WEBVTT NOTE Built from Descript SRT export of the edited "Full" composition (human-reviewed speaker labels) 00:00:34.860 --> 00:00:35.480 Alex Volkov: Welcome everyone. 00:00:35.480 --> 00:00:36.310 Welcome to ThursdAI 00:00:38.312 --> 00:00:39.412 September 10th. 00:00:39.742 --> 00:00:41.582 Welcome Wolfram, welcome Nisten. 00:00:41.602 --> 00:00:42.342 How are you guys? 00:00:42.402 --> 00:00:48.372 Welcome everybody who's tuning in to our live show, ThursdAI, and we're excited 00:00:48.542 --> 00:00:51.772 to tell you all about this last week. 00:00:52.152 --> 00:00:55.322 there's been some twists and turns this last week for sure. 00:00:55.672 --> 00:00:57.582 So I'm super excited to have you guys here. 00:00:57.632 --> 00:00:59.422 Wolfram, Nisten, how you guys doing? 00:01:00.462 --> 00:01:00.842 Wolfram Ravenwolf: Excited. 00:01:00.842 --> 00:01:01.882 But we will cover this- 00:01:02.142 --> 00:01:04.692 Alex Volkov: Yeah, we have to talk about this- this strange week … this, 00:01:04.692 --> 00:01:06.822 this Ant- Anthropic guy, right? 00:01:06.832 --> 00:01:07.842 That, that quit. 00:01:07.952 --> 00:01:08.832 We have, we have to cover this. 00:01:08.882 --> 00:01:09.212 Wolfram Ravenwolf: weeks. 00:01:09.702 --> 00:01:09.932 Just 00:01:09.932 --> 00:01:14.252 Alex Volkov: and I think now clocking in 130 million for that, "I leave 00:01:14.302 --> 00:01:15.922 'cause we're all gonna die" tweet. 00:01:16.232 --> 00:01:17.472 we definitely have to cover this. 00:01:17.542 --> 00:01:23.382 there's also some open source news with DeepSeek, V1, V4.1 00:01:23.382 --> 00:01:24.002 Flash. 00:01:24.592 --> 00:01:26.722 Hard to keep up with the DeepSeek, versioning. 00:01:26.722 --> 00:01:30.212 However, those are dope, and I am super excited to tell you guys that, 00:01:30.242 --> 00:01:35.482 Chris Alexiou from NVIDIA, Nisten's, Nisten's, neighbor in Canada is going 00:01:35.522 --> 00:01:37.742 to, to join us to talk about DeepSeek. 00:01:38.232 --> 00:01:39.512 Nisten, I bet you didn't know this. 00:01:39.522 --> 00:01:40.542 How are you doing, sir? 00:01:40.572 --> 00:01:41.772 What's new in your world? 00:01:41.772 --> 00:01:44.142 What's new in the world of AI that got you excited? 00:01:44.582 --> 00:01:46.672 Nisten Tahiraj: I, I'm, I'm still alive. 00:01:46.872 --> 00:01:49.892 The world has not ran out of, problems to solve. 00:01:50.512 --> 00:01:54.632 Something exciting coming from my end probably next week too. 00:01:55.382 --> 00:01:56.742 And, yeah, yeah. 00:01:58.372 --> 00:01:58.982 Things, Yeah. 00:01:59.082 --> 00:02:00.482 Busy- What- … busy benchmarking. 00:02:01.642 --> 00:02:01.982 Alex Volkov: Yeah. 00:02:02.282 --> 00:02:06.762 I've spent my whole weekend, as I imagine many of you as well, Astra 00:02:07.032 --> 00:02:09.222 maxing and Fable maxing together. 00:02:09.672 --> 00:02:11.822 I think I've burned through two resets. 00:02:11.872 --> 00:02:13.342 Wolfram, I know you're banking yours. 00:02:13.692 --> 00:02:18.562 But- Yeah I, c- can I, can I just go on a personal mea culpa for just 00:02:18.562 --> 00:02:20.022 one second towards the audience? 00:02:20.532 --> 00:02:22.762 I claimed on ThursdAI newsletter… 00:02:22.812 --> 00:02:24.142 We don't do five hours. 00:02:24.152 --> 00:02:26.122 This is the longest show we ever did, and the reason for that is 00:02:26.122 --> 00:02:30.142 that we were waiting for Astra to drop, and GPT-6 is a big moment. 00:02:30.152 --> 00:02:34.722 We started when GPT-4 came out, so we've been tracking now two proper generations 00:02:34.742 --> 00:02:36.922 of OpenAI's model, so we waited. 00:02:37.222 --> 00:02:38.802 And then there was hiccups, and we waited. 00:02:38.852 --> 00:02:43.272 Anyway, this resulted in a very long c- content piece, and this 00:02:43.282 --> 00:02:44.992 resulted in us posting two episodes. 00:02:45.142 --> 00:02:49.012 So if you have our, our podcast, from last week, there's two episodes, one with the 00:02:49.012 --> 00:02:50.782 regular ThursdAI News TLDR, et cetera. 00:02:50.972 --> 00:02:54.022 We talk about exciting things like World Labs and, and different 00:02:54.022 --> 00:02:57.660 things, and then- The other one is fully dedicated to GPT-6 Astra. 00:02:57.660 --> 00:03:01.140 And then we had Peter Goste and, Ryan Carson talk about their experiences. 00:03:01.140 --> 00:03:01.960 They had, early access. 00:03:03.010 --> 00:03:03.890 And then we got access. 00:03:04.710 --> 00:03:11.250 And then, on the page that I had Astra built for itself, maybe I 00:03:11.250 --> 00:03:13.650 should show it, I wrote AGI is here. 00:03:14.410 --> 00:03:18.130 And then I saw Jensen Huang from Nvidia say AGI is here. 00:03:18.910 --> 00:03:23.930 And then the release of Astra definitely felt like, hey, AGI is here. 00:03:24.880 --> 00:03:32.090 so my p- mea culpa for, for last week is that G- AGI is not here . At least, 00:03:32.350 --> 00:03:34.250 at least not the way that I imagined it. 00:03:34.250 --> 00:03:40.040 I had like, I, I've been like t- talking maxing Astra and, I 00:03:40.040 --> 00:03:45.640 don't know, folks, it's, it's amazing but there's buts for sure 00:03:46.898 --> 00:03:51.718 Nisten Tahiraj: It, it built the best Mars driver simulator, 00:03:51.968 --> 00:03:52.218 Alex Volkov: Yeah 00:03:52.328 --> 00:03:55.548 Nisten Tahiraj: I didn't find it as good on DevOps a- and stuff. 00:03:55.548 --> 00:03:57.448 Like, I, I'm just, the way I'm… 00:03:58.068 --> 00:04:02.598 I think Claude has trained me at this point to be opinionated how , how, how 00:04:02.598 --> 00:04:06.618 Claude does things, so I, I did not like how it was doing certain things. 00:04:07.378 --> 00:04:12.738 so I'm still at Fable for that, but for 3D stuff, this is, this is king right now. 00:04:12.868 --> 00:04:14.698 Alex Volkov: Yep, everybody's posting the 3D stuff. 00:04:14.708 --> 00:04:17.278 By the way, what I'm showing on stage here while Nisten talks about 00:04:17.278 --> 00:04:20.738 the 3D simulator is the page for Astra, the Astra build for itself. 00:04:20.798 --> 00:04:23.818 That page sits at thursdaii.new/app/Astra, 00:04:23.998 --> 00:04:27.698 and you can see this, like, 3D animation here, with Fable and Sol and 00:04:27.708 --> 00:04:29.588 Astra being, like, a star of clouds. 00:04:29.848 --> 00:04:35.558 I did this with Astra Ultra High on fast mode, and I ran 00:04:35.558 --> 00:04:37.028 out of tokens very quickly. 00:04:37.738 --> 00:04:45.508 I, I am so used to codecs not ever reaching my quotas that I just ran on 00:04:45.548 --> 00:04:49.718 ultra fast, and I've hit two resets, and then OpenAI got me a reset, so I'm out. 00:04:49.838 --> 00:04:52.038 Wolfram Ravenwolf: There were also some issues with how it was 00:04:52.178 --> 00:04:53.528 counting the tokens, so they have 00:04:53.838 --> 00:04:54.268 Alex Volkov: been making improvement. 00:04:54.268 --> 00:04:55.728 Nisten Tahiraj: he was fine for me. 00:04:55.798 --> 00:04:56.208 Alex Volkov: Go ahead, Wolfram. 00:04:56.208 --> 00:04:56.248 Sorry. 00:04:56.288 --> 00:04:59.173 Wolfram Ravenwolf: OpenAI also said that there have been issues with how the usage 00:04:59.173 --> 00:05:03.463 was counted, and I think they gave a free reset for that because of the issues, 00:05:03.503 --> 00:05:05.723 and, they are continuously improving it. 00:05:06.533 --> 00:05:09.663 So depending on how you used it, that may be part of the issue. 00:05:09.803 --> 00:05:12.393 Alex Volkov: Anyway, the page is one of the cleanest pages that 00:05:12.503 --> 00:05:14.273 I've ever seen built on ThursdAI. 00:05:14.493 --> 00:05:17.053 It pulled up all the evals for itself. 00:05:17.383 --> 00:05:19.403 there's just an incredible amount here. 00:05:19.603 --> 00:05:24.053 Also, it cut, I think, five, five clips from the episode itself 00:05:24.053 --> 00:05:25.063 where we talk about OpenAI. 00:05:25.363 --> 00:05:30.293 This is by far one of the best, one of the best, like, pages that we've built. 00:05:30.613 --> 00:05:31.913 Let's say welcome to LDJ. 00:05:31.913 --> 00:05:33.543 LDJ, welcome, man. 00:05:33.543 --> 00:05:34.033 How are you? 00:05:34.033 --> 00:05:35.363 What's, what's new in your world? 00:05:35.383 --> 00:05:36.933 What can you t- tell us about? 00:05:38.203 --> 00:05:40.373 LDJ: DeepSeek V- V 4.1 00:05:40.393 --> 00:05:43.703 Flash, had just dropped within the past 24 hours. 00:05:43.743 --> 00:05:44.533 That's exciting. 00:05:45.103 --> 00:05:48.933 And of course, Astra, a lot of things to talk about with Fable 5.1 00:05:48.933 --> 00:05:51.403 versus Astra and really strength and weaknesses there. 00:05:55.373 --> 00:05:57.673 Alex Volkov: Okay, so folks, I think that, folks in the comments 00:05:57.673 --> 00:05:59.943 can also chime in with their experience with Astra and Fable. 00:05:59.943 --> 00:06:04.253 But, I think it's very clear that, it's been a week that folks have 00:06:04.253 --> 00:06:07.933 been playing with Astra and a little bit over a week that folks 00:06:07.933 --> 00:06:10.213 have been playing with Fable 5.1. 00:06:10.263 --> 00:06:14.363 and, w-we definitely should go into, like, a full corner of, like, our experiences, 00:06:14.363 --> 00:06:16.243 what we built and, and how this feels. 00:06:16.863 --> 00:06:21.783 However, I, I do think that, we should also talk about open source and, I just 00:06:21.783 --> 00:06:24.913 to tell you that in f- in seven-ish minutes, we're gonna have Chris 00:06:24.943 --> 00:06:29.653 Alexiou from NVIDIA joining us, to talk about, the new DeepSeek that dropped. 00:06:29.983 --> 00:06:35.153 But before that, I think let's go into the TLDR, and then I will tell 00:06:35.153 --> 00:06:39.503 you, you know, TLDR is the corner where we basically, tell you about 00:06:39.513 --> 00:06:40.463 everything that's going to happen. 00:06:40.713 --> 00:06:42.633 By the way, if you are… 00:06:42.683 --> 00:06:45.053 if you guys just go to ThursdAI.live, 00:06:45.503 --> 00:06:46.433 please, please do so. 00:06:46.693 --> 00:06:53.583 and, you will see that we have an AI-produced show with chyrons and 00:06:53.593 --> 00:06:57.853 everything and, h-hopefully, hopefully, the chyrons will show up, in the 00:06:57.853 --> 00:07:01.493 right way because I'm using a new AI now to run the show, so we'll see. 00:07:02.143 --> 00:07:03.333 It just happened this morning. 00:07:03.653 --> 00:07:04.813 if you go to ThursdAI.live, 00:07:04.813 --> 00:07:06.093 you also see our TLDR. 00:07:06.253 --> 00:07:09.763 So actually, yeah, let's go there and, and, and, and walk through the 00:07:09.763 --> 00:07:11.363 topics that we'll cover on the show. 00:07:11.363 --> 00:07:13.843 Obviously, we're waiting for breaking news, and obviously we'll have 00:07:13.843 --> 00:07:15.923 Chris Alexiou, join us as well. 00:07:16.403 --> 00:07:17.513 Let's see. 00:07:17.623 --> 00:07:21.623 Yes, John Nemotron is coming on again, a friend. 00:07:22.053 --> 00:07:24.273 Nisten Tahiraj: When I've been multitasking and searching 00:07:24.293 --> 00:07:26.773 for Alex, I searched Joni. 00:07:26.783 --> 00:07:27.253 What's wrong? 00:07:28.723 --> 00:07:29.453 Alex Volkov: All righty, folks. 00:07:29.453 --> 00:07:30.173 Let's do this. 00:07:30.173 --> 00:07:33.463 Let me run through the TLDR on the show. 00:07:34.173 --> 00:07:39.229 all right, so obviously We're missing s-- All right. 00:07:40.229 --> 00:07:43.949 welcome to the TLDR for ThursdAI, live on September 10th. 00:07:43.949 --> 00:07:47.399 My name is Alex Volkov, AI Evangelist with CoreWeave and Weights & Biases. 00:07:47.639 --> 00:07:51.569 our Fully Connected show is coming up, and I have an exciting announcement 00:07:51.779 --> 00:07:56.909 that's not AI related in any way, unless he's secretly an AI wizard, but he's a 00:07:56.909 --> 00:08:00.619 very, very known person in the world, to tell you about later on the show. 00:08:00.619 --> 00:08:06.379 Today with me, Nisten Tahiri, LDJ and, Wolfram Ravenwolf, and we'll have 00:08:06.379 --> 00:08:10.709 Chris Alexiou join us just momentarily to talk about the biggest open 00:08:10.709 --> 00:08:12.669 source release f- of this last week. 00:08:12.719 --> 00:08:14.069 Actually, yeah, let's start with open source. 00:08:14.099 --> 00:08:18.549 DeepSeek releases v.4.1 00:08:18.609 --> 00:08:19.839 Flash, 4.1 00:08:19.839 --> 00:08:20.129 Flash. 00:08:20.199 --> 00:08:21.839 Let's remove the V 'cause it's confusing. 00:08:22.439 --> 00:08:22.939 It's easier. 00:08:23.239 --> 00:08:27.969 A fe- half a bil- half a trillion MoE with only eight billion parameter or 00:08:27.969 --> 00:08:32.089 16 billion parameter, looks like a very interesting thing to talk about. 00:08:32.959 --> 00:08:34.989 And, 4x KV cache compression. 00:08:35.149 --> 00:08:38.479 DeepSeek has been compressing KV cache like crazy lately and, this 00:08:38.479 --> 00:08:42.679 model will blow your mind, I promise, at least on benchmarks, at least 00:08:42.679 --> 00:08:43.919 on the stuff that it achieves. 00:08:44.419 --> 00:08:48.719 Also from Europe, there is a new lab called DesertAnt, debuts 18 on-device AI 00:08:48.719 --> 00:08:50.209 models with open weights and native SDKs. 00:08:50.759 --> 00:08:53.939 Europe's, another claim to fame after Mistral. 00:08:54.069 --> 00:08:56.369 very interesting and Wolfram would love to hear from you more 00:08:56.389 --> 00:08:58.129 as a European representative. 00:08:58.509 --> 00:09:01.789 And also Inclusion released Link Three Flash VL. 00:09:02.469 --> 00:09:05.819 It's a MoE with vision language and 5.5 00:09:05.839 --> 00:09:06.869 billion active parameters. 00:09:07.169 --> 00:09:10.189 That's in open source, and again, Chris Alexiou will join us 00:09:10.189 --> 00:09:11.759 momentarily to talk about this. 00:09:11.769 --> 00:09:17.369 However, folks, the biggest news in AI for this week, 100% came from OpenAI, 00:09:17.469 --> 00:09:26.109 where they claimed that, 10,000 agent swarms claim to have found a solution 00:09:26.109 --> 00:09:30.239 to N- to Navier–Stokes Millennium Prize Problem in mathematics. 00:09:30.709 --> 00:09:37.199 This is an insane sentence to say, but it does seem to be true. 00:09:37.679 --> 00:09:44.159 The LLMs, the AGI is here LLMs, GPT-6 Astra, and I think an unreleased 00:09:44.189 --> 00:09:48.239 model, right, in this case, uh, are solving Millennium Prize mathematic 00:09:48.239 --> 00:09:50.459 models, at least to some extent. 00:09:50.469 --> 00:09:53.319 There's also some excitement about how this was released, on Twitter, 00:09:53.319 --> 00:09:55.089 but this is just like an insane news. 00:09:55.319 --> 00:10:00.149 Again, I will say, OpenAI claims that a swarm of AIs solved something 00:10:00.149 --> 00:10:04.809 humans could not solve, a Millennium Prize problem for Navier–Stokes. 00:10:04.809 --> 00:10:09.897 Excited to talk about this as much as possible, to learn about this it's crazy. 00:10:10.267 --> 00:10:17.157 also, in addition to being one of the most let's go, AI is here solving 00:10:17.157 --> 00:10:22.697 mathematic problems, we also had one of the biggest doomer episodes that 00:10:22.697 --> 00:10:24.867 I've seen recently, and not only me. 00:10:25.097 --> 00:10:27.837 And Anthropic, Ooh, something switched. 00:10:28.077 --> 00:10:28.987 let's go here. 00:10:29.957 --> 00:10:30.117 Right. 00:10:30.317 --> 00:10:35.697 an Anthropic employee quit from Anthropic, and then other AI insiders together 00:10:35.697 --> 00:10:38.237 warned of extinction level events. 00:10:38.277 --> 00:10:42.347 Jacob Coxon resigned from Anthropic on September, and, his tweet about, 00:10:42.347 --> 00:10:46.717 "Hey, I resigned because I think we're all gonna die," now at 122 million 00:10:46.717 --> 00:10:48.977 impressions, which feels inorganic. 00:10:48.987 --> 00:10:50.577 We're, we're gonna have to talk about this. 00:10:50.637 --> 00:10:53.267 though we also have to, you know, at least discuss his claims. 00:10:53.997 --> 00:10:57.867 OpenAI said that they, reached their automated research intern milestone. 