[00:00.000 - 00:03.000] Also, I take a little bit of offense to [00:02.000 - 00:05.000] this idea that like [00:03.000 - 00:07.600] >> Umbrage? You take umbrage? I do take [00:05.000 - 00:09.720] umbrage to this idea that like, "Oh, you [00:07.600 - 00:10.960] know, it's like so bad to be distilling [00:09.720 - 00:12.240] model weights. Like, how can you do [00:10.960 - 00:12.680] that? Like, that's so illegal and blah [00:12.240 - 00:13.720] blah blah." [00:12.680 - 00:14.160] >> Well, how did you get your training [00:13.720 - 00:15.360] data? [00:14.160 - 00:17.120] >> Well, where did that come from? You [00:15.360 - 00:19.640] distilled all of human knowledge, huh? [00:17.120 - 00:22.080] Who made all of human knowledge? [00:19.640 - 00:22.080] Us. [00:25.360 - 00:29.320] Welcome to round two. [00:27.120 - 00:30.760] We're going to start with Anthropic and [00:29.320 - 00:32.920] OpenAI. [00:30.760 - 00:35.280] It seems like the vibes have shifted a [00:32.920 - 00:37.480] little bit in the last week or two. [00:35.280 - 00:39.760] Used to be everyone was using Claude [00:37.480 - 00:43.000] code. Now, Codex is emerging. They just [00:39.760 - 00:45.480] announced 4 million weekly active users [00:43.000 - 00:47.880] for Codex. I think also Greg Brockman [00:45.480 - 00:49.520] just like took over all of product, [00:47.880 - 00:51.280] unifying everything kind of around [00:49.520 - 00:52.960] Codex. What are you seeing on the [00:51.280 - 00:54.440] ground? [00:52.960 - 00:56.680] Yeah, I mean [00:54.440 - 00:58.680] I switched over to using Codex for a [00:56.680 - 01:00.160] whole 5 minutes this week cuz Anthropic [00:58.680 - 01:02.225] was down and so I was like, "Oh, I I [01:00.160 - 01:02.280] should probably set it up." Uh [01:02.225 - 01:04.000] >> [clears throat] [01:02.280 - 01:05.320] >> but I do think that I mean, part of the [01:04.000 - 01:07.280] reason was that I wanted to see, you [01:05.320 - 01:08.400] know, okay, what's new? Like how does it [01:07.280 - 01:10.280] do in different tests and start [01:08.400 - 01:12.520] comparing them? Um but the other part of [01:10.280 - 01:14.080] the reason was Anthropic's [01:12.520 - 01:17.240] uh one of the thing that they did this [01:14.080 - 01:20.920] week was change the way they're talking [01:17.240 - 01:23.720] about the limits for using Claude code [01:20.920 - 01:27.080] programmatically. So, using Claude-P or [01:23.720 - 01:28.480] using the Claude agent SDK. Uh both of [01:27.080 - 01:30.080] those are now going to come with a [01:28.480 - 01:31.720] credit, a $200 credit. [01:30.080 - 01:32.160] >> Wow, that sounds so good, right? No. [01:31.720 - 01:34.280] Yes. [01:32.160 - 01:35.080] >> Not actually a credit. Actually, [01:34.280 - 01:36.760] you're getting you're [clears throat] [01:35.080 - 01:38.280] getting screwed on the the pricing is [01:36.760 - 01:40.080] what that really is. It was a [01:38.280 - 01:42.720] like kind of [01:40.080 - 01:45.960] both ingeniously and disgustingly, you [01:42.720 - 01:47.520] know, well-marketed change to the thing [01:45.960 - 01:48.560] where it sounded so good, but then [01:47.520 - 01:50.400] people were like, "Wait, no, this is [01:48.560 - 01:52.440] terrible." So, that made me think, "Oh, [01:50.400 - 01:54.200] maybe I should be, you know, just making [01:52.440 - 01:56.480] sure everything that I have currently [01:54.200 - 01:58.160] running works on both Codex and Claude [01:56.480 - 01:59.640] code." So, I did start setting it up. I [01:58.160 - 02:01.080] didn't end up using it that much, but I [01:59.640 - 02:02.280] did use it for a little while. And I do [02:01.080 - 02:03.800] think that that's part of why people are [02:02.280 - 02:05.840] shifting is that [02:03.800 - 02:07.720] like cost performance, I think it's [02:05.840 - 02:09.800] probably a little bit better on that [02:07.720 - 02:11.120] for, you know, 5.5. [02:09.800 - 02:13.320] >> Yep. [02:11.120 - 02:15.080] Possibly even availability. They did do [02:13.320 - 02:17.280] some like, "Oh, you know, we'll like [02:15.080 - 02:18.920] sponsor you if you switch over" things. [02:17.280 - 02:20.040] They even had a like Open AI had like, [02:18.920 - 02:21.200] "Oh, like if you're a company and you [02:20.040 - 02:21.600] switch over, we'll give you 2 months for [02:21.200 - 02:23.040] free." [02:21.600 - 02:24.440] >> 2 free months, yep. Yeah, yeah. And I [02:23.040 - 02:25.440] looked at that and I And then it's like [02:24.440 - 02:27.320] it's just a form you fill out and [02:25.440 - 02:28.480] they're like, "2 months free if you [02:27.320 - 02:30.800] switch something something." And I was [02:28.480 - 02:32.640] like, "Okay, come on." Yeah, so I didn't [02:30.800 - 02:35.000] even bother submitting it, but [02:32.640 - 02:37.560] you know, in theory, they're doing [02:35.000 - 02:38.840] something there. Mostly on both sides [02:37.560 - 02:40.000] just a bunch of smoke and mirrors and [02:38.840 - 02:41.760] marketing nonsense, but [02:40.000 - 02:44.240] >> Yeah. I mean, it is a little bit [02:41.760 - 02:46.160] cheaper. So, yeah. [02:44.240 - 02:50.000] I mean, it part of it is just Open AI [02:46.160 - 02:52.520] using its advantage, which is this huge [02:50.000 - 02:54.160] war chest of compute. And Anthropic is [02:52.520 - 02:56.840] catching up and making deals left and [02:54.160 - 02:59.800] right, but that is just an edge that [02:56.840 - 03:02.160] Open AI has and it's been using. Yep, [02:59.800 - 03:03.600] exactly. So, I expect to see more of [03:02.160 - 03:05.760] this and I'm happy with them to keep [03:03.600 - 03:07.360] competing and hopefully keep competing [03:05.760 - 03:09.720] with Gemini as well. [03:07.360 - 03:12.200] Uh yeah, I think if either like if any [03:09.720 - 03:13.800] of them try to [03:12.200 - 03:15.160] make it too expensive, this is what [03:13.800 - 03:16.519] happens. You can see people just switch [03:15.160 - 03:18.000] to the other one. I hope it stays that [03:16.519 - 03:19.120] way. [03:18.000 - 03:21.080] Yeah, you mentioned Gemini, which I [03:19.120 - 03:22.600] hadn't mentioned yet. How does Gemini [03:21.080 - 03:25.440] fit into all of this? [03:22.600 - 03:27.239] I think Gemini's long-term strategy is [03:25.440 - 03:29.200] is this, basically. If like Google can [03:27.239 - 03:31.239] afford to burn more money than either [03:29.200 - 03:33.360] Open AI or Anthropic, cuz they have [03:31.239 - 03:34.480] actual revenue and they