the localhost:0002 | the $25 token is dead 的完整文字稿

00:03:50,738 Chauncey: do they just make money? 73 00:03:50,738 --> 00:03:52,455 'Cause that might be also the other thing, right? 74 00:03:57,583 --> 00:03:59,493 Frankcx: I think it's a double down, right? 75 00:03:59,543 --> 00:04:02,673 I mean, I think it's a double down into the idea of what an open weight model 76 00:04:02,693 --> 00:04:04,663 is and why it's important to NVIDIA. 77 00:04:04,753 --> 00:04:09,143 I mean, NVIDIA is known as the company that essentially builds cloud-based 78 00:04:09,203 --> 00:04:14,113 AI, but they're also, with their announcements and deployments of DGX 79 00:04:14,113 --> 00:04:18,833 Spark and RTX Spark coming, you know, the idea would be is that now you have 80 00:04:18,833 --> 00:04:22,973 models that can run wherever you want intelligence to be, whether that's on 81 00:04:22,973 --> 00:04:25,113 your device or in your data center. 82 00:04:25,133 --> 00:04:29,223 And Hugging Face is the tool that deploys those different solutions. 83 00:04:29,233 --> 00:04:33,293 So to me, it aligns really well with what NVIDIA is doing. 84 00:04:33,743 --> 00:04:36,434 But, you know, I think, time will tell if they change 85 00:04:36,734 --> 00:04:36,924 Jacob: Well I 86 00:04:36,924 --> 00:04:39,464 mean, there's just a risk they lose their users, like the 87 00:04:39,554 --> 00:04:41,354 community forks and like… 88 00:04:41,911 --> 00:04:42,261 Chauncey: Yeah 89 00:04:42,364 --> 00:04:44,784 Jacob: I think they've just got the users and the models, 90 00:04:44,884 --> 00:04:46,314 and so that's like the market. 91 00:04:47,554 --> 00:04:48,854 Neil, you were gonna say something? 92 00:04:48,910 --> 00:04:53,820 Neil: I saw an analogy that NVIDIA's chips, right? 93 00:04:53,830 --> 00:04:59,200 These are the ovens, and Hugging Face, they have the library of recipes, 94 00:04:59,400 --> 00:05:02,700 and if you control the oven and the recipes, you can start, really 95 00:05:02,700 --> 00:05:04,810 delivering outsized value to the market. 96 00:05:05,100 --> 00:05:09,890 I think it's also important to note that this isn't, NVIDIA's first attempt, to, 97 00:05:10,303 --> 00:05:10,683 Chauncey: Hmm 98 00:05:10,820 --> 00:05:15,810 Neil: they actually invested, in Hugging Face back in 2023, alongside 99 00:05:15,810 --> 00:05:18,160 Salesforce, Google, Amazon, and IBM. 100 00:05:18,510 --> 00:05:22,380 Last year, NVIDIA wanted to follow on with another 500 million. 101 00:05:22,410 --> 00:05:25,280 Hugging Face said, "No, we don't wanna give outsized control to a single 102 00:05:25,280 --> 00:05:31,110 investor." Potential hot take here would be, did the, breach of Hugging Face, 103 00:05:31,140 --> 00:05:32,760 lead potentially to this acquisition? 104 00:05:33,060 --> 00:05:33,770 Frankcx: Oh 105 00:05:33,809 --> 00:05:35,539 Neil: need, we do need support from 106 00:05:35,974 --> 00:05:36,244 Hmm 107 00:05:36,369 --> 00:05:36,389 to, 108 00:05:37,284 --> 00:05:37,924 Chauncey: Interesting 109 00:05:38,279 --> 00:05:41,259 Neil: some of the security and, and maximize, the, the value 110 00:05:41,259 --> 00:05:42,429 of this entity we've created"? 111 00:05:42,449 --> 00:05:44,129 So couple talking points there. 112 00:05:44,129 --> 00:05:45,409 This isn't, out of the blue. 113 00:05:45,479 --> 00:05:49,679 NVIDIA actually has, quite a extensive history, including being on Hugging 114 00:05:49,879 --> 00:05:51,279 Face's cap table, three years ago 115 00:05:52,080 --> 00:05:55,340 Robert: Does that become a hotspot for the regulators if they 116 00:05:55,340 --> 00:05:57,020 own the recipes and the ovens? 117 00:06:00,360 --> 00:06:02,930 Chauncey: I mean there's still-- Yeah, but there's still a decent 118 00:06:02,930 --> 00:06:03,910 amount of competition here. 119 00:06:03,910 --> 00:06:05,700 Maybe not great competition, right? 120 00:06:05,700 --> 00:06:07,540 But there are still other platforms out there. 121 00:06:07,540 --> 00:06:11,270 So what grounds do the hawks have to stand on in this case? 122 00:06:11,367 --> 00:06:13,747 Jacob: there's nothing preventing those users from going to another 123 00:06:13,864 --> 00:06:14,284 Chauncey: Yeah 124 00:06:15,417 --> 00:06:16,157 Jacob: And that's why I think, 125 00:06:16,157 --> 00:06:17,027 Chauncey: Or even starting one 126 00:06:17,912 --> 00:06:21,452 Jacob: especially for things like, uncensored models that 127 00:06:21,462 --> 00:06:23,302 strip away some of the guardrails. 128 00:06:23,602 --> 00:06:28,502 Like for that portion of the local AI community, I think they will go elsewhere, 129 00:06:29,382 --> 00:06:30,872 away from the corporate entities 130 00:06:32,763 --> 00:06:35,783 Frankcx: Well, leaning into what Neil was saying around they have a history 131 00:06:35,783 --> 00:06:38,663 of doing this, I used to work for this company called Silicon Graphics 132 00:06:38,663 --> 00:06:42,533 that was making big time, you know, graphics engines, and then NVIDIA 133 00:06:42,593 --> 00:06:45,443 comes along and literally scales it so that they give it to everybody. 134 00:06:45,533 --> 00:06:49,533 So it's like they have a history of finding out ways to scale their 135 00:06:49,533 --> 00:06:53,713 business to bring everybody into the market, and not just somebody 136 00:06:53,713 --> 00:06:54,873 that can own a data center. 137 00:06:54,963 --> 00:06:59,003 So I think that that's a big play for an NVIDIA here is like their customer 138 00:06:59,023 --> 00:07:02,943 isn't just the big giant data centers, it's literally you and me, and, you know, 139 00:07:02,993 --> 00:07:08,803 anybody that can afford a 3090 card that's used or buy a new RTX Spark device, you 140 00:07:08,803 --> 00:07:10,063 know, that's who they're playing for. 141 00:07:10,103 --> 00:07:11,693 So I like it personally. 142 00:07:11,763 --> 00:07:13,693 I think it helps our story, for local AI, 143 00:07:13,993 --> 00:07:14,423 Robert: hopefully it'll

144 00:07:14,505 --> 00:07:14,895 Chauncey: fingers 145 00:07:14,903 --> 00:07:16,303 Robert: the prices of RAM down 146 00:07:17,399 --> 00:07:17,949 Frankcx: Right. 147 00:07:18,329 --> 00:07:18,839 Robert: Maybe they can 148 00:07:18,975 --> 00:07:19,515 Chauncey: Mm-hmm. 149 00:07:21,257 --> 00:07:21,797 Frankcx: You're right. 150 00:07:21,877 --> 00:07:25,617 Yeah, I mean, that is an interesting… Maybe that's a show coming up of like, is 151 00:07:25,617 --> 00:07:29,147 local AI gonna take some of the heat and pressure off a data center deployment? 152 00:07:29,267 --> 00:07:31,077 You know, like, how do we balance that? 153 00:07:31,197 --> 00:07:36,767 But there will always be, in my mind, cloud AI, but, what level of it balances 154 00:07:36,767 --> 00:07:41,267 in the future is up to be seen, So moving on to the next topic of news that I'm 155 00:07:41,267 --> 00:07:46,057 bringing to the table today is around Qwen, which is our favorite Chinese model. 156 00:07:46,057 --> 00:07:47,037 At least it's one of mine. 157 00:07:47,037 --> 00:07:48,277 It's pretty capable. 158 00:07:48,537 --> 00:07:50,137 It's got amazing intelligence. 159 00:07:50,407 --> 00:07:54,387 It scales or quantizes, if I'm saying that right. 