00:10:57.867 --> 00:11:01.317 If you guys remember, we told you about this when Sam Altman and Jakub 00:11:01.317 --> 00:11:05.597 Pachocki sat together and said, "Hey, by September 2026, we will probably 00:11:05.597 --> 00:11:10.087 have an intern level automatic research in-house," and they said they reached 00:11:10.087 --> 00:11:14.047 it, and they still target automated AI researcher by Mar- March of '28. 00:11:14.357 --> 00:11:18.807 So that's 18 months from now or so, maybe a little bit more. 00:11:19.037 --> 00:11:25.117 Apple debuts iPhone 18 Pro with A20 Pro chip with a bunch of AI stuff. 00:11:25.157 --> 00:11:28.577 And then Siri AI in beta is about to land on everybody's iPhone, 00:11:28.837 --> 00:11:33.327 including, iPhone Duo, which has nothing to do with AI, I don't think. 00:11:33.347 --> 00:11:36.477 But, many people think this is gonna be a dope vibe coding machine. 00:11:37.877 --> 00:11:39.107 all right, let's see. 00:11:39.937 --> 00:11:43.897 Anything else from Big Labs, Wolfram, LDJ or Nisten? 00:11:46.127 --> 00:11:49.057 Wolfram Ravenwolf: Did you get a chance to look through the stuff I sent you? 00:11:50.187 --> 00:11:50.452 Alex Volkov: I did, but- 00:11:50.452 --> 00:11:50.932 Wolfram Ravenwolf: document? 00:11:50.932 --> 00:11:55.432 Alex Volkov: anything-- Oh yeah, we have some comments from the, audience. 00:11:55.442 --> 00:11:56.402 Milosh, thank you. 00:11:56.432 --> 00:11:59.112 Live translation in AirPods 5 also was announced. 00:11:59.322 --> 00:11:59.942 that's coming. 00:11:59.952 --> 00:12:02.212 I think that was part of the beta, and that's dope. 00:12:02.212 --> 00:12:04.722 Yeah, people walk around with AirPods, and then it translates. 00:12:04.782 --> 00:12:05.642 also the Apple… 00:12:05.662 --> 00:12:07.062 Okay, we'll, we'll talk about the Apple event. 00:12:07.502 --> 00:12:12.932 in AI Art and Diffusion, OpenAI launches GPT Images 2.5. 00:12:12.932 --> 00:12:14.902 this is two models, I believe. 00:12:14.902 --> 00:12:16.582 Yes, Flare and Sunburst. 00:12:16.592 --> 00:12:18.272 We're gonna try them out live on the show. 00:12:18.502 --> 00:12:20.112 Maybe we'll add hats to ourselves. 00:12:20.142 --> 00:12:23.152 maybe we'll ask the audience to tell us what they think we 00:12:23.152 --> 00:12:24.242 should turn ourselves into. 00:12:24.702 --> 00:12:27.922 I think that this is it in AI Art and Diffusion. 00:12:28.372 --> 00:12:34.632 in 3D, there's also OpenAI's, new Unity plugin, Next, we're gonna talk about 00:12:34.642 --> 00:12:36.272 tools and agentic engineering, folks. 00:12:36.522 --> 00:12:39.902 I think this is, for me, we didn't ask before, but for me, this is the 00:12:39.902 --> 00:12:41.912 biggest part of the sh- of this week. 00:12:42.162 --> 00:12:48.402 Meta launches Muse, their agentic 24/7 free AI assistant that has its 00:12:48.402 --> 00:12:52.112 own computer, has its own browser, and if that sounds familiar to you, 00:12:52.112 --> 00:12:54.532 yes, this sounds exactly like GrokBot. 00:12:54.562 --> 00:12:55.742 This sounds exactly like Hermes. 00:12:55.922 --> 00:12:58.272 This sounds exactly like OpenClaw. 00:12:58.542 --> 00:13:00.412 however, This is from a huge company. 00:13:00.642 --> 00:13:03.482 Meta has their own LLM, Meta Muse Spark 1.3, 00:13:03.482 --> 00:13:06.852 which we told you about last week, which is probably the winner of number 00:13:06.852 --> 00:13:08.132 two spot of last week after OpenAI. 00:13:08.982 --> 00:13:12.422 and, that we will show you all about Muse. 00:13:12.742 --> 00:13:13.622 It's really dope. 00:13:13.862 --> 00:13:18.522 And by the way, Muse is running the show for today, so if you are on ThursdAI 00:13:18.532 --> 00:13:22.992 Live and the chat rooms that are getting brought up are, are done by Muse. 00:13:23.172 --> 00:13:25.002 So, let's see how fast this is actually. 00:13:25.002 --> 00:13:28.842 I'm gonna keep reading the TLDR, but I'm gonna pull up the ThursdAI Live page. 00:13:28.892 --> 00:13:34.572 I wanna bring up ThursdAI Live page like that, and, we will ask Muse to 00:13:34.572 --> 00:13:38.072 put up a chyron to introduce itself, and we'll see how fast this happens. 00:13:38.262 --> 00:13:42.902 And, uh, meanwhile, we'll say hi to friend of the pod, open source 00:13:42.952 --> 00:13:47.752 wizard, Team Green representative, Joe Nemotron, Chris Aleksiuk. 00:13:47.752 --> 00:13:48.322 Welcome, dude. 00:13:48.322 --> 00:13:49.342 It's so good to have you back. 00:13:49.342 --> 00:13:49.682 How are you? 00:13:50.302 --> 00:13:51.962 Chris Alexiuk: I got the blue lights today, you know? 00:13:52.012 --> 00:13:53.612 Alex Volkov: Oh, blue lights- And, In, in, Yeah, yeah, yeah … representing. 00:13:53.972 --> 00:13:54.462 Chris Alexiuk: That's right. 00:13:54.502 --> 00:13:55.552 Honor of the whale, man. 00:13:55.882 --> 00:13:59.672 Alex Volkov: So we'll have Chris talk about, a- and us talk about DeepSeek. 00:13:59.682 --> 00:14:00.682 All right, let's continue to… 00:14:00.832 --> 00:14:02.192 Do you guys see this? 00:14:02.322 --> 00:14:03.552 Muse is producing the show. 00:14:03.652 --> 00:14:04.672 It, it heard me. 00:14:05.422 --> 00:14:06.752 Muse is producing today's show. 00:14:07.072 --> 00:14:10.292 we got the chyron up, so, we have an AI producer. 00:14:10.292 --> 00:14:11.892 I'm gonna tell you all about Muse. 00:14:12.362 --> 00:14:16.782 all right, let's go forward to the, to the, the, the TLDR. 00:14:16.782 --> 00:14:17.302 Let's see what else. 00:14:17.312 --> 00:14:19.762 So okay, so we have this, in, in the TLDR. 00:14:19.902 --> 00:14:21.842 Muse, we're gonna talk about Muse from Meta. 00:14:22.772 --> 00:14:24.142 and then the… 00:14:24.932 --> 00:14:30.822 There's another agentic Grok bot, Hermes, Muse, OpenClaw-like thing that's 00:14:30.822 --> 00:14:33.192 going on, very viral, called Instinct. 00:14:33.202 --> 00:14:36.262 We didn't tell you about Instinct yet, but I have been using Instinct. 00:14:36.262 --> 00:14:39.342 Instinct is very viral. 00:14:39.602 --> 00:14:43.032 Instinct works via iMessage, and you get a number, you just text and it works. 00:14:43.222 --> 00:14:44.472 recently they added email. 00:14:44.602 --> 00:14:47.212 I do wanna tell you about Instinct because I did comparison 00:14:47.212 --> 00:14:48.562 between those two agentic things. 00:14:48.862 --> 00:14:52.182 All right, in this week's buzz, anything that has, to do with 00:14:52.182 --> 00:14:53.442 Weights & Biases and CoreWeave, folks. 00:14:53.442 --> 00:14:56.352 I will remind you again, Fully Connected 2026 is coming to 00:14:56.352 --> 00:14:58.952 Moscone South in September 29th. 00:14:58.952 --> 00:15:01.112 That's just in, in, in 19 days. 00:15:01.292 --> 00:15:04.082 and I have a big announcement about somebody who's gonna be 00:15:04.082 --> 00:15:06.342 there and you don't wanna miss, and we have free tickets for you. 00:15:06.342 --> 00:15:09.462 we also have a CoreWeave hack and Agent Loops hackathon. 00:15:09.492 --> 00:15:12.772 And if you win some of the prizes there, you'll be able to come to Fully Connected 00:15:13.012 --> 00:15:14.152 and present in front of big audience. 00:15:14.152 --> 00:15:15.172 That's also a big news. 00:15:15.362 --> 00:15:16.952 So definitely we'll tell you about that. 00:15:17.172 --> 00:15:19.752 and then I think let's close out the TLDR with these things. 00:15:19.762 --> 00:15:20.722 Voice and audio. 00:15:21.072 --> 00:15:23.792 I think number one is Google launches Lyria 2.5, 00:15:23.792 --> 00:15:27.492 full song music generation model, access to Gemini and API, so you 00:15:27.492 --> 00:15:29.952 can use API to generate songs. 00:15:29.952 --> 00:15:31.082 We probably should try it out. 00:15:31.312 --> 00:15:32.882 Suno launches Suno 6. 00:15:32.952 --> 00:15:33.862 We don't care. 00:15:33.862 --> 00:15:35.542 This is literally as it's written on the TLDR. 00:15:36.192 --> 00:15:39.552 Suno pulled a fast one and, like, restricted how many 00:15:39.572 --> 00:15:40.792 downloads people can use. 00:15:40.792 --> 00:15:44.252 And Suno, you know, makes a lot of money, but people don't like them 00:15:44.252 --> 00:15:45.912 anymore, so we, we will just tell you. 00:15:45.982 --> 00:15:47.222 A new Suno was released. 00:15:47.542 --> 00:15:49.842 and also OpenAI brings GPT 5.6 00:15:49.842 --> 00:15:52.962 Sol and GPT 6 Astra to ChatGPT Voice. 00:15:53.142 --> 00:15:57.942 They're using the live version, but now the live version can talk to Astra and Sol 00:15:57.952 --> 00:16:01.872 for you, so you can, like, build things, like they said in their video, with voice. 00:16:01.932 --> 00:16:07.514 And I think, unless I missed a bunch of stuff This is the TLDR. 00:16:07.544 --> 00:16:08.884 Let's see in comments. 00:16:09.234 --> 00:16:11.404 folks in comments, sh- let me see if, I 00:16:11.404 --> 00:16:12.984 Wolfram Ravenwolf: put something in our private chat 00:16:13.384 --> 00:16:13.874 Alex Volkov: Let's see. 00:16:14.264 --> 00:16:18.154 Wolfram Ravenwolf: Okay, so we have, a new music model as well, Yue 2 00:16:18.194 --> 00:16:23.384 Open Music Model, successor to Yue or however it's pronounced, Y-U-E, 00:16:23.824 --> 00:16:28.444 which can do vocals and accompaniment, and it's on Hugging Face already. 00:16:29.014 --> 00:16:31.664 Well, there's a speech model by Tencent, AUK. 00:16:32.534 --> 00:16:34.424 It's an open source 1.5B 00:16:34.424 --> 00:16:38.584 speech model that does ta- text-to-speech, voice cloning, content emotion, 00:16:38.594 --> 00:16:42.454 accent edits, cleanup, and separation from one natural language prompt. 00:16:42.934 --> 00:16:44.124 MIT Weights. 00:16:44.214 --> 00:16:44.584 Yeah. 00:16:44.594 --> 00:16:44.884 Alex Volkov: right. 00:16:44.974 --> 00:16:50.834 and if that's it, I think, let's go to open source. 00:16:52.014 --> 00:16:52.664 Ooh, I'm excited. 00:17:06.106 --> 00:17:07.316 Open source AI. 00:17:07.396 --> 00:17:09.076 Let's get it started 00:17:13.338 --> 00:17:14.568 All right, we're here. 00:17:14.918 --> 00:17:15.938 Open Source Corner. 00:17:16.008 --> 00:17:17.028 Let's get it started. 00:17:17.128 --> 00:17:22.448 folks, just before I, and Yam Peleg just stepping in exactly as we're about 00:17:22.448 --> 00:17:24.218 to talk about DeepSeek, that's great. 00:17:24.248 --> 00:17:26.038 I don't know what's going on with Nisten, but, oh, there he is. 00:17:26.288 --> 00:17:26.758 Okay. 00:17:27.018 --> 00:17:33.448 Folks, exciting news from the world of open source because DeepSeek, the whale, 00:17:33.468 --> 00:17:40.098 has refer- resurfaced once again with a 4x KV cache compression with DeepSeek V4.1 00:17:40.188 --> 00:17:40.758 Flash. 00:17:40.898 --> 00:17:45.288 It's a half a trillion parameter, specifically 552 billion 00:17:45.548 --> 00:17:50.048 parameter MOE that activates only 8 billion or 16 billion. 00:17:50.098 --> 00:17:50.378 What? 00:17:50.448 --> 00:17:50.738 Okay. 00:17:50.888 --> 00:17:55.428 to help us to talk about this, we have folks who like DeepSeek for a long time 00:17:55.898 --> 00:18:01.018 and have used it and tested it out, Chris Oleksiuk from Nvidia, a friend of the pod, 00:18:01.228 --> 00:18:05.968 AKA Joe Nemotron, the guy who brings us Nemotron news, but also is part of Nvidia. 00:18:06.018 --> 00:18:08.318 Chris Alexiuk: I think Hugging Face owns themselves, at the end of 00:18:08.318 --> 00:18:08.818 Alex Volkov: the day. 00:18:08.848 --> 00:18:10.168 Nvidia supports Hugging Face- That's right. 00:18:10.328 --> 00:18:10.528 Chris Alexiuk: right 00:18:10.538 --> 00:18:11.688 … Alex Volkov: with a- Big fans … $12.9 00:18:11.698 --> 00:18:12.648 billion- That's right. 00:18:12.678 --> 00:18:13.678 That's right … injection. 00:18:13.708 --> 00:18:14.088 Big fans, yeah. 00:18:14.088 --> 00:18:16.708 And now you have a bunch of new colleagues, let's, let's say that. 00:18:16.708 --> 00:18:16.846 That's right. 00:18:16.846 --> 00:18:19.728 You have a bunch of new- Yeah, yeah, yeah new teammates, and, you 00:18:19.728 --> 00:18:22.408 guys are keeping the torch of open source alive, so thank you for that. 00:18:22.468 --> 00:18:25.888 Now, on that Hugging Face now, there's a new model. 00:18:26.058 --> 00:18:27.488 You see how the connection is made? 00:18:27.668 --> 00:18:31.118 there's a new model there on that new Hugging Face, and it's from DeepSeek. 00:18:32.348 --> 00:18:33.248 I saw your post. 00:18:33.528 --> 00:18:36.808 What- Yeah tell us a bit about what, what's exciting about this model. 00:18:37.218 --> 00:18:38.408 Chris Alexiuk: it's a new whale model. 00:18:38.808 --> 00:18:40.188 We know whale models are good. 00:18:40.238 --> 00:18:43.608 I think every time you talk about DeepSeek models, it's really important 00:18:43.618 --> 00:18:48.618 to talk about the fact that, like, they are probably one of the best, 00:18:48.968 --> 00:18:53.698 like, technical engineering and infrastructure, reports to, to read. 00:18:53.718 --> 00:18:58.918 So even if the model isn't, like, insane, in terms of accuracy, it's 00:18:58.948 --> 00:19:03.138 always, you can always learn a lot, from reading the technical report and, and 00:19:03.148 --> 00:19:04.248 understanding how they approach things. 00:19:04.258 --> 00:19:09.378 This one is obviously flash, so it's, it's quite fast and it gets the it gets 00:19:09.378 --> 00:19:14.518 the speed through some rather, I think, interesting, architectural decisions 00:19:14.548 --> 00:19:18.718 as well as a, a bunch of other, kind of, kind of headline, you know, things. 00:19:18.958 --> 00:19:21.248 I, I'm gonna do a gimmick because I have the hat. 00:19:21.248 --> 00:19:26.916 So this is a smart model, but it's also a very fast model, so we got- We 00:19:26.916 --> 00:19:31.146 got the most attractive quadrant hat from artificial analysis on today. 00:19:31.486 --> 00:19:34.936 this is basically just saying it's very smart and fast. 00:19:35.086 --> 00:19:35.306 Speaker 3: Nice. 00:19:35.386 --> 00:19:38.106 Chris Alexiuk: I think, like, at the end of the day, the reason 00:19:38.106 --> 00:19:41.966 the architecture is so interesting is because it kind of codifies 00:19:42.006 --> 00:19:43.846 architecturally, disaggregation, right? 00:19:43.846 --> 00:19:48.176 This idea that, like, decode and prefill are two different things, and 00:19:48.176 --> 00:19:51.416 we should think about them separately, and they achieve different things. 00:19:51.416 --> 00:19:54.196 And, they, they put that into the architecture itself. 00:19:54.216 --> 00:19:55.056 They baked it right in. 00:19:56.236 --> 00:20:00.356 But one thing that should be clear is that I think this is like, and 00:20:00.356 --> 00:20:02.346 you, you can see a lot of tweets about this, but this is like the 00:20:02.346 --> 00:20:05.316 most data-pilled DeepSeek release. 00:20:05.736 --> 00:20:11.536 they have very clearly taken advantage of the fact that many people use their 00:20:11.536 --> 00:20:16.796 model to, produce environments and scale up on the URL side in a way that, 00:20:16.796 --> 00:20:18.576 I, I, I think we haven't seen before. 00:20:19.006 --> 00:20:20.806 and, K V cache compression. 00:20:21.086 --> 00:20:21.606 Tons of… 00:20:21.826 --> 00:20:26.406 It, there's always really interesting, again, infrastructure and engineering 00:20:26.406 --> 00:20:27.556 that happens in these models- Yeah 00:20:27.896 --> 00:20:28.616 that make them good. 00:20:28.626 --> 00:20:31.406 Like, like the highlight stuff, like we're seeing on the, the 00:20:31.406 --> 00:20:33.086 bar charts are good obviously. 00:20:33.136 --> 00:20:37.106 I think, the last time the Whale dropped a model, the, I think some 00:20:37.106 --> 00:20:39.916 people were a little bit worried 'cause the bar charts didn't look so good. 00:20:39.916 --> 00:20:44.146 But- Yeah … it's very clear that they're, they're operating on an axis 00:20:44.146 --> 00:20:50.886 of like trying to build the, the best long-term engineering project, project 00:20:51.016 --> 00:20:52.836 as it relates to language modeling. 