have, you know, [03:33.360 - 03:37.480] they need to protect that revenue [03:34.480 - 03:39.680] stream, too. So, they're happy to like [03:37.480 - 03:41.800] sell Gemini at a loss for way longer [03:39.680 - 03:44.920] than Open AI and Anthropic can afford [03:41.800 - 03:46.640] to. Do they have a Codex or Cloud Code [03:44.920 - 03:47.880] equivalent? Like does Antigravity do [03:46.640 - 03:50.880] that? [03:47.880 - 03:54.320] They have Antigravity and Jewels and [03:50.880 - 03:56.480] Gemini CLI. So, they have three. [03:54.320 - 03:57.760] Yeah, why not? [03:56.480 - 03:59.000] Okay. [03:57.760 - 04:00.720] Yeah, but I think I would probably use [03:59.000 - 04:02.560] Gemini CLI. Right now, it's just a [04:00.720 - 04:04.160] little bit janky, but it'll catch up. [04:02.560 - 04:05.480] Like, it'll get there. So, no one uses [04:04.160 - 04:08.200] it today cuz it's not as good as the [04:05.480 - 04:09.880] other two, but I'm sure that give it in [04:08.200 - 04:11.000] a few quarters, like it'll it'll get [04:09.880 - 04:12.280] there. [04:11.000 - 04:15.600] Yeah, do you think people will continue [04:12.280 - 04:17.000] to use Claude code and Codex apps [04:15.600 - 04:18.840] though, even as they're using the Gemini [04:17.000 - 04:21.680] CLI? [04:18.840 - 04:22.720] Um maybe uh like the desktop apps and [04:21.680 - 04:24.040] stuff. I'm not sure. [04:22.720 - 04:26.840] >> lot of people are really liking the code [04:24.040 - 04:27.760] app the Codex desktop app. [04:26.840 - 04:29.000] Yeah, I think the thing that's [04:27.760 - 04:31.400] interesting about this is that I'm not [04:29.000 - 04:34.040] sure that it really matters which app [04:31.400 - 04:36.800] you use because I can already tell [04:34.040 - 04:39.160] Claude to run Gemini or Gemini to run [04:36.800 - 04:41.800] Anthropic to run OpenAI. Like, it [04:39.160 - 04:43.480] doesn't The thing that matters is like [04:41.800 - 04:45.280] which of the models and which of the [04:43.480 - 04:47.080] harnesses is doing all the compute and [04:45.280 - 04:48.800] all the inference. And so, it matters [04:47.080 - 04:50.560] like which of these are the core for the [04:48.800 - 04:51.919] agency you're running. And for the [04:50.560 - 04:53.600] people that are actually running lots of [04:51.919 - 04:58.000] agents, they're probably not running [04:53.600 - 05:01.280] Opus 4.7 in fast mode or even, you know, [04:58.000 - 05:03.000] GPT 5.5 because it's just too expensive. [05:01.280 - 05:04.640] So, that's those are the ones that [05:03.000 - 05:05.880] really matter over time. So, we'll see [05:04.640 - 05:07.120] where that shakes out. It like it [05:05.880 - 05:08.240] doesn't matter You can only type so [05:07.120 - 05:09.840] much. You can only read so much. You [05:08.240 - 05:12.000] could have it generate all the tokens [05:09.840 - 05:13.600] for you into the future and it like that [05:12.000 - 05:14.840] won't matter. Whatever interface you [05:13.600 - 05:16.160] have doesn't matter. It's the agents [05:14.840 - 05:17.280] that matter. [05:16.160 - 05:18.800] That's true. Though, I guess I'd push [05:17.280 - 05:21.400] back a little bit of like the interface [05:18.800 - 05:22.840] you're using becomes the default [05:21.400 - 05:24.680] model you use. [05:22.840 - 05:25.680] Especially if you are on a max plan [05:24.680 - 05:27.880] though, they're kind of getting rid of [05:25.680 - 05:29.400] that. That's right. But, it's going to [05:27.880 - 05:30.919] be pretty easy to tell it like, "Oh, you [05:29.400 - 05:32.200] got a lot of big work to do? Like, [05:30.919 - 05:33.840] here's a skill for calling into this [05:32.200 - 05:35.040] much cheaper agent, right?" Or like [05:33.840 - 05:36.600] doing this in this other programmatic [05:35.040 - 05:38.440] way. We're going to see a lot of that [05:36.600 - 05:40.320] because the costs are just going to [05:38.440 - 05:41.680] get a lot higher as you're running many [05:40.320 - 05:43.320] more agents. [05:41.680 - 05:44.919] If you were Anthropic, would you be [05:43.320 - 05:47.320] worried right now? [05:44.919 - 05:49.040] Yeah, but I've always said that. I I [05:47.320 - 05:51.200] still think that you know as much as I [05:49.040 - 05:52.360] should on Gemini like Google is the one [05:51.200 - 05:54.880] that's in the best position. So I'm [05:52.360 - 05:56.360] actually you know sometimes I [05:54.880 - 05:57.840] you know I might say something like oh [05:56.360 - 05:59.160] yeah why isn't Anthropic doing this but [05:57.840 - 06:00.760] I'm always sort of on their side. I want [05:59.160 - 06:02.120] Anthropic and OpenAI to do as well as [06:00.760 - 06:04.680] they possibly can so that we have any [06:02.120 - 06:06.000] competition for Google. Like that so I [06:04.680 - 06:07.200] don't actually mind too much. It's just [06:06.000 - 06:09.440] a little bit annoying when they change [06:07.200 - 06:11.360] the billing like that you know. [06:09.440 - 06:12.840] Along those lines [06:11.360 - 06:15.200] uh the information actually just posted [06:12.840 - 06:18.280] something about how Anthropic and OpenAI [06:15.200 - 06:21.480] now capture 89% of the roughly 80 [06:18.280 - 06:24.320] billion in annualized revenue across 34 [06:21.480 - 06:25.800] top AI startups. Mhm. [06:24.320 - 06:28.120] That's a lot. [06:25.800 - 06:29.360] Sounds right. I think about a year back [06:28.120 - 06:30.520] we were talking about when we talked [06:29.360 - 06:31.880] about models we were all like oh the [06:30.520 - 06:33.480] model layer is commoditized you know [06:31.880 - 06:35.400] like they're in a bad spot. That's not [06:33.480 - 06:37.240] really the zeitgeist anymore. How do you [06:35.400 - 06:38.840] think about that that trade-off and at [06:37.240 - 06:41.520] this point [06:38.840 - 06:45.000] do we know that the model layer value [06:41.520 - 06:46.520] capture crushes the app layer value [06:45.000 - 06:48.160] capture long-term? [06:46.520 - 06:49.880] I don't know. Um [06:48.160 - 06:51.600] I still think that that's long-term [06:49.880 - 06:53.919] where it goes [06:51.600 - 06:56.360] that it'll matter a little bit [06:53.919 - 06:57.880] less which of the like front-ends or or [06:56.360 - 06:58.400] models that you're using [06:57.880 - 07:00.680] but [06:58.400 - 07:01.840] >> Sorry you're saying which you believe in [07:00.680 - 07:04.280] long-term? [07:01.840 - 07:06.120] Oh long-term I think [07:04.280 - 07:07.880] that Google