00:11:36 · John Coogan
[00:11:36] S0: But the question is no longer is it artificial? We know it is. The question is no longer is it intelligent? We know it is at a lot of different tasks. The question is how general is it? Because Blender is open source, can be downloaded and copied a million times into reinforcement learning environments. We know that OpenAI bought quote tens of thousands of Mac minis and Mac studios to do computer use training on. And so the question is, is it only good at Blender? Will it work in Cinema four d? Will it work in Houdini? Will it work in a small piece of software that does three d modeling that has been fall has fallen by the wayside? Now there's an interesting dynamic here where I was in the boat of I could see that Blender was going to be the new hot thing, the best tool that I should have moved, but I already had so many plug ins and textures and rendering and I knew the layout that I downloaded Blender, opened it up and I was like, I don't really don't want to have to get over this learning curve even though I know in five years Blender is going be way better than Cinema four d. There's this switching cost. Now, you're in this weird world where it's pretty good if you're a software company where the Frontier Labs bullseye lands on you. And I'm I'm thinking about Slack, like Alex Wang at Meta just moved them to Slack off of Teams where they had like an internal workspace thing. And specifically he said For Benioff. Give it up for Benioff, truly. And specifically he said Slack is better for work with AI agents. And why is that? Well, it's because the Frontier Labs all use Slack and so they have Slack as a reinforcement learning environment and so it got really good at Slack. And I think that even though there are open source Slacks and you can vibe code a Slack competitor, as we have tested, Slack was the logical tool to pull off the shelf for the AI companies, for the AI agents, for the models and Blender seems to be that. So the interesting thing is that I I do wonder if people will wind up shifting from Cinema four d to Blender because the models are better at Blender? Or will the models generalize and say, yeah, this looks a little bit different because it's a slightly different piece of software, but I've trained so many hours in Blender that I can also do Cinema four d or also do Houdini.
00:13:51 · John Coogan, Jordi Hays
Because some of those other softwares are better at certain things. Some of them are better at smoke simulations. Some of them are better at water simulations. There are advantages Yeah. To each piece. I don't want to degrade any particular piece of three d software. I love them all equally. [00:14:08] S1: But I think There are some reports that it's that it's quite effective at at video editing.
00:14:14 · John Coogan
[00:14:14] S0: Cool. So I'd be interested to to to take it for a spin. I wonder the the interesting thing is that most video is still edited in Adobe Premiere. It's a closed source Adobe, you know, they they have their guardrails. They probably they might not want to, you know, say, oh yeah, sure. Take 10,000 licenses. No problem. Pay us $10 a month for that. They might they might have a different opinion. They may want an AI agent that's internal. There there might be all sorts of different things. But then again, there are pro there I mean, I know that there are open source video editing suites out there. I think that they lag a little bit, but there are some that are free or or or open source at least. The the other interesting thing, I mean, video editing, like all these things, like taste is so important in these. It's it's it's often less about the the the exact the exact instructions and more about the, like, telling the right story, using the right tools, not creating slop whether you're doing it yourself or with an AI agent. But I I think that gets to the core question of like how do you evaluate these model launches. We were discussing the idea that like the benchmark era might be coming to a close in some in some way because as soon as the benchmark's created, it's hill climbed very, quickly and then it's at 99 or a 100% pretty quickly and everyone sort of their eyes start glazing over in benchmark discussions. And it just goes back to like the vibe check. I think the the the best thing to do when a model comes out is just take it for a spin. And mostly, point it at something that you know really well because you don't want to be in the you you want to expose the models. What what what's the what's the gal man amnesia? That's the one I'm thinking of. You basically want to test if gal man amnesia is triggered by you using the model. So point it at something that you intimately understand and see if you're impressed. And then Yeah. Because you pointed at something What? Yeah. Yeah. We got to do the horse bench again. We have a couple couple more. Let's see comedy. I wanna I wanna see some new jokes. Shrimp fried rice bench. Also interesting, Gell Mann Amnesia not created by Murray Gell Mann, the physicist who was at Caltech who lived in my hometown and his house is actually for sale right now.
