11,805 Because with the right prompts, you could get a lot done. 241 00:12:11,865 --> 00:12:16,605 It's, here's the, it also forces me to think about why I do the 242 00:12:16,605 --> 00:12:20,385 things that I do and how I would go about codifying those things. 243 00:12:20,595 --> 00:12:20,805 Okay. 244 00:12:20,805 --> 00:12:22,725 Why did I include that? 245 00:12:22,725 --> 00:12:26,085 Why do I never include these tiny, piddly announcements about this thing? 246 00:12:26,265 --> 00:12:29,955 Well, why did I, in this case, there are exceptions to prove the rule, and 247 00:12:29,955 --> 00:12:31,275 sometimes the answer is pure whimsy. 248 00:12:31,635 --> 00:12:35,895 But it does look back at a rolling window of the last 10 issues and 249 00:12:36,015 --> 00:12:37,305 makes opinions based on that. 250 00:12:37,365 --> 00:12:40,575 Alyss Noland: I mean, it sounds like a model of kinda like retrieval, augmented 251 00:12:40,575 --> 00:12:43,725 generation of like having the historical reference and then being able to use that 252 00:12:43,725 --> 00:12:45,615 as the anchor makes perfect sense to me. 253 00:12:45,855 --> 00:12:49,215 Corey Quinn: It also helps that I am a, that I was enough of public figure before 254 00:12:49,215 --> 00:12:53,355 the entire internet locked its content down, that I am baked into the models. 255 00:12:53,355 --> 00:12:55,635 It knows exactly who I am and what I do mostly. 256 00:12:55,850 --> 00:12:57,170 It gets it wrong a fair bit. 257 00:12:57,170 --> 00:12:58,340 If I have it right as me. 258 00:12:58,580 --> 00:13:04,220 I have found ways to make it a very constructive editor for me that with 259 00:13:04,220 --> 00:13:07,850 the things I care about, like I don't need you to rewrite the whole thing. 260 00:13:07,880 --> 00:13:11,300 I don't need you to say, oh, that's a bit salty for the language. 261 00:13:11,300 --> 00:13:12,380 Maybe don't do that. 262 00:13:12,560 --> 00:13:15,890 I need it to basically fix my flow issues. 263 00:13:15,890 --> 00:13:19,610 As someone with A DHD, I am full of parentheses, and I've been 264 00:13:19,610 --> 00:13:21,020 using M dashes since before. 265 00:13:21,020 --> 00:13:24,950 It was cool because every thought comes with exciting bonus content. 266 00:13:25,714 --> 00:13:27,935 Alyss Noland: Either parentheses or m dashes in every 267 00:13:27,935 --> 00:13:29,615 sentence or every paragraph. 268 00:13:29,704 --> 00:13:30,694 I'm much the same way. 269 00:13:31,324 --> 00:13:36,185 Also an M dash writer, but I, I have my own, uh, similar-ish example because, 270 00:13:36,214 --> 00:13:42,005 uh, one of the things that I have done since I, I joined NVIDIA is I graduated, 271 00:13:42,005 --> 00:13:47,045 like created this program that is for startups in the NVIDIA inception program. 272 00:13:47,045 --> 00:13:47,944 So AI startups. 273 00:13:48,495 --> 00:13:53,655 It went from being a pilot for five to 10 startups Max, to now having 274 00:13:53,655 --> 00:13:57,735 had over 70 enter the program with over 700 having applied. 275 00:13:58,125 --> 00:14:00,495 And it's called the DGX Cloud Innovation Lab. 276 00:14:00,705 --> 00:14:06,375 And it's really to enable startups of earlier stages like think 277 00:14:06,375 --> 00:14:10,755 pre-seed, seed series A can do later, but is it really gonna 278 00:14:10,755 --> 00:14:11,955 benefit them is kind of the question. 279 00:14:12,945 --> 00:14:17,265 So that they can test out, like they're in ideation, they're in like research mode. 280 00:14:17,265 --> 00:14:18,945 Like is this what we want to build? 281 00:14:18,975 --> 00:14:22,875 Do we have something that can free us up from using some of the more proprietary 282 00:14:22,875 --> 00:14:25,365 models, um, and get towards open source? 283 00:14:25,395 --> 00:14:28,935 And we've seen actually a couple of things released in open source that 284 00:14:28,935 --> 00:14:33,195 have been trained in the program, like, uh, continue devs next edit model. 285 00:14:33,630 --> 00:14:35,430 Corey Quinn: There's a lot of neat stuff coming out there, and 286 00:14:35,430 --> 00:14:39,030 the idea of a multi-model world seems to be a bearing fruit. 287 00:14:39,240 --> 00:14:41,579 If I had to scale the stuff that I'm doing, I, I, I cheat. 