1 00:00:00,440 --> 00:00:02,640 This episode is brought to you by Forethought. 2 00:00:03,520 --> 00:00:06,200 Most companies build their customer experience in pieces. 3 00:00:06,560 --> 00:00:09,480 Sales in one system, support in another, and onboarding 4 00:00:09,480 --> 00:00:11,640 somewhere else. Forethought brings it all 5 00:00:11,640 --> 00:00:14,560 together. Forethought is an AI system 6 00:00:14,560 --> 00:00:17,680 made-up of advanced agents that handle sales, onboarding, 7 00:00:17,680 --> 00:00:21,160 support, and retention. Each team manages its own 8 00:00:21,160 --> 00:00:23,480 agents. The customer sees one unified 9 00:00:23,480 --> 00:00:26,360 experience. Forethought powers over a 10 00:00:26,360 --> 00:00:29,720 billion interactions every month for brands like Scale AI, 11 00:00:29,720 --> 00:00:34,640 Cohere, Air Table, and Upwork. Learn more at Forethought dot 12 00:00:34,640 --> 00:00:41,680 AI. Hey, it's Eric newcomer. 13 00:00:41,680 --> 00:00:44,880 Welcome to the Cerebral Valley podcast. 14 00:00:44,880 --> 00:00:50,560 Our occasional detour from the newcomer podcast with me are are 15 00:00:50,560 --> 00:00:55,520 now three time hosts Max Child and James Wilsterman, the Co 16 00:00:55,520 --> 00:00:58,920 founders of Volley and the Co hosts of the Struble Valley AI 17 00:00:58,920 --> 00:01:00,840 Summit. Welcome to the podcast, guys. 18 00:01:00,840 --> 00:01:04,160 Thank you, glad to be here. Eric, happy to be back, excited 19 00:01:04,160 --> 00:01:06,160 about the conference coming up in London. 20 00:01:06,280 --> 00:01:08,480 Yeah. So we, you know, always sort of 21 00:01:08,480 --> 00:01:11,680 jump into the Struble Valley podcast ahead of our Tribal 22 00:01:11,680 --> 00:01:15,360 Valley AI summits. And we've got one June 25th in 23 00:01:15,360 --> 00:01:17,680 London. People are like, oh, isn't it 24 00:01:17,680 --> 00:01:19,400 hard to do an international conference? 25 00:01:19,400 --> 00:01:22,960 And like, like everything in startups, it's easiest to do 26 00:01:22,960 --> 00:01:25,320 hard things when you underestimate how difficult it 27 00:01:25,320 --> 00:01:27,520 is until you're like, oh, yeah, we're just doing it. 28 00:01:27,840 --> 00:01:31,360 Someone is figuring out taxes. We have an event team there. 29 00:01:31,680 --> 00:01:34,680 I saw we have the ticket prices in pounds on the website so I. 30 00:01:34,880 --> 00:01:37,240 Think exactly. Like someone was on top of that. 31 00:01:37,320 --> 00:01:40,720 So a couple fun things for this episode for the long time 32 00:01:40,720 --> 00:01:43,080 listeners. At the end of the podcast, we 33 00:01:43,080 --> 00:01:48,640 will return to the startup draft, perhaps my favorite part 34 00:01:48,640 --> 00:01:51,720 of the show. We have overtime accumulated 35 00:01:51,720 --> 00:01:55,560 startups in our imaginary portfolio and get to score keep 36 00:01:55,840 --> 00:01:58,080 how we're doing working backwards. 37 00:01:58,080 --> 00:02:01,520 Before that, we're going to dig into some of our predictions 38 00:02:01,560 --> 00:02:04,200 from the last series. This is sort of a mid year 39 00:02:04,360 --> 00:02:07,520 snapshot, right guys, in terms of the predictions that we made? 40 00:02:07,680 --> 00:02:10,000 Yeah. We made about 10 predictions 41 00:02:10,000 --> 00:02:14,680 last year in November with the idea that they would come to 42 00:02:14,680 --> 00:02:17,640 fruition within a year. So this is the mid year check 43 00:02:17,640 --> 00:02:22,280 in, no official scorekeeping needed, but we will obviously be 44 00:02:22,280 --> 00:02:25,080 competitive on our mid year check insurance as well. 45 00:02:25,320 --> 00:02:28,560 In the latter half of the show, we'll be doing our, our games. 46 00:02:28,800 --> 00:02:32,200 But I just wanted to start off, you know, we've been doing this. 47 00:02:32,480 --> 00:02:36,160 We launched the first Cerebral Valley in March 2023. 48 00:02:36,480 --> 00:02:40,040 ChatGPT had just come out and everyone was getting their heads 49 00:02:40,040 --> 00:02:42,880 around models Max and James, if you want to sort of walk me 50 00:02:42,880 --> 00:02:46,600 through, how do you think of sort of the thematic evolution 51 00:02:46,880 --> 00:02:48,320 of AI in that period? And where? 52 00:02:48,320 --> 00:02:51,640 Where are we now as we're programming for Cerebral Valley 53 00:02:51,640 --> 00:02:55,360 London? Last time we chatted we it was 54 00:02:55,360 --> 00:02:57,680 the end of last year, which feels like a long time ago in 55 00:02:57,680 --> 00:02:59,200 the AI world. But we were all having this 56 00:02:59,200 --> 00:03:02,200 discussion of if we were hitting a scaling wall and if the models 57 00:03:02,200 --> 00:03:05,560 would stop getting better. And I would say it feels like 58 00:03:05,560 --> 00:03:08,520 every two weeks since then, something amazing has happened 59 00:03:08,520 --> 00:03:11,040 in AI. Is pretty trivial. 60 00:03:11,600 --> 00:03:13,840 Has been a great leap forward. I actually think the theme of 61 00:03:13,840 --> 00:03:17,360 hitting the scaling wall aged really poorly, which we somewhat 62 00:03:17,360 --> 00:03:19,440 predicted at the time. But you know, maybe we didn't 63 00:03:19,440 --> 00:03:21,880 know how wrong we were. I mean, just some highlights. 64 00:03:21,880 --> 00:03:25,600 We had, you know, GBT 01, the sort of first thinking model. 65 00:03:25,600 --> 00:03:29,160 We had O3, which has brought thinking models into the tool 66 00:03:29,160 --> 00:03:31,000 world. We've had Gemini really take 67 00:03:31,000 --> 00:03:33,720 great leaps forward and kind of become a state-of-the-art model 68 00:03:33,720 --> 00:03:36,160 system. We just had Claude 4-3 weeks 69 00:03:36,160 --> 00:03:37,880 ago, which a lot of people consider the greatest model 70 00:03:37,880 --> 00:03:41,080 built so far, maybe the best coding model ever made. 71 00:03:41,360 --> 00:03:44,920 We've had incredible Leafs in image generation from mid 72 00:03:44,920 --> 00:03:49,280 journey, Gemini Flux, a bunch of other folks, and then of course 73 00:03:49,280 --> 00:03:52,160 we're going to this episode 2. Video generation has also been 74 00:03:52,160 --> 00:03:55,080 unbelievable. VO3 in particular, which is 75 00:03:55,080 --> 00:03:59,080 Googles new video model I think is that the first sort of truly 76 00:03:59,080 --> 00:04:02,320 realistic seeming video generation model in my opinion 77 00:04:02,320 --> 00:04:04,040 and and kind of a great leap forward there. 78 00:04:04,040 --> 00:04:08,480 You know, Alexander Wang at our conference in November was 79 00:04:08,480 --> 00:04:12,520 probably the most prominent person arguing for a scaling 80 00:04:12,520 --> 00:04:16,040 wall or a potential issue. I mean, he was talking his own 81 00:04:16,040 --> 00:04:17,279 book. You know, he's in the post 82 00:04:17,279 --> 00:04:20,440 training business. I do think there's an argument 83 00:04:20,440 --> 00:04:23,640 that a lot of the progress, some of the progress, some of the 84 00:04:23,640 --> 00:04:26,560 progress we've seen has been about post training. 85 00:04:26,600 --> 00:04:30,200 The models are a certain level of smart, but then they sort of 86 00:04:30,440 --> 00:04:33,280 talk to each other, they're corrected in certain ways and 87 00:04:33,280 --> 00:04:35,160 that's where they get more intelligence. 88 00:04:35,160 --> 00:04:40,480 That seems to be the improvement we're seeing with O3 and the 89 00:04:40,560 --> 00:04:41,800 Chain of Thought models. Yeah. 90 00:04:41,800 --> 00:04:44,600 I mean, I think, I think Alexander Wang, I give him a lot 91 00:04:44,600 --> 00:04:47,280 of credit for what he was saying at our conference in November 92 00:04:47,280 --> 00:04:51,760 because he was arguing that maybe we would see a scaling 93 00:04:51,760 --> 00:04:55,400 wall to some degree in pre training, but we wouldn't see 94 00:04:55,400 --> 00:04:59,000 performance levelling at all. And I think that's exactly 95 00:04:59,000 --> 00:05:00,920 what's happened. And to your point, like a lot of 96 00:05:00,920 --> 00:05:05,360 the gains from thinking models have come from new post training 97 00:05:05,360 --> 00:05:08,640 techniques, reinforcement learning and new a whole new 98 00:05:08,640 --> 00:05:12,080 scaling paradigm I guess. Yeah, I mean, in the end, the 99 00:05:12,080 --> 00:05:13,920 models have gotten really, really effing good. 100 00:05:14,040 --> 00:05:16,880 And whether it's pre training or post training, they go good. 101 00:05:16,920 --> 00:05:17,840 Yeah. Yeah. 102 00:05:17,840 --> 00:05:20,200 I feel like the vibe was a little more like maybe this is 103 00:05:20,200 --> 00:05:22,880 the end of AI progress, not just like pre training is over. 104 00:05:22,880 --> 00:05:25,000 And to your point, there was a distinction drawn by some of the 105 00:05:25,000 --> 00:05:27,080 people at the conference, but I think there was a maybe a more 106 00:05:27,080 --> 00:05:30,240 negative tenor going forward from a lot of folks on stage. 107 00:05:30,240 --> 00:05:33,440 And instead, I think it's been maybe the craziest 6 months in 108 00:05:33,440 --> 00:05:34,600 the history of AI. I don't know. 109 00:05:34,600 --> 00:05:37,240 It's felt like that to me. The interesting thing is, I 110 00:05:37,240 --> 00:05:40,760 guess like in January, we had that DeepSeek moment, right? 111 00:05:40,760 --> 00:05:42,040 We haven't talked about that. I forgot. 112 00:05:42,040 --> 00:05:42,920 About DeepSeek. Too. 113 00:05:42,960 --> 00:05:43,720 Yeah, yeah. Wow. 114 00:05:44,920 --> 00:05:48,520 So to some degree that was maybe echoing some of these scaling 115 00:05:48,600 --> 00:05:54,200 wall concerns because if you can get such high performance out 116 00:05:54,200 --> 00:05:58,360 of, you know, open source models that are effectively competing 117 00:05:58,360 --> 00:06:01,440 with the frontier U.S. companies, like maybe there is 118 00:06:01,680 --> 00:06:04,880 an argument there. But then I guess, you know, 119 00:06:05,200 --> 00:06:09,320 that's kind of died down a bit as we've seen both anthropic and 120 00:06:09,320 --> 00:06:12,400 open AI come out with superior models to some degree. 