160 00:07:54,657 --> 00:07:57,947 But the interesting thing that Qwen, it had released a model 161 00:07:57,947 --> 00:07:59,677 3.8 a little over a week ago. 162 00:07:59,677 --> 00:08:00,517 We benchmarked it. 163 00:08:00,807 --> 00:08:04,427 But it also released last week or this week this interesting model 164 00:08:04,457 --> 00:08:10,197 called Qwen3.8-Flash.NEXT, which is actually a preview towards Qwen4. 165 00:08:10,537 --> 00:08:14,557 But it has some really interesting memory capabilities, and I know Jacob, you've 166 00:08:14,557 --> 00:08:17,367 played around a little bit with this, so maybe you can kind of tell us what you 167 00:08:17,424 --> 00:08:20,774 Jacob: I'm very, very passionate and interested in just like experimentation, 168 00:08:20,794 --> 00:08:24,144 'cause at this point I'm just learning, trying to be a sponge as much as possible. 169 00:08:24,144 --> 00:08:29,994 So the moment that it dropped, I had my agents working on the best way to figure 170 00:08:29,994 --> 00:08:34,594 out how to install it, 'cause I knew that it was 170, 180 billion parameters, 171 00:08:34,594 --> 00:08:38,454 and I was like, that seems like too much for what I can fit on a consumer card. 172 00:08:39,284 --> 00:08:40,664 it seems like it's just too big. 173 00:08:40,904 --> 00:08:44,104 I've got that external 3090 that I was thinking about running it on. 174 00:08:44,934 --> 00:08:48,964 And, I realized that there's a lot that I'm still learning, about this lookup 175 00:08:48,994 --> 00:08:53,914 table, these Ngram tables that are essentially like available for being 176 00:08:53,914 --> 00:08:57,734 able to… Well, they just don't need the same availability as the rest of the 177 00:08:57,734 --> 00:08:59,654 model does, is the way I think of it. 178 00:08:59,944 --> 00:09:05,294 They can use slower memory, they can use system RAM, and I heard some 179 00:09:05,294 --> 00:09:09,154 folks in the community talking about being able to use SSD storage for it. 180 00:09:09,984 --> 00:09:13,744 so that was what was cool for me, and I had some success with running the 181 00:09:13,744 --> 00:09:19,194 bulk of those model weights on SSD in the Ngram tables so that it could look 182 00:09:19,204 --> 00:09:20,824 up and then a small amount on VRAM 183 00:09:22,429 --> 00:09:24,319 Robert: Now, is this an MOE model, Frank? 184 00:09:26,209 --> 00:09:27,939 Is it a mixture of experts model 185 00:09:28,078 --> 00:09:30,118 Jacob: it is an MoE model technically. 186 00:09:30,118 --> 00:09:32,998 I think there are 6 billion parameters that are active at a time. 187 00:09:33,098 --> 00:09:33,988 I'll have to double-check. 188 00:09:34,068 --> 00:09:39,968 And then there's 51 billion parameters that are in that n-gram table, and then 189 00:09:39,968 --> 00:09:46,508 there's another 120 are available that should be on some fast RAM as well, 190 00:09:46,518 --> 00:09:51,238 but it doesn't need to be VRAM because they're not active all the time being MoE 191 00:09:52,861 --> 00:09:55,761 Frankcx: Yeah, it feels like even though you look at the size of the model and 192 00:09:55,761 --> 00:09:59,371 you're like, "That can't run on my device," the way it loads the portions 193 00:09:59,371 --> 00:10:02,181 of the model that you need are kind of in the chunks that you're asking 194 00:10:02,181 --> 00:10:04,011 for, not like everything all at once. 195 00:10:04,061 --> 00:10:04,779 So yeah, it's a, 196 00:10:05,079 --> 00:10:08,069 Robert: Which is the whole premise of a MoE model, right? 197 00:10:08,069 --> 00:10:11,057 It's a combination of a bunch of smaller models together, 198 00:10:11,357 --> 00:10:11,367 Chauncey: Mmhmm 199 00:10:11,479 --> 00:10:13,899 Robert: and it could pull from any one of them. 200 00:10:13,909 --> 00:10:16,789 there's a component that lives in there that says, "Oh, let's use this 201 00:10:16,789 --> 00:10:19,899 model for this task and this model for that one," which is awesome. 202 00:10:19,909 --> 00:10:20,829 It's super cool 203 00:10:21,958 --> 00:10:23,138 Jacob: Yeah, I think there's something like 204 00:10:23,201 --> 00:10:23,451 Frankcx: that, 205 00:10:23,751 --> 00:10:28,049 Jacob: 512 experts, and each one of those experts is a certain size in this model. 206 00:10:28,349 --> 00:10:28,689 Frankcx: Right 207 00:10:28,888 --> 00:10:31,138 Jacob: I'm still learning, but it's very interesting. 208 00:10:31,831 --> 00:10:35,281 Frankcx: I tried to run it on the CPU on my device, which is a Xeon-based 209 00:10:35,281 --> 00:10:38,621 device, and it was getting like four tokens per second, which is like 210 00:10:38,901 --> 00:10:41,891 barely… Like, if you think of a

token as like three quarters of a word, 211 00:10:42,191 --> 00:10:45,274 that's like somebody is slurring their words, like they're having trouble 212 00:10:45,524 --> 00:10:50,834 ran it on the 3090, it got up to 50 tokens per second, but because 213 00:10:50,834 --> 00:10:54,484 it was just chunking along, it had trouble finishing its tasks. 214 00:10:54,484 --> 00:10:57,204 So, there's definitely some tuning that has to be done. 215 00:10:57,514 --> 00:11:00,554 I think in the end, what I found is it's not built for the 3090. 216 00:11:00,594 --> 00:11:06,534 It has this weird lookup table that's 26 gigs and, a 3090 only has 24 gigs 217 00:11:06,534 --> 00:11:09,234 of VRAM, so it just doesn't fit right. 218 00:11:09,614 --> 00:11:15,374 It'll be interesting when, RTX Spark or DGX Spark or, other devices that ha- Like, 219 00:11:15,614 --> 00:11:19,884 Neil, I'd be interested if you tested it on a Mac that has unified memory, how 220 00:11:19,894 --> 00:11:23,624 that would run, because it seems like it's more geared towards that type of hardware 221 00:11:23,954 --> 00:11:26,604 than it is something like a 3090 card. 222 00:11:29,299 --> 00:11:30,999 Neil: Got my homework for next week, so 223 00:11:31,756 --> 00:11:32,196 Frankcx: All right 224 00:11:32,496 --> 00:11:35,479 Neil: received some feedback folks liked the, our ability to translate 225 00:11:35,489 --> 00:11:37,339 technical topics into business terms. 226 00:11:37,359 --> 00:11:40,919 I think pausing for a moment on Mixture of Experts, all the rage right now. 227 00:11:41,199 --> 00:11:43,169 Maybe this analogy fits, maybe it doesn't. 228 00:11:43,409 --> 00:11:48,039 I like to view Mixture of Experts as, analogous to the Encyclopedia Britannica 229 00:11:48,169 --> 00:11:50,209 bookshelf at my grandparents' house. 230 00:11:50,529 --> 00:11:50,799 You've got 231 00:11:50,992 --> 00:11:51,662 Frankcx: Oh, right 232 00:11:51,909 --> 00:11:52,679 Neil: of all 233 00:11:52,753 --> 00:11:53,263 Chauncey: All right 234 00:11:53,509 --> 00:11:54,149 Neil: topics. 235 00:11:54,589 --> 00:11:55,309 If you know 236 00:11:55,692 --> 00:11:57,902 Frankcx: B through C-A 237 00:11:57,959 --> 00:12:00,399 Neil: You don't need to open all of the books at the same time. 238 00:12:00,639 --> 00:12:03,602 So, just, again, wanna pause for a moment, define Mixture of Experts. 