seems like they're in a good [07:06.120 - 07:10.800] spot because they're the ones who can [07:07.880 - 07:12.440] afford to burn money on this stuff and [07:10.800 - 07:14.960] they're going to end up giving away the [07:12.440 - 07:17.680] models and the tokens etc. like for for [07:14.960 - 07:19.560] free-ish relative to Anthropic so it [07:17.680 - 07:20.919] doesn't really matter if Anthropic or [07:19.560 - 07:24.120] Google is like making a lot of money [07:20.919 - 07:25.800] today. It's about like long-term [07:24.120 - 07:27.760] what are the costs? Like right now we're [07:25.800 - 07:29.680] not in a cost optimization part of this [07:27.760 - 07:32.480] whole thing right? Definitely. Every I [07:29.680 - 07:33.840] mean it's it's what it's May right? So [07:32.480 - 07:35.360] how many months have people really been [07:33.840 - 07:36.880] using this stuff and most people in the [07:35.360 - 07:37.960] world still are not using this at all [07:36.880 - 07:39.240] right? You've seen that graph [07:37.960 - 07:41.560] >> We all we all love that graph. Yeah [07:39.240 - 07:43.240] exactly. It's like there's almost nobody [07:41.560 - 07:45.600] using this stuff. So as that expands and [07:43.240 - 07:46.600] lots more people are using it, then then [07:45.600 - 07:48.080] companies are going to start to think [07:46.600 - 07:49.480] like, oh, maybe I need to optimize this [07:48.080 - 07:51.600] a little bit. Even we are starting to [07:49.480 - 07:52.920] think about that as a startup, right? [07:51.600 - 07:54.680] In the beginning, if you're paying [07:52.920 - 07:55.960] $10,000 a month or whatever, like that's [07:54.680 - 07:56.880] that's a lot, but it's not the end of [07:55.960 - 07:58.360] the world cuz you're paying your [07:56.880 - 07:59.800] employees a lot more and like [07:58.360 - 08:01.920] collectively. But as you're starting to [07:59.800 - 08:03.360] pay like more than your employees for [08:01.920 - 08:05.520] tokens, which is where some of these [08:03.360 - 08:07.160] companies are probably getting to, then [08:05.520 - 08:08.520] you start to think like, huh, maybe I [08:07.160 - 08:10.520] should, you know, spend a few hours [08:08.520 - 08:12.400] thinking about how to like and make this [08:10.520 - 08:13.960] a little more efficient. [08:12.400 - 08:15.240] Then then the cost start to matter a [08:13.960 - 08:17.280] little bit more. We're just not in the [08:15.240 - 08:19.320] cost optimization part of this thing. [08:17.280 - 08:20.480] Like it's still in like exploding, like [08:19.320 - 08:21.640] I'm just trying to get stuff done. I'm [08:20.480 - 08:23.520] not really worrying about the cost [08:21.640 - 08:24.880] phase. [08:23.520 - 08:27.440] So, what does that mean for my initial [08:24.880 - 08:29.520] question of which do you think is like a [08:27.440 - 08:31.280] better spot to be in and captures more [08:29.520 - 08:32.599] value long-term, the model layer or the [08:31.280 - 08:34.200] app layer? [08:32.599 - 08:35.440] It's got to be the app layer, not the [08:34.200 - 08:37.719] model layer. I think the model layer [08:35.440 - 08:39.320] ends up commoditizing. Yeah. [08:37.719 - 08:41.440] Commoditizing and if you're not Google, [08:39.320 - 08:43.000] then you're not going to long-term [08:41.440 - 08:46.960] capture the value. [08:43.000 - 08:48.720] Uh yeah, basically. Like like Google, [08:46.960 - 08:50.400] it's about like for the infrastructure [08:48.720 - 08:52.680] model layer, it's about who has the best [08:50.400 - 08:54.160] cost structure. And so, Anthropic and [08:52.680 - 08:55.720] OpenAI do not have a better cost [08:54.160 - 08:57.160] structure than Google because Google has [08:55.720 - 08:59.040] their own chips and their own data [08:57.160 - 09:00.520] centers and their own everything. And [08:59.040 - 09:04.160] that just makes it a lot cheaper to [09:00.520 - 09:05.400] serve things in the long run. So, that's [09:04.160 - 09:07.360] that's who's going to win at the model [09:05.400 - 09:10.080] layer is my guess. [09:07.360 - 09:11.600] This is interesting. It seems like [09:10.080 - 09:13.200] you're going you believe in this [09:11.600 - 09:15.480] narrative, this kind of older narrative [09:13.200 - 09:16.880] that now feels a little outdated, of it [09:15.480 - 09:18.480] is going to be commoditized. Like we we [09:16.880 - 09:20.440] were always saying that and then now it [09:18.480 - 09:22.520] feels like it's not going to be because [09:20.440 - 09:24.640] everyone's like, cloud code and codex [09:22.520 - 09:27.280] and harnesses are critical. But you're [09:24.640 - 09:29.600] saying actually that long-term the [09:27.280 - 09:31.600] models are going to be commoditized. The [09:29.600 - 09:35.360] models are going to even be adapted to [09:31.600 - 09:37.320] their own harnesses. But [09:35.360 - 09:39.520] and maybe even have something sitting at [09:37.320 - 09:41.520] the like, you know, this harness has to [09:39.520 - 09:44.280] be used with this model layer, but maybe [09:41.520 - 09:46.400] you can use the Gemini model and and its [09:44.280 - 09:49.080] whole harness and model and everything. [09:46.400 - 09:51.560] But fundamentally, the problem is that [09:49.080 - 09:53.160] like an agent is just too flexible and [09:51.560 - 09:54.240] too general. At some point you get to [09:53.160 - 09:55.480] this higher level where it's like, well, [09:54.240 - 09:56.880] I can just tell it to do something. And [09:55.480 - 09:58.360] if I can just tell it to do something is [09:56.880 - 09:59.920] the interface that you have, well, I [09:58.360 - 10:01.720] mean, then who the [ __ ] cares? Like [09:59.920 - 10:03.160] anything can fit that interface, like [10:01.720 - 10:04.600] any kind of AI model. It doesn't really [10:03.160 - 10:06.080] matter what it is. And if it can't do [10:04.600 - 10:08.240] it, then it's not general intelligence. [10:06.080 - 10:10.400] And so, like you're kind of in this [10:08.240 - 10:13.720] weird catch-22 of like it kind of has to [10:10.400 - 10:14.840] be just cost per task kind of a thing. [10:13.720 - 10:16.200] That's all that really matters [10:14.840 - 10:17.440] eventually. [10:16.200 - 10:20.360] Yeah. [10:17.440 - 10:22.520] I have a theory that long-term the [10:20.360 - 10:25.320] reason that [10:22.520 - 10:28.000] OpenAI and Anthropic win and do so well [10:25.320 - 10:30.760] is because [10:28.000 - 10:33.240] the model and the harnesses are [10:30.760 - 10:35.520] a great wedge and entry point [10:33.240 - 10:38.240] into then doing all of this that we're [10:35.520 - 10:40.960] seeing already, forward deployed, huge, [10:38.240 - 10:41.720] basically like consulting agreements [10:40.960 - 10:44.520] um [10:41.720 - 10:46.560] with actual enterprises where not [10:44.520 - 10:47.960] developers that can just switch, uh but [10:46.560 - 10:49.560] actual enterprises and then like they [10:47.960 - 10:51.360] can kind of have lock-in there. Does [10:49.560 - 10:52.400] that seem right to you? Yeah, I think [10:51.360 - 10:53.800] that's what they're trying for and the [10:52.400 - 10:56.000] question is whether that succeeds or [10:53.800 - 10:57.560] not. And the place of the value would be [10:56.000 - 10:59.440] coming from there is like, oh, we have [10:57.560 - 11:00.920] this existing relationship or deployment [10:59.440 - 11:02.600] or a sales thing. Like that's where the [11:00.920 - 11:05.240] value value is coming from, not from the [11:02.600 - 11:07.240] model itself. Like the models in theory, [11:05.240 - 11:08.440] you can just even after this happens, [11:07.240 - 11:09.720] you can come along and Google can have [11:08.440 - 11:11.560] their own sales force and be like, hey, [11:09.720 - 11:12.640] we can swap it out and save you 30% and [11:11.560 - 11:15.040] you're like, yeah, okay. And they have [11:12.640 - 11:16.440] it happen slowly over time and it yeah. [11:15.040 - 11:18.160] Yeah, but I think it's it's like using [11:16.440 - 11:20.720] this lead where their models are the [11:18.160 - 11:22.640] best and there is like a real delta to [11:20.720 - 11:23.760] build the relationships because [11:22.640 - 11:26.839] companies have to build those [11:23.760 - 11:27.880] relationships right now. Mhm. [11:26.839 - 11:29.200] And then they're they have those [11:27.880 - 11:31.200] long-term partnerships and then it's [11:29.200 - 11:34.160] kind of easier. [11:31.200 - 11:37.280] Yeah, that's that's the hope. [11:34.160 - 11:40.480] Yeah. That's the hope. What would be the [11:37.280 - 11:41.720] the straw man the steel man of that? [11:40.480 - 11:43.280] Uh [11:41.720 - 11:45.960] the problem with making general [11:43.280 - 11:47.560] intelligence is that let's say that, you [11:45.960 - 11:49.520] know, let's say that today you had an [11:47.560 - 11:51.360] existing relationship with some company [11:49.520 - 11:52.520] that was a software provider, right? And [11:51.360 - 11:55.560] it's like, "Oh, yeah, we have this like [11:52.520 - 11:57.080] big, you know, medical records thing, [11:55.560 - 11:59.360] blah blah like [11:57.080 - 12:00.640] okay, well, if you was really expensive [11:59.360 - 12:01.600] it was on Epic or something and you [12:00.640 - 12:03.160] really wanted to migrate off of it. [12:01.600 - 12:03.960] Could you migrate? Yeah, but like oh, [12:03.160 - 12:05.840] it's going to like take a lot of [12:03.960 - 12:07.840] engineering time. Oh, engineering time, [12:05.840 - 12:09.720] huh? Hmm, what's this thing that's like [12:07.840 - 12:11.640] plummeting in cost over time? It's like, [12:09.720 - 12:13.640] "Oh, engineering time." So, switching [12:11.640 - 12:14.600] from one thing to another feels like the [12:13.640 - 12:16.040] kind of thing that for a lot of [12:14.600 - 12:17.520] companies might get easier and easier [12:16.040 - 12:19.360] over time. Now, are you going to switch [12:17.520 - 12:20.720] your banks and medical records? Probably [12:19.360 - 12:22.240] not. There's a lot of stakes [12:20.720 - 12:22.778] >> you mentioned Epic. Epic is like the [12:22.240 - 12:23.040] last thing [12:22.778 - 12:25.000] >> [laughter] [12:23.040 - 12:26.960] >> that will ever be switched off from. [12:25.000 - 12:29.040] >> one probably Well, people do actually [12:26.960 - 12:31.080] even. So, I would say banks are probably [12:29.040 - 12:32.560] the last ones. Medicals are the second [12:31.080 - 12:34.040] and then maybe there's like people or [12:32.560 - 12:35.240] something like that, you know, there's [12:34.040 - 12:36.720] something insurance and other things [12:35.240 - 12:38.240] where it's like and safety critical [12:36.720 - 12:39.760] things, infrastructure that's not really [12:38.240 - 12:42.560] worth switching. [12:39.760 - 12:44.040] Uh but for most businesses, you're not [12:42.560 - 12:45.960] locked in in that same way where it's [12:44.040 - 12:47.960] like, "Yeah, you know, we could switch [12:45.960 - 12:50.120] billing providers. It'd be a huge pain, [12:47.960 - 12:51.720] but like if we needed to or if we would [12:50.120 - 12:53.800] make a lot more money then we would do [12:51.720 - 12:54.800] it." And so, I think we're just not at [12:53.800 - 12:56.680] the [12:54.800 - 12:58.240] place where [12:56.680 - 13:00.440] people have to optimize this kind of [12:58.240 - 13:01.560] stuff yet, so they don't. Uh but we [13:00.440 - 13:03.880] might start to get there once it starts [13:01.560 - 13:05.000] to get a little more competitive. [13:03.880 - 13:06.160] You know, once anyone can do the same [13:05.000 - 13:08.360] thing cuz they all have access to the [13:06.160 - 13:09.640] same models, well, okay, now it's about [13:08.360 - 13:12.160] the like things that you're doing that [13:09.640 - 13:14.000] other people can't see. So, that 3% that [13:12.160 - 13:16.880] you're giving away to Stripe, well, you [13:14.000 - 13:18.240] know, if your margin is 6% then that 3% [13:16.880 - 13:19.560] makes a big difference. So, now you're [13:18.240 - 13:20.760] like, "Oh, maybe we switch to doing our [13:19.560 - 13:22.760] own kind of payment rails or something [13:20.760 - 13:24.440] something." So, that kind of pressure I [13:22.760 - 13:27.680] think is going to happen as you get like [13:24.440 - 13:29.640] more and more advanced AI systems. [13:27.680 - 13:30.920] When do you think that starts to happen? [13:29.640 - 13:32.240] A few years from now. It's not right [13:30.920 - 13:33.600] away. [13:32.240 - 13:36.120] Yeah. The question is how much you can [13:33.600 - 13:37.880] sell and then how [13:36.120 - 13:39.280] like easy it will be to switch how many [13:37.880 - 13:41.440] companies off. I don't know. We'll see [13:39.280 - 13:42.760] where it lands. Let's talk about one [13:41.440 - 13:43.880] more thing or one or two more things on [13:42.760 - 13:46.400] Anthropic. [13:43.880 - 13:47.240] So, Anthropic is now trading at like $1 [13:46.400 - 