121 00:06:12,680 --> 00:06:16,520 What models are you guys using? Applaud for Gemini 2/5 for 122 00:06:16,520 --> 00:06:19,440 coding and then for my own personal sort of research and 123 00:06:19,680 --> 00:06:21,240 note taking and stuff. Probably O3. 124 00:06:21,400 --> 00:06:23,840 That's the same for me. I've been using a lot of deep 125 00:06:23,840 --> 00:06:27,240 research through ChatGPT. I think that's my favorite 126 00:06:27,240 --> 00:06:29,680 product maybe of the last few months. 127 00:06:29,680 --> 00:06:32,200 I don't have time. I use O3 a lot. 128 00:06:32,200 --> 00:06:34,840 I use deep research some. I don't want to blow anyone up 129 00:06:34,840 --> 00:06:37,160 here, but I I had deep research, right? 130 00:06:37,160 --> 00:06:40,000 A whole like political consulting memo for somebody I 131 00:06:40,000 --> 00:06:42,240 was trying to get to run for office. 132 00:06:42,280 --> 00:06:44,000 You know, it's like it's amazing. 133 00:06:44,040 --> 00:06:47,320 I mean, there is a world where I would have gone and paid a 134 00:06:47,320 --> 00:06:50,520 consultant to be like draw out for me, like when races will be 135 00:06:50,520 --> 00:06:51,920 available and when, when they could run. 136 00:06:51,920 --> 00:06:54,840 And it's like you just like, oh, in the morning you're having a 137 00:06:54,840 --> 00:06:57,520 manic fit and you're like, oh, let's, let's see what Chachi BT 138 00:06:57,520 --> 00:06:59,040 can do. It's it's insane. 139 00:06:59,160 --> 00:07:02,040 It's crazy. This is a very nerdy, lame use 140 00:07:02,040 --> 00:07:06,720 case, but I like to buy cheap wine that is still good. 141 00:07:06,800 --> 00:07:10,520 And there was a secret wine from a local provider and they said 142 00:07:10,520 --> 00:07:14,120 it's this anonymous secret wine that we're selling for 1/5 of 143 00:07:14,120 --> 00:07:16,160 market price. But you know, if you bought it 144 00:07:16,160 --> 00:07:19,520 at market price, it would be a 250, three, $100 wine. 145 00:07:19,800 --> 00:07:22,240 And they were like, it comes from these amazing vineyards and 146 00:07:22,240 --> 00:07:24,640 the West side of Napa and the foothills, blah, blah, blah, 147 00:07:24,640 --> 00:07:26,280 blah, blah. And I just pasted the 148 00:07:26,280 --> 00:07:28,680 description into deep research and I was like, figure out what 149 00:07:28,680 --> 00:07:32,280 the secret wine is like 20 and 20 minutes later comes back. 150 00:07:32,280 --> 00:07:37,720 It was like, obviously this is like this BV Latour 2022 cab or 151 00:07:37,720 --> 00:07:40,720 whatever, and I was like, what? I've been super impressed with 152 00:07:40,920 --> 00:07:45,120 ChatGPT multimodal for shopping. The three of us are all going to 153 00:07:45,120 --> 00:07:49,600 the same wedding in France in a month or so, and I don't know if 154 00:07:49,600 --> 00:07:53,200 you guys have looked at the required attire, but they're 155 00:07:53,240 --> 00:07:55,080 half of. It Oh my God, my wife is very 156 00:07:55,080 --> 00:07:58,400 concerned. Yeah, this hot tip is paste that 157 00:07:58,400 --> 00:08:00,880 whole thing in a ChatGPT, ask it to shop for you. 158 00:08:00,880 --> 00:08:02,400 It'll be great. My wife and I. 159 00:08:03,200 --> 00:08:05,640 This was a good idea. I mean, it's not perfect, I 160 00:08:05,640 --> 00:08:07,520 won't lie. Like how like maybe one out of 161 00:08:07,720 --> 00:08:11,240 10 suggestions are like completely off base, but the 162 00:08:11,240 --> 00:08:14,080 shopping integration, it just, it's kind of showing where 163 00:08:14,080 --> 00:08:15,600 things are headed like, you know? 164 00:08:15,600 --> 00:08:19,000 What do you mean integration? Well, because it's actually like 165 00:08:19,640 --> 00:08:23,520 searching the web, you know, doing an agentic workflow of, of 166 00:08:23,520 --> 00:08:26,240 looking for these items and then it's pulling that information 167 00:08:26,240 --> 00:08:29,680 back into Chachi BT in line in the chat. 168 00:08:29,840 --> 00:08:31,720 There's links that link out, right? 169 00:08:31,760 --> 00:08:34,280 Eventually I'm sure you'll be able to just like add that to 170 00:08:34,280 --> 00:08:37,919 your cart within Chachi BT It's a full shopping experience. 171 00:08:37,919 --> 00:08:40,840 It's not just researching. Yeah, interesting. 172 00:08:40,919 --> 00:08:43,360 He used the word. He used the word agentic. 173 00:08:43,360 --> 00:08:44,080 Yeah. OK. 174 00:08:44,400 --> 00:08:46,480 I mean, yeah, we're going to talk a lot, you know, over the 175 00:08:46,480 --> 00:08:50,880 next couple episodes in terms of how AI can get put to use. 176 00:08:51,080 --> 00:08:55,000 The topic I really wanted to get into this week before reviewing 177 00:08:55,160 --> 00:08:58,000 our games and our scorekeeping is agents. 178 00:08:58,000 --> 00:09:02,320 Like I feel like agents have been at once sort of the 179 00:09:02,320 --> 00:09:06,600 buzziest thing in the backdrop of a couple events, but and I 180 00:09:06,600 --> 00:09:10,560 never quite here. And so I guess the first direct 181 00:09:10,560 --> 00:09:15,000 question I want to ask is, is deep research an agent? 182 00:09:15,080 --> 00:09:17,440 Is O3 an agent? Like what, what? 183 00:09:17,440 --> 00:09:21,160 What is an agent these days if it's just delivering you a 184 00:09:21,160 --> 00:09:25,280 report? I buy the agent definition that 185 00:09:25,280 --> 00:09:29,520 an agent is, you know, an AI tool that can actually do stuff 186 00:09:29,520 --> 00:09:31,520 for you, right? That can go through some sort of 187 00:09:31,520 --> 00:09:36,380 series of steps involving actions and you know, quote UN 188 00:09:36,380 --> 00:09:39,920 quote tool use is sort of one of the popular phrases these days 189 00:09:39,920 --> 00:09:41,720 of, you know, using different tools. 190 00:09:42,120 --> 00:09:45,120 Currently the only tools these agents can use really are 191 00:09:45,320 --> 00:09:49,960 essentially like web search and you know, maybe pulling shopping 192 00:09:49,960 --> 00:09:53,040 links and showing pictures to James's discussion. 193 00:09:53,280 --> 00:09:55,640 But I still fundamentally think like there is a big difference 194 00:09:55,640 --> 00:10:00,040 between, you know, ChatGPT of 6 to 12 months ago where right, 195 00:10:00,120 --> 00:10:01,960 you know, you ask it a question, it gives you an answer. 196 00:10:01,960 --> 00:10:05,400 Essentially it's you know it's a text in text out engine, right? 197 00:10:05,400 --> 00:10:07,200 You put text in one side, it gives you text out the other 198 00:10:07,200 --> 00:10:09,280 side, but it doesn't go do stuff that's. 199 00:10:09,280 --> 00:10:12,520 Part of that kind of web it's. Like, yeah, yeah, yeah, exactly 200 00:10:12,520 --> 00:10:14,320 right. Even the deep research that we 201 00:10:14,320 --> 00:10:16,480 were talking about, I mean, I used to be a management 202 00:10:16,520 --> 00:10:21,760 consultant for two years, 2 horrible years, and it can do 203 00:10:21,800 --> 00:10:25,520 much better versions of what I did as a management consultant, 204 00:10:26,040 --> 00:10:29,600 you know, in a matter of minutes, it could do 5 days of 205 00:10:29,880 --> 00:10:32,720 world class management consultant level research on a 206 00:10:32,720 --> 00:10:35,560 topic, right? And I guess if you don't think 207 00:10:35,560 --> 00:10:40,080 that's agentic, I think you're a little crazy, like you're too 208 00:10:40,080 --> 00:10:41,960 high, right? It's clearly running around. 209 00:10:42,040 --> 00:10:46,320 I I've seen some people use like time is it's like if it takes 210 00:10:46,320 --> 00:10:48,840 time, how long it takes, you know, if it's all going and 211 00:10:48,840 --> 00:10:51,440 doing things and interacting with the world and coming back, 212 00:10:51,760 --> 00:10:54,280 I want to throw down a gauntlet to me. 213 00:10:55,320 --> 00:10:58,320 And this could come soon. We'll we'll be in the world of 214 00:10:58,320 --> 00:11:02,760 agents once people are letting them run wild with their own 215 00:11:02,880 --> 00:11:06,440 credit cards once, once agents are spending money without a 216 00:11:06,440 --> 00:11:10,280 human check in, that's when we've got sort of real agents. 217 00:11:10,280 --> 00:11:11,080 What do you think? So does it. 218 00:11:11,080 --> 00:11:15,480 Not count in your mind if like my agent finds me a dope pair of 219 00:11:15,480 --> 00:11:18,160 shoes and, you know, text me, hey, can I buy this? 220 00:11:18,160 --> 00:11:20,680 And I'm like, yeah, go for it. No, it needs to transact. 221 00:11:20,920 --> 00:11:22,800 It needs. It needs to do without you. 222 00:11:22,920 --> 00:11:25,880 Oh, truly? It'll be like this. 223 00:11:25,880 --> 00:11:27,880 My great. Thing that's like that's an 224 00:11:28,040 --> 00:11:30,240 agent. I mean, I understand like you 225 00:11:30,240 --> 00:11:33,280 know, you, you book flights and an agent would have like asked 226 00:11:33,280 --> 00:11:36,000 you before that. Definition I I feel like if it 227 00:11:36,000 --> 00:11:38,560 does a restaurant reservation, maybe no money changes hands. 228 00:11:38,560 --> 00:11:40,440 That's still an agent. I mean, I I just think this. 229 00:11:40,440 --> 00:11:43,240 Is it's definitely an agent. I I'm just saying that's like a 230 00:11:43,240 --> 00:11:46,240 great employee solves the problem, right? 