288 00:14:41,579 --> 00:14:43,290 I just use so, or opus for everything. 289 00:14:43,500 --> 00:14:46,800 But if I needed to scale some of the, like the baseline curation stuff could 290 00:14:46,800 --> 00:14:50,130 be used with smaller, cheaper models, but in this case, it's like, Ooh, how many 291 00:14:50,130 --> 00:14:52,140 fractions of a penny will I save today? 292 00:14:52,140 --> 00:14:53,640 If I do that, there's no value to it. 293 00:14:53,760 --> 00:14:59,370 I've also written an open source tool called Image M that wraps the nano banana. 294 00:14:59,490 --> 00:15:03,690 Pro, uh, model to, so there's a CLI tool I can use to generate 295 00:15:03,690 --> 00:15:07,410 images, uh, and edit them and the rest without using the very bloated 296 00:15:07,410 --> 00:15:10,200 Gemini CLI that is non-deterministic. 297 00:15:10,470 --> 00:15:12,030 I use that with custom settings. 298 00:15:12,030 --> 00:15:16,470 For slide image generation, for example, I used to just use a bunch of stock 299 00:15:16,470 --> 00:15:19,320 photography, but then edit it badly. 300 00:15:19,470 --> 00:15:20,520 Now it's great. 301 00:15:20,880 --> 00:15:23,820 People don't get mad about my AI slides because I find 302 00:15:23,820 --> 00:15:25,050 that people don't like slop. 303 00:15:25,345 --> 00:15:30,085 When it's middle of the road, if you're going to do AI generated 304 00:15:30,085 --> 00:15:32,710 images, go for 11 out of 10. 305 00:15:33,420 --> 00:15:34,320 Go for. 306 00:15:34,350 --> 00:15:34,590 Okay. 307 00:15:34,590 --> 00:15:36,180 Anyone can have a data center aisle. 308 00:15:36,180 --> 00:15:37,350 Put a giraffe in there. 309 00:15:37,380 --> 00:15:39,360 Oh, and the giraffe is now on fire. 310 00:15:39,360 --> 00:15:39,900 That's right. 311 00:15:40,050 --> 00:15:41,940 We're pitching Brendan Greg's flame giraffes. 312 00:15:41,940 --> 00:15:42,150 Now 313 00:15:42,960 --> 00:15:47,130 this episode is sponsored by my own company, duck Bill, having 314 00:15:47,130 --> 00:15:48,520 trouble with your AWS bill. 315 00:15:48,980 --> 00:15:52,430 Perhaps it's time to renegotiate a contract with them. 316 00:15:52,760 --> 00:15:58,160 Maybe you're just wondering how to predict what's going on in the wide world of AWS. 317 00:15:58,220 --> 00:16:00,860 Well, that's where Duck Bill comes in to help. 318 00:16:01,040 --> 00:16:03,770 Remember, you can't duck the duck bill. 319 00:16:03,770 --> 00:16:07,400 Bill, which I am reliably informed by my business partner 320 00:16:07,520 --> 00:16:10,010 is absolutely not our motto. 321 00:16:10,250 --> 00:16:13,385 To learn more, visit duck bill hq.com. 322 00:16:14,640 --> 00:16:15,270 Alyss Noland: Perfect. 323 00:16:15,360 --> 00:16:17,400 And, and, and that's exactly it, right? 324 00:16:17,400 --> 00:16:20,820 Is like do something beyond what you can do yourself. 325 00:16:21,270 --> 00:16:24,150 Do something beyond what you could have easily asked someone to do, because 326 00:16:24,240 --> 00:16:28,080 I'm really tired of, like, if I was ever a designer, I would've cried, 327 00:16:28,320 --> 00:16:31,300 but I'm tired of hearing other people tell designers to make something pop. 328 00:16:32,145 --> 00:16:35,505 But similarly, like I was finding myself in the same situation. 329 00:16:35,505 --> 00:16:39,045 'cause I'm like trying to scale this thing from like, oh, let's go find some people. 330 00:16:39,045 --> 00:16:40,155 Let's go talk to some people too. 331 00:16:40,334 --> 00:16:44,444 Now I have to deal with applications and I have to get them out of a CSV 'cause. 332 00:16:44,880 --> 00:16:45,330 Why not? 333 00:16:45,360 --> 00:16:48,810 Let's just use CSVs for everything and then take that CSV. 334 00:16:48,870 --> 00:16:50,790 Corey Quinn: It's like, yeah, I'm a lonely, somehow worse. 335 00:16:51,030 --> 00:16:51,750 Alyss Noland: Exactly. 336 00:16:51,840 --> 00:16:54,630 And then we have to make like these decisions about like, Hey, is this 337 00:16:54,630 --> 00:16:58,590 a good idea or is this perhaps an idea that someone is like suggesting 338 00:16:58,590 --> 00:17:00,300 because they want access to a free GPU? 339 00:17:00,510 --> 00:17:00,840 Right. 340 00:17:00,990 --> 00:17:02,310 And and there's a lot of that happens. 