239 00:12:03,902 --> 00:12:04,592 Frankcx: love it 240 00:12:04,679 --> 00:12:09,339 Neil: this is going to allow us to leverage larger parameter models 241 00:12:09,589 --> 00:12:12,749 without, running into some of the memory bandwidth constraints that we've 242 00:12:12,749 --> 00:12:16,699 seen with, previous, models before Mixture of Experts became, top of mind 243 00:12:17,616 --> 00:12:19,406 Robert: And I'll share some stuff later on, 244 00:12:19,416 --> 00:12:19,726 Chauncey: Oh, nice 245 00:12:19,934 --> 00:12:21,594 Robert: this week with MoE models 246 00:12:22,828 --> 00:12:23,608 Frankcx: Oh, nice. 247 00:12:23,848 --> 00:12:27,618 Neil, given that you're talking about unified memory and kind of memory 248 00:12:27,618 --> 00:12:30,678 management, I think you've been doing some stuff with Apple as well. 249 00:12:30,678 --> 00:12:34,608 So what can you tell us about some of the Mac stuff or the news 250 00:12:34,608 --> 00:12:36,048 that you're excited about for Mac? 251 00:12:36,807 --> 00:12:41,097 Neil: So big announcements this week from the Apple world. 252 00:12:41,147 --> 00:12:47,917 They announced M5, M6 processors, Mac Minis, which became all the rage with the 253 00:12:47,947 --> 00:12:50,857 OpenCL moment earlier this calendar year. 254 00:12:51,127 --> 00:12:57,867 We're continuing to see Apple double down on this local AI moment. 255 00:12:57,887 --> 00:13:01,077 The twenty twenty-six was supposed to be the year of agents. 256 00:13:01,387 --> 00:13:05,197 I think one can make an argument that it's becoming the year of local and hybrid AI. 257 00:13:05,197 --> 00:13:05,702 Um, Apple continues to, um, innovate at the silicon layer, uh, dropping 258 00:13:05,702 --> 00:13:09,772 down to a two you look at some of the models like, Kimi V3, right, and three 259 00:13:09,772 --> 00:13:13,792 hundred and fifty gigs, you see almost a direct parallel between the, the 260 00:13:13,802 --> 00:13:17,522 type of silicon and, and these types of MoE models that we're seeing released. 261 00:13:17,772 --> 00:13:22,662 Ran some tests on an M3 Max, versus the thirty ninety. 262 00:13:22,662 --> 00:13:24,702 So Frank, thanks for sharing those benchmark tests. 263 00:13:24,992 --> 00:13:27,592 I, I think it boils down to what are you trying to get 264 00:13:27,612 --> 00:13:29,542 out of these local AI models? 265 00:13:29,772 --> 00:13:35,062 Ran those five scenarios, and the consistent takeaway was anything 266 00:13:35,062 --> 00:13:40,362 that's bandwidth bound, memory bound, Mac is outpacing the competition. 267 00:13:40,672 --> 00:13:45,122 When it comes to raw compute, NVIDIA's, you know, taking the cake, right? 268 00:13:45,372 --> 00:13:47,412 Apple is two to six X slower. 269 00:13:47,412 --> 00:13:52,202 So again, that's a, a couple generations back, but you're seeing Apple continue 270 00:13:52,202 --> 00:13:56,372 to double down on these, bandwidth, memory bound constraints, and you're 271 00:13:56,372 --> 00:14:00,422 seeing NVIDIA to, you know, continuing to outpace the competition when 272 00:14:00,422 --> 00:14:02,102 it comes to, to raw compute power. 273 00:14:02,102 --> 00:14:06,142 So it'll be interesting to see the shift in paradigm and, again, lots of, 274 00:14:06,142 --> 00:14:08,882 excitement for Mac enthusiasts, this week. 275 00:14:09,192 --> 00:14:12,812 Curious any other takes from the crew here on, Apple's recent announcements? 276 00:14:13,431 --> 00:14:16,261 Chauncey: I mean, I would say as a, fellow hardware nerd and just tracking 277 00:14:16,261

--> 00:14:19,461 these other guys, like, this is only just beneficial to the entire ecosystem, right? 278 00:14:19,461 --> 00:14:21,881 If Apple's going this way, Windows has to follow. 279 00:14:21,911 --> 00:14:25,711 Like, as an ecosystem, of course, if you think about Surface and Dell and 280 00:14:25,711 --> 00:14:28,871 Lenovo and everything else that's out there, like, it's so critical that 281 00:14:28,871 --> 00:14:30,771 we all continue to match this space. 282 00:14:30,781 --> 00:14:34,321 So I'm always excited when new things come out because that means that there's 283 00:14:34,321 --> 00:14:37,541 gonna be new things on the horizon for everyone else and we get some, I mean, 284 00:14:37,541 --> 00:14:38,961 it's just gonna continue to get better. 285 00:14:39,091 --> 00:14:42,481 Just imagine the amount of power we have now and what we're able to achieve, 286 00:14:42,491 --> 00:14:45,391 which all you guys are talking about so far of, like, testing locally. 287 00:14:45,531 --> 00:14:47,771 What we'll be able to do in just a year or two years it's 288 00:14:47,771 --> 00:14:49,221 just exponential from here out. 289 00:14:50,100 --> 00:14:50,150 Robert: I 290 00:14:50,150 --> 00:14:52,830 wouldn't exactly say Windows is following, right? 291 00:14:52,830 --> 00:14:55,910 But just to be clear, I mean, there's some pretty amazing things that are 292 00:14:55,910 --> 00:15:00,533 happening in the Windows ecosystem Look, Neil and I have had these conversations. 293 00:15:00,583 --> 00:15:05,373 It was just cloud, then it was cloud and device, and now it's cloud, edge, and 294 00:15:05,373 --> 00:15:07,593 device in sort of a three-tiered model. 295 00:15:08,283 --> 00:15:10,713 once you get to the phone, you 296 00:15:10,713 --> 00:15:11,913 know, bets are off 297 00:15:12,811 --> 00:15:15,561 Frankcx: I think the thing that I see is, look at what happened this past week. 298 00:15:15,581 --> 00:15:19,341 Two things: NVIDIA looking at Hugging Face, that's a double down on local 299 00:15:19,341 --> 00:15:24,461 AI and open-weight models, and then Apple, another multi-trillion 300 00:15:24,461 --> 00:15:28,371 dollar company, doubles down on local AI themselves by releasing 301 00:15:28,371 --> 00:15:30,061 these Macs that are in that space. 302 00:15:30,101 --> 00:15:34,971 So I would argue, and I'm not with Microsoft anymore, but I would argue 303 00:15:34,971 --> 00:15:40,041 that the governance that Microsoft is gonna bring to these local AI spaces, I 304 00:15:40,041 --> 00:15:43,121 think that's why a lot of people should be excited about what RTX Spark could 305 00:15:43,121 --> 00:15:47,881 bring, 'cause you're bringing this capability of what CUDA and speeds are, 306 00:15:48,131 --> 00:15:51,701 but you're bringing the governance of what you need, because many enterprise 307 00:15:51,721 --> 00:15:53,533 companies are scared to death of running, 308 00:15:53,833 --> 00:15:55,183 Robert: And they