231 00:11:46,240 --> 00:11:48,560 It's not like, oh, they come back to you and want all this 232 00:11:48,560 --> 00:11:50,840 feedback. It's like do the thing like pull 233 00:11:50,840 --> 00:11:53,480 the trigger like once we can trust to do that. 234 00:11:53,520 --> 00:11:56,280 I'm just saying that would be a landmark moment that I don't 235 00:11:56,280 --> 00:11:58,600 think is so far away. Or do you think that is far 236 00:11:58,600 --> 00:12:01,080 away? No, I, I honestly think that if, 237 00:12:01,400 --> 00:12:05,040 if you had the capability to do that already in ChatGPT, like 238 00:12:05,040 --> 00:12:07,400 James did his clothing research, probably some people would be 239 00:12:07,400 --> 00:12:09,480 like, yeah, fine, go do it. Like don't buy too many things 240 00:12:09,480 --> 00:12:11,600 without asking me. But like, I think people would 241 00:12:11,600 --> 00:12:13,560 already be down. I honestly think the company's 242 00:12:13,640 --> 00:12:16,280 reluctance is probably just mostly like they don't want it 243 00:12:16,280 --> 00:12:18,880 to go off the rails and spend, you know, thousands of dollars 244 00:12:18,880 --> 00:12:21,800 of people's money or whatever. But it's it's possible already 245 00:12:21,800 --> 00:12:23,600 in my opinion. I don't know if it would work 246 00:12:23,600 --> 00:12:25,200 yet. Like, and I think a lot of this 247 00:12:25,200 --> 00:12:29,040 ties into context how much Chachi BT knows about me. 248 00:12:29,040 --> 00:12:32,000 I mean it it's obviously, you know, starting to build that 249 00:12:32,000 --> 00:12:34,920 memory, but I don't think it knows enough without me 250 00:12:34,920 --> 00:12:38,400 prompting it or, you know, having very targeted list of 251 00:12:38,400 --> 00:12:41,920 shopping ideas for this wedding to go just start buying me 252 00:12:41,920 --> 00:12:43,840 stuff. I that'd be super interesting 253 00:12:43,840 --> 00:12:46,480 once we get there, but I don't think it's ready for that yet. 254 00:12:47,000 --> 00:12:48,760 Max just was sort of getting at this. 255 00:12:49,040 --> 00:12:51,320 I mean, stop limits I think are key, right? 256 00:12:51,360 --> 00:12:54,000 I mean, there's a degree to which algos, you know, I was 257 00:12:54,000 --> 00:12:55,880 talking to a former banker about this the other day. 258 00:12:55,880 --> 00:12:59,440 Like, you know, it's not crazy that we would let a machine, a 259 00:12:59,440 --> 00:13:02,320 computer make payment decisions on its own. 260 00:13:02,320 --> 00:13:05,440 Traders do it all the time. You just create some limits and 261 00:13:05,440 --> 00:13:07,720 checks and hopefully have people hovering. 262 00:13:07,720 --> 00:13:10,640 But like the algorithms have to move before a person could 263 00:13:10,640 --> 00:13:13,520 react. And so that's happening there. 264 00:13:13,720 --> 00:13:16,880 And so you can see with LLMS, it's like, OK, you build up a 265 00:13:17,000 --> 00:13:19,480 trust up to the, you know, $500 limit. 266 00:13:19,480 --> 00:13:22,400 And you're like, you can, yeah, it'll be interesting. 267 00:13:22,600 --> 00:13:25,560 I mean, how many chargebacks are our chargebacks can go up with 268 00:13:25,560 --> 00:13:27,640 everybody releasing power to agents? 269 00:13:27,640 --> 00:13:29,800 And you're like, oh man, we need to do more refunds when the 270 00:13:29,800 --> 00:13:33,240 agents do something crazy. I mean, I think that kind of 271 00:13:33,240 --> 00:13:36,000 gets into some of this like quote UN quote agentic web 272 00:13:36,000 --> 00:13:38,280 discussion, which is like, can we redesign the web in a way 273 00:13:38,280 --> 00:13:39,880 that enables more of this behavior? 274 00:13:39,880 --> 00:13:43,080 Because I do actually think if you're a clothing retailer, 275 00:13:43,480 --> 00:13:47,960 there probably is some kind of business model in which you let 276 00:13:47,960 --> 00:13:50,240 people buy way too many clothes and then they send most of them 277 00:13:50,240 --> 00:13:52,600 back or they cancel most of them before they ship or whatever. 278 00:13:52,600 --> 00:13:54,560 Right? Like, which obviously is along 279 00:13:54,560 --> 00:13:57,640 the lines of like a Stitch Fix or Trunk Club or some of these 280 00:13:57,640 --> 00:13:59,960 other companies which, you know, weren't wildly successful 281 00:13:59,960 --> 00:14:03,360 because I think the return fees are pretty punitive or, you 282 00:14:03,360 --> 00:14:05,640 know, and you end up with a lot of fraud and, you know, damaged 283 00:14:05,640 --> 00:14:07,560 clothes and stuff like that. But I do think there's an 284 00:14:07,560 --> 00:14:10,320 interesting question of like, could you build some sort of 285 00:14:10,320 --> 00:14:13,960 consumer commerce website where people's agents can like buy way 286 00:14:13,960 --> 00:14:17,760 too much stuff, like basically and then cancel it or return it 287 00:14:17,760 --> 00:14:20,680 or limit it or some in some way because people buying too much 288 00:14:20,680 --> 00:14:22,680 stuff by accident. And it's it's probably good if 289 00:14:22,680 --> 00:14:25,200 you're selling things like even if you have to find a way to let 290 00:14:25,200 --> 00:14:27,920 them cancel it or return it. My first reaction to what you 291 00:14:27,920 --> 00:14:31,320 were saying is, oh, this is like a workaround to allow agents to 292 00:14:31,320 --> 00:14:34,760 spend when we're really going to unwind it, but it makes the 293 00:14:34,760 --> 00:14:37,960 default spending versus not. It's like, oh, the sugar is 294 00:14:38,160 --> 00:14:40,000 good, which is which is really good for if. 295 00:14:40,000 --> 00:14:41,960 You're selling stuff. Yeah, yeah, yeah, yeah. 296 00:14:42,160 --> 00:14:44,840 James, what's the agent use case you're most excited about? 297 00:14:45,240 --> 00:14:47,560 I can't ignore coding. We haven't talked about coding, 298 00:14:47,560 --> 00:14:51,480 which I think is like the most actually valuable the. 299 00:14:51,520 --> 00:14:52,880 Real one. That's why we want to talk about 300 00:14:52,880 --> 00:14:53,720 it's boring. It's happening. 301 00:14:54,400 --> 00:14:56,160 And that's. Something actually works? 302 00:14:56,480 --> 00:15:00,960 Oh yeah, yeah, yeah, OK. AI podcasts are only about the 303 00:15:00,960 --> 00:15:04,400 future, you know. Yeah, I think, well, I just 304 00:15:04,400 --> 00:15:07,440 think that there's a lot to unpack about the future of the 305 00:15:07,440 --> 00:15:11,240 of coding. I mean, I am a CTOI code when I 306 00:15:11,240 --> 00:15:16,760 can and this agentic world has dramatically changed what I can 307 00:15:16,760 --> 00:15:21,680 do in terms of prototyping and participating in the coding at 308 00:15:21,680 --> 00:15:24,160 Volley and learning faster, right? 309 00:15:24,160 --> 00:15:27,440 I mean, you just learn so much faster about different 310 00:15:27,440 --> 00:15:31,000 technologies and tech stacks. And yeah, I think it's like if 311 00:15:31,000 --> 00:15:33,840 you're not an engineer, you maybe understand this a little 312 00:15:33,840 --> 00:15:36,240 bit or you've played around with things that are a little bit 313 00:15:36,240 --> 00:15:37,840 more. No code like lovable. 314 00:15:37,840 --> 00:15:39,840 Or figma or. Figma right? 315 00:15:39,840 --> 00:15:41,920 I mean I. Think both Lovable and Figma are 316 00:15:41,920 --> 00:15:45,360 speaking our event and they are now competitors in the the no 317 00:15:45,360 --> 00:15:48,560 code world. Yeah, both non coders and 318 00:15:48,560 --> 00:15:52,760 coders, if you've dabbled with any of this, you are, you know, 319 00:15:52,960 --> 00:15:55,800 receiving the future of like what could happen to all types 320 00:15:55,800 --> 00:15:59,000 of computer work, white collar work, if you want to call it, 321 00:15:59,000 --> 00:16:04,360 but is happening first in in the engineering space and it's it's 322 00:16:04,360 --> 00:16:06,800 pretty remarkable. I think Nat Friedman had a 323 00:16:06,800 --> 00:16:09,520 really good analogy that I can't get out of my head about agent 324 00:16:09,520 --> 00:16:11,880 tick coding. And he talks about the idea that 325 00:16:11,880 --> 00:16:16,120 you have, you know, you have this room full of interns who 326 00:16:16,120 --> 00:16:18,440 are all like junior engineers basically, right? 327 00:16:18,720 --> 00:16:22,520 And you assign each of them a task and they go off and try to 328 00:16:22,520 --> 00:16:24,720 do it. And then when they get stuck, 329 00:16:24,720 --> 00:16:27,080 they raise their hand and they say, hey, I need help here, 330 00:16:27,080 --> 00:16:29,320 come, come help me. I'm, I'm stuck with this bug or 331 00:16:29,320 --> 00:16:31,920 or this this, you know, issue the, the app's not working, 332 00:16:31,920 --> 00:16:35,680 whatever. And he podcasted, I think about 333 00:16:35,680 --> 00:16:37,360 six months ago and he was like, right now basically these 334 00:16:37,360 --> 00:16:40,160 interns raise their hands like every 5 minutes in like human 335 00:16:40,160 --> 00:16:42,160 time. Like so they do like 5 minutes 336 00:16:42,160 --> 00:16:43,720 of work and they raise their hand and you've got to go help 337 00:16:43,720 --> 00:16:44,760 them. And then they raise their hand 5 338 00:16:44,760 --> 00:16:46,480 minutes later and you know, over and over and over again, he's 339 00:16:46,480 --> 00:16:48,120 like, so you can't really have met that many of these, you 340 00:16:48,120 --> 00:16:50,920 know, imaginary interns going because you can just become, be 341 00:16:50,920 --> 00:16:52,600 running around fixing their problems all the time. 342 00:16:52,920 --> 00:16:54,840 You know, maybe only one really like because you're just 343 00:16:54,840 --> 00:16:56,160 constantly having to give them feedback. 