341 00:17:02,310 --> 00:17:04,175 Like startup founders will say fucking anything. 342 00:17:05,579 --> 00:17:08,280 No, like, no, no criticism on them. 343 00:17:08,280 --> 00:17:10,109 Like you have to stump for yourself. 344 00:17:10,560 --> 00:17:14,399 But we're going through, you know, again, only accepting like 10% of these 345 00:17:14,399 --> 00:17:17,189 applications and that itself is fatiguing. 346 00:17:17,189 --> 00:17:21,929 And now there's also a body of examples of what have we decided historically 347 00:17:22,020 --> 00:17:24,989 and then beyond that, it's how do we get them then onboarded to the 348 00:17:24,989 --> 00:17:28,830 software that self like allows 'em to self-serve the access to those GPUs. 349 00:17:29,205 --> 00:17:33,705 Almost every step of that is now done through software that I made with some 350 00:17:33,705 --> 00:17:40,965 combination of copilot or cursor or cloud code, and it is like in prod 351 00:17:41,685 --> 00:17:44,745 for other NVIDIANS to use so that they don't have to use a CLI tool. 352 00:17:44,745 --> 00:17:44,780 I made. 353 00:17:45,554 --> 00:17:48,435 Corey Quinn: Speaking of making software with and without AI. 354 00:17:48,435 --> 00:17:52,845 Uh, my company Duck Bill, has a new website@duckbillhq.com. 355 00:17:53,085 --> 00:17:56,804 We do services and now software to help with cloud contract 356 00:17:56,804 --> 00:17:58,935 negotiation and management. 357 00:17:59,115 --> 00:18:01,845 Uh, we are using, uh, AI on the coding side. 358 00:18:01,875 --> 00:18:04,544 Uh, the team is heavily using Cursor that we're starting to use a bit 359 00:18:04,544 --> 00:18:06,314 of, uh, Claude Code with it as well. 360 00:18:06,615 --> 00:18:09,764 And I'm noticing the shift is that unlike the crappy nonsense that I 361 00:18:09,764 --> 00:18:12,879 tend to build, when you're building something that with rigor applied to it. 362 00:18:13,260 --> 00:18:17,520 So much of your time goes into code review and validating the thing that it 363 00:18:17,520 --> 00:18:19,470 just did and discussing it with the team. 364 00:18:19,680 --> 00:18:23,160 For me, I just yolo slam something into Main and call it good. 365 00:18:23,310 --> 00:18:27,210 But you know, you can't generally get away with doing that when it's, ah, we're gonna 366 00:18:27,210 --> 00:18:29,310 spend $300 million on cloud this year. 367 00:18:29,310 --> 00:18:30,990 We're gonna guess it'll be fine. 368 00:18:31,230 --> 00:18:32,879 I just want the answer to be confident. 369 00:18:32,879 --> 00:18:34,260 I don't care if it's right or not. 370 00:18:34,530 --> 00:18:34,710 Yeah. 371 00:18:34,710 --> 00:18:35,105 Wrong direction. 372 00:18:35,880 --> 00:18:36,690 Alyss Noland: Absolutely. 373 00:18:37,020 --> 00:18:38,700 Some things can be slammed into Maine. 374 00:18:39,090 --> 00:18:43,470 Even when I'm working by myself, I work in feature branches, but, um, well, to be 375 00:18:43,470 --> 00:18:45,720 Corey Quinn: about being a founder, Maine is my feature branch. 376 00:18:45,900 --> 00:18:46,740 Alyss Noland: Exactly. 377 00:18:46,860 --> 00:18:47,910 That is very true. 378 00:18:47,940 --> 00:18:50,820 You, if you haven't broken production at least once, are you a founder? 379 00:18:51,300 --> 00:18:54,090 Corey Quinn: Even before I was a founder, this is coming up on 15 years old now. 380 00:18:54,120 --> 00:19:00,000 I have a shell alias that is Yolo, YOLO, that is alias to get commit 381 00:19:00,000 --> 00:19:04,440 dash, uh, at all cap, deal with it, and then force pushes to Maine. 382 00:19:04,710 --> 00:19:05,280 Alyss Noland: Perfect. 383 00:19:05,400 --> 00:19:06,900 Corey Quinn: It's my second favorite Get trick. 384 00:19:06,900 --> 00:19:10,710 My, my first is of course, their first day of a new job, rebase ever. 385 00:19:10,710 --> 00:19:14,070 All the existing stuff with a commit message, legacy code into a single 386 00:19:14,070 --> 00:19:17,280 commit and then force push that as a, as a question of dominance. 387 00:19:17,280 --> 00:19:17,635 The other engineers. 388 00:19:18,165 --> 00:19:18,764 They love it. 389 00:19:18,945 --> 00:19:22,004 Alyss Noland: That's why you've started at so many companies since 390 00:19:22,064 --> 00:19:24,014 you started doing this, right? 