should be, yeah. 309 00:15:55,193 --> 00:15:56,053 Well, not Linux, but 310 00:15:56,063 --> 00:15:57,363 Of running agents for sure 311 00:15:57,873 --> 00:15:58,223 Chauncey: Yeah 312 00:15:58,666 --> 00:15:59,406 Jacob: I was talking to a customer 313 00:15:59,571 --> 00:16:02,341 Frankcx: being able to have that governance that Microsoft can bring over 314 00:16:02,341 --> 00:16:07,061 top of what local AI can, bring from an intelligence standpoint, there's some 315 00:16:07,061 --> 00:16:09,001 exciting things just around the corner 316 00:16:09,070 --> 00:16:14,310 Jacob: I was talking to a company in Japan this week, and they were pretty frank 317 00:16:14,310 --> 00:16:18,840 with me, like: "Hey, right now we can't run agents at all in our organization." 318 00:16:19,060 --> 00:16:20,990 And they are very AI forward. 319 00:16:20,990 --> 00:16:23,860 So AI forward, they actually have a partnership now to run local 320 00:16:23,860 --> 00:16:25,960 models on all of their laptops. 321 00:16:26,030 --> 00:16:29,380 So they're running local models everywhere, but those local models 322 00:16:29,420 --> 00:16:33,100 are not oper-- They're-- it's all just connecting to the cloud or to 323 00:16:33,100 --> 00:16:35,060 specific outputs and not agented tasks. 324 00:16:35,060 --> 00:16:40,950 So they're like: "We want to run like Scout or things like OpenClaw or, you 325 00:16:40,950 --> 00:16:46,080 know, a whole host of other agents." And, I think there's a continued story 326 00:16:46,080 --> 00:16:50,040 that, Windows will continue to add value with Microsoft Execution Containers, 327 00:16:50,280 --> 00:16:52,740 Agent 365, all of that goodness. 328 00:16:54,041 --> 00:16:55,671 Frankcx: Well, you guys will get a kick out of this story. 329 00:16:55,671 --> 00:16:58,281 You may not know this, but I've been playing around with doing a little bit 330 00:16:58,281 --> 00:17:00,719 of Uber driving myself, just on like a, 331 00:17:01,019 --> 00:17:02,059 Robert: In the sprinter? 332 00:17:02,081 --> 00:17:05,571 Frankcx: I want, I wanna get down into Cap Hill and see what the 333 00:17:05,571 --> 00:17:09,261 cool kids are doing, so I'm driving them around, dad in his Volvo. 334 00:17:09,651 --> 00:17:13,741 I picked up this guy, and we had this conversation, and I'm like: "Hey, what 335 00:17:13,741 --> 00:17:17,651 do you do?" And he's like: "Oh I run an aerospace company." I'm like: "Oh, 336 00:17:17,651 --> 00:17:21,221 that's pretty cool." I was like: "Are you guys using much AI?" He's like: 337 00:17:21,481 --> 00:17:25,211 "I use it to write emails, but I can't use it for engineering because we're 338 00:17:25,211 --> 00:17:29,231 doing contracts with Boeing and the defense side that I can't share the IP 339 00:17:29,231 --> 00:17:30,881 of what we're doing with the cloud." 340 00:17:30,961 --> 00:17:34,311 And I was like: "Well, do you know much about local AI?" And he's like: "Well, 341 00:17:34,311 -->

00:17:37,931 tell me." And I'm like, here's this Uber driver telling him about local AI. 342 00:17:38,311 --> 00:17:41,581 He used to be in this band called Fifth Angel, and he was the lead 343 00:17:41,581 --> 00:17:44,861 singer for it, and it's this Seattle hair metal band that came out 344 00:17:44,891 --> 00:17:46,821 Robert: may- maybe his next band will be called Air Gap 345 00:17:47,679 --> 00:17:48,699 Chauncey: Ooh, nice. 346 00:17:48,749 --> 00:17:49,449 That's good. 347 00:17:49,519 --> 00:17:50,039 Robert: Hair, you got a 348 00:17:50,369 --> 00:17:51,019 Chauncey: That's good. 349 00:17:51,229 --> 00:17:51,419 Robert: gap 350 00:17:51,549 --> 00:17:51,999 Frankcx: gap. 351 00:17:52,569 --> 00:17:54,439 Chauncey: Robert coming up with the band names, man 352 00:17:55,199 --> 00:17:55,519 Frankcx: Yeah. 353 00:17:56,179 --> 00:17:58,239 Does anyone else have any news they wanna share? 354 00:17:59,739 --> 00:18:02,509 All right, if not, it feels like we should do a baseline. 355 00:18:02,559 --> 00:18:06,009 Baseline to me is like, Chauncey, I know you're interested in making sure we're all 356 00:18:06,009 --> 00:18:08,019 talking speak that everybody understands. 357 00:18:08,019 --> 00:18:09,489 Maybe you can give us a little bit of a clue 358 00:18:09,547 --> 00:18:12,347 Chauncey: This is also referred to as the section of the podcast where 359 00:18:12,347 --> 00:18:14,097 it's like, I have no idea what you guys are actually talking about. 360 00:18:14,097 --> 00:18:14,907 Can you please explain it? 361 00:18:14,957 --> 00:18:16,607 So can you do a couple things? 362 00:18:16,667 --> 00:18:17,807 One, I know some of it. 363 00:18:18,077 --> 00:18:20,467 Like what's the difference when we're-- Well, first of 364 00:18:20,467 --> 00:18:21,967 all, like what is quantization? 365 00:18:21,977 --> 00:18:23,717 So you kind of mentioned quantizing earlier. 366 00:18:23,717 --> 00:18:27,167 Like, I think it's probably really worth diving a little bit into the 367 00:18:27,167 --> 00:18:28,967 details of what quantization means. 368 00:18:29,227 --> 00:18:32,207 And then I would actually love to know your thoughts on the 369 00:18:32,207 --> 00:18:35,317 difference in cost per token and how we're thinking about that too. 370 00:18:37,595 --> 00:18:41,275 Frankcx: Yeah, I mean, I hear a lot of people talking about, the $25 token or the 371 00:18:41,275 --> 00:18:43,785 $6 token or, like, you know, what is that? 372 00:18:43,795 --> 00:18:46,225 Nobody's paying $6 for a token. 373 00:18:46,235 --> 00:18:49,285 Well, it comes down to it's $6 per million tokens. 374 00:18:49,315 --> 00:18:54,745 And what the funny part is that this past week, you know, Grok, which I think is 375 00:18:54,745 --> 00:18:58,965 Neil's favorite AI, like it's, I called it last week your drunk uncle, but your 376 00:18:58,965 --> 00:19:03,955 drunk uncle sobered up and became a pretty capable, model at $6 per million tokens, 377

00:19:03,955 --> 00:19:09,335 which honestly is like a fifth of the cost of what, OpenAI and Claude charge 378 00:19:09,345 --> 00:19:12,315 for their, frontier models at $25 a token. 379 00:19:12,935 --> 00:19:16,345 But I think the point is, is that because of the ability to start 380 00:19:16,345 --> 00:19:20,875 running these models locally, it's starting to drive down, you know… 381 00:19:20,875 --> 00:19:25,495 And when Grok comes out with a $6 per million, you know, token cost, that 382 00:19:25,495 --> 00:19:28,665 really makes everybody else look at it like, "Hey, what are you doing 383 00:19:28,665 --> 00:19:32,395 there? And how come you're not charging as much as we are?" Well, you know, 384 00:19:32,395 --> 00:19:35,815 they're not the premier, you know, AI, but they are showing themselves 385 00:19:35,815 --> 00:19:37,345 to have just as much capability. 386 00:19:37,345 --> 00:19:40,965 And Neil, maybe you can talk a little bit about Grok, but it'd be like, you know, 387 00:19:40,965 --> 00:19:45,005 it feels like there are models out there, because they're all sharing algorithms, 388 00:19:45,005 --> 00:19:47,015 they're all becoming kind of equally smart 389 00:19:47,736 --> 00:19:51,706 Neil: Yeah, it's becoming more of a commodity value-based discussion. 390 00:19:51,736 --> 00:19:54,286 The days of unlimited tokens are over, right? 391 00:19:54,326 --> 00:20:00,046 As we mentioned, currently work at Microsoft and we have set token limits. 392 00:20:00,076 --> 00:20:06,596 And so continuing to route every prompt through frontier models, Opus 5, GPT 393 00:20:06,596 --> 00:20:09,156 Sol 5.6 no longer becomes practical. 394 00:20:09,166 --> 00:20:12,946 So, shifting some of the development needs to Groq 4.6 395 00:20:13,166 --> 00:20:15,126 stretches my token budget further. 396 00:20:15,336 --> 00:20:19,466 And if it can do, as good or close to as good of a job, if I can get more 397 00:20:19,466 --> 00:20:23,746 done with my existing token budget, that's why, Groq, the drunken uncle, 398 00:20:23,746 --> 00:20:27,246 that sobered up, is now becoming, one of my favorite models to work with. 399 00:20:27,306 --> 00:20:31,306 not all tokens are equal, dependent upon which models you choose. 400 00:20:31,306 --> 00:20:34,776 And I think Groq offers, great value at this moment in time 401 00:20:35,237 --> 00:20:37,857 Chauncey: I'm really interested to see where we go with, model 402 00:20:37,857 --> 00:20:39,147 choice and making it easier. 