344 00:16:56,760 --> 00:17:00,280 I do think like the frontier for like how much quote UN quote 345 00:17:00,280 --> 00:17:04,440 human work these agentic coding tools can do now without you 346 00:17:04,440 --> 00:17:06,400 having to run over and help them when they raise their hand is 347 00:17:06,400 --> 00:17:10,400 like probably somewhere in the like 15 to 30 minute range now 348 00:17:10,400 --> 00:17:13,400 where like, you know, they obviously come back to you very 349 00:17:13,400 --> 00:17:14,839 quickly. Like they iterate through their 350 00:17:14,839 --> 00:17:17,800 work very fast because they can type at a superhuman speed, 351 00:17:17,800 --> 00:17:18,760 right? They can put out hundreds of 352 00:17:18,760 --> 00:17:21,000 lines of code. But the sort of amount of human 353 00:17:21,000 --> 00:17:24,119 level work I would say in my testing that they get stuck is 354 00:17:24,119 --> 00:17:26,800 probably like, you know, some of that 15 to 30 minute range. 355 00:17:26,960 --> 00:17:29,440 I'm confused, like are you saying 15 to 30 minutes of like 356 00:17:29,440 --> 00:17:31,320 what it would take for a human intern? 357 00:17:31,320 --> 00:17:34,360 What it would take like a good human engineer like, you know, 358 00:17:34,480 --> 00:17:36,640 Yeah, I got it. You know, mid to senior software 359 00:17:36,640 --> 00:17:38,640 engineer just just hammering away at code, right? 360 00:17:38,720 --> 00:17:39,920 Yeah. Like how much code do you get 361 00:17:39,920 --> 00:17:41,880 out between them getting stuck? Right. 362 00:17:42,000 --> 00:17:43,880 Like, yeah. I would say yeah, maybe half an 363 00:17:43,880 --> 00:17:45,400 hour of like human, human work, but. 364 00:17:45,400 --> 00:17:47,320 You get it. You get it in 3 minutes or two 365 00:17:47,320 --> 00:17:49,200 minutes you get. It in a minute, you know, or 30 366 00:17:49,200 --> 00:17:51,560 seconds usually, which which is which is mind boggling. 367 00:17:51,720 --> 00:17:55,480 But like the dream is you get, you know, 4 hours of human human 368 00:17:55,480 --> 00:17:58,760 work or 8 hours of human work or eventually, you know, weeks or 369 00:17:58,760 --> 00:18:00,920 just starts fixing itself and you never go in there, right? 370 00:18:01,240 --> 00:18:05,400 I mean, you know, I spent a fair bit of time vibe coding a month 371 00:18:05,400 --> 00:18:09,360 or so ago. I do think there are some self 372 00:18:09,360 --> 00:18:15,760 driving car aspects in the sense that like the last 5% or 10% of 373 00:18:15,760 --> 00:18:19,320 a problem is very important. And like it's like, oh, it looks 374 00:18:19,320 --> 00:18:21,320 close. It looks close, but it's like, 375 00:18:21,320 --> 00:18:25,640 sure, it's really good when I'm like copying the sub stack 376 00:18:25,640 --> 00:18:28,440 design, dropping a lovable like, oh, rebuild that. 377 00:18:28,440 --> 00:18:31,400 But like at some point I feel like you've just been working 378 00:18:31,400 --> 00:18:35,000 long enough and you have some minor tweak you want to make and 379 00:18:35,000 --> 00:18:37,320 it just starts getting stuck and has no idea. 380 00:18:37,320 --> 00:18:41,280 I mean, partially that's why I'm trying to do no code and no code 381 00:18:41,680 --> 00:18:44,800 knowledge. But I do think these programs 382 00:18:44,800 --> 00:18:47,880 are really good at being enticing in the beginning, and 383 00:18:47,880 --> 00:18:50,240 then they sort of get overwhelmed as the project 384 00:18:50,240 --> 00:18:53,600 starts to expand. Maybe this project that you were 385 00:18:53,600 --> 00:18:56,840 doing, you won't be able to ship because of that last five mile 386 00:18:56,840 --> 00:18:59,440 problem or whatever. The fact that you even started 387 00:18:59,440 --> 00:19:02,760 it is kind of amazing that you did that. 388 00:19:02,760 --> 00:19:05,360 Like you wouldn't have been able to even start the project three 389 00:19:05,360 --> 00:19:06,720 months ago or six months ago or something. 390 00:19:06,720 --> 00:19:07,440 Sure. Yeah. 391 00:19:07,440 --> 00:19:10,200 So I don't know, it's just kind of interesting, like where that 392 00:19:10,200 --> 00:19:12,720 heads is. You know, I think eventually you 393 00:19:12,720 --> 00:19:15,160 get to the point where it's a finished product and then 394 00:19:15,200 --> 00:19:17,680 there's just like way more coding projects happening in the 395 00:19:17,680 --> 00:19:19,880 world, right? And then, you know, I think 396 00:19:19,880 --> 00:19:22,640 within companies that already have an engineering workforce 397 00:19:22,640 --> 00:19:25,680 like it actually can get across that that last mile. 398 00:19:25,960 --> 00:19:30,600 I think another important theme that always comes up is like if 399 00:19:30,600 --> 00:19:34,960 we froze this moment in time, how much value is there that 400 00:19:34,960 --> 00:19:40,120 people are still understanding versus depending on continued 401 00:19:40,120 --> 00:19:42,800 progress in the models that's that's revolutionary. 402 00:19:42,800 --> 00:19:46,000 Where are you guys on how much harvesting could be done with 403 00:19:46,000 --> 00:19:49,520 what would already exist versus what we're waiting on? 404 00:19:50,360 --> 00:19:53,960 I would say a lot like a lot of harvested progress. 405 00:19:54,440 --> 00:19:56,720 Like I can't even think like trillions. 406 00:19:56,720 --> 00:19:58,680 I don't know. I haven't decided of 10s of 407 00:19:58,680 --> 00:20:00,200 trillions, but let's say trillions. 408 00:20:00,440 --> 00:20:01,680 What's the measures? In value. 409 00:20:01,960 --> 00:20:05,760 Oh, like GDP. GDP value of just harvesting the 410 00:20:05,760 --> 00:20:09,120 stuff that we've done so far. But it's amazing you still see 411 00:20:09,120 --> 00:20:12,000 people who are like, skeptic. I mean, I don't know. 412 00:20:12,000 --> 00:20:14,760 I just don't want to be, you know, deluding myself here. 413 00:20:15,080 --> 00:20:16,360 Yeah. I mean, I think like 414 00:20:16,360 --> 00:20:19,400 fundamentally we've all had the experience, right? 415 00:20:19,400 --> 00:20:25,320 Where something that used to take 8 hours now takes less than 416 00:20:25,320 --> 00:20:27,640 a minute, right? I mean, like we've all had that 417 00:20:27,640 --> 00:20:29,320 experience, right? And so if you just do the math 418 00:20:29,320 --> 00:20:31,920 on that, you say, OK, I took something that used to be a 500 419 00:20:31,920 --> 00:20:34,520 minute problem and now it's a one minute problem. 420 00:20:34,920 --> 00:20:38,240 Unless you believe that thing had no value whatsoever, how 421 00:20:38,240 --> 00:20:41,600 could you not see a 500X in knowledge work in front of you 422 00:20:41,600 --> 00:20:45,040 and say, hey, maybe we haven't harvested all the value on this 423 00:20:45,040 --> 00:20:47,800 500X improvement, though this thing I just did. 424 00:20:48,000 --> 00:20:50,640 How could you not believe that that's going to have insane 425 00:20:50,640 --> 00:20:53,640 ramifications throughout our world and and insane amounts of 426 00:20:53,640 --> 00:20:55,520 value that eventually can be created. 427 00:20:55,520 --> 00:20:58,320 You know, so I don't know. I just to me, I just look at 428 00:20:58,320 --> 00:21:00,840 that, you know, 2 orders of magnitude differential we're 429 00:21:00,840 --> 00:21:03,800 already seeing on labor and think it's got to be huge. 430 00:21:04,000 --> 00:21:06,080 Here's a question for you guys. Like, I have a friend who's in 431 00:21:06,080 --> 00:21:09,560 PE and he, you know, doesn't really, he's not really in the 432 00:21:09,560 --> 00:21:13,280 tech world that much, but he, you know, tried chat GPD didn't 433 00:21:13,280 --> 00:21:15,320 really, you know, find it super valuable. 434 00:21:15,520 --> 00:21:19,160 He had never tried O3. And I just showed him what was 435 00:21:19,160 --> 00:21:21,960 possible and he, like, you know, kind of changed his whole 436 00:21:21,960 --> 00:21:24,440 opinion of what types of companies he should buy in the 437 00:21:24,440 --> 00:21:24,920 future. Wow. 438 00:21:26,600 --> 00:21:29,360 Well, I do think that's a real danger that there are, there are 439 00:21:29,360 --> 00:21:33,120 things that I'm changing about my life and medical decisions 440 00:21:33,120 --> 00:21:37,400 and lots of stuff off O3 BS. It's so persuasive that we'll 441 00:21:37,400 --> 00:21:39,840 never really be able to back out the psychological. 442 00:21:39,840 --> 00:21:41,880 And, you know, it's all right. It's like, is it having an 443 00:21:41,880 --> 00:21:43,360 effect? It's like, yeah, it's deeply 444 00:21:43,360 --> 00:21:45,920 affecting big decisions in my life. 445 00:21:46,640 --> 00:21:49,080 You know, just because it's like the thought partner, it's there 446 00:21:49,080 --> 00:21:51,600 just like, you know, if you were to have the friend that's there 447 00:21:51,600 --> 00:21:54,360 while you're soundboarding an idea, like that's going to be 448 00:21:54,360 --> 00:21:56,200 dramatic. And I think just the fact that 449 00:21:56,200 --> 00:21:58,480 it's in the loop. Wouldn't you say overall that 450 00:21:58,480 --> 00:21:59,840 it's in person giving good advice? 451 00:21:59,840 --> 00:22:01,120 That's why I keep it in the loop. 452 00:22:01,120 --> 00:22:03,800 Yeah, right. It's not like, but but it will 453 00:22:03,800 --> 00:22:06,480 be hard to know if it if it hallucinated. 454 00:22:06,560 --> 00:22:08,440 I don't know. It's like, is it gonna? 455 00:22:08,480 --> 00:22:10,880 Yeah, hopefully in a couple of years I'll chase after me in 456 00:22:10,880 --> 00:22:13,080 this shot. GPT will be. 457 00:22:13,160 --> 00:22:15,360 I told you a couple of key things that I think you took to 458 00:22:15,360 --> 00:22:18,120 heart and and now that I'm a little smarter, I'm gonna go 459 00:22:18,120 --> 00:22:20,320 back on that. Yeah, we'll see. 460 00:22:20,440 --> 00:22:23,600 Having lived through the iPhone experience of of having the 461 00:22:23,600 --> 00:22:27,360 first iPhone in 2007 and having to spend the next four years 462 00:22:27,360 --> 00:22:30,120 explaining to my friends why they had to get an iPhone and 463 00:22:30,120 --> 00:22:31,440 why they. Should probably. 