13:49.240] trillion [13:47.240 - 13:52.160] on the secondary market. [13:49.240 - 13:55.040] How high do you think it can go? [13:52.160 - 13:57.240] Diamond hands. I mean to the moon. I [13:55.040 - 13:59.360] don't know. It's like Anthropic I mean [13:57.240 - 14:01.920] obviously not. Eventually it goes down. [13:59.360 - 14:06.000] And my cynical take is that [14:01.920 - 14:07.680] you're not allowed to have uh new [14:06.000 - 14:09.360] companies like this in like a tech [14:07.680 - 14:11.280] bubble. They're not allowed to go down [14:09.360 - 14:13.400] until after they IPO. Like you got to [14:11.280 - 14:15.520] have somebody else holding the bag. So, [14:13.400 - 14:16.800] you have to have like the general public [14:15.520 - 14:18.160] Oh, great. Now like there's pension [14:16.800 - 14:20.400] funds and everything there. Great. Now [14:18.160 - 14:21.840] you can zero it out. Because then then [14:20.400 - 14:23.520] you can take like then you can short it, [14:21.840 - 14:24.760] too. So, you can't short it beforehand. [14:23.520 - 14:25.920] So, this way you can make money on up [14:24.760 - 14:27.600] you make money when it's coming down. [14:25.920 - 14:29.360] Like you just Yeah. So, that that's my [14:27.600 - 14:30.920] cynical take that [14:29.360 - 14:32.760] you know it can't can't go down until [14:30.920 - 14:35.440] it's public. The VCs have to get paid [14:32.760 - 14:36.320] back and then we can see it fall. [14:35.440 - 14:37.320] Yeah. [14:36.320 - 14:39.680] Yeah. And then hedge funds can make [14:37.320 - 14:42.720] money as it's falling. [14:39.680 - 14:44.480] Yes. And then there's like real pressure [14:42.720 - 14:45.720] uh where there's not right now when it's [14:44.480 - 14:47.160] just private. Yeah. [14:45.720 - 14:49.160] >> And not initially either. Like it's [14:47.160 - 14:50.280] going to have a huge pop I'm sure and [14:49.160 - 14:52.160] How much do you think it'll go out for [14:50.280 - 14:53.320] when it IPOs? [14:52.160 - 14:56.640] I mean I would have said a trillion [14:53.320 - 14:57.560] before, but I don't know. Yeah. I mean I [14:56.640 - 15:00.320] mean I don't think it's going to be a [14:57.560 - 15:02.200] down round. Let's say that. [15:00.320 - 15:04.760] Yeah. I mean there's so much [15:02.200 - 15:06.560] pent-up demand. Yeah. [15:04.760 - 15:10.000] Who do you think has a higher market cap [15:06.560 - 15:11.480] in a year? Anthropic or OpenAI? [15:10.000 - 15:13.600] I don't know. [clears throat] Yeah, that [15:11.480 - 15:15.360] one's that one's tough. Yeah. I don't [15:13.600 - 15:16.640] know. Where is OpenAI now even? I I [15:15.360 - 15:18.280] don't know. It's it's all a little bit [15:16.640 - 15:19.800] speculative. I basically I expect them [15:18.280 - 15:22.720] both to IPO [15:19.800 - 15:24.360] similar kind of ish things and then [15:22.720 - 15:25.680] after that we'll see. We don't know. I [15:24.360 - 15:28.040] they could it could still go up a little [15:25.680 - 15:29.680] bit um but I would be a lot more worried [15:28.040 - 15:32.680] after they IPO'd. [15:29.680 - 15:35.600] Last topic within Anthropic. [15:32.680 - 15:37.880] Our friend Richard Noe had a incendiary [15:35.600 - 15:39.600] tweet earlier today. [15:37.880 - 15:41.200] Saying that the biggest effect that EA [15:39.600 - 15:43.120] has had on the world is empowering [15:41.200 - 15:45.480] Anthropic via narratives, funding, and [15:43.120 - 15:47.560] talent. Whether good or bad, that's so [15:45.480 - 15:49.520] different from EA's anticipated paths to [15:47.560 - 15:51.880] impact that indicates a massive blind [15:49.520 - 15:54.040] spot. [15:51.880 - 15:56.120] Not even just blind spot but it it's [15:54.040 - 16:00.080] extremely ironic. Right? I think [15:56.120 - 16:02.520] actually both Anthropic and sorry, [16:00.080 - 16:04.200] Anthropic, OpenAI, and DeepMind I think [16:02.520 - 16:05.440] were all started by people who were [16:04.200 - 16:06.720] like, "Oh, yeah, AI seems really [16:05.440 - 16:08.240] dangerous. Like we got to do something [16:06.720 - 16:10.720] about that. We got to make it safe." I'm [16:08.240 - 16:13.400] pretty sure that DeepMind was funded by [16:10.720 - 16:16.360] Peter Thiel I think a long time ago. [16:13.400 - 16:17.400] Yeah, because yeah, I think probably cuz [16:16.360 - 16:18.640] Demis was like, "Oh, yeah, you know, [16:17.400 - 16:20.320] there we we should probably do something [16:18.640 - 16:21.480] about that." And then started the thing [16:20.320 - 16:23.320] and then after that like it sort of [16:21.480 - 16:24.640] inspired OpenAI and they're like, "Oh, [16:23.320 - 16:25.920] yeah, we got to do something with with [16:24.640 - 16:28.040] you know, AI and safety and everything. [16:25.920 - 16:29.680] We got to make this nonprofit." Okay. [16:28.040 - 16:31.000] Yeah, I think you know, well, you know, [16:29.680 - 16:33.600] AI matters a lot so we can't just let [16:31.000 - 16:35.360] Google do it. Yeah, yeah, exactly. We [16:33.600 - 16:36.480] have to be the good the good actors. [16:35.360 - 16:39.200] Yeah, we got we got to do it good, you [16:36.480 - 16:40.920] know? Uh and like all of these all of [16:39.200 - 16:42.960] this growth all of it came from people [16:40.920 - 16:45.560] being afraid of doing exactly the thing [16:42.960 - 16:47.600] that they ended up doing. Like had you [16:45.560 - 16:48.880] just not started those companies, not [16:47.600 - 16:51.080] put together all of that capital and all [16:48.880 - 16:53.280] those resources a much slower takeoff [16:51.080 - 16:55.160] probably. You know, Gary Marcus might [16:53.280 - 16:57.080] still be relevant. Like there would be a [16:55.160 - 16:58.960] lot of you know, the world would look a [16:57.080 - 17:01.760] little bit different. [16:58.960 - 17:03.760] Yeah. Richard said in a later tweet, [17:01.760 - 17:05.480] "Cynically, EA's purpose is to funnel [17:03.760 - 17:07.040] resources to power seekers who [17:05.480 - 17:09.680] self-deceive enough to consider [17:07.040 - 17:12.291] themselves altruists." [17:09.680 - 17:12.400] That's accurate. You agree? [17:12.291 - 17:14.160] >> [laughter] [17:12.400 - 17:16.199] >> Uh well, [17:14.160 - 17:17.880] it's accurate in that some like that's [17:16.199 - 17:19.280] an existence statement so it does that [17:17.880 - 17:20.959] for some people. I would not say that [17:19.280 - 17:22.600] all EAs do that, but I would definitely [17:20.959 - 17:24.319] say that