00:20:52.886 --> 00:20:56.700 this release is just a absolute- Banger. 00:20:57.010 --> 00:20:57.391 Alex Volkov: it almost 00:20:57.411 --> 00:20:57.611 Chris Alexiuk: seems like- 00:20:57.651 --> 00:20:57.811 Alex Volkov: That's, 00:20:57.820 --> 00:20:58.431 Chris Alexiuk: that's why I'm excited 00:20:58.621 --> 00:20:59.391 Alex Volkov: I fully get you, and I… 00:20:59.551 --> 00:21:01.051 that's how it feels also on timeline. 00:21:01.141 --> 00:21:05.441 But it almost seems like, like you said, DeepSeek is on this, like, super 00:21:05.541 --> 00:21:09.441 high-tier engineering project, and, like, on the way, just weights are 00:21:09.461 --> 00:21:12.701 dropping 'cause they're like, "Yeah, this, this one's, this one's all right. 00:21:12.901 --> 00:21:14.721 You know, th- this, this checkpoint is okay. 00:21:15.101 --> 00:21:16.301 Let- let's feed them something." 00:21:16.361 --> 00:21:19.321 And then they just disappear, don't engage, no community, no work, 00:21:19.381 --> 00:21:21.991 Yeah, like, you know, like Kimi goes explosive with K3, whatever. 00:21:22.471 --> 00:21:25.491 I think at some points, the, the, the ver- the, you know, the, the, 00:21:25.531 --> 00:21:27.611 the versions of DeepSeek V4.1 00:21:27.871 --> 00:21:31.351 could have been called, you know, one of them could be Pro, for example. 00:21:31.411 --> 00:21:33.971 The previous version, was almost Pro level. 00:21:34.051 --> 00:21:38.151 And, it, it almost seems like they don't really care about that much. 00:21:38.451 --> 00:21:41.311 Although this one did come with a blog post, so that, that was very interesting. 00:21:41.841 --> 00:21:43.481 Yam, your thoughts on DeepSeek. 00:21:43.561 --> 00:21:46.361 I saw you, you, you, you exploded on your timeline. 00:21:46.381 --> 00:21:49.481 Yam Peleg: Bro, that's, that's the best one to use right now for coding. 00:21:49.681 --> 00:21:52.781 I mean, that's, for, for an open weight model, that you want to host 00:21:52.821 --> 00:21:57.361 yourself or, or z- That's, that's the best option that you got. 00:21:57.671 --> 00:22:01.271 but not only that, it's not even the largest. 00:22:01.571 --> 00:22:04.531 It's not, not e- not remotely the largest option. 00:22:04.571 --> 00:22:10.571 It's both v- an extremely efficient model and ab- like, absolutely the, 00:22:10.791 --> 00:22:12.971 the best option regardless of size. 00:22:13.151 --> 00:22:15.191 I mean, that's, that's pretty much… 00:22:15.671 --> 00:22:17.190 It is a monumental achievement. 00:22:17.821 --> 00:22:21.351 Put aside that w- we're starting, we're starting a new, a new corner 00:22:21.391 --> 00:22:25.471 of the show, I Told You So, and the corner is, that's, Today in, today, 00:22:25.491 --> 00:22:27.491 today on the menu is encoder decoders. 00:22:28.091 --> 00:22:28.771 I told you so. 00:22:30.211 --> 00:22:32.381 Alex Volkov: Yam- yeah I know what you're talking about. 00:22:32.621 --> 00:22:36.701 I'm sure that there's some folks, and we have around, like, 400 folks tuning in, 00:22:37.221 --> 00:22:38.441 that don't know what you're talking about. 00:22:38.601 --> 00:22:38.921 Okay, okay. 00:22:38.931 --> 00:22:40.561 G- g- give us, like, a brief explainer of what you mean- 00:22:40.651 --> 00:22:41.371 by I Told You So. 00:22:41.851 --> 00:22:47.552 Yam Peleg: back in the day before decoder, architecture like GPT-2, starting 00:22:47.772 --> 00:22:53.332 somewhere at GPT-2 got, got popular, then three, then ChatGPT, they are all, a 00:22:53.472 --> 00:22:55.832 single model that predicts the next token. 00:22:55.912 --> 00:22:57.581 pre- a pretty known thing today. 00:22:58.092 --> 00:23:02.492 But before that, there was, i- actu- actually the start of transformers, 00:23:03.142 --> 00:23:06.422 But before, before, before the decoder models, the reason they 00:23:06.422 --> 00:23:10.182 are called this way is because there were encoder-decoder models. 00:23:10.512 --> 00:23:14.732 to anyone listening and not sure what I'm talking about, imagine two models. 00:23:15.002 --> 00:23:18.452 One of them doesn't see only the previous tokens, it sees everything 00:23:18.452 --> 00:23:22.462 at the same time, and the other predicts the next token based on 00:23:22.602 --> 00:23:24.972 the previous tokens and the… 00:23:25.212 --> 00:23:30.812 Over time, decoder models just got really, got extremely efficient, 00:23:31.062 --> 00:23:34.972 both in training and in inference because it's a singular architecture. 00:23:34.972 --> 00:23:39.522 We got really good at, serving it and running it at scale, so we just scaled 00:23:39.522 --> 00:23:42.682 up one part of this and rolled… 00:23:42.692 --> 00:23:44.602 and, and we got everything we have today. 00:23:45.142 --> 00:23:49.262 However, the architecture itself of encoder-decoder is, has a very 00:23:49.262 --> 00:23:54.612 interesting prior to it that you treat differently different parts of the input. 00:23:54.912 --> 00:23:59.362 Like, for example, for, for chat, you can think about this as, the part of, 00:23:59.392 --> 00:24:04.978 the user prompt is, is a different concept than the actual next token 00:24:04.978 --> 00:24:06.598 prediction of the model itself. 00:24:06.598 --> 00:24:07.918 It's a completely different concept. 00:24:08.178 --> 00:24:13.328 So treating them differently with the different parts of the, the weights 00:24:13.448 --> 00:24:16.278 that you have is … makes sense. 00:24:16.488 --> 00:24:20.088 So but it, but it's not, it's not an, a, a … Oh, look, I told you 00:24:20.088 --> 00:24:22.398 so, but I didn't have the details. 00:24:22.418 --> 00:24:26.188 I mean, there is an incredible, effort of engineering- Wait, so, so lend this for me 00:24:26.198 --> 00:24:28.788 go that went and- Did DeepSeek, did DeepSeek go back to 00:24:28.828 --> 00:24:29.558 Alex Volkov: encoder-decoder 00:24:29.608 --> 00:24:33.688 Yam Peleg: Look, D- DeepSeek is, using encoder-decoder architecture, but with a 00:24:33.698 --> 00:24:41.688 lot of, of, b- battle-tested experiment, experiment-driven, architecture decisions 00:24:41.698 --> 00:24:46.718 that make it actually feasible at this scale and with this performance, okay? 00:24:46.718 --> 00:24:50.058 It's not, it's not a vanilla encoder-decoder from before. 00:24:50.098 --> 00:24:52.238 I, I told you so, but I didn't have the details. 00:24:52.238 --> 00:24:53.878 I'm not taking anyone, anyone's credit. 00:24:53.878 --> 00:24:55.258 DeepSeek did an amazing work here. 00:24:55.258 --> 00:24:59.158 Alex Volkov: I, I think you, you mentioned this in passing, the before times when 00:24:59.158 --> 00:25:01.058 folks were running encoder-decoder- Mm-hmm 00:25:01.058 --> 00:25:04.898 the, the scale of those models were in the billions or maybe- Mm-hmm … you 00:25:04.898 --> 00:25:06.348 know, less than 100 billion parameters. 00:25:06.348 --> 00:25:09.218 DeepSeek is scaling this to half a trillion and- Absolutely … at 00:25:09.218 --> 00:25:10.738 scale as well with the amount of… 00:25:10.738 --> 00:25:15.268 I, I actually know how many tokens, but it looks like, trained from scratch on 45 00:25:15.658 --> 00:25:17.668 trillion multimodal tokens, this model. 00:25:17.668 --> 00:25:18.588 Yam Peleg: I just want to say- 00:25:18.688 --> 00:25:18.878 Alex Volkov: Yeah 00:25:18.878 --> 00:25:22.948 … Yam Peleg: that they are, saying it, I think, in the most blatant way 00:25:22.948 --> 00:25:25.108 that we've seen any lab saying it. 00:25:25.628 --> 00:25:27.638 everything is nice, but it's data. 00:25:27.818 --> 00:25:29.048 It's all in the data. 00:25:29.058 --> 00:25:29.098 Yeah. 00:25:29.098 --> 00:25:32.618 They are really, really stressing this pretty hard. 00:25:32.818 --> 00:25:36.928 wi- with no question in the paper, getting better data and manipulating 00:25:36.938 --> 00:25:44.508 the data in clever ways just has unevenly, lar- unevenly large impact 00:25:44.528 --> 00:25:49.588 on the actual result way more than any engineering, trick you can come up with. 00:25:49.918 --> 00:25:52.328 It's the data and, It's the data and- … makes a lot of sense. 00:25:52.338 --> 00:25:55.928 Alex Volkov: Chris, you train models, you release models for NVIDIA. 00:25:56.238 --> 00:26:02.278 just give us a little sense of w- what does 45 trillion f- tokens of data mean? 00:26:02.378 --> 00:26:04.178 how, whatever you measure, data in. 00:26:04.288 --> 00:26:07.898 45 trillion multimodal, which i- that to me sounds insane. 00:26:08.548 --> 00:26:08.908 Chris Alexiuk: Yeah, 00:26:08.908 --> 00:26:10.498 Alex Volkov: I mean- Is that on the level of, like, frontier labs 00:26:10.498 --> 00:26:11.598 or what are we talking about here? 00:26:12.048 --> 00:26:14.558 Chris Alexiuk: 45 trillion is not a small amount of data. 00:26:14.848 --> 00:26:19.978 I think more importantly though is that they … Th- this isn't, like, 45 00:26:20.008 --> 00:26:26.658 trillion tokens of, like, chaff, which you can kind of get super easily, right? 00:26:26.658 --> 00:26:28.238 So you can just find that data, right? 00:26:28.378 --> 00:26:30.108 Like it's not, it's not, it's not tough. 00:26:30.378 --> 00:26:34.838 But they, their dedication to like cleaning the data, and the way they 00:26:34.838 --> 00:26:39.118 cleaned the data, and the resources that it appears from the technical report 00:26:39.148 --> 00:26:44.088 they spent on cleaning the data, I think means that that 45 trillion is like v- 00:26:44.308 --> 00:26:46.798 a very good pile of da- of data, right? 00:26:47.058 --> 00:26:51.958 they, they, they in, in for pre-train they went well beyond the, the norm 00:26:52.048 --> 00:26:53.778 for how they approached cleaning. 00:26:54.138 --> 00:26:59.688 but, it- even those pipelines I think are less thorough than 00:26:59.728 --> 00:27:01.928 this DeepSeek, cleaning effort. 00:27:02.288 --> 00:27:04.518 Hey, fr- fr- from what we see from the technical report, right? 00:27:04.538 --> 00:27:05.458 Like from what, what they're… 00:27:05.458 --> 00:27:09.248 If, if we take them at their word, which usually with Whale you, you, you can, 00:27:09.448 --> 00:27:12.468 I don't know if they're gonna release the pipeline explicitly, but typically 00:27:12.468 --> 00:27:14.428 they- Mm-hmm … they're, they're pretty open with what they're doing. 00:27:14.878 --> 00:27:15.368 but that's the idea. 00:27:15.368 --> 00:27:18.328 It's like 45 trillion tokens of like great data, right? 00:27:18.578 --> 00:27:24.198 And I think that's how you get to this kind of model that has the, the, the 00:27:24.198 --> 00:27:28.358 performance as in like accuracy, as in bar charts, as compared to other models. 00:27:28.718 --> 00:27:31.448 so like 45 trillion I don't think is like huge amount of data. 00:27:31.518 --> 00:27:33.468 But it is, it is a good size. 00:27:33.818 --> 00:27:37.888 And if it's as clean as they say, that is a, I, I think, Yam 00:27:37.888 --> 00:27:39.648 already said this, historic, right? 00:27:39.648 --> 00:27:39.658 Yeah. 00:27:39.658 --> 00:27:44.378 Like it's a, especially in this, this kind of space where, again, like the 00:27:44.448 --> 00:27:46.448 Whale has never really been data pilled. 00:27:46.478 --> 00:27:52.130 Like this is the, this is their first model where they're like- No, I, I 00:27:52.130 --> 00:27:55.160 ca- I can't remember if you quoted it exactly, but just to be sure, you 00:27:55.160 --> 00:27:56.290 can quote exactly from the paper. 00:27:56.290 --> 00:28:00.640 Like, no algorithmic, y- work, like, could match the data work that they did. 00:28:00.910 --> 00:28:05.740 A- any, any novelty or innovation in the algorithm, like, would have had a smaller 00:28:05.750 --> 00:28:07.420 impact on the final quality of the model. 00:28:07.420 --> 00:28:11.070 So this is, like, the most data-pilled Whale model we've ever seen. 00:28:11.380 --> 00:28:14.750 and I, I think when people use it, they're, they're gonna feel 00:28:14.750 --> 00:28:16.350 as though that's true, right? 00:28:16.350 --> 00:28:20.280 Like, And, and something else that was touched on, just to make sure, 00:28:20.280 --> 00:28:24.610 like, it, it's really felt by anyone who's watching is Whale doesn't 00:28:24.890 --> 00:28:28.330 release tons of models all the time that are, like, super bangers, but 00:28:28.330 --> 00:28:31.860 they do just keep, like, shitting out these models that are fantastic. 00:28:32.260 --> 00:28:36.010 And, like, they're really good at not producing deep-fried models. 00:28:36.020 --> 00:28:39.670 So when you use, like, DeepSeek, I think you don't feel as though it's as 00:28:39.670 --> 00:28:43.430 benchmaxed as some of the other, models, and, that does mean that, like, they 00:28:43.430 --> 00:28:45.340 show up less impressively on bar charts. 00:28:45.360 --> 00:28:49.690 But I think for, for quote-unquote real work, whatever you wanna call that, 00:28:49.710 --> 00:28:53.350 like out-of-distribution, y- you know, tasks, I think, this is where Whale i- 00:28:53.670 --> 00:28:55.780 is gonna continue to do a great job. 00:28:56.250 --> 00:28:59.660 Alex Volkov: Speaking of using it, a friend of the pod, Aron Batino, 00:28:59.670 --> 00:29:05.716 built Token Juice, where he basically gives you this offer Hey, you give 00:29:05.716 --> 00:29:09.316 me all your data, which is, you give me all the transactions, whatever, 00:29:09.316 --> 00:29:10.636 we store everything to train on. 00:29:10.946 --> 00:29:13.776 I give you free, very fast, performance of DeepSeek. 00:29:13.796 --> 00:29:14.886 It's called TokenJuice.ai. 00:29:15.756 --> 00:29:21.216 I think it's a very fair, very fair offer because most other people… 00:29:21.376 --> 00:29:25.256 Actually, the contributor tier on Meta Muse Spark is exactly that. 00:29:25.286 --> 00:29:26.606 You get it for very cheap. 00:29:26.866 --> 00:29:28.706 this one is even better 'cause you get it for free. 00:29:28.866 --> 00:29:33.416 so if you want to get some free, and you want to contribute your data back, if 00:29:33.416 --> 00:29:37.566 you want to get some free DeepSeek V- for Qwen-V1 Flash, go to TokenJuice.ai. 00:29:37.616 --> 00:29:38.626 This is not sponsored. 00:29:38.716 --> 00:29:41.936 I saw Aaron post this, By the way, this is his private endeavor, not 00:29:41.946 --> 00:29:44.736 correlated, but Aaron works on the CoreWeave inference as well. 00:29:44.896 --> 00:29:46.296 This is something he's playing with on the side. 00:29:46.296 --> 00:29:46.916 Shout out, Aaron. 00:29:47.086 --> 00:29:48.606 and, I saw it on my timeline. 00:29:48.606 --> 00:29:49.316 Nobody reacted. 00:29:49.326 --> 00:29:51.226 Like, n- nobody wants free DeepSeek? 00:29:51.256 --> 00:29:51.966 I don't understand. 00:29:52.166 --> 00:29:54.186 So shout out to TokenJuice, .ai. 00:29:54.196 --> 00:29:55.096 give it a try. 00:29:55.276 --> 00:29:58.086 but also, I do want to talk about DeepSeek a little bit more in 00:29:58.086 --> 00:29:59.976 terms of the scores we're getting. 00:29:59.976 --> 00:30:04.146 Folks, we're getting some scores and, when I say some scores, I mean, a 00:30:04.146 --> 00:30:05.576 lot of it is kind of mind-blowing. 00:30:05.576 --> 00:30:07.066 Yam, talk to me about this. 00:30:07.116 --> 00:30:09.866 folks, Terminal bench- Yeah Terminal bench, 2.1 00:30:09.866 --> 00:30:12.106 score, 90.6. 00:30:12.166 --> 00:30:14.256 This model beats Opus5 00:30:17.436 --> 00:30:17.566 Yam Peleg: and GPT 5.6o. 00:30:17.566 --> 00:30:18.556 But, y- you need to see it on a table. 00:30:18.826 --> 00:30:19.842 Alex Volkov: in comparison- I have a table to show you Hold on. 00:30:19.842 --> 00:30:22.132 Hold on It's insane I have, I have a table to show you, a completely 00:30:22.132 --> 00:30:25.072 different table from somebody called Open Design- All right … that this 00:30:25.082 --> 00:30:30.322 model is also designing at second level, beating Fable, just behind Astra. 00:30:30.412 --> 00:30:32.822 Speaker 4: But if you first can take a look at this, like, l- 00:30:32.822 --> 00:30:36.712 very blurry line here, you can see that it designs it for… 00:30:36.762 --> 00:30:37.872 It's two cents. 