464 00:22:31,840 --> 00:22:34,280 Consider getting an iPhone. I mean James, you had a 465 00:22:34,280 --> 00:22:36,440 BlackBerry until what, 2011 or something like that? 466 00:22:36,440 --> 00:22:42,520 Like 10/20/11 like. James, I'm sorry, I was the 467 00:22:42,600 --> 00:22:44,600 lager to. I had a BlackBerry for a long 468 00:22:44,640 --> 00:22:45,960 time. Yeah, and people would be like, 469 00:22:45,960 --> 00:22:47,200 well, Max, what? Why? 470 00:22:47,200 --> 00:22:49,160 Why do you, I need an iPhone. What do you do with an iPhone? 471 00:22:49,240 --> 00:22:51,320 And I'm like, well, you can like browse the Internet. 472 00:22:51,320 --> 00:22:53,360 And they're like, well, you know, I browse the Internet on 473 00:22:53,360 --> 00:22:54,160 my phone. It's fine. 474 00:22:54,360 --> 00:22:56,000 And I'd be like, well, these apps are pretty cool. 475 00:22:56,000 --> 00:22:58,280 And they'd be like what? Like the the app that makes it 476 00:22:58,280 --> 00:22:59,760 look like you're drinking a beer and I'm. 477 00:22:59,760 --> 00:23:04,160 Like, well, I'm not. Using that every day, but just 478 00:23:04,160 --> 00:23:05,920 the last point I have to make on this BlackBerry thing. 479 00:23:05,920 --> 00:23:09,120 I'm sorry, BlackBerry, which obviously is a dead company 480 00:23:09,120 --> 00:23:10,000 today. No one you know has a 481 00:23:10,000 --> 00:23:14,640 BlackBerry, right? Their sales continue to grow for 482 00:23:14,640 --> 00:23:17,280 four years after the iPhone was released right? 483 00:23:17,520 --> 00:23:21,880 So O 708-0910 eleven BlackBerry sales continue to increase even 484 00:23:21,880 --> 00:23:25,360 though we all today look back and even in the movie about 485 00:23:25,360 --> 00:23:27,960 BlackBerry, they pretended that the day the iPhone came out it 486 00:23:28,280 --> 00:23:31,600 was like P BlackBerry. Sorry guys, you missed the 487 00:23:31,600 --> 00:23:34,080 future. And I'm just pointing out that 488 00:23:34,080 --> 00:23:37,200 in the real world, sales continue to grow for the next 4 489 00:23:37,200 --> 00:23:38,840 years. And so I'm just saying that when 490 00:23:38,840 --> 00:23:42,360 you're in one of these sort of Roadrunner, Wiley Coyote chases 491 00:23:42,360 --> 00:23:45,560 the Roadrunner over the ledge moments where you take a second 492 00:23:45,560 --> 00:23:47,640 before you look down and realize that gravity is pulling you 493 00:23:47,640 --> 00:23:50,240 there. I think similarly with AI, we're 494 00:23:50,240 --> 00:23:52,720 all like, holy crap, we just went off the ledge. 495 00:23:52,720 --> 00:23:55,640 Like shit is about to get really dramatically different here. 496 00:23:56,000 --> 00:24:00,360 And you can still stand there in the air for a year, two years or 497 00:24:00,360 --> 00:24:03,800 three years or four years before gravity really hits, right. 498 00:24:03,800 --> 00:24:07,600 And and I think to your initial question, like we are in that 499 00:24:07,600 --> 00:24:11,960 period where even if nothing changed, like we're already off 500 00:24:11,960 --> 00:24:14,600 the ledge, you know, stuff, stuff is going to be 501 00:24:14,600 --> 00:24:16,280 dramatically different. No matter what happens from 502 00:24:16,280 --> 00:24:18,840 here, which I know we all believe, you know, the models 503 00:24:18,840 --> 00:24:19,960 are going to continue to get better. 504 00:24:20,160 --> 00:24:22,440 So the pace of acceleration is going to be even higher. 505 00:24:22,680 --> 00:24:27,000 It's a perfect endnote. Let's move the conversation to 506 00:24:27,000 --> 00:24:30,280 our predictions from six months ago. 507 00:24:30,280 --> 00:24:33,480 Basically, we're not going to get 2 in the weeds. 508 00:24:33,520 --> 00:24:37,160 It was a fun discussion. James, do you want to take us 509 00:24:37,160 --> 00:24:40,560 through them, What the question was, where we each landed and 510 00:24:40,560 --> 00:24:43,640 then we'll give a quick reaction and then go to the next one. 511 00:24:43,840 --> 00:24:44,800 Sure. Sounds good. 512 00:24:45,280 --> 00:24:49,880 Just to clarify also we back in November, we asked Claude and 513 00:24:49,880 --> 00:24:53,560 Chachi BT to generate these predictions for us including 514 00:24:53,560 --> 00:24:57,080 providing probability estimates of how likely they were. 515 00:24:57,080 --> 00:24:59,920 And then so our job was to take the over or the under on each 516 00:24:59,920 --> 00:25:02,720 prediction. The first one was open AI shifts 517 00:25:02,720 --> 00:25:07,640 GPT 5 with a greater than 10 trillion parameter model Max was 518 00:25:07,640 --> 00:25:10,120 the over, Eric under and I took the over. 519 00:25:10,520 --> 00:25:11,480 Thoughts. Where is it? 520 00:25:12,160 --> 00:25:13,840 GPT 5 here. Yeah, so far. 521 00:25:14,000 --> 00:25:16,360 So far I'm correct. Six months. 522 00:25:16,640 --> 00:25:20,000 I mean, I think we much hung up on they wouldn't use the name or 523 00:25:20,000 --> 00:25:22,920 maybe there was a high chance that they would abandoned the 524 00:25:22,920 --> 00:25:24,800 name. It seems like they're going to 525 00:25:24,800 --> 00:25:27,640 do it honestly right now. What the vibes are that they're 526 00:25:27,640 --> 00:25:32,240 going to kill the O series and just make 5 the overall? 527 00:25:32,400 --> 00:25:35,720 I feel super good about the name and it's shipping this year. 528 00:25:35,720 --> 00:25:37,440 At this point. I think the only thing that 529 00:25:37,440 --> 00:25:40,200 could hit the under would be the 10 trillion parameters because 530 00:25:40,200 --> 00:25:42,960 didn't Brad, like at the CEO literally say they were going to 531 00:25:42,960 --> 00:25:45,720 ship GPT 5 this year and it was going to be called GPT 5? 532 00:25:45,760 --> 00:25:47,240 Like I'm pretty sure he said that like last. 533 00:25:47,240 --> 00:25:50,240 Seems like a. So I would still smash the over 534 00:25:50,240 --> 00:25:52,280 the 10 trillion parameter. Thing is, I guess the one. 535 00:25:52,480 --> 00:25:54,280 Certainly I would buy the over now. 536 00:25:55,040 --> 00:25:57,760 I might take the under because of that 10 trillion like I oh. 537 00:25:59,040 --> 00:26:01,880 You're I'm sticking to my bet. Whatever I we're the best locked 538 00:26:01,880 --> 00:26:02,840 in. Yeah, I can't be wrong. 539 00:26:03,240 --> 00:26:06,920 I was trying to figure out like so Claude 4 Opus just launched 540 00:26:06,920 --> 00:26:10,880 and there's no comment on the parameters that I can find, but 541 00:26:11,160 --> 00:26:15,080 the best estimate I could get from Claude was 2 to 3 trillion, 542 00:26:15,480 --> 00:26:18,560 so I don't know if that's nice. Nice. 543 00:26:18,840 --> 00:26:22,520 I mean a lot of the improvements seem to be post training and 544 00:26:22,520 --> 00:26:24,040 chain. Of thought and if they're. 545 00:26:24,040 --> 00:26:28,160 Going to merge, you know, the O series with GPD 5. 546 00:26:28,160 --> 00:26:30,480 It's partially because they think they need part of the 547 00:26:30,480 --> 00:26:33,840 improvement off these reasoning models, and so maybe it's a sign 548 00:26:33,840 --> 00:26:35,760 that the parameters aren't getting as big as we thought. 549 00:26:36,160 --> 00:26:39,120 Also also it seems like adding that many parameters you know 550 00:26:39,400 --> 00:26:42,520 creates huge issues with inference cost and just serving 551 00:26:42,520 --> 00:26:43,920 them it's. Expensive. 552 00:26:44,280 --> 00:26:48,120 Yeah. Anyway, the next one was three 553 00:26:48,120 --> 00:26:52,520 or more countries enact national rules regarding AI medical 554 00:26:52,520 --> 00:26:56,480 diagnosis. Max, you took the under, Eric 555 00:26:56,480 --> 00:26:58,800 took the over, and I took the over. 556 00:26:58,800 --> 00:27:02,360 We had it at a 70% probability. You guys are crazy. 557 00:27:02,360 --> 00:27:04,680 You guys are crazy. There's no way you think in the 558 00:27:04,680 --> 00:27:08,000 next 6 months three countries are gonna enact regulations. 559 00:27:08,000 --> 00:27:09,800 That happened so far, medical diagnosis. 560 00:27:10,440 --> 00:27:14,080 Here's what I had. Oh, you. 561 00:27:14,080 --> 00:27:16,000 Found something chachi BT is saying. 562 00:27:16,000 --> 00:27:19,000 Yes, EU AI Act. Exactly, Yeah. 563 00:27:19,440 --> 00:27:24,640 UKSMHRFUS to regs online chachi BT things we're like in good 564 00:27:24,640 --> 00:27:25,440 shape all. Right. 565 00:27:25,440 --> 00:27:28,160 Well you took 70% so never forget that I get 2 to one. 566 00:27:28,160 --> 00:27:29,280 Yeah. I don't know how you're going to 567 00:27:29,280 --> 00:27:31,200 do the overall math at the end, but I'm sure whatever. 568 00:27:31,880 --> 00:27:34,040 Onward and upward. All right, Tesla full staff 569 00:27:34,040 --> 00:27:37,960 driving approved for unsupervised driving in one or 570 00:27:37,960 --> 00:27:43,800 more US state, 40% probability. Max had the over, Eric had the 571 00:27:43,800 --> 00:27:46,880 over, and I had the over. I think we're in good shape 572 00:27:46,880 --> 00:27:49,560 there, right? Isn't Texas gonna happen like 573 00:27:49,560 --> 00:27:52,440 yeah in June, supposedly. Like, yeah, in two weeks, 574 00:27:52,520 --> 00:27:54,520 correct? But it does seem like right? 575 00:27:54,520 --> 00:27:57,760 Chachibi says not yet Pilot cars on private roads in Texas. 576 00:27:57,760 --> 00:28:00,480 No public permit. It's coming to Austin in June. 577 00:28:00,560 --> 00:28:02,240 Now. I think this gets into the 578 00:28:02,240 --> 00:28:03,920 letter of the prediction, though. 579 00:28:04,600 --> 00:28:08,600 Is it a state law or city law, whatever. 580 00:28:08,600 --> 00:28:10,360 OK, we'll come back to that in a few months. 