some do fit that [17:22.600 - 17:26.400] characterization. [17:24.319 - 17:27.800] Uh and you know, [17:26.400 - 17:29.560] maybe it's a minority even, but it [17:27.800 - 17:31.240] doesn't take that many in order to go [17:29.560 - 17:33.960] make some huge company and then have a [17:31.240 - 17:37.520] big impact on the world and yeah. [17:33.960 - 17:38.760] That's fair. I will also say [17:37.520 - 17:40.440] I mean, I guess [17:38.760 - 17:42.880] like you can sit with the cynical take [17:40.440 - 17:46.200] like it's been an unbelievably [17:42.880 - 17:48.080] powerful recruiting and company building [17:46.200 - 17:49.760] strategy for Anthropic. [17:48.080 - 17:51.760] >> Oh, it's been amazing for Anthropic, [17:49.760 - 17:54.400] yeah. I don't think Anthropic would be [17:51.760 - 17:56.680] Anthropic without it. No, it's been a [17:54.400 - 17:58.400] masterclass in like how to [17:56.680 - 18:00.760] do a good job with some sort of mission [17:58.400 - 18:02.720] like this. Yes. And I think the people [18:00.760 - 18:04.680] at the top, the founders, many of them [18:02.720 - 18:06.800] like really do believe in it and that's [18:04.680 - 18:08.400] how they were able to build this company [18:06.800 - 18:10.400] out of it. But like you know, you even [18:08.400 - 18:12.480] today, despite the scale that [18:10.400 - 18:15.120] Anthropic's at, [18:12.480 - 18:18.560] people there are so mission-oriented. [18:15.120 - 18:21.000] They're so aligned. And they care so [18:18.560 - 18:22.920] much and that has all sorts of amazing [18:21.000 - 18:24.320] downstream effects. It means they stay. [18:22.920 - 18:25.880] We have not seen I think all the [18:24.320 - 18:27.480] founders are still there at Anthropic [18:25.880 - 18:30.720] compared to OpenAI. [18:27.480 - 18:32.000] Uh you know, Anthropic can pay less. Uh [18:30.720 - 18:36.080] you know, early before they were just [18:32.000 - 18:38.640] like at 40 billion plus in in ARR, you [18:36.080 - 18:40.640] know, they could attract top talent from [18:38.640 - 18:43.080] OpenAI and elsewhere without having to [18:40.640 - 18:45.120] pay them top dollar. Like it had all of [18:43.080 - 18:47.000] these amazing impacts and has like still [18:45.120 - 18:49.760] to this day. Mhm. [18:47.000 - 18:51.280] Yeah, exactly. [18:49.760 - 18:54.840] All right, we'll move off of Anthropic [18:51.280 - 18:56.600] and OpenAI. Let's talk about China. [18:54.840 - 18:58.120] Also related to Anthropic. I mean, [18:56.600 - 18:58.960] everything's related to That's true. [18:58.120 - 19:00.760] We're going to talk about Anthropic [18:58.960 - 19:02.160] again here. I forgot. [19:00.760 - 19:04.280] Uh and also everything is related to [19:02.160 - 19:05.920] Anthropic and OpenAI and [19:04.280 - 19:09.600] all of that. Okay. [19:05.920 - 19:12.160] So, there was this insane trip that [19:09.600 - 19:15.480] President Trump just took to China [19:12.160 - 19:17.520] and he had like 14 CEOs on board [19:15.480 - 19:20.480] representing roughly like 16 trillion [19:17.520 - 19:22.560] plus in market cap. [19:20.480 - 19:25.200] All going to China [19:22.560 - 19:27.640] to kind of make deals. At the same time [19:25.200 - 19:29.760] Anthropic put out this blog post [19:27.640 - 19:32.040] with a very different take. [19:29.760 - 19:34.040] Let me see if I can [19:32.040 - 19:36.840] pull it up. Oh, was Anthropic's blog [19:34.040 - 19:40.680] post in reaction to the Trump trip? [19:36.840 - 19:42.040] I don't think it was a coincidence. Hm. [19:40.680 - 19:44.040] It definitely didn't mention that [19:42.040 - 19:46.218] explicitly. [19:44.040 - 19:48.000] But interesting timing if not. [19:46.218 - 19:50.360] >> [laughter] [19:48.000 - 19:53.120] >> So it was called 2028 two scenarios for [19:50.360 - 19:54.640] global AI leadership. [19:53.120 - 19:56.920] The core thesis was that democracies [19:54.640 - 20:00.160] need to maintain a clear AI lead over [19:56.920 - 20:01.880] the CCP by 2028 because frontier AI will [20:00.160 - 20:04.040] be so powerful [20:01.880 - 20:06.160] militarily, economically, everything [20:04.040 - 20:07.640] that we need to make sure that the US [20:06.160 - 20:10.920] and democratic [20:07.640 - 20:12.280] companies are in the lead. [20:10.920 - 20:13.640] And so because of that they're saying [20:12.280 - 20:16.000] that the thing that will determine that [20:13.640 - 20:18.040] is compute. [20:16.000 - 20:20.000] And so because of that we need to be [20:18.040 - 20:22.960] really careful with compute and like [20:20.000 - 20:24.880] tighten up regulatorily [20:22.960 - 20:27.560] from a regulatory perspective uh for [20:24.880 - 20:30.040] compute, close compute loopholes [20:27.560 - 20:32.280] defend US model innovations [20:30.040 - 20:34.400] and export American AI were some of the [20:32.280 - 20:36.360] things that they talked about. [20:34.400 - 20:38.160] To start off, do you buy this argument? [20:36.360 - 20:40.240] Do you agree? [20:38.160 - 20:41.800] No, it doesn't really it's not even [20:40.240 - 20:42.880] internally consistent in this particular [20:41.800 - 20:45.680] case. They're like, "Oh, we need to [20:42.880 - 20:47.800] like, you know, restrict access and like [20:45.680 - 20:50.000] export American AI." Well, what is [20:47.800 - 20:51.640] exporting American AI and restricting [20:50.000 - 20:52.960] access? Like how how do Like those two [20:51.640 - 20:56.280] things are literally saying the exact [20:52.960 - 20:58.000] opposite things, right? Like it's not it [20:56.280 - 20:59.880] doesn't even it's not even coherent in [20:58.000 - 21:01.240] that sense. No, I think what it means is [20:59.880 - 21:03.400] like we want people buying from us, [21:01.240 - 21:05.120] Anthropic, and also we want people not [21:03.400 - 21:06.560] stealing our, you know, trade secrets. [21:05.120 - 21:08.240] Okay, we know that you're a company. [21:06.560 - 21:09.120] Yes, we get it. We know that that's what [21:08.240 - 21:11.080] you want. You didn't need to make a [21:09.120 - 21:12.680] whole essay about that. But it has [21:11.080 - 21:13.680] nothing to do with America or democracy. [21:12.680 - 21:15.920] Those are just the things that you [21:13.680 - 21:17.640] wanted as a company. Fine. [21:15.920 - 21:21.280] Un- uninteresting though. [21:17.640 - 21:22.680] Uh yeah, so I think it's not a like [21:21.280 - 21:24.640] it doesn't strike me as particularly [21:22.680 - 21:26.240] serious thought