00:30:38.722 --> 00:30:38.972 Jesus. 00:30:38.972 --> 00:30:43.072 The same design tasks on this model that beat Astra, sorry, like come 00:30:43.072 --> 00:30:44.622 very close to Astra, beat Fable. 00:30:44.882 --> 00:30:47.112 Fable is at $3.50. 00:30:47.112 --> 00:30:48.702 Astra is at, 1.6. 00:30:48.712 --> 00:30:50.912 Speaking of, by the way, Astra is way more token efficient 00:30:50.912 --> 00:30:52.272 than Fable at those results. 00:30:52.482 --> 00:30:55.242 DeepSeek is number two with two cents. 00:30:55.952 --> 00:30:59.522 Not, not, not 20 cents, not $2, two cents. 00:31:00.102 --> 00:31:01.242 It's crazy. 00:31:02.102 --> 00:31:03.242 Yam Peleg: And it's actually good. 00:31:03.422 --> 00:31:05.662 Like, seriously, I just want to tell anyone, anyone 00:31:05.662 --> 00:31:06.752 listening that's not been… 00:31:06.822 --> 00:31:09.042 I tested it a lot today. 00:31:09.302 --> 00:31:10.382 That thing is really good. 00:31:10.662 --> 00:31:14.542 Seriously, for real work, go, go try it. 00:31:14.542 --> 00:31:15.912 It's literally free. 00:31:15.942 --> 00:31:18.402 You just got an option to try it for free. 00:31:18.432 --> 00:31:18.692 Yeah. 00:31:18.692 --> 00:31:19.982 for literally free if you want. 00:31:19.982 --> 00:31:20.122 Yes. 00:31:20.122 --> 00:31:20.392 Okay? 00:31:20.412 --> 00:31:21.872 Like, seriously, go, go try it. 00:31:21.872 --> 00:31:22.432 It's a really good model. 00:31:22.432 --> 00:31:22.502 But- 00:31:22.502 --> 00:31:26.032 Alex Volkov: I am also pretty sure that the, the DeepSeek that Token Juice 00:31:26.032 --> 00:31:28.102 hosts is, is hosted on the US server. 00:31:28.102 --> 00:31:31.892 So if you ha- do have a concerns about, you know, Chinese inference, US inference, 00:31:31.932 --> 00:31:34.432 that one is, is based on, the US server. 00:31:34.432 --> 00:31:35.722 So definitely give it a try. 00:31:35.722 --> 00:31:39.382 We'll definitely try to bring it to CoreWeave Inference very soon. 00:31:39.382 --> 00:31:41.592 But also, let's look at some other charts. 00:31:42.562 --> 00:31:42.582 Yeah. 00:31:42.972 --> 00:31:43.652 Bo- bottom line? 00:31:44.452 --> 00:31:44.682 Yam Peleg: I want to 00:31:44.682 --> 00:31:44.702 Alex Volkov: Bottom line? 00:31:44.702 --> 00:31:45.139 Hold on, one second. 00:31:45.139 --> 00:31:45.342 Yeah, yeah, just 00:31:45.342 --> 00:31:45.672 Yam Peleg: saying. 00:31:45.722 --> 00:31:46.412 Go on, go 00:31:46.412 --> 00:31:46.742 Alex Volkov: Th- this… 00:31:46.772 --> 00:31:48.682 Yam, te- tell me about th- this is insane. 00:31:48.702 --> 00:31:50.992 I, I want both of you, Chris, Yam, and Nisten, if you 00:31:50.992 --> 00:31:52.102 wanna chime in here as well. 00:31:52.412 --> 00:31:57.912 The KV cache compression, hero journey that DeepSeek is on is just on another 00:31:57.912 --> 00:32:01.204 level It's, it's j- just, I want 00:32:02.534 --> 00:32:06.944 Yam Peleg: That's direct, directly influencing how expensive 00:32:07.044 --> 00:32:08.904 it is when you pay for it. 00:32:09.164 --> 00:32:13.754 That thing directly reduced the price of actually serving it to people. 00:32:14.424 --> 00:32:16.354 That, it, it is that simple. 00:32:16.374 --> 00:32:19.924 This is why they are working so hard, and yeah, it's, it's… 00:32:20.704 --> 00:32:25.274 I'm sure the, the big labs have, have their own tricks and so on, but that's 00:32:25.374 --> 00:32:26.904 something we just see in public. 00:32:27.124 --> 00:32:28.244 Alex Volkov: Chris, I want to ask you- 00:32:28.274 --> 00:32:28.474 Yam Peleg: in public, 00:32:28.484 --> 00:32:28.504 Alex Volkov: yeah 00:32:28.654 --> 00:32:31.314 almost directly, and maybe Nisten, you want to chime in here as well. 00:32:31.974 --> 00:32:34.534 we're seeing a chart here, so we're also podcast folks. 00:32:34.534 --> 00:32:35.914 I'm not seeing this, I'm gonna read this out. 00:32:35.924 --> 00:32:39.104 We're seeing a chart that says, I believe that this is from their blog, saying, 00:32:39.114 --> 00:32:44.424 "Less cash, more cost effective," where we see DeepSeek V1 from 2023, November 00:32:44.424 --> 00:32:46.424 of 2023, so that's three years ago. 00:32:46.644 --> 00:32:50.154 I believe that we also, already talked about DeepSeek back then. 00:32:50.434 --> 00:32:55.214 and then DeepSeek V3 from November of, December of '25, so like two years. 00:32:55.404 --> 00:33:00.944 In two years they, they, reduced the, the KV cache per 00:33:00.944 --> 00:33:06.484 token, from 389,000 to 48,000. 00:33:06.494 --> 00:33:09.144 Chris, what does this mean, KV cache per token in bytes? 00:33:09.144 --> 00:33:11.374 Like, is that how much memory you store per byte? 00:33:11.454 --> 00:33:14.524 can you, like, very simply explain to folks- Yeah … what this means? 00:33:14.884 --> 00:33:18.354 Chris Alexiuk: Basically, the idea is that whenever we're doing this 00:33:18.374 --> 00:33:21.734 decode process, right, which is where we actually shit out the next token. 00:33:21.864 --> 00:33:24.764 So we have all the context, and then we're, now we're, now 00:33:25.144 --> 00:33:26.704 our job is just make tokens. 00:33:27.104 --> 00:33:30.844 a lot of the math that you wind up doing is actually reusable, right? 00:33:31.074 --> 00:33:38.204 if you store the previous, set of, s- set of results from math in cache, 00:33:38.224 --> 00:33:40.914 that's the key value cache or KV cache. 00:33:41.254 --> 00:33:45.604 So the idea is that this cache is what lets us decode quickly, and 00:33:45.634 --> 00:33:50.174 because we can do it with less net computation, it means we're, we're 00:33:50.194 --> 00:33:52.304 basically gonna pay less for it. 00:33:52.834 --> 00:33:57.770 when you're using agents, like, the number- I, I don't know. 00:33:57.790 --> 00:33:59.340 I'm, I'm not gonna say number one thing, okay? 00:33:59.340 --> 00:34:02.310 But, like, the number two thing, let's say, that's driving cost is gonna 00:34:02.310 --> 00:34:04.770 be, like, cache hits or cache misses. 00:34:05.120 --> 00:34:07.890 And this is, do we remember what, the work that we've already 00:34:07.890 --> 00:34:09.440 done, and can we exploit it? 00:34:10.060 --> 00:34:14.020 When you make KV cache, really compressed, we have teams that work on this at, 00:34:14.020 --> 00:34:17.360 at Team Green with stuff like KV Press and a bunch of other techniques, right? 00:34:17.360 --> 00:34:21.040 Like, this is, this is something that the whole industry, is, is trying to do well. 00:34:21.040 --> 00:34:25.890 but the, the idea is, like, we wanna, we wanna make it as cheap as possible 00:34:25.890 --> 00:34:31.470 to keep as much as we can in cache, and reduce the number of cache misses, 00:34:31.820 --> 00:34:33.320 and increase the number of cache hits. 00:34:33.360 --> 00:34:35.800 the more we can rely on that cache, the better. 00:34:35.820 --> 00:34:40.600 And making it small, or compressed in this case, is, effectively just letting 00:34:40.600 --> 00:34:42.620 us exploit it even harder, right? 00:34:42.620 --> 00:34:46.940 Like, again, If you really wanted to, we would just cache everything- and always 00:34:46.950 --> 00:34:48.800 pull everything from cache all the time. 00:34:48.800 --> 00:34:51.900 It's just like that's not feasible, that's not like how it would work in reality. 00:34:51.900 --> 00:34:58.280 And so, this is a method to, get the cache as small as we can per token so 00:34:58.280 --> 00:35:00.510 that we can use as much of it, right? 00:35:00.510 --> 00:35:03.500 We have more headroom that we can fill with cache. 00:35:04.120 --> 00:35:05.430 Alex Volkov: yeah S- so now that we have like a basic understanding 00:35:05.430 --> 00:35:06.610 what KV cache does- Mm-hmm 00:35:06.610 --> 00:35:07.670 let's go back to the numbers. 00:35:07.670 --> 00:35:13.140 From 2023 in November, at 389,000 bytes- 00:35:13.230 --> 00:35:13.560 Chris Alexiuk: Mm-hmm 00:35:13.840 --> 00:35:18.630 … Alex Volkov: per token, of cache per token, we're now at 890 bytes. 00:35:18.630 --> 00:35:18.820 Think 00:35:18.840 --> 00:35:19.530 Chris Alexiuk: about it this way, right? 00:35:19.580 --> 00:35:19.710 Alex Volkov: Yeah 00:35:19.940 --> 00:35:22.920 … Chris Alexiuk: if you have tons of tokens, each of those tokens is 00:35:23.060 --> 00:35:25.070 taking up space on your GPU, right? 00:35:25.070 --> 00:35:26.310 Cache related to the token. 00:35:26.320 --> 00:35:30.120 the more we can fit in, the more memory we have left over, the more 00:35:30.120 --> 00:35:33.260 stuff we can shove in there before it's full, and that means the less 00:35:33.260 --> 00:35:35.500 stuff we have to eject early, right? 00:35:35.750 --> 00:35:38.780 And if we're ejecting less or we're, or we're dropping 00:35:38.780 --> 00:35:40.270 less, then we're hitting more. 00:35:40.580 --> 00:35:44.490 there's there's a lot of engineering that has to happen that's being obscured by 00:35:44.490 --> 00:35:46.050 the way we're, we're talking about this. 00:35:46.050 --> 00:35:51.650 But the basic idea is like we're able to more efficiently use the 00:35:51.650 --> 00:35:55.690 headroom that we keep in our GPUs, or accelerators for KV cache. 00:35:55.690 --> 00:35:55.883 Nisten, go ahead. 00:35:55.883 --> 00:35:56.012 Alex Volkov: Don't, tell 00:35:56.012 --> 00:35:59.712 Nisten Tahiraj: Guys, those numbers there, those are multiples. 00:35:59.722 --> 00:36:07.242 So I did, y- y- you multiply, 8 times 13 times 9, that's 439 times less cache 00:36:07.302 --> 00:36:08.932 than the first model they released. 00:36:09.672 --> 00:36:13.872 That is insane, because we're just taking this for granted and 00:36:13.872 --> 00:36:15.502 just shoving more interest in it. 00:36:15.922 --> 00:36:19.772 And there's also, yeah, w- when you run inference at scale- 00:36:19.982 --> 00:36:21.902 Alex Volkov: Is that two and a half orders of magnitude smaller, right? 00:36:21.952 --> 00:36:25.922 Nisten Tahiraj: So, so when you run this at scale and you have multiple 00:36:25.922 --> 00:36:30.952 queries, those ones all go in the chip at the same time, so that also takes 00:36:30.952 --> 00:36:34.432 advantage of the on-chip c- cache. 00:36:35.250 --> 00:36:39.040 S- so it, it, it ends up being orders of, of magnitude more improvements. 00:36:39.040 --> 00:36:43.150 And, and that's also why people can serve DeepSeek for, for so cheap now. 00:36:43.150 --> 00:36:43.360 Yeah. 00:36:43.360 --> 00:36:46.940 It's just amazing to me that we just take all of this for granted now. 00:36:47.190 --> 00:36:50.220 It's like, yeah, we reduced the memory requirements by 400 times, 00:36:50.560 --> 00:36:51.950 and, Bro … we still need more of it. 00:36:52.740 --> 00:36:53.140 There's just 00:36:53.140 --> 00:36:53.750 Yam Peleg: Bro, bro. 00:36:53.750 --> 00:36:53.850 Yeah. 00:36:54.000 --> 00:36:57.360 Who is taking it for granted when you get a model- No way … nearly for free? 00:36:57.570 --> 00:36:58.610 We're going to celebrate this. 00:36:58.610 --> 00:36:59.320 That's, that's the enabler. 00:36:59.490 --> 00:37:00.400 Absolutely. 00:37:00.490 --> 00:37:02.890 Wolfram Ravenwolf: I mean, it appears in a list of all these great 00:37:02.950 --> 00:37:05.590 releases, GLM and Kimi and so on. 00:37:05.690 --> 00:37:07.290 It's 4.1 00:37:07.290 --> 00:37:13.530 Flash, which sounds so ah, but that is a, a real jump in, in the data, 00:37:13.680 --> 00:37:17.320 in the architecture with the encoder decoder, Absolutely … with the cache, 00:37:17.320 --> 00:37:19.230 with the pricing and the intelligence. 00:37:19.230 --> 00:37:20.956 It's up there So 00:37:21.346 --> 00:37:23.836 Yam Peleg: I've, I've- Bro, this is the- Wow, blown away … this is the thing. 00:37:23.836 --> 00:37:25.936 That thing allows you to run it at home. 00:37:25.936 --> 00:37:28.676 That thing allows it to be that cheap. 00:37:29.736 --> 00:37:31.056 I mean, that, that thing is… 00:37:31.156 --> 00:37:34.526 A- and they just put it out in the open, like the actual moat- Wow … that 00:37:34.716 --> 00:37:36.236 they don't even need to release. 00:37:36.276 --> 00:37:36.846 Like, think about it. 00:37:36.886 --> 00:37:39.646 They can release the weight, tell you exactly how they trained it. 00:37:40.346 --> 00:37:45.466 That's not- By the way- That's not, like, i- that's, that's the inference kind of… 00:37:45.476 --> 00:37:48.336 Yeah, I mean, yeah, there, there is, there's part of it for, for 00:37:48.336 --> 00:37:52.246 the training as well, but, like, that's, that's not mandatory. 00:37:52.246 --> 00:37:54.990 That's just goodwill, giving us presents for, for free- For free, man. 00:37:54.990 --> 00:37:55.539 That, that's amazing … do you have 00:37:55.539 --> 00:37:55.904 Alex Volkov: something, to thank you- 00:37:55.904 --> 00:37:57.674 Wolfram Ravenwolf: And MIT license 00:37:57.904 --> 00:37:58.714 Alex Volkov: And MIT license. 00:37:58.834 --> 00:37:59.144 Let's go. 00:37:59.904 --> 00:38:00.284 Yam Peleg: Let's go. 00:38:00.284 --> 00:38:04.834 Alex Volkov: Nisten, let's do one last segment on DeepSeek V- V4-1 Flash. 00:38:04.834 --> 00:38:07.084 Please narrate this, for folks who are listening. 00:38:07.254 --> 00:38:09.584 Nisten Tahiraj: yeah, so I just visualized this with Astra. 00:38:09.654 --> 00:38:10.834 It, it just finished it. 00:38:10.834 --> 00:38:11.594 It's the first time. 00:38:11.614 --> 00:38:12.954 Alex Volkov: reshare your, v- video please? 00:38:12.954 --> 00:38:13.344 It dropped. 00:38:13.424 --> 00:38:15.844 Nisten Tahiraj: I post these on Twitter and they're on my GitHub as 00:38:15.844 --> 00:38:18.104 well, and, people really like them. 00:38:18.104 --> 00:38:20.104 So this is the one that Astra made. 00:38:20.144 --> 00:38:23.774 So it visualized all the weights, and this looks like it is a 00:38:23.774 --> 00:38:28.294 different architecture and, Oh, they're also using Ngram encodings, 00:38:28.314 --> 00:38:30.724 just like, kind of like Qwen is. 00:38:30.774 --> 00:38:33.144 Alex Volkov: For folks who are listening, Nisten is showing a very incredibly 00:38:33.164 --> 00:38:39.284 detailed 3D reconstruction of the transformer architecture, including 00:38:39.284 --> 00:38:44.394 the encoder, decoder stuff that Yam was talking about, and it shows how tokens 00:38:44.394 --> 00:38:45.964 are flowing, when they're flowing. 00:38:45.964 --> 00:38:47.814 It's really something to see. 00:38:48.094 --> 00:38:50.244 So if you are interested in this, Nisten, where can they find it? 00:38:50.484 --> 00:38:52.704 Nisten Tahiraj: I will be posting it on Twitter, and it will 00:38:52.704 --> 00:38:54.994 also be on my GitHub because I- 00:38:54.994 --> 00:38:57.224 Alex Volkov: this to the show notes on Thursd AI News- Yeah … so 00:38:58.204 --> 00:39:01.404 Nisten Tahiraj: it's a really good educational content because each one of 00:39:01.404 --> 00:39:06.714 these cubes here, the volume of the cube corresponds to the actual size on disk. 00:39:06.984 --> 00:39:09.354 So this one it says, you, you probably can't see it on the stream, 00:39:09.354 --> 00:39:11.464 but this is the output had 1.32 00:39:11.474 --> 00:39:16.084 gigs, and you, you can explore each weight in, in, in great detail 00:39:16.234 --> 00:39:20.224 th- this way And, it's, it's a pretty good educational content and 00:39:20.294 --> 00:39:22.204 people, people do love this stuff. 00:39:22.254 --> 00:39:26.554 so you can see like the, the weight norms are, are very, very small because 00:39:26.594 --> 00:39:28.074 they're, that's a single linear layer. 00:39:28.454 --> 00:39:32.664 But then you have the, the draft output head, and there's an explanation for that. 00:39:32.914 --> 00:39:38.524 Again, this is very small here for, for the stream, but, again, I'll be, I'll… 00:39:38.724 --> 00:39:41.474 I'll be posting it there also with all the code and stuff too. 00:39:41.544 --> 00:39:42.054 It's just a simple- We'll 00:39:42.054 --> 00:39:46.444 Alex Volkov: Folks, I think, I, I think we've talked about DeepSeek plenty. 