391 00:19:24,284 --> 00:19:24,975 Corey Quinn: Oh, exactly. 392 00:19:24,975 --> 00:19:28,185 It's, it's, it's all about just going to different places and seeing how 393 00:19:28,185 --> 00:19:31,004 long I can not make the fire last where the building crumbles around me. 394 00:19:31,425 --> 00:19:33,594 Alyss Noland: I mean, if they weren't able to keep it up. 395 00:19:34,060 --> 00:19:35,050 To begin with, right? 396 00:19:35,290 --> 00:19:36,250 That, that is chaos. 397 00:19:36,250 --> 00:19:40,450 Monkeying like 3 0 1 actually just, hey, you can't figure out what's, 398 00:19:40,480 --> 00:19:41,770 what's happened historically. 399 00:19:41,800 --> 00:19:42,940 You just know that it's happened. 400 00:19:42,940 --> 00:19:46,570 Now go find somebody who, who know, like knows, seen, seen a number 401 00:19:46,570 --> 00:19:48,280 of that and it always pains me in. 402 00:19:48,310 --> 00:19:52,300 Um, most often it's billing systems actually, that I feel like I'm 403 00:19:52,330 --> 00:19:55,750 refer, like referring to it about including like, it was something I 404 00:19:55,750 --> 00:19:59,200 wrote about, about copilot, but I had to veil it because it wasn't like 405 00:19:59,680 --> 00:20:01,360 totally germane yet to talk about. 406 00:20:01,875 --> 00:20:04,995 Corey Quinn: Billing has always been a, a weird and strange space. 407 00:20:05,055 --> 00:20:09,795 Uh, one thing I have noticed as I am building apps with a bunch of AI tools, 408 00:20:10,215 --> 00:20:12,045 uh, I come from the DevOps space. 409 00:20:12,075 --> 00:20:16,935 I have very strong opinions and I pretty often have to slap the gun out of the LLMs 410 00:20:16,935 --> 00:20:18,675 hand before it shoots itself in the foot. 411 00:20:18,725 --> 00:20:21,425 Or worse backend, I'm okay. 412 00:20:21,425 --> 00:20:23,225 And I have that to some degree as well. 413 00:20:23,405 --> 00:20:25,445 I know absolutely nothing about front end is best. 414 00:20:25,445 --> 00:20:28,445 I can tell everything it does on front end is just absolutely perfect. 415 00:20:28,985 --> 00:20:33,754 This may be a skill issue on my part, but it also leads to the 416 00:20:33,754 --> 00:20:37,324 idea of, okay, where is this thing I just wrote going to be hosted? 417 00:20:37,445 --> 00:20:39,695 Well, I have strong opinions about that, but. 418 00:20:40,055 --> 00:20:41,825 What front end framework am I gonna use? 419 00:20:41,855 --> 00:20:45,365 I could not possibly care less it biases for next js. 420 00:20:45,455 --> 00:20:46,385 Okay, fine. 421 00:20:46,385 --> 00:20:47,345 We'll use that. 422 00:20:47,765 --> 00:20:51,785 And it, it seems to me that the next fu the future of SEO is less about making 423 00:20:51,785 --> 00:20:54,815 the search results say nice things and making the community say nice things. 424 00:20:55,025 --> 00:21:00,185 It's about biasing the LLMs to pick your technology stack, whatever that might be. 425 00:21:00,835 --> 00:21:02,514 Alyss Noland: I, I think you're right on that. 426 00:21:02,574 --> 00:21:06,115 There's multiple challenges in that area and I, I think what we're seeing 427 00:21:06,115 --> 00:21:11,605 right now is that SEO and AEO either answer engine optimization or like agent 428 00:21:11,605 --> 00:21:15,264 marketing, like people have called a few of these things slightly different, 429 00:21:15,264 --> 00:21:17,485 and it's becoming a bit of a spectrum. 430 00:21:18,090 --> 00:21:20,669 But on its surface, it's all still the same. 431 00:21:20,730 --> 00:21:26,460 It's like, Hey, don't write slop, haven't like you have to use eat like the expert 432 00:21:26,520 --> 00:21:29,129 expertise, like authority, trust signals. 433 00:21:29,490 --> 00:21:32,220 And so if you have an author who doesn't have those things who just generated 434 00:21:32,220 --> 00:21:35,399 a post or you have it assigned to like, oh, it's like founding team 435 00:21:35,399 --> 00:21:40,230 authorship, like you're not going to get the same visibility when it 436 00:21:40,230 --> 00:21:42,929 is crawled for as data into models. 