about this. It strikes [21:24.640 - 21:27.720] me as man, we have like a trillion [21:26.240 - 21:28.800] dollars at stake and I really want I [21:27.720 - 21:30.160] really want this to like, you know, I [21:28.800 - 21:32.000] want to IPO and I want this to be worth [21:30.160 - 21:33.720] something as opposed to like oh, I like [21:32.000 - 21:35.120] thought deeply. Like there aren't there [21:33.720 - 21:37.240] aren't people outside of Anthropic like [21:35.120 - 21:38.760] being from, you know, other think tanks [21:37.240 - 21:39.960] etc. being like oh, we really need to, [21:38.760 - 21:42.320] you know, do do these particular [21:39.960 - 21:42.400] policies etc. that are not paid by that [21:42.320 - 21:43.720] money. [21:42.400 - 21:45.840] >> back I'll push back on that. I think a [21:43.720 - 21:47.680] lot of people are very careful and [21:45.840 - 21:49.600] nervous and scared and it's been a whole [21:47.680 - 21:51.480] fight, right? And that's why Jensen has [21:49.600 - 21:53.240] had to lobby so much because there are a [21:51.480 - 21:55.960] lot of China hawks who are like they [21:53.240 - 21:58.480] can't buy, you know, we can't allow them [21:55.960 - 21:59.560] Nvidia products and everything else. [21:58.480 - 22:02.040] Um [21:59.560 - 22:03.880] sort of. I guess what I'm saying is the [22:02.040 - 22:05.720] the pressure for like, you know, no [22:03.880 - 22:07.520] distillation or whatever is not coming [22:05.720 - 22:09.040] from outside parties. This is like [22:07.520 - 22:11.360] Anthropic being like we want to maintain [22:09.040 - 22:13.280] our competitive edge, right? Sure. [22:11.360 - 22:14.920] >> The like the ideas about like chip [22:13.280 - 22:16.520] export controls are already covered [22:14.920 - 22:18.120] under and there that's actually like a [22:16.520 - 22:19.360] really complicated thing and has been [22:18.120 - 22:21.560] for a really long time and we've just [22:19.360 - 22:22.920] done a really bad job of doing neither [22:21.560 - 22:25.120] of the two policies that would have made [22:22.920 - 22:27.440] sense. Like one policy would have been [22:25.120 - 22:29.400] actually have export controls and not [22:27.440 - 22:30.840] send a bunch of, you know, stuff over to [22:29.400 - 22:32.280] China and then oh great, they wouldn't [22:30.840 - 22:33.600] have any chips. And then the other one [22:32.280 - 22:36.160] would be like no, no, no, just sell them [22:33.600 - 22:37.880] and and then less pressure for them to [22:36.160 - 22:39.320] go make their own industry. We kind of [22:37.880 - 22:40.680] picked an awkward middle which is like [22:39.320 - 22:42.240] kind of the worst of both worlds where [22:40.680 - 22:44.040] they have plenty of compute to train [22:42.240 - 22:47.480] their models that are almost as good as [22:44.040 - 22:49.240] the labs models here and yet they also [22:47.480 - 22:50.680] have enough incentive to like be really, [22:49.240 - 22:51.640] really serious about putting tons of [22:50.680 - 22:53.320] money towards developing their own [22:51.640 - 22:55.120] internal capacity. So [22:53.320 - 22:56.280] wow, hooray, we did a terrible, terrible [22:55.120 - 22:57.480] job. Now, there were definitely people [22:56.280 - 22:59.080] on both sides of that argument and it [22:57.480 - 23:00.920] was complicated, but we should have [22:59.080 - 23:02.760] picked and been consistent with our [23:00.920 - 23:04.840] policies. Pick a side, you know, flip a [23:02.760 - 23:06.360] coin would have been better than do half [23:04.840 - 23:08.880] of each of the things is where we ended [23:06.360 - 23:10.680] up. So, I expect us to do exactly the [23:08.880 - 23:12.400] same going forward of half of each of [23:10.680 - 23:13.560] the strategies of like, "Oh, yeah, we're [23:12.400 - 23:15.640] going to like try and sell stuff." Which [23:13.560 - 23:18.200] is literally what Anthropic thing said [23:15.640 - 23:21.061] is like, "Do Do Yes, we want them to buy [23:18.200 - 23:21.280] our stuff and also not use our stuff." [23:21.061 - 23:23.120] >> [laughter] [23:21.280 - 23:24.920] >> So, we'll just be doing that. Everyone's [23:23.120 - 23:26.240] trying to do that. So, that doesn't make [23:24.920 - 23:28.120] sense. It's a bad way forward, but [23:26.240 - 23:30.520] that's where we are. There's no question [23:28.120 - 23:32.600] that the Chinese models have Chinese [23:30.520 - 23:36.440] labs have done a lot of distillation of [23:32.600 - 23:38.880] the existing AI labs in the US, right? [23:36.440 - 23:41.200] So, that But, the problem is that that's [23:38.880 - 23:43.560] almost impossible to stop and it's not [23:41.200 - 23:46.120] really a coherent thing to even talk [23:43.560 - 23:50.520] about doing. It doesn't really make [23:46.120 - 23:52.120] sense. Like, if you make Mythos or you [23:50.520 - 23:53.920] make the next model and you sell it to [23:52.120 - 23:55.840] me and I'm like, "Hey guys, this was my [23:53.920 - 23:58.000] prompt and it gave me this essay." Like, [23:55.840 - 24:00.320] okay, I can I have to be able to publish [23:58.000 - 24:02.000] that. Like, how else would I even like [24:00.320 - 24:03.720] talk about what the heck I was doing? [24:02.000 - 24:05.400] But, doing that and there's no way that [24:03.720 - 24:07.000] can be against the terms of service, but [24:05.400 - 24:08.560] doing that alone is sufficient to [24:07.000 - 24:11.000] produce just more even just publishing [24:08.560 - 24:12.360] the essay is just making more data. And [24:11.000 - 24:13.320] then, they could back out like, "Oh, I [24:12.360 - 24:15.400] wonder what the prompt could have been." [24:13.320 - 24:17.520] Like, you're just making more data by [24:15.400 - 24:19.440] making AI able to do all these things so [24:17.520 - 24:21.840] cheaply. So, it's and that's going to [24:19.440 - 24:23.080] end up somewhere on the internet and [24:21.840 - 24:24.800] then it's going to be possible for [24:23.080 - 24:26.360] people to make way more, you know, we're [24:24.800 - 24:27.960] we're making so much more code, so many [24:26.360 - 24:30.280] more open-source projects, right? All [24:27.960 - 24:32.120] that code can be used to distill the [24:30.280 - 24:34.680] next version of the model, right? So, [24:32.120 - 24:36.360] it's or like to distill the the weights [24:34.680 - 24:38.320] effectively, right? It's Yeah, it's one [24:36.360 - 24:40.240] hop away, but it doesn't It doesn't It's [24:38.320 - 24:42.320] not really coherent. Also, I take a [24:40.240 - 24:44.480] little bit of um [24:42.320 - 24:47.320] offense to this idea that like, Umbrage? [24:44.480 - 24:49.480] You take umbrage? I do take umbrage [24:47.320 - 24:51.600] uh to this idea that like, "Oh, you [24:49.480 - 24:52.840] know, it's like so bad to be distilling [24:51.600 - 24:55.120] model weights. Like, how can you do [24:52.840 - 24:56.520] that? Like, how's that so illegal?" [24:55.120 - 24:58.320] Like, "Oh, where did that come from? You [24:56.520 - 25:00.800] distilled all of human knowledge, huh? [24:58.320 - 25:01.760] Who made all of human knowledge? [25:00.800 - 25:03.560] Us? [25:01.760 - 25:04.840] Yeah, okay, cool. So, like, why does it [25:03.560 - 25:06.720] stop right there when you're making [25:04.840 - 25:08.600] money and not like upstream of that? [25:06.720 - 25:09.520] When especially when literally all of [25:08.600 - 25:11.080] these companies, I think with the [25:09.520 - 25:14.440] exception of Google, [25:11.080 - 25:16.080] um have like in court admitted to like, [25:14.440 - 25:17.720] "Oh, yeah, we pirated a bunch of books." [25:16.080 - 25:19.400] Like, that that's pretty straight-up [25:17.720 - 25:21.280] illegal. Like, there's a bunch of stuff [25:19.400 - 25:23.040] that was, you know, it wasn't like, "Oh, [25:21.280 - 25:24.840] you know, well it was all public data [25:23.040 - 25:25.360] and like we were existing in existing [25:24.840 - 25:27.560] regime." No, no, it's straight-up [25:25.360 - 25:29.760] illegal. [25:27.560 - 25:31.240] Like, completely against the law. Like, [25:29.760 - 25:33.720] you were not supposed to do that. Like, [25:31.240 - 25:35.320] had you been fined at the maximum [25:33.720 - 25:36.480] amount, it would have put Anthropic out [25:35.320 - 25:38.120] of business. [25:36.480 - 25:40.840] Like, it just would have been [25:38.120 - 25:42.040] like No, it was like a lot a lot cuz [25:40.840 - 25:43.280] like if you look at the number of works [25:42.040 - 25:45.360] and the maximum penalty for a [25:43.280 - 25:47.200] copyrighted work from a punitive sense, [25:45.360 - 25:48.320] which is what it should be in in some [25:47.200 - 25:49.240] like you can argue that it should be [25:48.320 - 25:50.520] punitive because you would want to [25:49.240 - 25:54.000] prevent people from doing this type of [25:50.520 - 25:56.720] thing. That is like $100,000 per copy. [25:54.000 - 25:58.360] And so, it's like, "Okay, well, that's [25:56.720 - 26:00.040] like, you know, [25:58.360 - 26:01.480] way way more money than they had." So, [26:00.040 - 26:03.000] you could have zeroed them out from like [26:01.480 - 26:04.600] some judge could have been like, "Yep, [26:03.000 - 26:06.240] that's it. Like, we're we're zeroing you [26:04.600 - 26:09.200] out and returning all, you know, money [26:06.240 - 26:10.880] to you know, the government, I guess, or [26:09.200 - 26:13.040] or the people or something." Of course, [26:10.880 - 26:14.360] didn't happen. They settled for nothing, [26:13.040 - 26:16.000] right? As part of one of their earlier [26:14.360 - 26:18.320] fund raises. [26:16.000 - 26:20.000] But, still I think it's a little hard to [26:18.320 - 26:21.200] make this claim that like, "Oh, yeah, [26:20.000 - 26:23.080] like we did everything perfectly and [26:21.200 - 26:25.760] like now people can't, you know, filter [26:23.080 - 26:26.840] our weights." It's a little, yeah. Yeah. [26:25.760 - 26:29.880] So, [26:26.840 - 26:32.200] what is your recommended action plan [26:29.880 - 26:34.120] then? Like, nothing and just no [26:32.200 - 26:36.280] exploitations? Kind of like [26:34.120 - 26:37.680] distillation's going to happen? Just I [26:36.280 - 26:38.960] think accept that distillation is going [26:37.680 - 26:41.520] to happen. I don't know how you get away [26:38.960 - 26:42.800] around that one. Like, it's you as long [26:41.520 - 26:45.240] as you let people use it, it's going to [26:42.800 - 26:46.960] happen. And also, like if you let [26:45.240 - 26:48.560] someone use the model, how do you know [26:46.960 - 26:50.320] that person isn't just working for the [26:48.560 - 26:52.280] Chinese government? Like, [26:50.320 - 26:54.120] China could just pay people in the US to [26:52.280 - 26:55.760] just things for them. There's nothing [26:54.120 - 26:58.240] wrong with that. It's like part of [26:55.760 - 26:59.680] global trade, etc. Or China could pay [26:58.240 - 27:01.400] someone in Kenya who pays someone in, [26:59.680 - 27:02.760] you know, Liechtenstein to pay someone [27:01.400 - 27:04.120] like, okay, sure. Like, it doesn't [27:02.760 - 27:06.960] really matter how many hops there are. [27:04.120 - 27:08.920] Like, it's going to end up using AI to [27:06.960 - 27:10.960] do a lot of stuff for a lot of people. [27:08.920 - 27:12.400] So, maybe you could just give it up, [27:10.960 - 27:14.280] stop thinking about it like that, and [27:12.400 - 27:16.000] just go to the long-run end game of [27:14.280 - 27:17.840] like, okay, cool. So, like, tokens are [27:16.000 - 27:19.400] like electricity. Now, let's think about [27:17.840 - 27:20.600] like what where do we want to be as a [27:19.400 - 27:21.800] world as [27:20.600 - 27:23.600] Like, let's just think about how we want [27:21.800 - 27:26.320] this stuff to work instead of like some [27:23.600 - 27:28.200] weird, oh, there's two possible futures, [27:26.320 - 27:29.960] one with AI in China, and one with the [27:28.200 - 27:33.440] US. Like, no, there's a lot of possible [27:29.960 - 27:36.360] futures, and all of them end up in 2028, [27:33.440 - 27:38.520] maybe that's the case. In 2032, it's all [27:36.360 - 27:40.480] going to be [ __ ] super powerful, [27:38.520 - 27:43.120] uh and available to everybody. So, it [27:40.480 - 27:45.120] doesn't really matter. Yeah. It also [27:43.120 - 27:46.120] kind of caps out over time. I think one [27:45.120 - 27:48.280] of the other things that we talked [27:46.120 - 27:50.040] about, uh we'll we'll get into it to [27:48.280 - 27:51.760] that as well. I don't believe this like [27:50.040 - 27:53.880] oh, we're going to have this like, you [27:51.760 - 27:55.480] know, this edge of like democracies over [27:53.880 - 27:56.920] China, and it's going to be compounding. [27:55.480 - 28:00.440] That's not how these things go. It's [27:56.920 - 28:00.440] diminishing, not compounding.

Vibes & Benchmarks Ep 02: Will Google outlast Anthropic and OpenAI?