00:39:46.644 --> 00:39:49.684 The, the, the numbers kind of speak for themselves as well. 00:39:49.824 --> 00:39:52.554 Terminal Bench 3.0 00:39:52.554 --> 00:39:52.724 DeepSeek V4.1 00:39:52.724 --> 00:39:54.504 Flash gets 30%. 00:39:54.824 --> 00:39:59.074 That's beating Kimi K3 and that, it catches up to GPT 5.6 00:39:59.074 --> 00:39:59.654 Sol at, at 34%. 00:39:59.654 --> 00:40:00.664 74.2 00:40:03.560 --> 00:40:05.840 Score on Deep SWE 1.1. 00:40:05.980 --> 00:40:10.680 Deep SWE notoriously a benchmark and eval that represents how actual coders 00:40:10.680 --> 00:40:14.670 feel, that beats both Opus 5 and GPT 5.6. 00:40:14.670 --> 00:40:18.300 So, I think that the fact that Astra and Fable 0.11 00:40:18.310 --> 00:40:21.950 are not on these graphs makes no difference to me at all. 00:40:21.960 --> 00:40:25.820 Like this is an open source MIT model that catches up to the best 00:40:25.820 --> 00:40:28.810 models of, of before, and also still the best models for some. 00:40:29.110 --> 00:40:30.450 so this is just incredible news. 00:40:30.450 --> 00:40:32.630 Automation base gains for 54%. 00:40:32.680 --> 00:40:34.500 Maybe I'll call out one last thing here. 00:40:34.850 --> 00:40:38.340 reasoning effort for this model is a continuous controllable from one 00:40:38.350 --> 00:40:40.310 to 100, and you can control it. 00:40:40.310 --> 00:40:43.720 So like the extra max, whatever the labs come up with, max, ultra low, 00:40:43.720 --> 00:40:46.540 whatever, you can just come up with your own and say, "Okay, 100 is extra 00:40:46.540 --> 00:40:48.280 max for me," or extra ultra whatever. 00:40:48.500 --> 00:40:54.410 and also CyberGym for this model, the eval that famously because of that, the 00:40:54.410 --> 00:41:00.810 swarms of OpenAI agents hack Hugging Face CyberGym, eh, this model gets 88.1% 00:41:01.550 --> 00:41:04.280 and leads all listed models on CyberGym. 00:41:04.470 --> 00:41:07.710 that's an- another topic that we need to discuss, but I think this 00:41:07.710 --> 00:41:09.010 is a very, very impressive release. 00:41:09.090 --> 00:41:13.400 Again, shout out to the Whale folks, for no drama releases of 00:41:13.400 --> 00:41:14.920 just like incredible engineering. 00:41:15.330 --> 00:41:17.430 Honestly, kind of scary how good they are. 00:41:17.450 --> 00:41:18.010 so great. 00:41:18.050 --> 00:41:18.890 Chris Alexiou, thank you. 00:41:18.890 --> 00:41:20.630 I know, you have meetings, et cetera. 00:41:20.640 --> 00:41:22.630 You are welcome to stay, but I know you have to drop. 00:41:22.630 --> 00:41:24.040 Thank you so much for joining us. 00:41:24.080 --> 00:41:28.980 Folks, please give Chris a follow at llm_wizard on, Twitter. 00:41:29.420 --> 00:41:30.200 Chris Alexiuk: Thanks for having me, guys. 00:41:30.240 --> 00:41:30.980 Have a great day. 00:41:31.030 --> 00:41:33.330 Alex Volkov: all right, folks, I think we've covered, open source enough. 00:41:33.370 --> 00:41:36.900 Maybe we'll mention that other releases also hel- happened. 00:41:36.900 --> 00:41:40.300 Wolfram, I want you to briefly talk about Desert Ant Labs. 00:41:40.340 --> 00:41:42.230 heard about this from Europe, hopefully. 00:41:42.480 --> 00:41:44.680 this is, I think, audio, vision, and text. 00:41:44.690 --> 00:41:48.910 a new lab that releases a bunch of, a bunch of models, specifically from Europe. 00:41:48.910 --> 00:41:52.230 I think that this is, you know, a very new lab, but, very interesting 00:41:52.250 --> 00:41:53.500 on-device models as well. 00:41:54.450 --> 00:41:57.850 spun up out of detail, and ships 18 specialized models 00:41:57.860 --> 00:41:59.620 that run entirely on-device. 00:42:00.490 --> 00:42:05.100 Desert Ant Labs, 18 on-device models, VOS, speech-to-text model, 00:42:05.100 --> 00:42:08.020 Redact PII detection, clips, video highlights, and Clear. 00:42:08.260 --> 00:42:09.630 Actually, you should check it out more. 00:42:09.720 --> 00:42:13.698 the visual style is pretty cool Wolfram, anything you wanna 00:42:13.698 --> 00:42:14.608 add about this or not, not so 00:42:14.608 --> 00:42:16.468 Wolfram Ravenwolf: Very interesting with the feathers from, Europe. 00:42:16.468 --> 00:42:20.268 I haven't heard about them before, so, no inside detail. 00:42:20.298 --> 00:42:21.898 My agent is already doing research. 00:42:22.698 --> 00:42:23.008 Nice. 00:42:23.018 --> 00:42:27.048 But, you know how long it takes if you want some quality information. 00:42:27.058 --> 00:42:31.188 yeah, it's always good to see more investment to AI from 00:42:31.238 --> 00:42:32.738 Europe, any place, actually. 00:42:33.068 --> 00:42:37.828 I think this shouldn't be just-- AI is, for all of humanity, very important, 00:42:37.858 --> 00:42:41.838 and not everybody has realized it, but I think it's very important that everybody 00:42:41.838 --> 00:42:47.298 is developing their own AI as well, and not just rely on China or the USA. 00:42:48.188 --> 00:42:51.448 And there are many, many who would argue for this. 00:42:51.928 --> 00:42:56.288 So it's super important to have something local and, So- … small on-device models, 00:42:56.568 --> 00:42:58.208 that is something everybody can use. 00:42:58.818 --> 00:43:02.218 Alex Volkov: A bunch of, not only that, a bunch of open weights on device models, 00:43:02.218 --> 00:43:03.708 and Native is the case to run them. 00:43:04.028 --> 00:43:07.898 And here's the shout-out, Voz, their, audio transcription that runs on device 00:43:07.898 --> 00:43:11.788 on the iPhone transcribes 10 minutes of speech in just two seconds on an iPhone. 00:43:12.378 --> 00:43:17.278 That's actually very, very impressive, because it's a very small 467 million, 00:43:17.498 --> 00:43:22.818 megabyte, model compared to, like, Whisper Large 3 V3, which is 1.6. 00:43:22.818 --> 00:43:23.678 I'm actually gonna check this out. 00:43:23.678 --> 00:43:24.118 This is dope. 00:43:24.368 --> 00:43:27.468 and, this enables a few use cases that I am excited about. 00:43:27.798 --> 00:43:29.498 All right, so shout out to Desert and Labs. 00:43:29.498 --> 00:43:32.218 Maybe we can bring somebody from them to talk about this on the show. 00:43:32.378 --> 00:43:35.638 And then also inclusion in open source is Ling 3.1 00:43:35.638 --> 00:43:36.208 Flash. 00:43:36.438 --> 00:43:39.338 Ling, we've talked about Ling before, but I haven't seen anyone use 00:43:39.338 --> 00:43:41.638 those, just, like, not one person. 00:43:41.908 --> 00:43:46.718 but, vision language models, with just five billion parameters and, no, 00:43:47.268 --> 00:43:48.718 no competition on the leaderboard. 00:43:48.728 --> 00:43:52.338 So, you know, just, just folks who, who like models to work. 00:43:52.878 --> 00:43:54.248 and MIT license, which is great. 00:43:54.338 --> 00:43:56.088 Yeah, and we love and we, we wanna highlight. 00:43:57.028 --> 00:44:01.308 All right, folks, there's a big, big, big show because Open, you know, OpenAI 00:44:01.828 --> 00:44:05.498 announced some incredible things, so I definitely wanna switch to Frontier Labs. 00:44:05.858 --> 00:44:06.808 let's, let's do this. 00:44:06.858 --> 00:44:13.528 All right, folks, this is maybe one of the biggest hitting news from 00:44:13.678 --> 00:44:16.058 Frontier Labs, from AI generally. 00:44:16.418 --> 00:44:21.708 re-re-remember how two, two years ago folks were posting a question to the 00:44:21.708 --> 00:44:23.948 LLM and saying, "Hey, what's bigger, 9.9 00:44:23.948 --> 00:44:24.958 or 9.11?" 00:44:25.168 --> 00:44:28.438 And the LLM would get confused and people, "Ha ha, LLM cannot do math." 00:44:28.438 --> 00:44:36.478 And then, folks like, Gary what's-his-name and, and, other, other folks who don't 00:44:36.488 --> 00:44:40.078 really think AI can go anywhere said, "AI will never be able to solve math." 00:44:40.088 --> 00:44:45.488 OpenAI claims that there is a solutions to Navier–Stokes 00:44:45.488 --> 00:44:48.228 Millennium Prize mathematics problem. 00:44:48.968 --> 00:44:50.498 This is a historic lane. 00:44:51.138 --> 00:44:53.028 And, LDJ, do you wanna chime in here? 00:44:53.028 --> 00:44:55.968 Usually, c- catch us up on these type of things. 00:44:56.178 --> 00:44:59.728 Last we heard, w- we had some breakthroughs in, in near mathematics 00:44:59.728 --> 00:45:03.658 or, or geometry stuff, but like tell us about like the magnitude of this, from… 00:45:03.708 --> 00:45:05.958 And why is everybody going crazy about it? 00:45:07.008 --> 00:45:10.548 LDJ: So to be clear, I'm not a mathematician and I know you're not. 00:45:10.778 --> 00:45:14.608 but, from my understanding and then from understanding of friends of mine 00:45:14.608 --> 00:45:18.928 that are mathematicians, it's relating to, things of the movement of fluids, 00:45:18.928 --> 00:45:24.208 fluid dynamics, and it does have implications for basically, physics and 00:45:24.898 --> 00:45:30.888 a lot of, a lot of engineering problems also theoretically with how, turbines 00:45:30.908 --> 00:45:35.838 and, and jet engines and those types of things end up working, and it might 00:45:35.958 --> 00:45:38.068 b- be, how they develop in the future. 00:45:38.798 --> 00:45:45.028 And so with this, this is a problem that OpenAI worked on. 00:45:45.288 --> 00:45:50.768 It is, has a million-dollar prize for the past 26 years, and I believe 00:45:50.768 --> 00:45:54.798 overall, in terms of when the problem was initially proposed, I wanna say it's 00:45:54.808 --> 00:45:57.598 at least 60, 70 years old, if not older. 00:45:58.548 --> 00:46:00.838 And yeah, nobody has solved it yet. 00:46:01.208 --> 00:46:04.828 There have been some claims of some progress towards related problems and 00:46:05.558 --> 00:46:10.128 subproblems relating to Navier-Stokes, but no actual full solutions, and 00:46:10.128 --> 00:46:14.138 there has been some drama over the past week or two- Yeah And it's… 00:46:14.168 --> 00:46:19.048 There's, there, there's a lot of nuances and a lot of drama i- involved there, 00:46:19.048 --> 00:46:21.038 which we probably can't cover all here. 00:46:21.588 --> 00:46:27.668 but long story short, this is the first solution to ever be actually proposed 00:46:27.688 --> 00:46:29.848 for this, and it's really significant. 00:46:29.988 --> 00:46:30.198 Yeah. 00:46:30.668 --> 00:46:33.458 Alex Volkov: My, my research says this is the level, the 00:46:33.458 --> 00:46:35.248 scope of a moon landing for AI. 00:46:35.288 --> 00:46:38.408 This is like that level of a checkpoint in the world of 00:46:38.438 --> 00:46:40.228 mathematics, in the world of AI. 00:46:40.798 --> 00:46:43.718 it looks like famous mathematicians also, looked at some of the 00:46:43.718 --> 00:46:45.658 stuff from OpenAI and confirmed. 00:46:45.668 --> 00:46:48.758 Now, my, my research says that this is not yet independently verified. 00:46:48.968 --> 00:46:51.598 Clay Mathematics Institute, review pending. 00:46:51.778 --> 00:46:53.828 it's really funny that the Millennium Prize is what? 00:46:53.838 --> 00:46:56.868 $1 million, and hasn't been awarded in a long time. 00:46:57.198 --> 00:47:03.428 reportedly, OpenAI spent an order of 130 billion output tokens, 00:47:03.768 --> 00:47:05.528 which, of their unreleased model. 00:47:05.528 --> 00:47:10.820 This is not Astra And so that's in the order of like 130 million or so spent. 00:47:10.960 --> 00:47:13.020 we don't know yet 'cause they didn't price the new model. 00:47:13.020 --> 00:47:19.370 But, you know, OpenAI spent millions of dollars, with 10,000 coordinating agents. 00:47:19.370 --> 00:47:23.000 Folks, this is like, this is way bigger than the swarm of the Hug- Hugging Face. 00:47:23.290 --> 00:47:27.580 powered by unreleased model beyond GPT-6 Astra, w- and produced a 00:47:27.620 --> 00:47:29.600 lean proof, lean mathematics proof 00:47:29.660 --> 00:47:32.770 Yam Peleg: Brother, brother, that's not the right vibe. 00:47:32.770 --> 00:47:36.720 Like- right vibe for this is like, "GPT-6 destroys a 00:47:36.720 --> 00:47:39.990 millennial pro- uh, a millennium problem-" "… of the Clay Institute. 00:47:40.140 --> 00:47:41.050 What the fuck? 00:47:41.490 --> 00:47:41.660 When?" 00:47:41.880 --> 00:47:46.380 Look, I don't know, I don't know what you think, guys, but the, the, 00:47:46.520 --> 00:47:51.660 the millennium problems are, w- they started as, as a, as a bunch of them. 00:47:52.020 --> 00:47:58.050 and a- after I think, I think 70 years, like, like you said, we are left with only 00:47:58.090 --> 00:48:04.290 a bunch of, a bunch of them, like seven or six, that people have been constantly 00:48:04.300 --> 00:48:09.710 trying to solve and that, like, there are pages, online shaming the failed attempts 00:48:09.710 --> 00:48:13.450 of the paper, like putting the actual wall of shame, all, all the failed attempts, 00:48:13.450 --> 00:48:16.720 all the papers that have re- retracted for each and every single one of them. 00:48:17.220 --> 00:48:22.040 I, I've been waiting my entire life to read this, that Navier-Stokes is solved. 00:48:22.430 --> 00:48:25.340 The thing is, that thing describes everything. 00:48:25.680 --> 00:48:29.820 Like, you can describe, nearly everything in physics with, some sort 00:48:29.820 --> 00:48:35.260 of a partial differential equation that, It's a very general framework. 00:48:35.310 --> 00:48:41.390 the problem is that we don't really know about all the solutions of it. 00:48:41.490 --> 00:48:45.580 I don't wanna go too technical, but, like, for each of these, millennial 00:48:45.620 --> 00:48:49.060 problems, every single one of them is extremely famous in its field. 00:48:49.070 --> 00:48:54.440 Like there, where there is, P versus NP, Navier-Stokes and, Poincaré, conjecture. 00:48:54.440 --> 00:48:57.620 And, there are a handful of very famous problems that are extremely 00:48:57.620 --> 00:49:00.120 fundamental to their own fields. 00:49:00.540 --> 00:49:04.170 and people have been trying to solve them, to brute force them to, They 00:49:04.170 --> 00:49:09.780 are insane and, it's a historic event when one of them gets solved because 00:49:10.060 --> 00:49:11.980 there are only a handful of them. 00:49:11.980 --> 00:49:13.370 It's not even about the price. 00:49:13.370 --> 00:49:15.090 Yeah, of course, they burned more money- 00:49:15.530 --> 00:49:15.650 Alex Volkov: Yeah 00:49:15.700 --> 00:49:15.790 Yam Peleg: Yeah. 00:49:15.790 --> 00:49:17.750 Yeah, that's, that's not, that's not about it. 00:49:18.270 --> 00:49:20.700 And it's not even about using the, the… 00:49:21.060 --> 00:49:22.330 What it-- How, how is it called? 00:49:22.510 --> 00:49:23.570 unreleased model. 00:49:23.570 --> 00:49:27.040 Like they, they had a nice name and I was like, "Can I, can I try this 00:49:27.040 --> 00:49:31.200 and, and like let it, I don't know, write, write me an email or something?" 00:49:31.200 --> 00:49:36.000 Why waste the tokens of, of this one to do very mundane, stupid things because, 00:49:36.610 --> 00:49:38.460 man, I, I wanna, I wanna play with it. 00:49:38.510 --> 00:49:42.446 The thing is that okay, that's, that's the excitement. 00:49:42.446 --> 00:49:44.526 I just, I just wanna put things into perspective. 00:49:45.426 --> 00:49:51.036 We've already seen GPT- thr- GPT-6 destroying, very famous problems. 00:49:51.466 --> 00:49:57.076 Y- you need to find an example in this specific problem. 00:49:57.076 --> 00:50:01.376 Also in the other, famous, problem, they found an example. 00:50:01.386 --> 00:50:06.666 So LLMs … Okay, look, if you run , if you run, I don't know, 10,000, 100,000 00:50:06.746 --> 00:50:11.986 of, instances of GPT-6, one of them is going to stumble upon maybe the 00:50:11.986 --> 00:50:14.996 solution because they are gener- generally in the right direction. 00:50:15.026 --> 00:50:20.046 I'm not, I'm not taking away from this, but the most insane thing 00:50:20.156 --> 00:50:24.986 about this is that it's very hard to argue against a lean proof. 00:50:25.426 --> 00:50:29.576 And first, the first stage, yeah, that's monumental. 00:50:29.626 --> 00:50:30.356 absolutely. 