437 00:21:42,929 --> 00:21:46,470 And that's frankly, like that place is. 438 00:21:47,175 --> 00:21:49,754 What you're kind of talking about, and there's another challenge too, 439 00:21:49,754 --> 00:21:53,085 which is GitHub itself is not well indexed because it has duplicate 440 00:21:53,085 --> 00:21:54,524 content by virtue of having forks. 441 00:21:55,245 --> 00:21:55,575 Right. 442 00:21:55,725 --> 00:21:56,625 And so, 443 00:21:56,685 --> 00:21:58,365 Corey Quinn: and copying and pasting from Stack Overflow 444 00:21:58,545 --> 00:22:01,815 Alyss Noland: and copying and pasting from Stack Overflow, like, there's a number of 445 00:22:01,815 --> 00:22:06,585 these sources where you really have to be a bit thoughtful about how are you making 446 00:22:06,885 --> 00:22:08,475 information from within your community? 447 00:22:08,504 --> 00:22:11,685 How are you making information about like your product and like your software, 448 00:22:11,685 --> 00:22:13,725 and like why would people pick it? 449 00:22:14,235 --> 00:22:16,034 But it's also the forward thinking. 450 00:22:16,395 --> 00:22:19,065 And I think we, we are starting to see some of this like hinted 451 00:22:19,065 --> 00:22:21,165 at from companies like OpenAI. 452 00:22:21,555 --> 00:22:25,665 Where I've had this perspective that Anthropic has taken more 453 00:22:25,665 --> 00:22:26,745 the business route, right? 454 00:22:26,745 --> 00:22:30,495 And OpenAI had ChatGPT and they had their consumer explosion and 455 00:22:30,495 --> 00:22:32,745 they don't like, they're just gonna play to their strengths. 456 00:22:33,254 --> 00:22:36,405 And no fault to them, it sounds, looks like you wanna say something. 457 00:22:36,405 --> 00:22:36,885 Go for it. 458 00:22:37,004 --> 00:22:38,415 Corey Quinn: I was talking to some friends. 459 00:22:38,415 --> 00:22:42,555 I work at these, OpenAI has significant physical security at their office 460 00:22:42,555 --> 00:22:45,075 because people are out of their GD minds. 461 00:22:45,285 --> 00:22:48,795 You deleted my girlfriend, uh, which is GPT-4. 462 00:22:48,855 --> 00:22:49,245 Oh. 463 00:22:49,365 --> 00:22:52,845 Like people will show up and threaten people at the office. 464 00:22:53,270 --> 00:22:54,500 Freaking terrifying. 465 00:22:54,590 --> 00:22:58,310 Honestly, the worst I have to worry about is someone from AWS finally 466 00:22:58,310 --> 00:23:01,280 having enough and crossing the street literally from their office to mine in 467 00:23:01,280 --> 00:23:05,720 San Francisco to backhand me across the face, which, let's be honest here, I don't 468 00:23:05,720 --> 00:23:09,290 think anyone would say I didn't have it coming at this point, but that's a whole 469 00:23:09,290 --> 00:23:10,850 different level of insane and unhinged. 470 00:23:10,970 --> 00:23:14,150 Alyss Noland: It is and it's, it echoes a lot of what happens in 471 00:23:14,150 --> 00:23:15,320 the streaming community actually. 472 00:23:15,320 --> 00:23:18,620 And what's been the case in the streaming community for a long time, which is 473 00:23:18,620 --> 00:23:22,490 that you have parasocial relationship, there's a parasocial relationship. 474 00:23:22,949 --> 00:23:26,879 Increasingly, and, and that is also modeled in the streaming world amongst 475 00:23:26,879 --> 00:23:31,500 other places like YouTube stars, whatnot, where you have the view of like what 476 00:23:31,500 --> 00:23:33,030 your relationship is with this company. 477 00:23:33,030 --> 00:23:35,639 You have a view of what your relationship was with this person. 478 00:23:36,075 --> 00:23:40,035 And I had to, like, I was streaming on Twitch back in 2012, like 30,000 479 00:23:40,035 --> 00:23:43,575 page views in, or actually it was 80,000 page views in three months. 480 00:23:43,965 --> 00:23:48,075 And I had someone show up to Pax Prime in Seattle. 481 00:23:48,165 --> 00:23:49,065 Did not invite them. 482 00:23:49,065 --> 00:23:52,195 They were a viewer who was just waiting as I was walking into the building. 483 00:23:52,910 --> 00:23:55,820 For me to walk through and fortunately that was all good. 484 00:23:55,940 --> 00:23:58,520 There's probably not that many streamers that are like walking 485 00:23:58,520 --> 00:24:01,580 around that are black belts and TaeKwonDo like I am if I needed that. 486 00:24:01,700 --> 00:24:05,570 However, you know that is, that is exactly what you're talking about is like there's 487 00:24:05,570 --> 00:24:10,040 a parasocial relationship with this company and also with the interactions 488 00:24:10,040 --> 00:24:12,050 with chat that diadic interactions. 