403 00:20:39,147 --> 00:20:42,537 I mean, like the average person-- Your developers, obviously, they're gonna 404 00:20:42,537 --> 00:20:45,287 know, like, "Oh, I should probably choose something cheaper," and whatnot. 405 00:20:45,297 --> 00:20:48,697 But like the average person's just gonna go, "Oh, this says it works better. 406 00:20:48,707 --> 00:20:52,397 I'm gonna click on work better." Right? And, but ultimately, like I 407 00:20:52,397 --> 00:20:56,487 would say companies would value either Microsoft or whoever making this kind 408 00:20:56,487 --> 00:21:00,397 of orchestration layer that says, "Hey, for sure, I'm building a PowerPoint, 409 00:21:00,427 --> 00:21:02,457 I'm gonna have this level of model. 410 00:21:02,607 --> 00:21:05,207 But if I'm just answering a simple question, I should be able to like 411 00:21:05,247 --> 00:21:06,887 dynamically switch to another model." 412 00:21:06,887 --> 00:21:09,687 I think that's where I'm really excited to see some of this happen. 413 00:21:09,687 --> 00:21:12,547 And I know we have some auto settings in, Cowork and in 414 00:21:12,547 --> 00:21:14,377 Scout, but not quite there yet. 415 00:21:14,387 --> 00:21:18,307 It's still a pre-selected handful of models versus if it can really 416 00:21:18,307 --> 00:21:21,707 start dynamically going across whatever your organization says. 417 00:21:21,717 --> 00:21:24,197 Like, I want only these six models to ever run. 418 00:21:24,197 --> 00:21:26,691 I think that's when we're gonna start seeing some pretty cool things happen. 419 00:21:26,991 --> 00:21:29,471 Robert: it goes down a whole other layer that general 420 00:21:29,471 --> 00:21:30,821 users don't understand, right? 421 00:21:30,871 --> 00:21:31,431 Chauncey: Well, yeah. 422 00:21:31,672 --> 00:21:31,911 Yeah 423 00:21:31,972 --> 00:21:36,001 Robert: know, sonnet or whatever, then there's like medium, max, 424 00:21:36,301 --> 00:21:36,761 Chauncey: Yep. 425 00:21:37,061 --> 00:21:39,052 Robert: extra large, extra max, 426 00:21:39,190 --> 00:21:39,450 Frankcx: Right 427 00:21:39,682 --> 00:21:40,782 Robert: big max, 428 00:21:40,862 --> 00:21:41,302 Chauncey: Plus 429 00:21:41,360 --> 00:21:44,030 Frankcx: wondering, to bring it back to Chauncey's original question 430 00:21:44,040 --> 00:21:47,020 around quantization, I'm wondering who can take a shot at that. 431 00:21:47,030 --> 00:21:49,460 And when you quantize something, does it make it less 432 00:21:49,845 --> 00:21:53,785 Jacob: Well I'll do my first pass, and then I'll get corrected by Robert or 433 00:21:53,785 --> 00:21:57,195 Neil, is probably gonna be my choice here. 434 00:21:57,495 --> 00:22:00,903 The way I think of quants is I think of it like any other compression. 435 00:22:01,465 --> 00:22:04,905 Like you're making the model and the weights take up less space, 436 00:22:05,775 --> 00:22:11,165 a result, you are potentially cutting some of the detail. 437 00:22:11,945 --> 00:22:16,505 Just like when you have a JPEG image that's compressed, you may 438 00:22:16,505 --> 00:22:18,015 be cutting out some of the detail. 439 00:22:18,015 --> 00:22:20,645 It's doing some calculations of, "Hey, we're grouping these pixels 440 00:22:20,645 --> 00:22:24,255 together. They look the same." It's saving you space on disk. 441 00:22:24,255 --> 00:22:25,705 It's saving you space in RAM. 442 00:22:26,085 --> 00:22:32,015 And so if a model like was a seven

billion parameter model, like its full weights 443 00:22:32,185 --> 00:22:35,305 would probably be like double that. 444 00:22:35,365 --> 00:22:37,075 They would probably be like 14 gigs. 445 00:22:38,075 --> 00:22:44,305 if you, compress it down to Q8, then it might be seven gigs. 446 00:22:44,315 --> 00:22:47,415 So it might cut the space in half, that's kinda how I think about it, is like you 447 00:22:47,415 --> 00:22:51,035 may be trading off some precision, you may not be able to trust the outputs in the 448 00:22:51,035 --> 00:22:55,955 same way, in a world where you have models where you're running them to do a task, 449 00:22:55,995 --> 00:22:59,975 and then you have another model checking its work, then maybe it makes sense. 450 00:23:00,045 --> 00:23:00,285 Chauncey: So 451 00:23:00,381 --> 00:23:00,621 Robert: Yep, 452 00:23:00,645 --> 00:23:00,895 Chauncey: it 453 00:23:01,341 --> 00:23:02,841 Robert: The trade-off is accuracy. 454 00:23:03,251 --> 00:23:03,571 How do you 455 00:23:03,675 --> 00:23:04,035 Chauncey: Hmm 456 00:23:04,811 --> 00:23:08,341 Robert: more, you know, tighter without losing the accuracy? 457 00:23:08,341 --> 00:23:09,741 And that's the trick that goes with it. 458 00:23:10,001 --> 00:23:13,201 And look, we're even seeing some models get down to one-bit 459 00:23:13,301 --> 00:23:15,131 quantization, which is insane. 460 00:23:15,481 --> 00:23:17,401 But that's where the technology is going. 461 00:23:17,791 --> 00:23:20,981 But that, that's always, as I understand it, that's the challenge is how do you 462 00:23:20,981 --> 00:23:27,823 make it and use less, but still maintain the accuracy and the credibility of the, 463 00:23:28,123 --> 00:23:28,163 Chauncey: S- 464 00:23:28,171 --> 00:23:29,131 Robert: the answers you get? 465 00:23:29,303 --> 00:23:30,313 Chauncey: how do you choose then? 466 00:23:30,453 --> 00:23:34,923 Like, if I have this option of a frontier version that's gonna be obviously the most 467 00:23:34,923 --> 00:23:38,733 powerful, the most intelligent, and like, that's kind of like what I would think I 468 00:23:38,733 --> 00:23:42,433 would want all the time versus something that is more efficient, that is gonna be 469 00:23:42,433 --> 00:23:44,253 smaller, that I can maybe run locally. 470 00:23:44,533 --> 00:23:47,353 Like, if there is kind of that more gut reaction of just like, 471 00:23:47,363 --> 00:23:48,323 well, then how do I choose? 472 00:23:48,343 --> 00:23:49,743 Like, what would be your kind of answer then? 473 00:23:50,882 --> 00:23:55,492 Jacob: I would say the highest quant that you can the space that you have. 474 00:23:55,792 --> 00:24:00,142 So like the benchmarks for like 3.8, 27 billion parameters that we talked about 475 00:24:00,142 --> 00:24:03,432 last week, and, I think Robert, you might have some more to share about this week. 476 00:24:03,722 --> 00:24:08,492 Like those benchmarks were generally ran at full precision, like no