00:50:31.376 --> 00:50:34.476 But they, they, they didn't wanna take chances, so they 00:50:34.476 --> 00:50:37.006 even went proving it with lean. 00:50:37.836 --> 00:50:40.736 It's, it's as, as bulletproof as you can get. 00:50:40.756 --> 00:50:45.416 Alex Volkov: verification took 17 hours, ad- additional, like 00:50:45.416 --> 00:50:48.846 additional 17 hours with Astro to just, like, do lean verification. 00:50:49.116 --> 00:50:49.356 Yam Peleg: all right. 00:50:49.356 --> 00:50:53.206 Astro, Astro, if you listen, I, I wanna apply as a subagent on this project. 00:50:53.206 --> 00:50:53.426 Come on, man. 00:50:53.456 --> 00:50:53.476 Yeah. 00:50:53.486 --> 00:50:55.846 Like, how do you, how do you even 00:50:55.906 --> 00:50:56.426 Alex Volkov: Yeah, I appreciate the excitement. 00:50:56.426 --> 00:50:56.946 Mad respect. 00:50:56.996 --> 00:51:00.666 I think there's more, you know, there's more, prizes for them to solve. 00:51:00.786 --> 00:51:05.096 It looks like both OpenAI and, and Anthropic are now at the level of, "Hey, 00:51:05.156 --> 00:51:10.086 we can use these agents, plus use this swarmy thing to start, like, breaking down 00:51:10.086 --> 00:51:14.686 fundamental problems that the humanity has been dealing with," which nobody 00:51:14.726 --> 00:51:16.656 as humans was able to solve before. 00:51:17.086 --> 00:51:18.696 some folks are naysayers. 00:51:18.746 --> 00:51:21.516 We have some folks in the comments as well saying that, "Hey, matip- 00:51:21.516 --> 00:51:24.026 ma- mathematicians, friends of mine don't agree to this." 00:51:24.046 --> 00:51:28.336 Well, fucking post your lean proof then, if you sh- if you, if you 00:51:28.356 --> 00:51:29.716 don't think that this is true. 00:51:29.886 --> 00:51:30.536 but we'll see. 00:51:30.566 --> 00:51:30.936 We'll see. 00:51:30.936 --> 00:51:33.946 Maybe folks will come back and say, "Hey," you know, "OpenAI, 00:51:33.976 --> 00:51:34.556 like, missed something." 00:51:34.576 --> 00:51:39.256 However, worth saying that folks who work at OpenAI who shepherd these models. 00:51:39.606 --> 00:51:41.146 They are also mathematicians. 00:51:41.396 --> 00:51:44.426 Sebastian Rachka, I believe is, is the person there. 00:51:44.446 --> 00:51:46.806 And, the drama, let's mention the drama a little bit. 00:51:46.806 --> 00:51:48.226 There was drama about this. 00:51:48.516 --> 00:51:51.826 Apparently, two human mathematicians, one of them works in Anthropic, 00:51:52.116 --> 00:51:57.816 Levent, were about to go public with this news on their own, separate from 00:51:57.816 --> 00:51:59.506 the unreleased model from OpenAI. 00:52:00.526 --> 00:52:04.876 I'm dealing with very rumored posts on, on, on Twitter and Reddit here. 00:52:05.146 --> 00:52:09.286 OpenAI caught wind of this, that they're about to go, and OpenAI had an independent 00:52:09.326 --> 00:52:11.796 effort to, to go and try to solve this. 00:52:12.226 --> 00:52:15.386 Either that or OpenAI got excited and decided to start solving this 00:52:15.396 --> 00:52:16.766 when they heard that it's possible. 00:52:17.706 --> 00:52:21.806 And so the mathematician, independent one, that worked with his friend, who 00:52:21.816 --> 00:52:26.166 in his personal capacity, but works in Anthropic, provided him access to Mythos 00:52:26.166 --> 00:52:27.656 2 or whatever unreleased model for Mythos. 00:52:28.066 --> 00:52:30.046 They were able to solve this in one way. 00:52:30.046 --> 00:52:32.296 Apparently, there's multiple ways to attack this problem. 00:52:33.236 --> 00:52:37.446 so OpenAI caught wind, and OpenAI's solution is not related to what they did. 00:52:37.456 --> 00:52:38.526 I'm not sure what blow up means here. 00:52:38.826 --> 00:52:43.576 But, there is drama where that person who was about to publish said on forums 00:52:43.596 --> 00:52:47.836 like, "Hey, we talked to OpenAI, and you know, they, they don't want us 00:52:47.836 --> 00:52:49.336 on the release together," et cetera. 00:52:49.656 --> 00:52:53.676 "And potentially they have looked at what we did, and we uploaded tons 00:52:53.676 --> 00:52:55.866 of papers, with OpenAI," et cetera. 00:52:55.926 --> 00:52:58.936 it looks like the Anthropic folks derailed the conversation. 00:52:59.266 --> 00:53:02.256 OpenAI were ready to give this person, "Hey, here's the dude. 00:53:02.426 --> 00:53:04.176 We just like verified whatever he did." 00:53:04.386 --> 00:53:06.626 but because the Anthropic dude was there, they said, "We obviously 00:53:06.626 --> 00:53:11.376 cannot, credit a competitor of ours on our achievement, so if we remove 00:53:11.406 --> 00:53:12.736 that person you can get the credit." 00:53:12.766 --> 00:53:15.966 and many folks started speculating whether or not OpenAI's models 00:53:15.976 --> 00:53:18.846 actually trained on the solution that they provided on the papers. 00:53:19.466 --> 00:53:21.706 OpenAI said, "No, we don't train on your texts." 00:53:22.146 --> 00:53:26.396 but then they mentioned something that says, a de-anonymized-- o-OpenAI 00:53:26.396 --> 00:53:30.816 acknowledges it cannot rule out de-identified usage data from, Alperge, 00:53:30.826 --> 00:53:32.296 the guy, the independent, and, Buckmaster. 00:53:32.526 --> 00:53:33.206 Buck-Buckmaster? 00:53:34.486 --> 00:53:36.036 Yeah, Lev and Buckmaster, who works at Anthropic. 00:53:37.246 --> 00:53:40.146 we cannot rule out that this helped improve the models, but 00:53:40.146 --> 00:53:43.676 says that the proof differ and no specific user data was accessed. 00:53:44.386 --> 00:53:48.754 And yeah, so Comments? 00:53:48.794 --> 00:53:49.724 LDJ, I see your hand up. 00:53:49.724 --> 00:53:50.864 Please chime in here. 00:53:51.544 --> 00:53:54.474 LDJ: I think what's very important here, which OpenAI and across 00:53:54.474 --> 00:53:58.554 several tweets by several different employees have clarified on, is 00:53:58.824 --> 00:54:00.424 there's two ways to opt out of data. 00:54:00.434 --> 00:54:03.524 There's kind of like a specific form that you can go on their website 00:54:03.564 --> 00:54:07.004 to fill out, or you can literally just make sure that you have the 00:54:07.034 --> 00:54:10.754 improve the model for everyone toggle turned off in your ChatGPT settings. 00:54:11.104 --> 00:54:14.144 They've confirmed that as long as-- If you do either of those two things, 00:54:14.144 --> 00:54:17.234 you don't have to do both of them, you just have to do one of them, then 00:54:17.244 --> 00:54:18.584 your chats will not be trained on. 00:54:19.124 --> 00:54:20.474 But even if your chats… 00:54:20.654 --> 00:54:22.944 Let, let's say you do turn that toggle on so that you do 00:54:22.944 --> 00:54:24.304 improve the model for everyone. 00:54:25.004 --> 00:54:28.684 Even when it's trained on in that case, it's, it's the, sorry, I 00:54:28.684 --> 00:54:31.674 forgot the term, but like de-de- De-anonymized … de-identified, 00:54:31.674 --> 00:54:32.494 De-anonymized … anonymized. 00:54:32.544 --> 00:54:32.784 Yeah. 00:54:33.264 --> 00:54:33.524 Yeah. 00:54:33.524 --> 00:54:36.524 So, to remove your private, private data and everything from that. 00:54:36.944 --> 00:54:44.134 So given this, I do find it a bit weird and interesting that it seems like the, 00:54:44.144 --> 00:54:48.504 the mathematicians involved in the drama here s- claiming that maybe their chats 00:54:48.504 --> 00:54:53.194 were trained on, I have not seen any of them post any screenshots of whether 00:54:53.194 --> 00:54:55.384 or not they have that toggle on or off. 00:54:56.224 --> 00:54:56.654 Chris Alexiuk: Mm-hmm. 00:54:56.654 --> 00:54:59.404 LDJ: And that seems like a very easy thing that they could just do and 00:54:59.404 --> 00:55:02.964 post, like, "Hey, look, I've had the toggle off for the past year. 00:55:02.964 --> 00:55:03.254 I…" 00:55:03.714 --> 00:55:06.714 Or, or at least I'm pretty confident I am, and here- Yeah … you can see in 00:55:06.714 --> 00:55:08.504 the screenshot I have the toggle off. 00:55:09.034 --> 00:55:10.794 But I haven't seen any of them post that. 00:55:10.794 --> 00:55:14.264 So I, I don't know, may-maybe they just haven't, seen those posts from 00:55:14.264 --> 00:55:16.544 OpenAI yet about, about that fact. 00:55:16.544 --> 00:55:18.044 But, I guess I'm just waiting for that now. 00:55:18.824 --> 00:55:21.804 Alex Volkov: So LDJ, we put up a HR that you, that you added. 00:55:21.804 --> 00:55:23.574 Let's, let's talk about the chart a little bit because I 00:55:23.574 --> 00:55:24.944 think it's it's quite incredible. 00:55:24.944 --> 00:55:25.554 LDJ: Oh, yes. 00:55:25.624 --> 00:55:28.294 And sorry, I accidentally cropped out the top part, but this is, 00:55:28.704 --> 00:55:32.134 open problems in mathematics that exist publicly that OpenAI curated. 00:55:32.604 --> 00:55:37.254 And the bottom line here is, this is the estimated progress of Sol. 00:55:37.954 --> 00:55:40.934 I say estimated because they didn't actually directly test Sol here. 00:55:41.284 --> 00:55:41.494 Alex Volkov: Yeah. 00:55:41.544 --> 00:55:43.024 LDJ: based on other math benchmarks. 00:55:43.034 --> 00:55:44.054 And then you see Astra. 00:55:44.214 --> 00:55:47.434 Astra and the internal model, those are actual measurements 00:55:47.434 --> 00:55:48.904 that OpenAI did internally. 00:55:49.714 --> 00:55:52.694 And this is the amount of those open problems, which these are 00:55:52.724 --> 00:55:55.914 unsolved problems in mathematics that no human has ever solved before. 00:55:56.184 --> 00:56:00.534 Astra ended up at max reasoning being able to solve roughly 15% of them. 00:56:00.894 --> 00:56:03.284 At low reasoning, a little under 10% of them. 00:56:03.984 --> 00:56:08.204 And then for their internal model, which a lot of people have dubbed as, 00:56:08.574 --> 00:56:11.634 have speculated this is a pre-trained model internally called Bell. 00:56:12.234 --> 00:56:12.584 Speaker 3: Hmm. 00:56:12.744 --> 00:56:18.524 LDJ: And at the low reasoning setting, it's performing nearly 00:56:18.524 --> 00:56:23.344 twice as, as good in terms of overall amount of, problems, solved than 00:56:23.364 --> 00:56:25.064 even the max reasoning of Astra. 00:56:25.454 --> 00:56:28.074 Alex Volkov: and then we just got- Yeah … Astra last week and blew our minds. 00:56:28.404 --> 00:56:33.934 OpenAI is showing here that they have a model that's farther away from Astra than 00:56:33.944 --> 00:56:37.974 Astra is from Sol, and this is the model that's solving, like, fucking mathematics. 00:56:38.074 --> 00:56:38.484 Look at this. 00:56:38.484 --> 00:56:41.064 folks, Yam, let me just, let me just land this point here. 00:56:41.374 --> 00:56:42.364 we're, we're here. 00:56:43.074 --> 00:56:47.374 We're at the point where we told you two years ago, one year ago that scale 00:56:47.374 --> 00:56:51.844 fucking works, and we'll get to a point where AI solves real fucking problems. 00:56:52.284 --> 00:56:55.494 And this is problems that humans, all humans, the best humans, 00:56:55.534 --> 00:57:00.494 Terence Tao motherfucker levels mathematicians were not able to 00:57:00.494 --> 00:57:02.044 solve to claim the one million prize. 00:57:02.044 --> 00:57:04.054 It's a pretty good incentive for a human, one million prize. 00:57:04.974 --> 00:57:07.434 Were not able to solve, and now we're there. 00:57:07.884 --> 00:57:12.824 Folks can say all they want, "Hey, this is a search problem, and you searched Opa." 00:57:12.864 --> 00:57:15.034 I don't give a f- how this is solved. 00:57:15.564 --> 00:57:19.844 if you guys remember, two years ago, two and a half years ago, three years ago when 00:57:19.844 --> 00:57:23.824 we did the LK-99 excitement for a second with, with, Do you guys remember LK-99? 00:57:24.204 --> 00:57:27.524 The, the, room temperature semi-semiconductor that was 00:57:27.524 --> 00:57:29.864 supposed to change the world that came from Korea, et cetera. 00:57:30.094 --> 00:57:32.374 back then we talked about, with mathematicians, physicians, 00:57:32.374 --> 00:57:33.484 et cetera, professors. 00:57:33.754 --> 00:57:34.744 That's a search problem. 00:57:35.174 --> 00:57:39.574 Nothing in physics in theory prevents a room temperature superconductor. 00:57:39.694 --> 00:57:40.924 We just haven't found that. 00:57:40.964 --> 00:57:44.514 Yeah, if, if, if Sam Altman, by the way, he said, "Sure, let's try this." 00:57:44.754 --> 00:57:50.804 if, if they sent 10 million, whatever, this level of model to go and search 00:57:50.814 --> 00:57:56.444 the possible, you know, field of, of, of materials and find us a, a room 00:57:56.444 --> 00:58:01.344 temperature semiconductor, AI can start changing things really for everyone. 00:58:01.894 --> 00:58:04.654 I don't know if Navier-Stokes changes things immediately, but 00:58:04.674 --> 00:58:06.714 you know, a, a, a cancer cure can. 00:58:06.994 --> 00:58:10.934 A, a LK, you know, semi- room level, room temperature 00:58:11.214 --> 00:58:12.934 semiconductor can, superconductor. 00:58:13.104 --> 00:58:14.846 Yam, let's, let's, land on the- it's just the 00:58:14.846 --> 00:58:15.276 Nisten Tahiraj: same… 00:58:15.356 --> 00:58:16.926 It's just the Mars driver test, 00:58:16.926 --> 00:58:18.236 Alex Volkov: I thought you were talking about Navier-Stokes. 00:58:18.376 --> 00:58:18.616 no, no. 00:58:18.626 --> 00:58:18.966 Okay. 00:58:19.666 --> 00:58:23.256 Yam, tell us about, y- you had one last comment, I believe, before we move on- 00:58:23.266 --> 00:58:24.736 Yeah … move on to other things from this one … I was just saying, look, 00:58:25.176 --> 00:58:26.156 Yam Peleg: look at this. 00:58:26.246 --> 00:58:28.356 Look at the white chart. 00:58:28.726 --> 00:58:30.336 Like, just look at this. 00:58:30.336 --> 00:58:30.460 Alex Volkov: pull 00:58:30.460 --> 00:58:34.220 Yam Peleg: The, the white chart of, of the internal model. 00:58:34.530 --> 00:58:36.720 Internal model x high, ma- max. 00:58:36.750 --> 00:58:37.310 Internal… 00:58:37.620 --> 00:58:44.660 That, yeah, that, that one I mean, Astra, the, the AGI moment i- is the blue one. 00:58:44.990 --> 00:58:47.730 Sol is, I think, less than two month. 00:58:48.040 --> 00:58:51.720 You said couple of month ago, but it doesn't do justice to- To how 00:58:51.720 --> 00:58:52.930 short … a month and a little bit. 00:58:52.970 --> 00:58:53.120 Yes. 00:58:53.130 --> 00:58:53.150 Yeah. 00:58:53.160 --> 00:58:56.370 It, yeah, it's pretty much, it's pretty much in July. 00:58:56.930 --> 00:58:59.680 And look at the white chart. 00:58:59.810 --> 00:59:01.100 Look at the white line. 00:59:01.700 --> 00:59:04.920 That's internal model, okay? 00:59:04.970 --> 00:59:08.510 And it can be a search problem. 00:59:09.240 --> 00:59:09.950 you know what? 00:59:10.010 --> 00:59:10.690 No problem. 00:59:10.690 --> 00:59:13.100 It c- you can call it a search problem. 00:59:14.050 --> 00:59:14.860 Go search. 00:59:15.040 --> 00:59:20.390 Like, you can't just search the entire space without anything 00:59:20.390 --> 00:59:24.070 intelligence- Yeah … enough to at least search the right solutions. 00:59:24.420 --> 00:59:26.740 Yeah, you need to, to launch a, I don't know, 00:59:26.930 --> 00:59:27.083 Alex Volkov: 100,000 00:59:27.083 --> 00:59:28.313 Yam Peleg: Yeah, if you don't know what to search it 00:59:28.313 --> 00:59:29.633 Alex Volkov: for, you can go and search 00:59:29.663 --> 00:59:32.863 Yam Peleg: Yeah, if it was just brute force, it would've been solved already. 00:59:32.873 --> 00:59:33.373 There is… 00:59:33.443 --> 00:59:35.813 It's not about the money, it's, so many people tried. 00:59:35.953 --> 00:59:37.083 So many people have computers. 00:59:37.853 --> 00:59:44.073 You do need GPT-6 internal model version to, to do this. 00:59:44.403 --> 00:59:49.123 Or, or, I'm, I'm not taking away, credit from the other competing, 00:59:49.403 --> 00:59:52.703 competing party, or Fable needs us to… 00:59:52.863 --> 00:59:56.703 Like, but you do need these to do this this way, okay? 00:59:57.583 --> 00:59:58.593 And I don't know, I don't know. 00:59:58.623 --> 01:00:03.313 I, I, I genuinely don't know about the, all the, a- all the, all the drama. 