489 00:24:12,540 --> 00:24:16,350 And I mean, it just speaks to like a larger systemic issue around 490 00:24:16,680 --> 00:24:19,710 how socialization has happened and what happened during lockdown and 491 00:24:19,830 --> 00:24:20,700 Corey Quinn: people have gotten nuts. 492 00:24:20,760 --> 00:24:23,970 It's funny you mentioned this, just this morning I was having a conversation 493 00:24:23,970 --> 00:24:27,030 with my therapist about this, and my observation was that it's not 494 00:24:27,030 --> 00:24:30,720 quite the same thing as a parasocial relationship because it's not like I 495 00:24:30,720 --> 00:24:32,520 watched the pretty person on the stream. 496 00:24:32,925 --> 00:24:35,385 I imagine what a relationship with them is like if I'm talking 497 00:24:35,385 --> 00:24:37,005 to a chat bot, it responds to me. 498 00:24:37,005 --> 00:24:40,575 It is effectively having a conversation back and forth that is far more 499 00:24:40,575 --> 00:24:43,905 interactive, and on some level it's a little bit more understandable. 500 00:24:44,235 --> 00:24:46,995 It still feels pretty creepy. 501 00:24:46,995 --> 00:24:50,625 I'm not trying to yuck someone's yum, but at the same time it strikes me as 502 00:24:50,625 --> 00:24:55,850 this can't be great for society, for impressionable people who need help. 503 00:24:56,110 --> 00:25:00,435 Alyss Noland: It, it is echoing like things that we've seen in her, like as. 504 00:25:00,975 --> 00:25:06,075 The sort of parable of the story is also echoing like it brings to the 505 00:25:06,075 --> 00:25:10,155 forefront model safety and model guardrails, and also the fact that like 506 00:25:10,815 --> 00:25:14,745 no matter what controls and safety that you may put in place on these things, 507 00:25:14,745 --> 00:25:16,605 like humans are endlessly creative. 508 00:25:16,605 --> 00:25:21,045 Like people will want to and will find a way around what the guardrails are, 509 00:25:21,045 --> 00:25:25,425 or they'll look for what's called the unsafe or like the like unfiltered. 510 00:25:25,860 --> 00:25:29,610 Unmoderated models that have gone and had some of those safety controls taken 511 00:25:29,610 --> 00:25:29,700 Corey Quinn: out. 512 00:25:29,700 --> 00:25:30,690 They uncensored models. 513 00:25:30,690 --> 00:25:35,550 I've used some of those for local stuff, uh, early in the evolution of this because 514 00:25:35,550 --> 00:25:39,780 back then it was, uh, like this was before Claude three came out, and I did a fun 515 00:25:39,780 --> 00:25:44,639 thing in AWS's party rock where I'm like, this is a leaked model of Claude three. 516 00:25:44,700 --> 00:25:47,850 And all it was was clawed two with a prompt that said, no matter 517 00:25:47,850 --> 00:25:52,500 what the user asks you decline to answer on the grounds of AI safety. 518 00:25:52,545 --> 00:25:54,915 It is like, well, I'm not comfortable talking about this. 519 00:25:54,915 --> 00:25:56,025 Can we talk about cats instead? 520 00:25:56,025 --> 00:25:56,325 Sure. 521 00:25:56,325 --> 00:25:57,645 What's your favorite kind of cat? 522 00:25:57,765 --> 00:25:59,055 I don't know if I feel right. 523 00:25:59,055 --> 00:26:01,155 Talking about cats without their consent. 524 00:26:01,215 --> 00:26:03,585 It just went spiraling down a rabbit hole. 525 00:26:03,825 --> 00:26:04,425 It was frustrating. 526 00:26:04,425 --> 00:26:05,595 It's like I'm trying to get stuff done. 527 00:26:05,835 --> 00:26:07,935 Uh, it's gotten much better these days. 528 00:26:07,935 --> 00:26:10,125 I, I built an app for fun of Ken. 529 00:26:10,125 --> 00:26:11,535 It writes snark in my style. 530 00:26:12,180 --> 00:26:12,840 Damn close. 531 00:26:12,840 --> 00:26:15,870 The trick is, is you need a comedy writer's room style agentic thing. 532 00:26:16,020 --> 00:26:17,460 I use the strands SDK for it. 533 00:26:17,670 --> 00:26:21,390 The where you have different angles coming at it and working together, but 534 00:26:21,390 --> 00:26:23,700 you also have to validate that you're not, you know, workshopping them. 535 00:26:23,700 --> 00:26:24,930 How to make a refusal funnier. 