581 00:28:11,040 --> 00:28:16,920 AI will write the copy for greater than 50% of a major news 582 00:28:16,920 --> 00:28:22,600 outlets articles. 30% chance Max had the over, Eric had the over 583 00:28:22,600 --> 00:28:27,080 and I took the under. I just feel like my, my under is 584 00:28:27,080 --> 00:28:28,360 great here. I mean, nobody's. 585 00:28:28,360 --> 00:28:30,240 Nobody's. Claiming this nobody's first 586 00:28:30,240 --> 00:28:32,000 newsroom. Everybody is like, what's the? 587 00:28:32,720 --> 00:28:34,600 Eric's gonna do this by the end of the year. 588 00:28:35,080 --> 00:28:37,000 Just to go build. A new media company. 589 00:28:37,000 --> 00:28:39,600 I won't be major though. I don't think I'll I'll qualify. 590 00:28:40,000 --> 00:28:41,400 You could be major by the end of the year. 591 00:28:41,400 --> 00:28:43,800 Why not? Apparently, the Chicago Sun 592 00:28:43,800 --> 00:28:47,400 Times inadvertently ran an AI generated book list filled with 593 00:28:47,400 --> 00:28:50,800 errors, sparking backlash. Clearly not paying for 03 there 594 00:28:50,800 --> 00:28:52,600 I think. I just think we're gonna see so 595 00:28:52,600 --> 00:28:54,640 many of these things, like hallucinations, like, aren't we 596 00:28:54,640 --> 00:28:56,400 already seeing this from the Trump administration? 597 00:28:56,400 --> 00:28:59,760 Like just random things that don't make any sense. 598 00:29:00,080 --> 00:29:03,040 I certainly don't want, you know, AI to replace the 599 00:29:03,040 --> 00:29:05,560 newsrooms. I'm just expecting a lot of like 600 00:29:05,560 --> 00:29:09,320 stories over the next year about like academic papers and news 601 00:29:09,320 --> 00:29:11,600 articles just having obvious hallucinations, right? 602 00:29:11,600 --> 00:29:13,280 Because people are going to be using. 603 00:29:13,560 --> 00:29:15,600 But shouldn't we be more disturbed that the Trump 604 00:29:15,600 --> 00:29:18,840 administration is like, I feel like if Democrats were in charge 605 00:29:18,920 --> 00:29:21,080 and they were releasing government reports that were 606 00:29:21,080 --> 00:29:23,960 clearly written by AI, yeah, it would be like the biggest 607 00:29:23,960 --> 00:29:26,360 cultural story of the moment. It's like, well, the governments 608 00:29:26,360 --> 00:29:29,120 already phoning it in with Trump, it's like, oh, at least 609 00:29:29,120 --> 00:29:31,880 they're using it is. Literally what I was going to 610 00:29:31,880 --> 00:29:33,960 say. I was like, I would honestly 611 00:29:33,960 --> 00:29:37,920 prefer the AI to be doing this than the miscellaneous Trump 612 00:29:37,920 --> 00:29:40,480 administration employee. Like knows nothing about 613 00:29:40,480 --> 00:29:42,640 anything. I just want Sam Altman to give 614 00:29:42,640 --> 00:29:46,080 every administration staffer free O3 so that they have O3 615 00:29:46,080 --> 00:29:47,880 right. Reasonably intelligent fake 616 00:29:47,880 --> 00:29:51,200 reports instead of real ones. I do think this core take that 617 00:29:51,200 --> 00:29:55,160 the Internet split everybody's view and this is part of why 618 00:29:55,280 --> 00:29:59,080 originally Marc Andreessen was so pro crypto and anti AI. 619 00:29:59,200 --> 00:30:02,920 I do think AI is going to bring us potentially closer together 620 00:30:02,920 --> 00:30:05,880 where people are asking rock like is this bullshit true or 621 00:30:05,880 --> 00:30:07,840 false? I mean, it's possible people 622 00:30:07,840 --> 00:30:11,080 build crazier models, but for now, while the models are sort 623 00:30:11,080 --> 00:30:14,200 of generally in agreement with each other about how the world 624 00:30:14,200 --> 00:30:19,200 works, it could be a major force for cultural consensus over the 625 00:30:19,200 --> 00:30:22,760 next couple decades. It's basically the network TV of 626 00:30:22,760 --> 00:30:25,600 the Internet essentially, right? We're all watching the three 627 00:30:25,600 --> 00:30:28,960 major channels and they all broadcast kind of like exactly, 628 00:30:29,040 --> 00:30:32,800 but not family friendly content. And your take is that Marc 629 00:30:32,800 --> 00:30:35,160 Andreessen? Marc Andreessen doesn't want to 630 00:30:35,160 --> 00:30:37,560 bring us together. He was weird, crazy shit. 631 00:30:37,560 --> 00:30:39,040 Yeah, that's what crypto is all about. 632 00:30:39,040 --> 00:30:42,200 Like, do whatever you want, like, and AI is a source of 633 00:30:42,200 --> 00:30:44,040 conformity. But then Andreessen Horowitz 634 00:30:44,040 --> 00:30:47,240 basically saw what was winning, and it was like, we not need to 635 00:30:47,240 --> 00:30:49,840 go where the money is. And, you know, that's how I see 636 00:30:49,840 --> 00:30:52,120 the story playing out. I mean, they were resistant at 637 00:30:52,120 --> 00:30:55,360 first. Khosla led the open AI venture 638 00:30:55,360 --> 00:30:56,640 round. Anyway, keep going. 639 00:30:56,720 --> 00:31:00,200 Didn't Peter Thiel say like AI was communist or something? 640 00:31:00,200 --> 00:31:02,400 Yeah. Yeah, communist and crypto was 641 00:31:02,400 --> 00:31:03,800 libertarian or whatever. Right. 642 00:31:04,000 --> 00:31:08,280 OK, Fully AI scripted and AI rendered feature film gets a 643 00:31:08,280 --> 00:31:13,000 theatrical release 25% chance. Max with the under, Eric with 644 00:31:13,000 --> 00:31:15,520 the over and James with the under. 645 00:31:15,840 --> 00:31:19,480 This was this was Eric using his insider knowledge of pay to play 646 00:31:19,480 --> 00:31:23,160 tactics within the movie industry to to try to grab us on 647 00:31:23,160 --> 00:31:25,920 this that someones going to pay a theater chain to take their 648 00:31:25,920 --> 00:31:28,120 movie even if it sucks. Basically, make sense? 649 00:31:28,320 --> 00:31:30,800 I mean, we still have time. I had to double check this that 650 00:31:30,800 --> 00:31:35,560 it wasn't hallucination, but according to chat GVT there is a 651 00:31:35,800 --> 00:31:40,720 fully AI generated movie releasing soon, Pirate Queen 652 00:31:40,760 --> 00:31:44,120 Zheng Yi Sao, billed as the world's first fully AI generated 653 00:31:44,120 --> 00:31:45,160 feature film. Exactly. 654 00:31:45,240 --> 00:31:46,560 That's going to get a festival run. 655 00:31:46,720 --> 00:31:49,920 I gotta watch it. All right, moving on, more than 656 00:31:49,920 --> 00:31:55,560 three major smartphone OEMs chip phones with AI Co processors 657 00:31:55,560 --> 00:32:01,080 running 7 billion parameter plus models on device. 7 billion is 658 00:32:01,080 --> 00:32:03,480 pretty high. The rumor mill on Apple is 659 00:32:03,480 --> 00:32:05,560 saying they're going to be able to run 3 to 4 billion. 660 00:32:06,000 --> 00:32:09,360 So even if you believe in James's claim that the chip 661 00:32:09,360 --> 00:32:12,200 they've had in there for 12 years is an AI chip, it still 662 00:32:12,200 --> 00:32:14,480 might not be able to do a 7 billion parameter model. 663 00:32:14,760 --> 00:32:18,080 So we were specifying an Apple, Samsung, Google, Chami, 664 00:32:18,720 --> 00:32:23,680 dedicated AI Co processors that run 7 billion perimeter LLMS 665 00:32:23,680 --> 00:32:29,160 fully locally. We had the probability at 75% to 666 00:32:29,160 --> 00:32:32,960 unders Max and Eric and I took the over and I yeah, I was 667 00:32:33,520 --> 00:32:38,640 counting the the current technology as capable of running 668 00:32:38,640 --> 00:32:40,720 those types of models. But yeah, to your point, Max, I 669 00:32:40,720 --> 00:32:43,120 think 7 billion might be a bit high, right? 670 00:32:44,000 --> 00:32:47,200 3 to 4 is what the rumor mills saying for this year, but we'll 671 00:32:47,200 --> 00:32:48,200 know in a few more months I guess. 672 00:32:48,400 --> 00:32:51,440 Next. Anthropic releases a model 673 00:32:51,440 --> 00:32:57,720 scoring over 90% on U Bar, the unified benchmark for AI 674 00:32:57,720 --> 00:33:02,440 reasoning, which does not exist according to our own research. 675 00:33:02,640 --> 00:33:07,320 During the podcast recording last year, Complete Hallucinated 676 00:33:07,800 --> 00:33:10,960 prediction from Claude. How are you feeling about that 677 00:33:10,960 --> 00:33:13,360 one guys? Great. 678 00:33:13,800 --> 00:33:16,480 I should, yeah. Certainly raises red flags, 679 00:33:16,560 --> 00:33:22,360 yeah. Moving on to #8 / 4 Fortune 500 680 00:33:22,360 --> 00:33:28,280 firms 5 or more cut greater than 25% of their middle management 681 00:33:28,280 --> 00:33:34,800 roles by the end of this year, crediting AI explicitly with a 682 00:33:34,800 --> 00:33:38,320 25% probability. Actually interesting. 683 00:33:38,320 --> 00:33:44,120 Here we have an over and over from Max and over from Eric and 684 00:33:44,120 --> 00:33:47,320 on under from myself. I'm feeling pretty good about 685 00:33:47,320 --> 00:33:49,160 the under. Yeah, yeah, look, I'm. 686 00:33:49,160 --> 00:33:52,120 Gonna do the research on this. We're only asking for five firms 687 00:33:52,120 --> 00:33:54,680 to cut 25% of only middle management. 688 00:33:54,680 --> 00:33:56,440 So that's a pretty that's a low bar. 689 00:33:56,640 --> 00:33:59,160 According to my research with ChatGPT, this has not 690 00:33:59,160 --> 00:34:01,800 materialized. Many large companies are 691 00:34:01,800 --> 00:34:05,240 experimenting with AI and none have reported cutting 1/4 of 692 00:34:05,280 --> 00:34:08,880 their management. Yeah, maybe we overestimated, 693 00:34:08,880 --> 00:34:11,560 first of all how much we thought it would give them air cover for 694 00:34:11,560 --> 00:34:13,719 all sorts of things. But yes, exactly. 695 00:34:13,719 --> 00:34:17,000 Maybe we didn't want to get right into the AI narrative. 