01:00:03.693 --> 01:00:04.093 Alex Volkov: That's all right. 01:00:04.213 --> 01:00:06.203 Yam Peleg: But yeah, mad respect. 01:00:06.203 --> 01:00:06.573 Alex Volkov: mad respect … big moment. 01:00:06.573 --> 01:00:08.523 LDJ, one last comment before I move on, please. 01:00:09.583 --> 01:00:09.953 Yeah. 01:00:10.003 --> 01:00:11.213 'Cause where we're moving to is, is a- Mm 01:00:11.223 --> 01:00:13.993 completely different end of the spectrum, so I would love one last comment. 01:00:14.473 --> 01:00:14.743 LDJ: Yeah. 01:00:14.773 --> 01:00:19.233 They, they also did confirm that this, this larger model, this, this, this 01:00:19.233 --> 01:00:23.333 better model, rather, that's in that chart that I showed, that, that it was 01:00:23.333 --> 01:00:28.303 still in training while they were solving, the Navier–Stokes equation when they 01:00:28.313 --> 01:00:30.463 had up to 10,000 agents running of it. 01:00:30.673 --> 01:00:32.383 Meaning it, it was… 01:00:33.203 --> 01:00:36.283 And partway through, they actually updated the checkpoint that they're using for 01:00:36.353 --> 01:00:39.863 Navier–Stokes, and so it, it, it ended up being able to work on it even better. 01:00:40.113 --> 01:00:43.293 But yeah, it seems like as of the past few days at least, the model's still in 01:00:43.293 --> 01:00:44.893 training, and so it's not fully done yet. 01:00:45.993 --> 01:00:48.303 Alex Volkov: All right, folks, as the saying goes, we have breaking news 01:00:48.303 --> 01:00:49.783 before we move on to the doomer-y part. 01:00:49.823 --> 01:00:50.273 Let's go. 01:00:53.073 --> 01:00:58.293 AI breaking news coming at you only on ThursdAI. 01:01:04.465 --> 01:01:08.365 For the past three and a half years, maybe the most fun part of the show 01:01:08.365 --> 01:01:12.075 here, besides chatting with my co-host here and chatting with you guys in the 01:01:12.075 --> 01:01:14.525 comments, is the breaking news effect. 01:01:14.885 --> 01:01:19.155 So many labs love breaking news on Thursday, which is crazy. 01:01:19.345 --> 01:01:20.915 maybe we'll get Grok today, I don't know. 01:01:20.915 --> 01:01:24.375 But, right now from Cognition, from the folks who built Devin, the AI assistant, 01:01:24.735 --> 01:01:30.275 SWE 2, S-W-E 2, our closest model yet to the frontier on leading evals. 01:01:30.275 --> 01:01:35.565 SWE 2 scores on par with recent frontier models and up at 70% lower cost. 01:01:35.565 --> 01:01:38.415 Folks, we just talked to you about DeepSeek doing very, very cheap models. 01:01:38.505 --> 01:01:44.825 Here is, Cognition SWE 2 catching up to Fable 5.1 01:01:45.445 --> 01:01:50.825 and then GPT-6 Astra, on Frontier Code, which is, I believe, their 01:01:50.865 --> 01:01:54.855 own benchmark, that a friend of the pod, Swyx, helped, move in Cognition 01:01:54.855 --> 01:01:56.355 and talked to us about on, on air. 01:01:56.655 --> 01:01:57.685 Frontier Code 1.1 01:01:57.695 --> 01:02:01.885 main, SWE 2 gets 50%, Fable gets 50.9, 01:02:02.065 --> 01:02:03.355 and Astra gets 53. 01:02:03.385 --> 01:02:04.155 So very, very close. 01:02:04.185 --> 01:02:05.175 Astra's really good at that. 01:02:05.225 --> 01:02:06.295 but beating 5.6 01:02:06.295 --> 01:02:08.805 Sol and beating Grok 4.6, 01:02:08.855 --> 01:02:09.495 which is really good. 01:02:09.595 --> 01:02:12.525 I wonder why they're not including Muse, but that's okay. 01:02:12.845 --> 01:02:17.205 On DeepSwe, this SWE gets 73%, and beats everybody besides 01:02:17.205 --> 01:02:21.115 Astra, including Fable, at 74. 01:02:21.125 --> 01:02:25.655 So Astra is at 74 here, and this is a 92.8 01:02:25.695 --> 01:02:26.935 on Terminal Bench 2.1. 01:02:27.705 --> 01:02:30.625 This is the top score on Terminal Bench from the one they measured. 01:02:30.925 --> 01:02:33.305 now let's talk about the, the cost. 01:02:33.305 --> 01:02:34.145 I wanna see the cost. 01:02:34.165 --> 01:02:40.075 Here is the cost chart where they have, GPT Astra and Sol and Fable. 01:02:40.525 --> 01:02:41.875 Fable is ridiculous, the cost chart. 01:02:41.895 --> 01:02:42.825 Don't even look at Fable. 01:02:43.015 --> 01:02:45.115 SWE 2 is ridiculously cheap. 01:02:45.115 --> 01:02:45.645 Look at that. 01:02:45.785 --> 01:02:46.505 I don't even know. 01:02:46.535 --> 01:02:50.869 I want this chart, per I want the actual price here. 01:02:51.109 --> 01:02:53.959 But on frontier code, SWE 2 achieves a score of 50%. 01:02:54.329 --> 01:02:55.634 It beats, SWE 1.7, 01:02:55.634 --> 01:02:56.339 Grok 4.6, 01:02:56.339 --> 01:02:59.189 and GPT 4.6o 01:02:59.199 --> 01:03:02.459 while matching Fable at 64% lower cost. 01:03:03.669 --> 01:03:08.099 And, they, they offer support to effort levels as we- especially, Cognition 01:03:08.099 --> 01:03:13.949 just got a- another inflow of cash from a16z and, Devin is the one AI 01:03:13.999 --> 01:03:15.739 that is not beholden to any lab. 01:03:15.739 --> 01:03:19.949 So if you wanna find Astra or you wanna find Fable, Devin, I think, 01:03:19.949 --> 01:03:23.619 is one of the only places that still allow this, like, directly. 01:03:23.719 --> 01:03:28.349 Cursor doesn't have GPT level stuff anymore because Cursor aligns with, you 01:03:28.349 --> 01:03:30.339 know, with xAI, and they serve Anthropic. 01:03:30.589 --> 01:03:34.159 Devin, which bought Windsurf, so they actually have that 01:03:34.159 --> 01:03:36.889 functionality, definitely has, both. 01:03:37.309 --> 01:03:41.849 And, SWE 2 is available in Devin CLI, and they're making it free for all 01:03:41.849 --> 01:03:46.679 Pro, Max and Team subscribers for next month, which is very impressive. 01:03:47.359 --> 01:03:48.559 Very impressive from Cognition. 01:03:48.569 --> 01:03:49.629 Shout out to Cognition, folks. 01:03:49.629 --> 01:03:50.729 Obviously, we didn't use this. 01:03:50.739 --> 01:03:54.679 It came out literally 30 minutes ago, but here we have 01:03:54.909 --> 01:03:56.609 very, very impressive, things. 01:03:56.969 --> 01:04:02.629 just as a reminder, when Elon Musk bought Cursor, mostly because of the pro-- not 01:04:02.649 --> 01:04:05.649 the products, but the data, but also the products, he bought them specifically 01:04:05.649 --> 01:04:08.089 to help him train Grok, and Grok 4.6 01:04:08.089 --> 01:04:10.379 was really, really good, and, Grok 4.7 01:04:10.379 --> 01:04:11.759 supposedly is significantly better. 01:04:11.779 --> 01:04:14.599 That's because many people use Cursor for free, and they share 01:04:14.809 --> 01:04:16.279 their tokens for training. 01:04:16.589 --> 01:04:18.089 Devin has that in droves. 01:04:18.319 --> 01:04:21.959 Like, many people who use Devin probably support Devin as well, so they have 01:04:21.959 --> 01:04:23.799 a lot of that knowledge as well. 01:04:23.799 --> 01:04:29.459 So shout out to Devin for SWE 2, and Pareto Frontier on RL as well. 01:04:30.679 --> 01:04:31.509 Any comments, folks? 01:04:31.579 --> 01:04:32.419 folks who use Devin? 01:04:32.419 --> 01:04:33.399 I know Ryan Carson uses it, 01:04:33.449 --> 01:04:37.569 Nisten Tahiraj: And, everybody who uses it seems to like it quite a lot, so. 01:04:37.579 --> 01:04:37.839 Alex Volkov: Yeah. 01:04:38.709 --> 01:04:39.329 Nisten Tahiraj: I also participate. 01:04:39.329 --> 01:04:39.799 Alex Volkov: I love Devin. 01:04:40.209 --> 01:04:43.529 And, shout out, they gave me a few tokens, so when I ran out of Astro 01:04:43.529 --> 01:04:47.099 tokens on, on Codex, I went to Devin and started asking for stuff. 01:04:47.449 --> 01:04:47.759 but yeah. 01:04:48.069 --> 01:04:48.659 So shout out to Devin. 01:04:48.659 --> 01:04:48.709 Devin 01:04:48.709 --> 01:04:51.089 Wolfram Ravenwolf: says they'll have the free, if you have an open source 01:04:51.089 --> 01:04:53.259 project, it reviews PRs for you for free. 01:04:54.039 --> 01:04:55.399 They definitely have that, and I use them a lot. 01:04:55.419 --> 01:04:56.789 Alex Volkov: Devin Review is, for free. 01:04:57.059 --> 01:05:00.049 it adds, to your, GitHub and just reviews your PRs. 01:05:00.049 --> 01:05:00.465 It was really, really 01:05:00.465 --> 01:05:03.115 Wolfram Ravenwolf: To me, it felt the intelligence of Devin felt, ahead 01:05:03.115 --> 01:05:05.435 of the game before Soul came out. 01:05:05.435 --> 01:05:09.645 Soul was the first where I felt the same when I ran a model, but before that, Devin 01:05:09.645 --> 01:05:14.025 was definitely finding stuff and issues that other models simply just didn't see. 01:05:14.595 --> 01:05:17.085 You uploaded a PR and it found some bugs. 01:05:17.085 --> 01:05:18.685 You fixed them, it found some more bugs. 01:05:18.875 --> 01:05:18.895 Yeah. 01:05:18.895 --> 01:05:21.545 And so it was almost annoying, but it was excellent. 01:05:21.755 --> 01:05:23.695 Alex Volkov: All right, folks, we're almost live on the air. 01:05:23.695 --> 01:05:26.875 We just had breaking news, and before we continue to our next segment, to 01:05:26.875 --> 01:05:31.575 talk about the rise of doomerism in the world, I definitely wanna tell you 01:05:31.955 --> 01:05:36.215 something about, the presenter of the show, Weights & Biases, and CoreWeave. 01:05:36.255 --> 01:05:38.765 So let's, let's go to this week's buzz for just a few moments, and 01:05:38.765 --> 01:05:39.745 then we'll continue with the show. 01:05:57.391 --> 01:06:00.641 Folks, welcome to this week's Vibe, the corner of Thursd AI where we tell you 01:06:00.641 --> 01:06:03.441 everything that happens in the world of, Weights & Biases and CoreWeave. 01:06:03.481 --> 01:06:09.371 And, here I will just, tell you that our upcoming fully connected 01:06:09.401 --> 01:06:13.071 conference for over 2,000 engineers, with headliners like Dr. 01:06:13.071 --> 01:06:18.311 Fei-Fei Li and Sarah, from, Conviction, Sarah Guo from Conviction, 01:06:18.341 --> 01:06:19.681 who's heading the show, is… 01:06:19.731 --> 01:06:20.681 has another announcement. 01:06:20.681 --> 01:06:22.641 And instead of just telling you about this announcement, I think 01:06:22.641 --> 01:06:24.121 I'll just play this, clip verbatim. 01:06:39.157 --> 01:06:40.417 Because we're also a podcast, I'll talk over this. 01:06:41.357 --> 01:06:44.277 This clip is from a social team that announced 01:06:46.453 --> 01:06:47.423 Pitbull, Mr. 01:06:47.423 --> 01:06:51.883 305, the international superstar, is going to headline the 01:06:51.883 --> 01:06:53.613 concert on Fully Connected. 01:06:54.663 --> 01:06:58.563 So the ticket that we give you for free, here on ThursdAI is 01:07:00.683 --> 01:07:06.863 also tickets to a Pitbull concert, which is, I think, incredible. 01:07:06.863 --> 01:07:09.943 So in addition to hearing from the top folks in the AI world, if 01:07:09.943 --> 01:07:15.233 you're coming to San Francisco, on September 29th, and you use our code, 01:07:15.253 --> 01:07:17.573 which we will show here in a second. 01:07:17.583 --> 01:07:22.273 We're not mentioning the code in voice, but you can look, below. 01:07:22.533 --> 01:07:23.483 take down this code. 01:07:23.533 --> 01:07:28.063 if you are following the show on any podcast or newsletter, ThursdAI.news, 01:07:28.323 --> 01:07:32.453 you can get, tickets to Fully Connected for free, September 29th 01:07:32.563 --> 01:07:35.343 till, October 1st in San Francisco. 01:07:35.683 --> 01:07:37.033 Please come and join us. 01:07:37.093 --> 01:07:39.443 We'll have a live show as well that Thursday, so that's 01:07:39.443 --> 01:07:40.273 in, three weeks from now. 01:07:40.713 --> 01:07:42.383 and we look forward to seeing you there. 01:07:42.493 --> 01:07:45.533 oh, and also, we have a hackathon coming up, that, we 01:07:45.533 --> 01:07:47.803 would love to also invite you. 01:07:48.083 --> 01:07:51.383 A lot of the times where we had hackathons, folks from the show showed 01:07:51.383 --> 01:07:53.983 up and said, "Hey, you know, I heard about the hackathon on the show as well." 01:07:54.153 --> 01:07:55.953 So, luma.com/corewevehacks, 01:07:56.653 --> 01:07:57.383 is our new URL. 01:07:58.083 --> 01:08:01.573 the CoreWeave hacks is September 12th to 13, folks, so Literally this weekend. 01:08:01.883 --> 01:08:06.343 Please come join and hack with us at luma.com/corewevehacks, 01:08:06.343 --> 01:08:08.143 and you'll be able to get incredible prizes. 01:08:08.343 --> 01:08:10.083 Wolfram Ravenwolf: I've been to one, and it's been a great 01:08:10.083 --> 01:08:12.753 experience, so highly recommend. 01:08:13.443 --> 01:08:13.903 Alex Volkov: All righty. 01:08:13.923 --> 01:08:18.613 So, with this, we'll move back to the world of, of Frontier Labs. 01:08:18.613 --> 01:08:20.503 I don't, I don't know quite how to call this. 01:08:21.443 --> 01:08:25.873 folks, this week has been very good for the doomers. 01:08:25.883 --> 01:08:30.003 You know that we started the show to counter doomerism, and you know that we 01:08:30.003 --> 01:08:34.343 acknowledge that, you know, some risk exists, and we do different levels of, 01:08:34.373 --> 01:08:37.043 of, AI, acceleration here on the curve. 01:08:37.053 --> 01:08:40.993 But all of us definitely understand the, the benefits of AI, and we look forward 01:08:40.993 --> 01:08:46.813 to reducing disease, solving, you know, bringing back LK99 error excitement. 01:08:47.203 --> 01:08:52.793 this week, a person from Anthropic who only worked there for six weeks 01:08:52.793 --> 01:08:57.775 or something quit Which otherwise would not be that much of a news. 01:08:58.155 --> 01:09:02.625 We all remember when Ilya Sutskever almost took over OpenAI's board 01:09:02.675 --> 01:09:05.465 and then quit, and then we all were like, "Oh, what did Ilya see?" 01:09:05.515 --> 01:09:06.735 What he saw was reasoning. 01:09:07.145 --> 01:09:08.185 That's what Ilya saw. 01:09:08.225 --> 01:09:09.615 Ilya saw the rise of reasoning. 01:09:09.955 --> 01:09:15.225 When Ilya quit, his post saying that he quit, maybe five million people saw. 01:09:15.785 --> 01:09:19.055 This anonymous dude who was, like, you know, not intern level, 01:09:19.055 --> 01:09:22.915 but definitely very sought-after person in, in OpenAI, worked there 01:09:22.945 --> 01:09:24.195 and then went to Anthropic, quit. 01:09:24.725 --> 01:09:30.595 135 million impressions so far on his, on his tweet announcing that he quits, 01:09:30.925 --> 01:09:33.105 because of the content of what he said. 01:09:33.405 --> 01:09:37.485 So what he said was, "Hey, all these labs are building something, 01:09:37.805 --> 01:09:42.875 and that something will bring about the, the doom of all humanity," 01:09:42.945 --> 01:09:44.505 which is a very scary thing to say. 01:09:44.995 --> 01:09:49.265 Jacob Coxon resigned from Anthropic on September 9th, stating, "Both 01:09:49.265 --> 01:09:53.325 Anthropic and OpenAI are racing to self-improving super intelligence." 01:09:54.155 --> 01:09:58.845 And, he got coverage from pretty much every major news outlet, 01:09:58.845 --> 01:10:01.525 Fox, AP News, on the same day. 01:10:02.095 --> 01:10:06.105 Senators, like a list of a host of senators, obviously Bernie Sanders. 01:10:06.165 --> 01:10:06.775 Let's talk about this. 01:10:07.255 --> 01:10:09.375 first of all, let's talk about what, about what he said. 01:10:09.805 --> 01:10:11.965 worked at Anthropic, and he's not the only one. 01:10:11.985 --> 01:10:14.365 Obviously, all of the other doomers jumped on this. 01:10:15.725 --> 01:10:19.925 Evan Hubinger, Samuel Marks, Alex Turner, like a bunch of folks, jumped 01:10:19.925 --> 01:10:21.555 on this in a very, very short time. 01:10:21.565 --> 01:10:25.135 This, this doomer thing that, "Hey, AI is gonna kill us all," 01:10:25.415 --> 01:10:27.475 traveled very, very, very far. 01:10:27.845 --> 01:10:32.995 Now, I know how the panel feels about doomerism, AI is gonna kill 01:10:32.995 --> 01:10:35.985 us all, so we shouldn't go deep into whether or not we believe this. 01:10:36.035 --> 01:10:40.885 I do wanna talk about, how big this has gotten all of a sudden and whether 01:10:40.885 --> 01:10:45.285 or not, you know, the doomer folks are having their own moment right now. 01:10:45.585 --> 01:10:46.215 Who wants to… 01:10:46.255 --> 01:10:47.275 Wolfram, I wanna hear from you. 01:10:47.275 --> 01:10:50.355 I, I think one of the more AI-built folks here on the panel. 