00:32:23 · Jordi Hays, John Coogan
[00:32:23] S0: I I'm interested. [00:32:25] S1: You can see the YouTube video here, the preview of the nine six four turbo junior high performance version. It's an absolutely crazy, crazy video that I got organically served to me. And I was like, this is what we've been looking for, for sure. This is exactly what we were talking about. I need one. I think you do. [00:32:45] S0: I need one. [00:32:46] S1: Anyway, GPT five GPT Live, a new generation of voice models for natural human AI interaction launched today from OpenAI, rolling out to ChatGPT starting today, taking shots at that guy on Instagram who's unsatisfied with legacy voice models. It's a lot smarter. [00:33:05] S0: I have a feeling he's gonna find a way to keep using the old models. [00:33:09] S1: Oh, maybe. Or he'll just become who's the the Liberator? He's sort of the Pliney the Liberator of voice models in that I know he's taking shots and he's like very unhappy with the quality of the voice models, but at the same time, he does a great job of actually exposing and finding limitations, not necessarily jailbreaks, but he he by watching those videos, you can very clearly understand, okay, I understand what's going on here. It's not it doesn't have tool use, so it can't count. It can't set a timer. It doesn't have a long running context window that can stay with you for a long time, so it gets lost. You can't talk to each other. There's a whole bunch of different ways. And once you have all of those documented by just a he's basically like a white hat hacker. He's he's He found a
00:33:09 · John Coogan, Tyler, Will Mayer, Ben Thompson, Jordi Hays, Rodrigo Liang
[00:33:09] S1: Oh, maybe. Or he'll just become who's the the Liberator? He's sort of the Pliney the Liberator of voice models in that I know he's taking shots and he's like very unhappy with the quality of the voice models, but at the same time, he does a great job of actually exposing and finding limitations, not necessarily jailbreaks, but he he by watching those videos, you can very clearly understand, okay, I understand what's going on here. It's not it doesn't have tool use, so it can't count. It can't set a timer. It doesn't have a long running context window that can stay with you for a long time, so it gets lost. You can't talk to each other. There's a whole bunch of different ways. And once you have all of those documented by just a he's basically like a white hat hacker. He's he's He found a [00:33:53] S5: bunch of benchmarks. [00:33:55] S1: Oh, wait. What? [00:33:55] S5: He found a bunch [00:33:56] S7: of, like, [00:33:56] S8: good benchmarks. [00:33:56] S1: Yeah. Yeah. Yeah. Basically benchmarks, like if you can satisfy his his videos and run those experiments again in a satisfying way, you probably have a better product at the end of the day. So I don't know. I think it's I think it's cool. I'm excited. [00:34:09] S0: Very yeah. I'm super excited for this one. I I wonder if the yeah, the real time nature unlocks [00:34:17] S4: Mhmm. [00:34:17] S0: Like, under unlocks, like, almost like a new kind of, like, pod like, real time, like, podcast format where you can just actually have a conversation about a topic. [00:34:26] S1: Yeah. [00:34:27] S0: Right?
00:34:27 · Jordi Hays, John Coogan
[00:34:27] S0: Right? [00:34:27] S1: Yeah. We should have it sitting right here. It has a full duplex architecture, which means that it can listen and speak at the same time so it can jump in appropriately. [00:34:37] S0: And it can also route, like, certain questions to smarter models Yes. To run-in the background. You can still have a conversation. [00:34:44] S1: Yes. It's [00:34:45] S0: very cool.
00:34:45 · John Coogan
[00:34:45] S1: Which is great because my default flow, like, if I'm driving, which I don't think I should be doing, but I will tap the record button, not the voice mode, and I will dictate a really long prompt. And then I will click send and then I will wait for it to cook and then I will have it read that back to me. And that's all just like a couple extra clicks, couple extra steps to get what I really want, which is a conversation. I was interested about the idea of where this goes. Like, is there is there a world where you are talking to a model and you can have sort of multiple threads of conversations going within one conversation? So you can say, hey. I need you to go and do a deep research report on, you know, IBM Watson, and it goes and cooks on that. And then while it's doing that, you're still talking to it and you're still having a conversation. But then you also might say, Well, catch me up on the latest news, and it has that at the fingertips. So, it can talk to you about the latest news and then you can go and it can say, Hey, by the way, I'm done with the report, or I'm done with the code that I wrote, but you're talking to me about other problems in your organization, other decisions, other research projects that you have and it's all done in one continuous conversation like, you know, me interacting with you. Like, you know, we might you might delegate something and then continue the conversation and then move on. Oh, I got to text this person. Okay. I'll text them. They'll do that. They'll come back to me in twenty minutes with a revision to the thing that we're working on. But in the meantime, let's talk about whatever else we got going on today. So interesting to see where this goes. In other OpenAI news, GPT 5.6 Soul, along with Terra and Luna will will launch publicly this Thursday. Preview access is expanding globally. Lots of people are having fun with this. Some people got early access. Other people were joking about clearly not getting early access but wanting early access. But a lot of people are having fun. Tyler Cowen? What
00:36:46 · Jordi Hays, John Coogan
[00:36:46] S0: did John Palmer say? [00:36:47] S1: What did he say? He's in here? [00:36:49] S0: No, he had a [00:36:51] S1: He has one. He says, It seems I'm allowed to talk about Five Six now. I wasn't one of the testers, but I heard it's really good. Again, I haven't used it yet, but this will probably be my default model from here on when I get access. It's a good one. People are having a lot of fun. Kit Langton says, I finally talk about GPT 5.6. I've been using it for many months, if not decades. And what is glaringly obvious to me now as I look back on my life is that you did not have access to it. Oh, it's very fun. People are very happy with it. Interesting. There's a there's a little bit of a fork around, like, when how the the vibe around 5.6 is different than Fable five, when how interactive it is versus, somebody said, like, Fable is, like, light speed across the galaxy and, and Soul is like a Porsche, the Nurburgring, more engaging. I don't know. It seems like we're we're entering, some some fragmentation in, like, the right tool for the job.
00:37:51 · Jordi Hays, John Coogan
[00:37:51] S0: Yeah, Ethan says my big takeaway is that both Soul, Ethan had access. So he says both Soul and Fable represent jumps over previous models and have opened a large gap with the next best AIs. People will have preferences for one or the other, but if you're doing any work where better intelligence matter, those two models are your only choices. Dean says, I think for me, the main takeaway with Soul and Fable is that I can't remember a time when the leading models were A, so decidedly ahead of everything else, and b, so distinct from one another. So Yeah. So, yeah, a lot of people are talking about how they're using the models Yeah. Together. [00:38:26] S1: It's probably a function of just tighter tighter distillation protocols, like there's less pollution. Like, also, like, not everyone's training on the exact same data anymore because it used to just be everyone just trains on the entire web corpus. And now each company has different data sets, different data brokers, different usage patterns that they can train on, and then different sort of research tastes. So go give it a