696 00:34:17,360 --> 00:34:19,760 If we got a tariff induced recession, this actually might 697 00:34:19,760 --> 00:34:22,760 happen so. OK #9 Deepmine and Google 698 00:34:22,760 --> 00:34:27,000 discover a new drug that clears phase one trials within 2025. 699 00:34:27,719 --> 00:34:33,320 We gave that a 20% probability and Max took the under, Eric 700 00:34:33,320 --> 00:34:35,480 took the under, and I took the under. 701 00:34:35,480 --> 00:34:38,000 I think all looking pretty strong here. 702 00:34:38,040 --> 00:34:42,280 And D mine spun out isomorphic, which would be the start up. 703 00:34:42,280 --> 00:34:43,880 I think that would potentially do this. 704 00:34:43,880 --> 00:34:47,040 So there's a there's a chance that even if it happens, we can 705 00:34:47,040 --> 00:34:49,320 all claim technicality that it doesn't. 706 00:34:49,320 --> 00:34:52,800 But I I think it's, it's not looking likely, right? 707 00:34:52,800 --> 00:34:54,920 They have to be in phase one trials already. 708 00:34:55,159 --> 00:34:56,320 Yeah, I mean I. Think we? 709 00:34:56,320 --> 00:34:59,760 Yeah, I think that they are planning to enter trials by the 710 00:34:59,760 --> 00:35:04,080 end of this year, so unlikely to have completed phase one trial. 711 00:35:04,600 --> 00:35:09,360 And lastly, we have the international AI treaty with 712 00:35:09,360 --> 00:35:13,800 greater than or equal to 15 signatories, including three of 713 00:35:13,800 --> 00:35:19,320 the US, China, EU and UK. 50% probability we all took the 714 00:35:19,320 --> 00:35:22,200 under. Seems like a good bet so far. 715 00:35:22,320 --> 00:35:25,440 Yeah, the AI believes too much in human institutions. 716 00:35:25,640 --> 00:35:29,120 Right, because US, China both seem unlikely, right? 717 00:35:29,360 --> 00:35:35,600 Is Europe getting EUUK viable getting the US which under Trump 718 00:35:35,600 --> 00:35:40,040 is now like no AI regulation and China which is we do what we 719 00:35:40,040 --> 00:35:42,320 want. The fact that the EU and UK get 720 00:35:42,320 --> 00:35:46,720 separate credit here doing a lot of work, but three of the four 721 00:35:46,720 --> 00:35:48,280 seems high do. You know, do you know what the 722 00:35:48,280 --> 00:35:50,600 score is? So just pulling together the 723 00:35:50,600 --> 00:35:54,520 scores, I asked Chachi PT to create a scoring system. 724 00:35:54,560 --> 00:35:59,040 It came out with a formula inspired by Breyer style scoring 725 00:35:59,040 --> 00:36:01,920 which I had never heard of but. Yeah, that's how close you are 726 00:36:01,920 --> 00:36:03,280 to the probability. Yeah exactly. 727 00:36:03,280 --> 00:36:06,000 Yeah, seems like a good scoring system. 728 00:36:06,280 --> 00:36:12,120 It gives 3rd place to Max Dinged for betting against AI film 729 00:36:13,280 --> 00:36:16,520 better calibrated on his conservative bets like hardware 730 00:36:16,520 --> 00:36:21,920 and policy. 2nd place Eric solid instinct, slight overconfidence 731 00:36:21,960 --> 00:36:25,680 on a few misses and in first myself. 732 00:36:25,920 --> 00:36:28,080 Great balance of bold but accurate calls. 733 00:36:28,080 --> 00:36:30,400 Thank you, Chachi I. Love it, love it. 734 00:36:30,720 --> 00:36:33,800 I think I think this AI film take is complete bullshit. 735 00:36:33,800 --> 00:36:38,480 So I need to I need to score adjusted for that. 736 00:36:38,680 --> 00:36:44,040 This is just a check in, no medals awarded yet, but I will 737 00:36:44,040 --> 00:36:49,160 take the pole position and see you guys in a few months where 738 00:36:49,160 --> 00:36:53,800 we can do the final tally. All right, let's do our fantasy 739 00:36:53,800 --> 00:36:55,880 draft. Max, you want to talk us through 740 00:36:56,040 --> 00:36:58,000 the game and then we'll get into our picks. 741 00:36:58,280 --> 00:37:04,000 Yes, we invented an ingenious game based on fantasy football 742 00:37:04,440 --> 00:37:09,920 that allowed us to draft teams of startups into imaginary 743 00:37:09,920 --> 00:37:12,240 rosters. We've done two different drafts. 744 00:37:12,240 --> 00:37:14,440 We did one about a year and a half ago and one about six 745 00:37:14,440 --> 00:37:18,320 months ago. We restricted the draft list, 746 00:37:18,320 --> 00:37:21,320 the draft board as it were, to companies that had raised over 747 00:37:21,320 --> 00:37:25,480 $100 million at the time. So if there's obvious omissions 748 00:37:25,480 --> 00:37:27,200 that come up in your mind, it's probably because they hadn't 749 00:37:27,200 --> 00:37:29,880 raised 100 million at the. Cursor Cursor. 750 00:37:29,880 --> 00:37:32,240 Being the most. I think like there was some like 751 00:37:32,240 --> 00:37:34,360 where we've that we didn't include because. 752 00:37:34,360 --> 00:37:39,080 Yeah, We also excluded specific like chip based companies, 753 00:37:39,240 --> 00:37:42,720 Chinese companies, I don't know for robotics, I can't remember 754 00:37:42,720 --> 00:37:46,040 healthcare, we anything we thought we were even Dumber than 755 00:37:46,040 --> 00:37:47,560 normal about we left off the list. 756 00:37:47,960 --> 00:37:51,600 So we all drafted teams. We did a snake draft. 757 00:37:51,920 --> 00:37:57,160 Most notably, the first draft involved a discount that we had 758 00:37:57,440 --> 00:38:00,080 one person had to take for getting the first pick because 759 00:38:00,080 --> 00:38:01,960 the first pick was very obvious. It was Open AI. 760 00:38:02,560 --> 00:38:07,200 Eric paid, I believe, $75 billion in handicap to draft. 761 00:38:07,200 --> 00:38:09,800 Opening auction 1st and I and I really good about it. 762 00:38:10,080 --> 00:38:12,840 Aged pretty well since they're valued at 300 million right now 763 00:38:12,840 --> 00:38:15,000 3. 100 billion, would you say? 300 billion. 764 00:38:15,000 --> 00:38:16,480 I'm sorry. Yeah, like coughed. 765 00:38:16,480 --> 00:38:21,440 Yeah, it doesn't matter. And I will say before we say our 766 00:38:21,440 --> 00:38:25,240 teams, I have not yet had the first pick in any draft. 767 00:38:25,240 --> 00:38:27,160 So I just want everyone to remember that when making our 768 00:38:27,160 --> 00:38:32,120 teams, OK, I will go through my team first, which I will admit 769 00:38:32,120 --> 00:38:33,800 upfront is in last place right now. 770 00:38:34,280 --> 00:38:38,360 All right, my team, Databricks, my star worth $62 billion, 771 00:38:38,600 --> 00:38:42,240 Cohere AI, model company, modular AI, language company 772 00:38:42,240 --> 00:38:46,880 Scale AI, Sierra AI, Sakana AI and Hebia. 773 00:38:47,240 --> 00:38:50,880 And if you know all of those names, you are far too online. 774 00:38:51,000 --> 00:38:54,440 I'm bullish on Sierra. I think that's, I think scale. 775 00:38:54,440 --> 00:38:57,040 Has room to run as well. Oh yeah, yeah, and obviously 776 00:38:57,040 --> 00:39:01,200 Databricks isn't going anywhere. But Anthropic surpassing data 777 00:39:01,200 --> 00:39:04,600 bricks has been a bit of a sob story for my team because James 778 00:39:04,600 --> 00:39:07,680 got Anthropic at Crazy crazy money if I recall. 779 00:39:07,680 --> 00:39:10,560 The 4th pick that wasn't even your third pick because it's a 780 00:39:10,560 --> 00:39:13,560 snake draft. So it was me then Max the data 781 00:39:13,560 --> 00:39:17,040 bricks, then James with you'll say in a second, which doesn't 782 00:39:17,040 --> 00:39:20,200 make any sense, and then fourth with Anthropic. 783 00:39:20,840 --> 00:39:23,560 It's so embarrassing in retrospect, just. 784 00:39:23,640 --> 00:39:26,880 To just to clarify, like we drafted these teams originally 785 00:39:26,880 --> 00:39:30,680 in 2023 and then we did, I don't know, we drafted it. 786 00:39:30,880 --> 00:39:35,760 We, we had an ad drop waiver period last November and this 787 00:39:35,760 --> 00:39:39,240 again is a mid year check in. No, no ads, no drops, but 788 00:39:39,520 --> 00:39:43,960 checking in on the teams. So Max, I have your score so far 789 00:39:43,960 --> 00:39:48,440 right now at 93 billion, mostly because some of your teams have 790 00:39:48,440 --> 00:39:51,760 not raised or exited since you drafted them. 791 00:39:51,840 --> 00:39:54,880 I have no valuation on Sierra and scale is at 25. 792 00:39:55,080 --> 00:39:56,760 I'm somewhat optimistic on both of those. 793 00:39:57,240 --> 00:40:00,440 I managed to somehow pick the only foundation model company in 794 00:40:00,440 --> 00:40:02,720 the world that isn't wildly overvalued. 795 00:40:02,720 --> 00:40:08,880 Cohere SSI thinking machines, Anthropic. 796 00:40:08,920 --> 00:40:11,520 Like just throw a dart boarded foundation models. 797 00:40:11,520 --> 00:40:13,280 You've got a $40 billion company. 798 00:40:13,280 --> 00:40:16,520 But I'm. So Silicon Valley, the show said 799 00:40:16,520 --> 00:40:17,840 this from the beginning. No. 800 00:40:17,840 --> 00:40:19,680 Revenue is so much better than revenue. 801 00:40:19,680 --> 00:40:22,000 Cohere is a real business, so people can value. 802 00:40:22,040 --> 00:40:25,160 And. SSI and thinking machines are a 803 00:40:25,160 --> 00:40:29,360 dream. I have made a huge real. 804 00:40:29,480 --> 00:40:32,120 Revenue I I think SSI has real revenue, but anyway. 805 00:40:32,520 --> 00:40:35,400 Yeah, Max, what's your learning from this so far? 806 00:40:35,800 --> 00:40:38,720 I would say we already knew this, but you know the winners 807 00:40:38,720 --> 00:40:41,560 keep winning, right? Obviously Open AI swamps 808 00:40:41,560 --> 00:40:43,200 everything else that's happened in the entire draft. 809 00:40:43,200 --> 00:40:45,960 So we have a true power law which is nothing else matters 810 00:40:45,960 --> 00:40:49,080 comparison to Open AI even with the handicap which. 811 00:40:49,200 --> 00:40:51,840 Which we knew, which we knew was the risk when we came into it. 812 00:40:51,880 --> 00:40:54,600 Was happening, but regardless it still happened. 