quant at all. 477 00:24:08,892 --> 00:24:12,852 So you want it to be as precise as possible, but there's a 478 00:24:12,852 --> 00:24:16,402 trade-off in speed and efficiency and what can physically fit. 479 00:24:17,612 --> 00:24:18,222 That's my take. 480 00:24:19,797 --> 00:24:19,827 Frankcx: Yep. 481 00:24:21,607 --> 00:24:26,357 Well, getting to that, I think it's time to go into the tension of the week, and 482 00:24:26,357 --> 00:24:29,847 I know that Robert, you said you'd be willing to be on the hot seat, so I'm 483 00:24:29,955 --> 00:24:30,445 Robert: Why not, eh? 484 00:24:30,607 --> 00:24:31,289 Give it a go. 485 00:24:31,589 --> 00:24:31,689 Frankcx: it 486 00:24:31,877 --> 00:24:34,027 Robert: I'm always anxious to get into the fight. 487 00:24:34,447 --> 00:24:38,227 So listen, we've been talking about AI running it locally. 488 00:24:38,267 --> 00:24:39,907 Now we can run it ourselves. 489 00:24:40,217 --> 00:24:41,757 The models keep getting better. 490 00:24:42,047 --> 00:24:45,787 it, just obviously saves money than using the ones… But the big cloud 491 00:24:45,787 --> 00:24:47,537 models keep getting cheaper as well. 492 00:24:48,337 --> 00:24:51,647 so the big question is: where are the AI companies eventually 493 00:24:51,647 --> 00:24:52,617 gonna get their money? 494 00:24:52,637 --> 00:24:53,927 What are they gonna charge for? 495 00:24:54,337 --> 00:24:57,397 Wendell at Level 1 Tech says the answer is speed. 496 00:24:57,487 --> 00:25:00,907 Not smarter answers, just the same answer, but faster. 497 00:25:01,307 --> 00:25:04,987 the example I used, earlier off air was, you know, I'm gonna order something 498 00:25:04,997 --> 00:25:07,367 from Amazon, it'll come in two days. 499 00:25:07,447 --> 00:25:10,527 If I want it in a day, I'm gonna pay, two ninety-nine to get it. 500 00:25:10,537 --> 00:25:14,477 So we're already starting to see it, in the AI world. 501 00:25:14,537 --> 00:25:16,377 If you wanna wait, you pay less. 502 00:25:16,387 --> 00:25:18,097 If you want it right now, you pay more. 503 00:25:18,447 --> 00:25:20,297 Same AI, just a different clock. 504 00:25:20,377 --> 00:25:23,897 And so, you know, the twenty-five dollar, token is dead, right? 505 00:25:23,897 --> 00:25:28,227 I think within 18 months, the only thing that frontier labs are-- really 506 00:25:28,227 --> 00:25:31,857 gonna be able to sell is the speed, and not so much of the intelligence. 507 00:25:31,877 --> 00:25:35,267 I experienced some of this this week, in loading, a thirty billion 508 00:25:35,267 --> 00:25:39,907 parameter model, onto one of my devices, and I was blown away, at the 509 00:25:39,917 --> 00:25:44,157 speed and the accuracy and, just the capability of it and how far it's come. 510 00:25:44,177 --> 00:25:48,317 So, in plain English, soon being smarter will not be enough to 511 00:25:48,317 --> 00:25:51,507 charge more for, you're gonna have to be faster, and that's my take. 512 00:25:52,969 --> 00:25:53,299 Frankcx: Yeah. 513 00:25:54,529 --> 00:25:57,779 I mean, you know, for me, I think of it as like Neil's… He's Mr. 514 00:25:57,779 --> 00:26:01,839 Analogy, so I'll bring in the analogy of like, to me, like a cloud AI 515 00:26:01,839 --> 00:26:03,569 is like hiring a consulting firm. 516 00:26:03,589 --> 00:26:05,079 They can bring in a lot of people. 517 00:26:05,359 --> 00:26:06,299 They can do a lot of work. 518 00:26:06,319 --> 00:26:07,619 They can scale very well. 519 00:26:07,639 --> 00:26:08,779 They're expensive. 520 00:26:09,259 --> 00:26:11,909 They don't know you very well, so they have to take the time to learn 521 00:26:11,909 --> 00:26:13,419 you, you know, what you're doing. 522 00:26:14,139 --> 00:26:17,139 if you think about local AI, could literally go out and hire 523 00:26:17,139 --> 00:26:20,919 an MBA from business school and bring them in and train them. 524 00:26:21,119 --> 00:26:21,719 They're smart. 525 00:26:21,759 --> 00:26:24,689 They're just as smart as who you could bring in from a consulting group, 526 00:26:25,059 --> 00:26:26,489 but they just can't do as much work. 527 00:26:26,539 --> 00:26:28,529 They don't have the same volume that they can do. 528 00:26:28,539 --> 00:26:30,759 You could hire more and more of them, but then you have to take 529 00:26:30,759 --> 00:26:34,419 on the responsibility of managing them and taking care of them. 530 00:26:34,769 --> 00:26:38,979 for me, it's this, type of analogy that the AI is getting smart enough 531 00:26:38,999 --> 00:26:42,199 that you can run it locally, and then you're able to control privacy. 532 00:26:42,249 --> 00:26:44,569 That person isn't gonna take that information and share it with 533 00:26:44,569 --> 00:26:46,009 some other consulting agency. 534 00:26:46,379 --> 00:26:48,979 It's sovereign, meaning that, it's controlled as to where 535 00:26:48,999 --> 00:26:50,409 it goes and where it lives. 536 00:26:50,789 --> 00:26:55,079 It's local, so if something happened, like, there was a storm and, things got 537 00:26:55,079 --> 00:26:58,339 knocked out, that consulting company that, flies in every week would not be 538 00:26:58,339 --> 00:27:01,689 able to get to you, but that business person can come to work that day. 539 00:27:02,099 --> 00:27:04,699 So there's a lot of advantages of having local. 540 00:27:04,999 --> 00:27:07,259 But that doesn't mean that the cloud's ever gonna go away. 541 00:27:07,279 --> 00:27:10,719 It's gonna be there for when you need it and the tough problems. 542 00:27:11,039 --> 00:27:14,509 But I would argue that AI being balanced is the way we're 543 00:27:14,509 --> 00:27:15,699 looking at AI in the future 544 00:27:19,475 --> 00:27:20,535 Chauncey: To your point about, the analogy too, going back to speed, 545 00:27:20,975 --> 00:27:24,795 a bigger team of individuals may or may not come to that answer faster. 546 00:27:24,825 --> 00:27:28,955 We can maybe

throw more people at that project, so you get that ability 547 00:27:28,965 --> 00:27:33,325 to get either a more cohesive answer or a faster answer, because of 548 00:27:33,325 --> 00:27:34,425 that ability to throw more at it. 549 00:27:34,435 --> 00:27:35,745 Is that kinda the mindset here? 550 00:27:37,843 --> 00:27:38,085 Maybe? 551 00:27:38,143 --> 00:27:38,823 Frankcx: Yeah, I would agree 552 00:27:38,924 --> 00:27:43,064 Jacob: I don't know if I believe the assertion fully, Robert, because 553 00:27:43,089 --> 00:27:44,529 Robert: not say it with enough conviction? 554 00:27:45,459 --> 00:27:46,799 Chauncey: Ding, ding, let's get it on 555 00:27:46,946 --> 00:27:51,076 Jacob: I feel like everyone's just trying to acquire users, and so it's a race 556 00:27:51,076 --> 00:27:54,596 to the bottom temporarily while they're trying to prove that their models or 557 00:27:54,596 --> 00:27:57,366 their values, is where the user should be. 558 00:27:57,576 --> 00:28:01,646 it's to the point a couple weeks ago, there's this Ox Alpha, like kind of 559 00:28:01,646 --> 00:28:05,296 like a stealth model on OpenRouter, so you can choose if you're, running 560 00:28:05,296 --> 00:28:09,726 a cloud model via OpenRouter, you can choose this Ox Alpha and it gives you a 561 00:28:09,766 --> 00:28:14,266 anonymous model that you don't know what it is, but it's free to run and use. 562 00:28:14,286 --> 00:28:17,276 So you're, using a model, you don't know what it is, you don't have any trust. 563 00:28:17,276 --> 00:28:20,206 It's in the cloud, it's not local, but, it's free to use. 564 00:28:20,466 --> 00:28:22,286 And so they're trying to acquire 565 00:28:23,092 --> 00:28:24,932 Robert: Model roulette is what that's called. 566 00:28:25,316 --> 00:28:25,926 Jacob: What's that? 567 00:28:26,226 --> 00:28:27,696 Robert: That's called model roulette. 568 00:28:27,996 --> 00:28:28,806 Jacob: Exactly. 569 00:28:28,842 --> 00:28:29,742 Robert: Good Lord 570 00:28:29,956 --> 00:28:34,406 Jacob: but everyone wants users, so, like, they're gonna continue lowering the cost, 571 00:28:34,416 --> 00:28:36,376 but I don't think that that's sustainable. 