00:16:47 · John Coogan, Jordi Hays
But he was having coffee regularly with another physicist. I forget what his name was but he's really famous. And the other guy was like, we noticed this together, I'm naming it after you. Do you have the story? Did you look this up? The origin? [00:17:03] S1: It says it's named after Nobel Prize winning physicist, Gale Mann. [00:17:06] S0: Yes. It's named after him but it was coined by his buddy who was another physicist who was like [00:17:11] S1: Michael Crichton. [00:17:12] S0: Is that it? Okay. Yes. Yeah. So they were hanging out, having coffee, they realized this thing and it's like [00:17:19] S1: Tyler should name this specific situation. Yeah. Cosgrove Paradox. [00:17:23] S0: Where where Where where someone asks you to fact check something they don't really know anything about. Anyway, we can move on because there's a whole bunch more news. Let me tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real time experiences and new value with Cisco. So oh, there are other reactions to Astra we should go through in the timeline. The big one comes from Joe. He says, you're never gonna believe it, but someone one prompted the new model to make a completely new and immersive three d game. Yes. The bar has already been raised. The goalposts have been moved. It's not enough to make a game. It has to be funny, entertaining. And I don't know. I I know how I I feel like a lot of people are taking the easy way out with these web based games. Playing a game on a MacBook Pro in a Chrome tab, that's not real gaming. It's gotta be it's gotta be running
00:17:23 · John Coogan, Jordi Hays
[00:17:23] S0: Where where Where where someone asks you to fact check something they don't really know anything about. Anyway, we can move on because there's a whole bunch more news. Let me tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real time experiences and new value with Cisco. So oh, there are other reactions to Astra we should go through in the timeline. The big one comes from Joe. He says, you're never gonna believe it, but someone one prompted the new model to make a completely new and immersive three d game. Yes. The bar has already been raised. The goalposts have been moved. It's not enough to make a game. It has to be funny, entertaining. And I don't know. I I know how I I feel like a lot of people are taking the easy way out with these web based games. Playing a game on a MacBook Pro in a Chrome tab, that's not real gaming. It's gotta be it's gotta be running [00:18:25] S1: In Steam? [00:18:26] S0: PC, on Steam, or on a PS five, or on an Xbox. It's gotta it's gotta have more to it than just some fun game that you can play around within the browser. I do love that the browser games are coming back. Like, don't know if either of you ever experienced like the Flash era, but like Flash games were huge. Line Rider. Line Rider. Okay. What about you? [00:18:46] S1: In elementary school, we'd play on cool math games. [00:18:49] S0: Cool math games. Okay. What what game? Line Rider or something else? [00:18:53] S1: I don't know. I don't know. I don't remember. There's so many.
00:18:53 · Jordi Hays, John Coogan, Bar Winkler
[00:18:53] S1: I don't know. I don't know. I don't remember. There's so many. [00:18:56] S0: Was Sorry? Just like I paid You put any money on it ever? Didn't make any money? Let's see. There's also some updates. Token pricing is effectively meaningless now. The benchmarks are meaningless. The token pricing charts are meaningless. Everything's meaningless. No. So Steven Heidel is sharing the cost efficiency, how you should think about cost efficiency in the modern era. So 3.8 Flash looks 13 times cheaper when measured per token but Astra is cheaper per task since it's far more efficient. Measure your cost per task, not per token. So interesting. I will we see a decline in if models get more token efficient, will we see a decline I mean we probably won't see an actual decline in in token volumes. But we've seen Sundar Pichai from Google and Alphabet stand on stage and talk about the exponential increases in token volumes. But you might see a kink in that graph or like a deceleration to some extent if models just get way more token efficiency. Like if the next version of Gemini or 3.8 Flash is 100x more token efficient, you might see a little bit less growth in token volumes but they might be doing more valuable work than ever. So sort of an interesting nuance you got to adapt to the modern times, I suppose. So the intelligence index versus the output tokens per intelligence task, There's a whole bunch of charts you can dig through. Let me tell you about the New York Stock Exchange. Wanna change the world? Raise capital at [00:20:33] S2: the New York Stock Exchange.