813 00:40:55,320 --> 00:41:00,400 Secondly, I would say Databricks is, you know, a merely a $60 814 00:41:00,400 --> 00:41:05,720 billion company, but that looks lame compared to, you know, like 815 00:41:05,840 --> 00:41:08,560 XAI being valued 80 like. You know, all right, all right, 816 00:41:08,840 --> 00:41:10,200 let's not spoil. James, you want. 817 00:41:10,200 --> 00:41:13,520 To go next. My team with the first pick on 818 00:41:13,520 --> 00:41:16,440 my draft that you guys were making fun of just the moments 819 00:41:16,440 --> 00:41:21,440 ago hugging face, No value because they have not raised 820 00:41:21,440 --> 00:41:24,360 since 2023. I've drafted them because they 821 00:41:24,360 --> 00:41:27,000 had one of the highest valuations at the time of the 822 00:41:27,000 --> 00:41:28,680 draft. They were valued, I think, over 823 00:41:28,680 --> 00:41:30,840 a billion dollars. I thought they were valued at 4 824 00:41:30,840 --> 00:41:32,320 or 4 billion. Dollars at the time, yeah, 825 00:41:32,320 --> 00:41:33,280 something like that, yeah. Yeah. 826 00:41:33,560 --> 00:41:42,040 OK So Anthropic giving me 61 1/2 billion value replit hasn't 827 00:41:42,040 --> 00:41:47,440 raised since the original draft. I exited Adept at 1 billion. 828 00:41:47,520 --> 00:41:51,880 I snagged XAI with the first pick last November. 829 00:41:52,360 --> 00:41:53,360 I'm. So jealous of that. 830 00:41:53,360 --> 00:41:57,680 Locked in 80 billion of value right there because they raised 831 00:41:57,680 --> 00:42:02,840 earlier this year and runway also raised $4 billion 832 00:42:02,840 --> 00:42:08,640 valuation. 11 Labs also raised at a $3.3 billion valuation and 833 00:42:08,960 --> 00:42:14,200 poolside no raise recently. Also one of those foundation 834 00:42:14,200 --> 00:42:18,440 model companies that we have yet to really see appear on the 835 00:42:18,440 --> 00:42:23,360 draft board, but I am happy with my overall team and my score of 836 00:42:23,360 --> 00:42:27,920 close to $150 billion currently. You know, something I just 837 00:42:27,920 --> 00:42:33,440 thought about, You're extremely lucky that XAI purchased X and 838 00:42:33,440 --> 00:42:37,240 not the opposite way, because if it had been X, it purchased XAI. 839 00:42:37,480 --> 00:42:40,960 We'd be able to like, force you to disown whatever growth, but 840 00:42:40,960 --> 00:42:45,440 now you get to benefit from this combined monstrosity, which 841 00:42:45,720 --> 00:42:48,360 kudos to you. I would just say had I been able 842 00:42:48,360 --> 00:42:51,200 to draft first, I would be in James's position of being in 843 00:42:51,200 --> 00:42:55,600 second place with XAI, so I don't personally think that a 844 00:42:55,600 --> 00:42:58,640 coin flip should be dictating my performance right now. 845 00:42:59,640 --> 00:43:01,400 What is you? I know. 846 00:43:01,400 --> 00:43:05,360 What was me I I will say to give James credit here, I I truly 847 00:43:05,360 --> 00:43:08,440 believe that anthropic is is the pick of the draft or, you know, 848 00:43:08,520 --> 00:43:12,400 so far I think that just getting anthropic at the fourth position 849 00:43:12,640 --> 00:43:16,320 in retrospect looks insane. And so I think that that is the 850 00:43:16,840 --> 00:43:18,400 the greatest. I don't even know if I call it a 851 00:43:18,400 --> 00:43:21,000 sleeper is sort of a semi sleeper pick, but that that 852 00:43:21,000 --> 00:43:23,240 clearly to me has had the most appreciation. 853 00:43:23,320 --> 00:43:25,600 Eric, why don't you all? Right. 854 00:43:25,600 --> 00:43:29,480 So yeah, I picked Open AI with a $75 billion handicap. 855 00:43:29,480 --> 00:43:33,880 Now it's worth 300 billion. So I'm getting basically 225 856 00:43:33,880 --> 00:43:37,960 billion for that. Inflection sold for 1.43 857 00:43:37,960 --> 00:43:41,440 billion. Character sold to Google for 2.5 858 00:43:41,440 --> 00:43:45,720 billion. Glean we're scoring at 4.6 859 00:43:45,720 --> 00:43:49,280 billion, but rumored to be raising at 7 billion. 860 00:43:49,560 --> 00:43:54,240 Miss Straw AI worried about that one. 6 billion right now. 861 00:43:54,480 --> 00:43:56,800 Perplexity. Oh man, I'm getting no credit 862 00:43:56,800 --> 00:43:58,280 for that. That's going to be a good 10 863 00:43:58,280 --> 00:44:02,240 right now, but it will be 14 is apparent according to the rumor, 864 00:44:02,240 --> 00:44:05,240 so we'll see. Safe super intelligence. 865 00:44:05,240 --> 00:44:08,320 I knew this was buzzy, but I don't even know if I could have 866 00:44:08,320 --> 00:44:13,920 seen this one raised at $32 billion already. 867 00:44:13,920 --> 00:44:17,960 It's more perplexity like That's insane. 868 00:44:18,640 --> 00:44:22,880 Kodium sold for 3 billion do. You want to explain that it's 869 00:44:23,240 --> 00:44:24,840 they renamed to Windsor for their name. 870 00:44:25,080 --> 00:44:28,240 Kodium is windsurf. Yeah, they sold to Open AI and 871 00:44:28,240 --> 00:44:31,360 then Harvey. No credit right now but rumored 872 00:44:31,360 --> 00:44:38,560 to be raising at 5 billion. Total value $274.5 billion. 873 00:44:39,640 --> 00:44:41,480 Yeah. I feel really good about this. 874 00:44:41,480 --> 00:44:44,360 I mean, hysterically, as I think I mentioned on the last episode, 875 00:44:44,360 --> 00:44:49,840 I I wrote a bear case about open AI after this at 157 billion, I 876 00:44:49,840 --> 00:44:52,160 think, but whatever. So I'm getting it narratively 877 00:44:52,160 --> 00:44:55,320 both ways. But yeah, I mean, I'm proud of 878 00:44:55,320 --> 00:44:58,360 all my picks. I think even my sort of singles 879 00:44:58,360 --> 00:45:04,880 are selling and I'm bullish on basically everything except I 880 00:45:04,880 --> 00:45:06,720 would like to hear what's going on with Mistral. 881 00:45:07,160 --> 00:45:10,040 But yeah, it's I mean, it's a power law business. 882 00:45:10,040 --> 00:45:13,000 It's crazy that like I'm like, oh, glean that's that's a good 883 00:45:13,000 --> 00:45:14,320 company. I was totally right that that 884 00:45:14,320 --> 00:45:16,960 would be a good company, but it it doesn't really matter for my 885 00:45:17,000 --> 00:45:19,400 my performance. Yeah, I mean I think you are 886 00:45:19,400 --> 00:45:22,680 consistently hitting singles and double s, but you could have 887 00:45:22,680 --> 00:45:25,160 nothing on your team except opening eye and be beating us by 888 00:45:25,160 --> 00:45:26,840 100 plus billion dollars at this point. 889 00:45:26,840 --> 00:45:28,080 So it. Doesn't, no. 890 00:45:28,080 --> 00:45:31,440 Which is why I made a that we could not randomly assign the 891 00:45:31,440 --> 00:45:34,280 first one and I I said we know. You were right. 892 00:45:34,640 --> 00:45:36,960 So I'm complaining, I I'm not complaining. 893 00:45:36,960 --> 00:45:38,600 You made the right decision 100%. 894 00:45:38,600 --> 00:45:42,560 It's just it, it is remarkable. It's like the whole game is just 895 00:45:42,560 --> 00:45:44,800 like, open. AI who drafted Open AI who 896 00:45:44,920 --> 00:45:47,760 drafted one Open AI go back. To that episode. 897 00:45:47,760 --> 00:45:49,680 Yeah, exactly. Yeah, No, we talked about it. 898 00:45:49,680 --> 00:45:52,560 I mean, I think it was it would they were raising at 90 at the 899 00:45:52,560 --> 00:45:54,640 time. And so you ended up with a $75 900 00:45:54,640 --> 00:45:57,480 billion handicap on a $90 billion company, which seemed 901 00:45:57,480 --> 00:46:00,320 like a reasonable deal to us. But you know, we were. 902 00:46:00,320 --> 00:46:02,520 We were all wrong, obviously. Or at least we're wrong with us. 903 00:46:02,520 --> 00:46:04,600 Check in. These do have to last, right? 904 00:46:04,600 --> 00:46:07,080 It's like 5 years or something. We're we're like, yeah. 905 00:46:07,240 --> 00:46:10,760 Five years, yeah. Yeah, we had a good shot with 906 00:46:10,760 --> 00:46:13,040 Sam Hoffman getting fired of your team. 907 00:46:13,040 --> 00:46:16,200 Going up sync. But but then he came back in 908 00:46:16,200 --> 00:46:17,240 force. All right. 909 00:46:17,360 --> 00:46:19,160 What's come on the market that you think we'll be looking at at 910 00:46:19,160 --> 00:46:21,200 the end of this year? Thinking machines. 911 00:46:21,200 --> 00:46:24,760 Cursor for sure. Manus. 912 00:46:24,960 --> 00:46:27,040 Oh yeah, yeah, yeah. Thinking Machines, cursor and 913 00:46:27,040 --> 00:46:28,560 manus are the ones that come to mind for me. 914 00:46:28,680 --> 00:46:30,760 Yeah, I mean, it's, it's going to be a tight band because we're 915 00:46:30,760 --> 00:46:33,080 we're picking them up and we're only interested in ones What 916 00:46:33,080 --> 00:46:34,520 that. They have to have raised $100 917 00:46:34,520 --> 00:46:36,760 million. All right, well, that's that's 918 00:46:36,760 --> 00:46:39,880 basically our episode. We're gonna have two more before 919 00:46:40,040 --> 00:46:43,120 the Cerebral Valley AI Summit in London on June 25th. 920 00:46:43,120 --> 00:46:47,040 And then at some point, once we've gathered ourselves, gone 921 00:46:47,040 --> 00:46:50,080 to that wedding in France we mentioned and relaxed a little, 922 00:46:50,080 --> 00:46:53,440 we'll come back to you and give you our thoughts from the event. 923 00:46:53,440 --> 00:46:56,440 I'm super excited about next week. 924 00:46:56,440 --> 00:47:00,880 No pressure James, our our game master over here, but we're 925 00:47:01,160 --> 00:47:04,960 trying to come up with some good concepts, but we'll be talking 926 00:47:05,440 --> 00:47:10,400 about voice and video and certainly in light of what VO3 927 00:47:10,400 --> 00:47:14,440 Googles new video creation model, it's an exciting time in 928 00:47:14,440 --> 00:47:16,280 video. So see you next week. 929 00:47:16,560 --> 00:47:18,000 Thanks guys. See ya. 930 00:47:18,320 --> 00:47:18,640 Thank you.