572 00:28:36,856 --> 00:28:39,896 I'm not into the economics of it with these companies, but, like, it just 573 00:28:39,896 --> 00:28:40,976 doesn't feel like that's sustainable. 574 00:28:40,976 --> 00:28:44,246 At some point, they're gonna, like, entrench that they've got their 575 00:28:44,246 --> 00:28:46,566 users and the prices go back up 576 00:28:48,445 --> 00:28:51,835 Neil: I'll offer a perspective in terms of speed. 577 00:28:52,225 --> 00:28:57,145 Robert, the argument that speed is going to be what these frontier labs 578 00:28:57,225 --> 00:29:03,525 offer, how does that shift when we get into asynchronous workflows? 579 00:29:03,525 --> 00:29:09,915 If I can craft my prompt to build the code I need overnight, why would I 580 00:29:09,935 --> 00:29:12,035 pay a premium for a frontier model? 581 00:29:12,045 --> 00:29:15,645 Why would I pay the outsource consultant, leveraging the example? 582 00:29:15,925 --> 00:29:21,585 Yes, maybe my MBAs take, much longer to execute that task, but if I can set 583 00:29:21,585 --> 00:29:26,405 them up proactively and achieve the desired result, certainly under some 584 00:29:26,405 --> 00:29:30,375 time constraint pressures, yeah, maybe I need to outsource and execute a sprint. 585 00:29:30,775 --> 00:29:33,925 Does speed become less of a factor when we start talking about, 586 00:29:34,205 --> 00:29:39,610 orchestration and autopilot workflows that can accomplish work when 587 00:29:39,910 --> 00:29:40,230 Robert: But is, 588 00:29:40,425 --> 00:29:41,135 Neil: lid is shut? 589 00:29:41,400 --> 00:29:44,430 Robert: is that a builder perspective or is that a general 590 00:29:44,430 --> 00:29:46,000 everyday user perspective? 591 00:29:47,561 --> 00:29:49,961 Neil: I would say right now it's a builder perspective. 592 00:29:50,051 --> 00:29:53,671 I'm seeing AI from a knowledge work perspective, a more one in 593 00:29:53,671 --> 00:29:57,121 one out, a more sophisticated Ask Jeeves, Google search, if you will. 594 00:29:57,351 --> 00:30:00,961 Once we start talking asynchronous workflows, which developers are, are 595 00:30:01,126 --> 00:30:01,416 Robert: Yeah. 596 00:30:01,536 --> 00:30:01,856 Yep 597 00:30:02,180 --> 00:30:02,540 Jacob: Yeah 598 00:30:02,641 --> 00:30:03,481 Neil: Right to that, 599 00:30:03,918 --> 00:30:04,168 Jacob: You 600 00:30:04,181 --> 00:30:06,761 Neil: think it's the developer workflows where speed doesn't matter. 601 00:30:06,761 --> 00:30:11,611 I'm used to crafting a prompt, letting my device cook for 10 minutes, 15 minutes. 602 00:30:11,801 --> 00:30:14,101 My benchmark test took an hour and a half this morning, 603 00:30:14,101 --> 00:30:14,351 Hmm Hmm 604 00:30:14,454 --> 00:30:15,134 acceptable. 605 00:30:15,174 --> 00:30:15,454 So, 606 00:30:15,842 --> 00:30:16,142 Jacob: Yeah. 607 00:30:16,142 --> 00:30:16,582 It's like, 608 00:30:16,744 --> 00:30:17,802 Neil: may not matter for developer 609 00:30:18,102 --> 00:30:21,582 Jacob: Frank, if you're tossing, one of those tasks like, you know, build an 610 00:30:21,602 --> 00:30:24,702 app, build whatever it is that you're building these days, like your four tokens 611 00:30:24,712 --> 00:30:28,812 per second example of that, 180 billion parameter model that you're running. 612 00:30:28,812 --> 00:30:31,572 Like, could theoretically just say, "Hey, I'm gonna be gone for the 613 00:30:31,572 --> 00:30:33,047 weekend," like build me something 614 00:30:33,347 --> 00:30:33,667 Frankcx: Right 615 00:30:34,072 --> 00:30:37,842 Jacob: come back and you've got something that's smarter and better and free so 616 00:30:38,061 --> 00:30:40,631 Frankcx: Especially these days when you're saying, like, you're kicking off 617 00:30:40,631 --> 00:30:44,931 seven

different Claude Costa sessions and you're like, "Well, just go to 618 00:30:45,231 --> 00:30:45,525 Robert: Fleet. 619 00:30:45,605 --> 00:30:46,915 You've got a fleet, yeah 620 00:30:47,011 --> 00:30:48,831 Frankcx: You're like, "Can you keep working on 621 00:30:49,059 --> 00:30:49,309 Robert: Does, 622 00:30:49,491 --> 00:30:50,021 Frankcx: please?" 623 00:30:50,129 --> 00:30:57,099 Robert: Does this equation change based on what we saw from NVIDIA and Hugging Face? 624 00:30:57,109 --> 00:31:02,249 Now, maybe not in that exact scenario, but thinking about 625 00:31:02,559 --> 00:31:09,449 consolidation in the industry does that change your on any of this? 626 00:31:10,173 --> 00:31:12,403 Chauncey: Wait, so just kind of rehashing that thought. 627 00:31:13,143 --> 00:31:18,733 Does NVIDIA potentially acquiring Hugging Face affect this speed versus like 628 00:31:18,773 --> 00:31:19,403 They're only gonna 629 00:31:19,413 --> 00:31:20,293 be building for speed 630 00:31:20,593 --> 00:31:23,131 Robert: Yeah, but not that specifically, but just in 631 00:31:23,131 --> 00:31:24,691 general industry consolidation. 632 00:31:24,711 --> 00:31:29,061 I mean, look at all the model providers we have, the hardware 633 00:31:29,061 --> 00:31:31,071 manufacturers, the data centers. 634 00:31:31,121 --> 00:31:36,961 Like for example, Microsoft makes models, build data centers, they provide services. 635 00:31:36,961 --> 00:31:39,181 I mean, are we gonna see more of that? 636 00:31:39,221 --> 00:31:42,311 And is that gonna affect your answer? 637 00:31:42,613 --> 00:31:44,943 Frankcx: there's some tools out there that are like these routers 638 00:31:44,963 --> 00:31:48,403 now that can essentially look at… They're like intelligence routers. 639 00:31:48,403 --> 00:31:51,173 They look at the question that's being asked, and then they determine what 640 00:31:51,173 --> 00:31:53,873 model or what tool, this can run on. 641 00:31:53,883 --> 00:31:58,203 And I feel like within an industry, if you had your own ability to route to 642 00:31:58,203 --> 00:32:02,043 local because you knew that job that you were requesting based on the prompt 643 00:32:02,063 --> 00:32:07,713 could be handled by a local AI or a cheaper model, then you would, right? 644 00:32:08,033 --> 00:32:12,823 So I feel like router, types of, intelligence is gonna be involving 645 00:32:12,823 --> 00:32:14,623 local AI in the future as well 646 00:32:15,269 --> 00:32:15,749 Robert: And does 647 00:32:15,959 --> 00:32:15,969 Chauncey: Hmm 648 00:32:15,999 --> 00:32:18,179 Robert: equation change, I'm gonna throw another curve ball 649 00:32:18,179 --> 00:32:20,179 here, with a three-tier model? 650 00:32:20,949 --> 00:32:26,029 If you have your device running a local model, you have a server a 651 00:32:26,029 --> 00:32:31,449 server or blade in your house that can run close to a frontier model, 652 00:32:32,259 --> 00:32:34,459 that change things even further? 653 00:32:36,883 --> 00:32:38,743 Chauncey: from a revenue perspective, like 654 00:32:38,863 --> 00:32:41,333 Robert: a, from a, what are they gonna charge for? 655 00:32:41,491 --> 00:32:41,851 Chauncey: Yeah. 656 00:32:42,071 --> 00:32:43,431 Yeah, well that's, that's, yeah. 657 00:32:44,193 --> 00:32:44,411 Yeah 658 00:32:44,493 --> 00:32:47,043 Robert: I could probably get some pretty serious speed off of it 659 00:32:47,946 --> 00:32:48,156 Jacob: You're, 660 00:32:48,405 --> 00:32:50,685 Chauncey: I mean it's, wonder even, like, how we ended up making 661 00:32:50,685 --> 00:32:55,825 a decision, if you think about even analogous to Office, right? 