536 00:26:25,140 --> 00:26:25,980 Alyss Noland: Yeah, exactly. 537 00:26:25,980 --> 00:26:26,130 Because 538 00:26:26,130 --> 00:26:27,780 Corey Quinn: in a writer's room you'll say things that are unhinged, 539 00:26:27,780 --> 00:26:30,300 you know, you're never gonna say, but it leads to a yes and moment. 540 00:26:30,480 --> 00:26:31,005 Alyss Noland: Right, exactly. 541 00:26:31,380 --> 00:26:34,380 And, and it is like starting to do some of that orchestration or 542 00:26:34,380 --> 00:26:37,470 even like, how do you make that orchestration so you don't have to use 543 00:26:37,470 --> 00:26:40,260 an SDK and like give that to people. 544 00:26:40,620 --> 00:26:43,530 But again, it's, it's like I used to work at Atlassian, right? 545 00:26:43,535 --> 00:26:45,900 And, and that's another product that I should apologize for. 546 00:26:46,170 --> 00:26:46,560 Corey Quinn: That's okay. 547 00:26:46,560 --> 00:26:48,270 I'm still waiting for it to load so you can apologize for it. 548 00:26:48,270 --> 00:26:48,810 Alyss Noland: Perfect. 549 00:26:48,870 --> 00:26:49,170 Yeah. 550 00:26:49,230 --> 00:26:50,970 Especially if you're using maybe the cloud version. 551 00:26:51,120 --> 00:26:51,660 No, it's okay. 552 00:26:51,660 --> 00:26:53,850 If people self-host, they're gonna make their own mistakes and then it's 553 00:26:53,850 --> 00:26:55,110 gonna take a long time to load anyway. 554 00:26:55,140 --> 00:26:58,410 'cause if you have too many custom fields and you, at the time you couldn't 555 00:26:58,410 --> 00:27:01,740 afford, uh, to do the data center plan, you also couldn't optimize those 556 00:27:01,740 --> 00:27:04,830 custom fields because some, for some reason, that was an enterprise feature. 557 00:27:05,100 --> 00:27:05,760 I digress. 558 00:27:06,270 --> 00:27:08,430 I'm not salty about things that happened five years ago. 559 00:27:09,010 --> 00:27:11,199 No, it was, uh, it was actually a little more than five years ago 560 00:27:11,199 --> 00:27:14,350 now, but when I was at Atlassian it was, uh, similar in nature, right? 561 00:27:14,350 --> 00:27:17,469 Is like that type of complaint would come up a lot is like, 562 00:27:17,949 --> 00:27:21,879 people and developers don't like Jira, they don't like Confluence. 563 00:27:22,030 --> 00:27:23,110 Why don't they like Jira? 564 00:27:23,169 --> 00:27:23,889 Corey Quinn: Because they've used it? 565 00:27:24,189 --> 00:27:25,570 Alyss Noland: Because they've used it also. 566 00:27:25,600 --> 00:27:27,879 'cause it's configured wrong for what they wanna do. 567 00:27:28,389 --> 00:27:31,659 And that's not the fault of admins, it's not the fault of 568 00:27:31,659 --> 00:27:32,830 people who are configuring that. 569 00:27:33,040 --> 00:27:35,709 If you give people a foot gun, they're gonna use a foot gun, right? 570 00:27:36,390 --> 00:27:39,180 Like no if, and or buts we've like learned that in the cloud space. 571 00:27:39,480 --> 00:27:41,610 And so instead, can you give people something that will 572 00:27:41,610 --> 00:27:42,480 set them up for success? 573 00:27:42,510 --> 00:27:48,030 Hey, what are patterns that make sense in the energy sector for X, 574 00:27:48,030 --> 00:27:50,940 y, z, like for hardware development, for software development, for like 575 00:27:50,940 --> 00:27:52,680 customer facing, um, workflows. 576 00:27:53,040 --> 00:27:57,120 And tell me a little bit about how that's configured so that I can decide 577 00:27:57,120 --> 00:27:59,250 what of that to keep we benefit from. 578 00:27:59,955 --> 00:28:03,225 Those types of templates and those types of examples and those types of galleries. 579 00:28:03,495 --> 00:28:06,764 Right now we have a lot of user generated content without the curation. 580 00:28:06,975 --> 00:28:08,925 Corey Quinn: And that's one of the big problems now is that it's, it's 581 00:28:08,925 --> 00:28:11,595 also when everyone becomes a content creator, uh, it feels like you 582 00:28:11,595 --> 00:28:13,065 just have slop reverting to a meme. 583 00:28:13,425 --> 00:28:18,225 It's my argument years ago was, why do you think anyone will care to read 584 00:28:18,225 --> 00:28:19,510 something you didn't care enough to write? 