662 00:32:55,855 --> 00:33:00,565 At one point, we had Office running locally, and, everyone questioned why the 663 00:33:00,565 --> 00:33:05,675 heck we moved, we being Microsoft, moved to going to a cloud subscription-based 664 00:33:05,675 --> 00:33:09,355 model for, and now that that subscription model is king for everyone, right? 665 00:33:09,355 --> 00:33:10,835 That is not just, a Microsoft thing. 666 00:33:10,885 --> 00:33:12,665 Subscription is king across the board. 667 00:33:12,915 --> 00:33:15,605 And so how do we start looking at models in that same fashion? 668 00:33:15,605 --> 00:33:16,605 Are we reverting? 669 00:33:16,745 --> 00:33:17,365 I don't think so. 670 00:33:17,365 --> 00:33:21,495 I think there's still gonna be a play of individuals that are gonna want access to 671 00:33:21,495 --> 00:33:24,805 the newest, the latest, the greatest, the fastest of what the cloud has to offer 672 00:33:24,805 --> 00:33:27,635 and to that provides that scale, right? 673 00:33:27,685 --> 00:33:28,085 Frankcx: Yeah. 674 00:33:28,385 --> 00:33:29,328 Chauncey: where can the cloud grow? 675 00:33:29,328 --> 00:33:30,038 It can grow 676 00:33:31,675 --> 00:33:31,965 Frankcx: Yep. 677 00:33:32,635 --> 00:33:35,285 No, I think that's another show, because I feel like there's this 678 00:33:35,285 --> 00:33:39,075 whole idea of, like, butts in seats equals subscriptions times the number 679 00:33:39,075 --> 00:33:42,345 of people you have, and that doesn't equal an AI model when it comes to 680 00:33:42,524 --> 00:33:42,944 Chauncey: Right 681 00:33:43,045 --> 00:33:46,285 Frankcx: But, hey, let's, let's do a little vote around the horn. 682 00:33:46,345 --> 00:33:47,465 I'll jump in first. 683 00:33:47,785 --> 00:33:52,445 So I'm gonna support, Robert's, theory that the $25 token is, a 684 00:33:52,445 --> 00:33:56,075 thing of the past, especially since Grok and others are already leading. 685 00:33:56,135 --> 00:33:59,185 I don't know what that means towards the economy and all the investment 686 00:33:59,195 --> 00:34:03,235 banking that takes place around AI. 687 00:34:03,235 --> 00:34:05,345 Hopefully they can pay their bills, but, but that's the way I feel about it. 688 00:34:05,465 --> 00:34:05,835 Jacob? 689 00:34:07,131 --> 00:34:11,671

Jacob: I think that it's gonna get expensive again before it gets cheaper. 690 00:34:12,461 --> 00:34:17,221 Local is here, we all think it, but I think the market's not gonna be fully 691 00:34:17,221 --> 00:34:23,161 there and the costs are gonna go up again, especially as licenses, like open 692 00:34:23,161 --> 00:34:27,841 source licenses are in the mix of like what local models are able to do legally. 693 00:34:28,981 --> 00:34:32,921 I think that some of the Chinese AI firms are going to… Well, we're seeing them 694 00:34:32,941 --> 00:34:36,271 add more, and not as open of licensing. 695 00:34:36,591 --> 00:34:41,241 So you're gonna see more commercial customers want to-- they'll be paying 696 00:34:41,241 --> 00:34:44,471 more for it because it's in way higher demand before it goes down. 697 00:34:44,811 --> 00:34:47,181 So I don't agree with you yet, Robert, yet. 698 00:34:48,167 --> 00:34:48,277 Frankcx: All 699 00:34:48,441 --> 00:34:48,941 Robert: On the side 700 00:34:48,997 --> 00:34:49,737 Frankcx: Chauncey? 701 00:34:50,062 --> 00:34:51,912 Chauncey: Yeah, I'm on the same boat as Jacob. 702 00:34:52,232 --> 00:34:55,332 Demand's gonna be high enough to continue to maintain a higher price for now. 703 00:34:55,452 --> 00:34:56,362 Robert: Must be in marketing 704 00:34:57,710 --> 00:34:58,620 Frankcx: Yeah, they want the 705 00:34:58,920 --> 00:34:59,862 Chauncey: Or, yeah. 706 00:35:00,300 --> 00:35:01,150 Frankcx: And then, Neil 707 00:35:03,781 --> 00:35:06,481 Neil: Speed matters when we're talking security. 708 00:35:06,981 --> 00:35:10,851 If we have bad actors using these frontier models, we need access to 709 00:35:10,851 --> 00:35:12,671 these frontier models to combat. 710 00:35:12,871 --> 00:35:17,081 I'm going to agree with the stance that frontier models deliver speed. 711 00:35:18,371 --> 00:35:20,301 The use cases for speed vary. 712 00:35:20,321 --> 00:35:22,931 I think there will continue to be a premium for frontier models 713 00:35:22,941 --> 00:35:27,151 specifically to combat rising threat actors leveraging the same models, us 714 00:35:27,987 --> 00:35:28,967 Jacob: And people will pay for that, all right? 715 00:35:29,318 --> 00:35:31,558 Frankcx: All right, Robert, you get the last word. 716 00:35:31,598 --> 00:35:31,968 Robert: I'm 717 00:35:31,972 --> 00:35:32,452 Frankcx: know where your 718 00:35:32,558 --> 00:35:33,378 Robert: I'm holding firm. 719 00:35:33,378 --> 00:35:37,688 Now, what we might see is different tiers of speed you pay for. 720 00:35:37,888 --> 00:35:41,658 I think regardless of where it goes, and I do think speed is gonna 721 00:35:41,658 --> 00:35:44,088 be, the first, shoe to drop here. 722 00:35:44,458 --> 00:35:49,208 Over the next six to 12 months, I think we're gonna see some significant changes 723 00:35:49,218 --> 00:35:53,268 in how this looks, is good for us 'cause it gives us stuff to talk about 724 00:35:54,470 --> 00:35:54,710 Frankcx: Yep. 725 00:35:55,440 --> 00:35:57,350 right, let's, wrap up with on my device. 726 00:35:57,350 --> 00:35:59,540 Anybody have anything cool that they're working on? 727 00:35:59,632 --> 00:36:00,482 Robert: Here's the short story. 728 00:36:00,782 --> 00:36:01,272 Frankcx: problem 729 00:36:01,572 --> 00:36:06,962 Robert: week I talked about my, struggle to try and find out how we can get, 730 00:36:07,012 --> 00:36:09,762 on-device coding, a code agent running. 731 00:36:10,122 --> 00:36:14,392 And I of tripped into it or backed into it by mistake. 732 00:36:14,732 --> 00:36:17,932 The short version of the story is I loaded up a twenty-seven billion 733 00:36:17,932 --> 00:36:22,802 parameter model, connected it in, thank you Jacob, into Copilot CLI. 734 00:36:23,502 --> 00:36:27,162 I typed in, "Let's make something," and then I got distracted. 735 00:36:27,652 --> 00:36:31,772 I turned on autopilot, which essentially gave it full, permission to do 736 00:36:31,772 --> 00:36:34,132 whatever it wanted, and I walked away. 737 00:36:34,132 --> 00:36:37,272 And I came back, and it said, " Well, you didn't respond in ten minutes, so 738 00:36:37,272 --> 00:36:42,902 I built you something." And it built an online coding agent, and it said, " You 739 00:36:42,902 --> 00:36:47,592 should use a-- this MoE thirty billion parameter three point eight Qwen 740 00:36:47,592 --> 00:36:50,882 model." and, and the rest is history. 741 00:36:51,282 --> 00:36:52,536 There's a lot more to it, but it 742 00:36:52,836 --> 00:36:53,776 Chauncey: Oh, that's awesome. 743 00:36:53,876 --> 00:36:57,546 Frankcx: All right, gentlemen, I hope you all have a terrific weekend. 744 00:36:57,556 --> 00:37:05,036 So is no place like 127.0.0.1. 745 00:37:05,100 --> 00:37:05,820 Robert: hang with you fellas 746 00:37:05,886 --> 00:37:06,106 Frankcx: now 747 00:37:06,343 --> 00:37:06,703 Chauncey: See you. 748 00:37:07,120 --> 00:37:07,620 Jacob: Thanks, team 749 00:37:07,683 --> 00:37:08,093 Chauncey: Thanks guys