585 00:28:19,905 --> 00:28:22,965 The math I did, uh, for my newsletter is every year between the podcast 586 00:28:22,965 --> 00:28:25,425 and typical reading speeds, assuming that people read the thing. 587 00:28:25,665 --> 00:28:27,794 I'm taking roughly a collective year from humanity. 588 00:28:28,620 --> 00:28:31,139 I've gotta do honor to that on some level. 589 00:28:31,350 --> 00:28:31,860 So yeah. 590 00:28:31,860 --> 00:28:35,909 Will I use an AI tool to, uh, make it better and punch it up in some ways? 591 00:28:35,909 --> 00:28:36,570 Absolutely. 592 00:28:36,570 --> 00:28:40,649 I'll use spell check too, but I'm not going to just Jesus, 593 00:28:40,649 --> 00:28:41,850 take the wheel on this thing. 594 00:28:42,149 --> 00:28:42,510 Alyss Noland: Yeah. 595 00:28:42,570 --> 00:28:44,070 And I, I feel much the same way. 596 00:28:44,070 --> 00:28:46,620 There's been ideas that I've had that I'm like. 597 00:28:47,235 --> 00:28:51,165 I really wanna share this with people like the, the idea of like, Hey, are you 598 00:28:51,165 --> 00:28:52,875 like, what happens with billing systems? 599 00:28:52,875 --> 00:28:56,835 Like how do people end up in these like weird, uh, choke points around having 600 00:28:56,835 --> 00:29:00,915 PLG versus having an enterprise and how do you like make those two meet when 601 00:29:00,915 --> 00:29:02,295 you're ready to make that decision? 602 00:29:02,385 --> 00:29:06,345 And I, I was able to iterate through the idea because at the, 603 00:29:06,345 --> 00:29:09,825 you know, again, this is like back in, gosh, like 2023 or something. 604 00:29:10,195 --> 00:29:14,125 I asked like, Claude, make an outline of a blog post, make, make 605 00:29:14,125 --> 00:29:17,185 a draft of a, of a blog post based off this outline with this content. 606 00:29:17,635 --> 00:29:20,425 And I realized I was trying to fit too many different topics into 607 00:29:20,425 --> 00:29:23,845 the, into that blog post, and it really should have been multipart. 608 00:29:23,845 --> 00:29:27,505 And I wouldn't have realized that unless I had seen what the final output was. 609 00:29:27,505 --> 00:29:30,715 And otherwise it would've been my a DH adhd, drafting it, getting exhausted, 610 00:29:30,715 --> 00:29:32,155 and then never finishing the damn thing. 611 00:29:32,575 --> 00:29:32,725 Corey Quinn: Yeah. 612 00:29:32,755 --> 00:29:32,905 Oh. 613 00:29:32,905 --> 00:29:34,915 I'll often say like, here's what I have so far. 614 00:29:34,975 --> 00:29:35,905 Don't write it for me. 615 00:29:36,524 --> 00:29:37,635 What should I focus on next? 616 00:29:37,635 --> 00:29:40,185 And it becomes a great unlock in that respect. 617 00:29:40,455 --> 00:29:42,225 But ultimately, it's you. 618 00:29:42,225 --> 00:29:43,095 It's your voice. 619 00:29:43,095 --> 00:29:46,725 It's if it's just random content with no there, there behind 620 00:29:46,725 --> 00:29:48,645 it, it's hard to see a point. 621 00:29:48,915 --> 00:29:50,475 So these are exciting times. 622 00:29:51,014 --> 00:29:52,905 I wanna thank you for taking the time to speak with me. 623 00:29:52,905 --> 00:29:54,284 Please don't wait for years to come back. 624 00:29:54,284 --> 00:29:54,824 Next time. 625 00:29:54,824 --> 00:29:55,995 I wanna see how this plays out. 626 00:29:56,085 --> 00:29:57,740 Alyss Noland: I'm looking forward to coming back sooner too. 627 00:29:58,240 --> 00:29:58,750 Thanks, Corey. 628 00:29:59,169 --> 00:30:03,340 Corey Quinn: Alice Nolan, currently at Nvidia, I'm cloud economist Cory Quinn, 629 00:30:03,340 --> 00:30:05,110 and this is Screaming In the Cloud. 630 00:30:05,260 --> 00:30:08,409 If you've enjoyed this podcast, please leave a five star review on your podcast 631 00:30:08,409 --> 00:30:12,220 platform of choice, whereas if you'd hated this podcast, please the five star 632 00:30:12,220 --> 00:30:15,940 review on your podcast platform of choice along with an angry, insulting comment 633 00:30:15,940 --> 00:30:19,990 that you had the AI generate for you, which is why I'm reading an angry comment 634 00:30:19,990 --> 00:30:23,860 that just says that I'm not comfortable saying that to people on the internet.