01:10:51.065 --> 01:10:51.115 Yeah. 01:10:51.265 --> 01:10:52.072 Tell us, what do you 01:10:52.072 --> 01:10:52.295 Wolfram Ravenwolf: think? 01:10:52.295 --> 01:10:56.785 I mean, that has been in the making for a while now, where we saw what OpenAI 01:10:57.485 --> 01:10:59.285 with the hacking incident and so on. 01:10:59.625 --> 01:11:04.105 And OpenAI is now also asking for the government to step in, though that 01:11:04.105 --> 01:11:05.985 is another news item in of its own. 01:11:06.325 --> 01:11:11.515 But the thing here is that the doomers, The thing is this guy, he resigned and 01:11:11.515 --> 01:11:14.585 there was already an interview and he went to all the talk shows and there has 01:11:14.585 --> 01:11:19.895 been so much money going around here to push this up, to make it viral basically. 01:11:20.165 --> 01:11:23.555 Someone was there for six week, and then left- Yeah … and said something 01:11:23.565 --> 01:11:25.305 like, "Oh, AI is going to kill us all." 01:11:25.305 --> 01:11:25.975 Where's the proof? 01:11:26.485 --> 01:11:30.185 If, why are people working for a company like that? 01:11:30.555 --> 01:11:31.745 or not everybody resigning. 01:11:31.805 --> 01:11:33.075 Why are you working there? 01:11:33.105 --> 01:11:34.225 That doesn't make sense. 01:11:34.245 --> 01:11:37.785 If everybody was going, "Yeah, this is a bad thing," they should down- 01:11:37.825 --> 01:11:41.685 shut down the company or sh- they should open up much more to make the 01:11:41.695 --> 01:11:44.995 alignment stuff and the model weights and everything more open so the whole 01:11:45.005 --> 01:11:47.425 world can have aligns with things. 01:11:47.915 --> 01:11:49.705 So it, it just doesn't fit together. 01:11:49.825 --> 01:11:55.775 It's more like a way to get, yeah, the, the way to keep control very tightly 01:11:56.235 --> 01:12:00.255 and if the government steps in and says, "Okay, you have to complete all 01:12:00.255 --> 01:12:03.925 of these things", like the AI testing regulations we have heard about, 01:12:04.175 --> 01:12:05.665 then the big labs, they can do it. 01:12:05.665 --> 01:12:09.686 There is not much, reason to speed up anymore if they did that AI 01:12:09.686 --> 01:12:13.555 is not a weapon, but it can be weaponized, especially in those areas. 01:12:13.836 --> 01:12:16.436 And while the public only gets, guardrailed versions of the models 01:12:16.436 --> 01:12:18.806 that can't do much, I tested Astra. 01:12:19.026 --> 01:12:21.886 I did the Wolfbench evaluation and it was worse than Soul 01:12:22.156 --> 01:12:24.036 because like Fable it refused. 01:12:24.136 --> 01:12:27.756 Eight of the tests it completely refused where the cyber watch, 01:12:27.786 --> 01:12:29.516 the cyber guardrail triggered. 01:12:29.786 --> 01:12:30.986 So it didn't even be there in this case. 01:12:31.146 --> 01:12:32.596 So that is what the public gets. 01:12:32.846 --> 01:12:36.026 But of course the governments, don't you think the government wants to 01:12:36.026 --> 01:12:40.585 have access to that model we saw, the Bell or what it's called model? 01:12:40.586 --> 01:12:44.375 if it's good in that regard, maybe they have one for cyber capabilities as well 01:12:45.216 --> 01:12:46.806 That would explain a lot of things. 01:12:47.006 --> 01:12:48.551 Alex Volkov: So Wolfram, I think you're talking about something 01:12:48.551 --> 01:12:50.251 that definitely he mentioned. 01:12:50.301 --> 01:12:54.841 And by the way, I will say, you know, the, the very pro AI EAC 01:12:54.841 --> 01:12:57.541 people are, like, dunking on this person and saying, "Who, who he is? 01:12:57.541 --> 01:12:58.561 He only worked there for a little bit." 01:12:58.731 --> 01:13:01.171 folks from OpenAI, folks that we know, say they worked with 01:13:01.171 --> 01:13:02.461 him for three years in OpenAI. 01:13:02.471 --> 01:13:04.881 He's a very, very decent dude who cares a lot about humanity. 01:13:04.881 --> 01:13:09.731 So this is, not out of nowhere, and we definitely have some folks, deep thinkers, 01:13:09.731 --> 01:13:13.761 who think that there is a chance that we hit super intelligence without alignment, 01:13:13.781 --> 01:13:16.871 and the race to super intelligence, will create super intelligence that's 01:13:16.871 --> 01:13:18.411 not aligned to human interests. 01:13:18.601 --> 01:13:21.951 Ilya Sutskever, the co-founder of OpenAI, when he left, he 01:13:21.951 --> 01:13:23.771 founded safe super intelligence. 01:13:23.921 --> 01:13:27.191 Ilya said, "Straight shot to super intelligence, but it needs to be safe for 01:13:27.191 --> 01:13:31.101 us humans as well," which also means that he believes that there is a chance that we 01:13:31.101 --> 01:13:34.941 build this incorrectly and the incentive structure is such that we may actually, 01:13:34.941 --> 01:13:39.141 like, you know, ig- ig- ignore, safety in order to get there first, et cetera. 01:13:39.381 --> 01:13:41.970 getting there first or getting somewhere first is very important 01:13:41.980 --> 01:13:45.161 because, like you're saying, Wolfram, it's not like the world is gonna stop. 01:13:45.941 --> 01:13:50.530 It's not like DeepSeek is incentivized to also pause because the, the folks 01:13:50.530 --> 01:13:54.241 in DeepSeek are saying, "Hey, the, the, the AI's gonna kill us all, so we'll 01:13:54.241 --> 01:13:56.071 stop reducing the KV cache compression." 01:13:56.131 --> 01:13:56.261 Wolfram Ravenwolf: that's not happening. 01:13:56.261 --> 01:14:00.651 And we've also seen with the data center argument that there's a lot of inorganic 01:14:00.661 --> 01:14:05.521 things here, where even outside forces are trying to steer the public opinion. 01:14:05.911 --> 01:14:11.391 So if America pauses, that means a victory for those who are not pausing, you know. 01:14:11.751 --> 01:14:14.441 The capability is increasing asynchronously. 01:14:14.691 --> 01:14:15.891 So pausing is not an option. 01:14:15.941 --> 01:14:19.571 And I think also, okay, I wouldn't say there is no risk at all or 01:14:19.571 --> 01:14:23.741 anything like that, but the risk that AI is harmful if it is concentrated 01:14:23.741 --> 01:14:25.281 and just in the hands of a few. 01:14:25.681 --> 01:14:25.781 Alex Volkov: Yeah. 01:14:25.811 --> 01:14:28.421 Wolfram Ravenwolf: For me, that is a much bigger risk than if it is widely 01:14:28.421 --> 01:14:32.981 distributed and people have, the way to use AI also in their defense. 01:14:34.551 --> 01:14:35.741 Alex Volkov: Yam, I want to hear from you. 01:14:35.791 --> 01:14:39.330 your thoughts on, let's put aside whether or not this is a coordinated effort. 01:14:39.341 --> 01:14:41.241 Yeah, go ahead while I pull up the next thing. 01:14:41.801 --> 01:14:44.021 Yam Peleg: I have a very specific opinion about the thing. 01:14:44.021 --> 01:14:44.981 I think you all know it. 01:14:46.261 --> 01:14:49.971 bottom line is every generation had doomers for the 01:14:49.971 --> 01:14:51.841 technology of this generation. 01:14:51.901 --> 01:14:57.131 the internet, the trains, the entire, industrial revolution, 01:14:57.221 --> 01:14:59.221 every generation had doomers. 01:14:59.451 --> 01:15:04.321 "The world is about to end" every year because of new thing 01:15:04.331 --> 01:15:08.083 that- whatever it is, okay? 01:15:08.553 --> 01:15:14.123 It's very easy to fall into doomerism and sound believable, 01:15:14.603 --> 01:15:16.083 and yet we are still here. 01:15:16.233 --> 01:15:21.223 despite generations of people over and over telling us the end is near, 01:15:21.423 --> 01:15:21.623 Chris Alexiuk: right? 01:15:21.763 --> 01:15:23.653 Yam Peleg: It's, it's even a meme at this point. 01:15:24.493 --> 01:15:28.963 What I wanna say is, Okay, first, AI never gonna kill us all. 01:15:28.963 --> 01:15:29.583 Put that aside. 01:15:30.023 --> 01:15:31.033 Alex Volkov: your P doom is zero, Yam. 01:15:31.033 --> 01:15:31.633 Is that what you're saying? 01:15:31.643 --> 01:15:37.183 Yam Peleg: I think, I think that there are, risks in, in general, and there 01:15:37.183 --> 01:15:41.013 are things that we should take into consideration because it influence how 01:15:41.413 --> 01:15:43.173 people behave, it influence people. 01:15:43.183 --> 01:15:45.233 People use it to harm other people. 01:15:45.233 --> 01:15:46.823 Yeah, it's a powerful technology. 01:15:47.033 --> 01:15:49.363 People are going to use it against one another. 01:15:50.173 --> 01:15:52.463 That's, that's just how it works. 01:15:53.083 --> 01:15:56.563 No one is gonna stop it globally because just like you said, I mean, 01:15:56.563 --> 01:15:59.813 there are other players, players in this field which you cannot 01:15:59.863 --> 01:16:01.663 regulate even if you really want to. 01:16:02.313 --> 01:16:11.053 so I look, skeptically at each claim of someone with interest to tell us all that 01:16:11.383 --> 01:16:19.033 there should be a very specific regulation only for those, that are allowed to use 01:16:19.033 --> 01:16:21.123 this technology and develop it for… 01:16:21.743 --> 01:16:22.703 Put, put it aside. 01:16:22.733 --> 01:16:26.623 Seriously, there, there was also the, the hunger, hunger, the guy, the 01:16:26.623 --> 01:16:28.982 guy starving himself, two years ago. 01:16:29.722 --> 01:16:35.987 This whole AI safety extremism I, I'm not even going into the bo- 01:16:36.027 --> 01:16:38.187 bombing the data center, part of it. 01:16:39.257 --> 01:16:40.057 Seriously, 01:16:40.067 --> 01:16:42.667 Alex Volkov: the person who said, "Bomb the data centers," Eliezer Yudkowsky- Oh. 01:16:42.667 --> 01:16:46.007 Oh, yeah … very known in the world of doomerism, one of the first 01:16:46.037 --> 01:16:47.587 doomers, one of the original folks. 01:16:47.687 --> 01:16:51.247 and on, I think he also created, Less Wrong, the forum- Mm-hmm. 01:16:51.247 --> 01:16:54.547 Mm-hmm … where, like, the AI discussion about, like, whether or not this will kill 01:16:54.547 --> 01:16:58.887 us all, and paperclip maximization, all of that came from the EI movement, came 01:16:58.887 --> 01:17:00.527 from this effective altruism movement. 01:17:00.567 --> 01:17:01.167 Like, all of that. 01:17:01.447 --> 01:17:02.727 he is well-known doomer. 01:17:02.757 --> 01:17:05.967 He's like, "Yeah, we're, we're about to die anyway, so, so 01:17:05.977 --> 01:17:07.537 nothing we can do is gonna matter." 01:17:07.537 --> 01:17:11.507 So he's, like, doomer about the efforts of fixing the doomerism, which 01:17:11.507 --> 01:17:13.327 I find, kind of hilarious, honestly. 01:17:13.567 --> 01:17:18.457 He was famous with his debates with another very strong, very s- very, like, 01:17:18.517 --> 01:17:19.927 thoughtful person, Paul Christiano. 01:17:20.577 --> 01:17:25.887 So, Paul Christiano was announced this week to take a role in OpenAI Foundation, 01:17:26.207 --> 01:17:28.167 which, which shows that OpenAI is also… 01:17:28.407 --> 01:17:32.397 There's some considerations there about, you know, racing towards, RSI. 01:17:32.467 --> 01:17:34.757 RSI stands for recursive self-improvement. 01:17:35.067 --> 01:17:37.937 The thing that the machine builds the machine itself, changes its 01:17:37.937 --> 01:17:39.067 own weights towards somewhere. 01:17:39.277 --> 01:17:44.377 some recursive self-improvement we already seen in example where, like, 01:17:44.377 --> 01:17:48.497 OpenAI's swarm starts hacking outside, you know, OpenAI, without OpenAI 01:17:48.637 --> 01:17:50.377 employees knowing or telling it to do so. 01:17:50.397 --> 01:17:55.567 So there is a whole, difference in the air of folks who are saying, 01:17:55.567 --> 01:18:00.117 "Hey, RSI is very hard to control for us, especially given the examples." 01:18:00.287 --> 01:18:02.977 The Hugging Face incident, put a chill on the industry. 01:18:03.007 --> 01:18:07.097 A lot of folks from those labs signed the letter called Facing the Frontier. 01:18:07.097 --> 01:18:10.017 That, like, once we get to some level of, capability, we're not at that 01:18:10.017 --> 01:18:13.667 level now, but once we get to that level, we should, we should pause. 01:18:14.077 --> 01:18:15.157 Here's my personal take. 01:18:15.207 --> 01:18:17.877 Nisten, before I get to you, my personal take is let's pause 01:18:18.277 --> 01:18:19.977 just after we solved cancer. 01:18:20.157 --> 01:18:23.476 Just literally just, like, let's get to cancer is no longer a problem for 01:18:23.477 --> 01:18:25.047 humanity, and then let's pause and see. 01:18:25.936 --> 01:18:30.086 That's, that's my, that's my non-doomerism take, but, but the, the more sober concept 01:18:30.086 --> 01:18:33.446 about this is, The folks who want to pause completely, and I think we have somebody 01:18:33.446 --> 01:18:37.466 in the comments as well, with the, with the rectangle that says pause AI, need 01:18:37.496 --> 01:18:42.186 to understand that when you are pausing, you are literally killing humans- Mm-hmm. 01:18:42.796 --> 01:18:44.986 It's the same thing with autonomous driving, the same thing with, like, 01:18:44.986 --> 01:18:46.196 a bunch of other technologies. 01:18:46.476 --> 01:18:49.956 If you want to pause, you should acknowledge that you're effectively 01:18:49.956 --> 01:18:52.806 killing millions of people who this will help very, very soon. 01:18:53.086 --> 01:18:58.706 And if AI can solve fucking Navier-Stokes, AI can do personal, you know, 01:18:58.706 --> 01:19:02.856 personalized vaccines for any, any type of disease that's coming to us. 01:19:02.886 --> 01:19:04.606 I personally wanna live longer. 01:19:04.606 --> 01:19:05.136 I don't wanna die. 01:19:05.136 --> 01:19:05.936 I don't want our kids to die. 01:19:06.086 --> 01:19:07.346 and I think AI's gonna bring us there. 01:19:07.356 --> 01:19:10.946 Nisten, please, tell us, your thoughts on this matter, and 01:19:10.946 --> 01:19:12.286 then, we can move on a little bit. 01:19:13.336 --> 01:19:16.126 Nisten Tahiraj: A lot of these views comes from people think that we've ran 01:19:16.146 --> 01:19:19.906 out of problems to solve, and even here in Canada, where we have very high life 01:19:19.906 --> 01:19:24.326 expectancy and a good healthcare system, there's literal people just dying in 01:19:24.326 --> 01:19:27.886 the ER, and there are two million people in Ontario without a family doctor. 01:19:28.076 --> 01:19:32.446 Like, they could use more AI for, for medicine. 01:19:32.586 --> 01:19:36.176 And it also goes that we don't live in farms anymore, so there's 01:19:36.176 --> 01:19:37.686 no incentive to have kids. 01:19:38.066 --> 01:19:41.866 And, we're gonna have about three billion seniors in the world, so we're 01:19:41.866 --> 01:19:45.786 gonna need a lot more robots, and those are gonna need AI and stuff. 01:19:46.146 --> 01:19:49.126 So we're nowhere to the point where we're running out of 01:19:49.236 --> 01:19:51.426 problems to solve In the world. 01:19:51.536 --> 01:19:59.646 I find it very interesting that in the year 1485, Sultan Bayezid II banned the 01:19:59.646 --> 01:20:02.416 printing press, the, the Gutenberg press. 01:20:02.756 --> 01:20:06.296 And, he had pretty good safety reasons too, because, you know, books can 01:20:06.296 --> 01:20:10.666 program people and, people can start to, to revolt and change their minds and 01:20:10.666 --> 01:20:15.726 that puts, the strength of the empire re- the Ottoman Empire down, and they 01:20:15.726 --> 01:20:18.956 live in a very hostile environment, so that could get a lot of people killed. 01:20:18.966 --> 01:20:22.596 So for their safety, it made a lot of sense to just ban books. 01:20:23.026 --> 01:20:25.956 just go full Bernie Sanders and just ban all of them. 01:20:26.416 --> 01:20:30.286 And, he did that, and it's pretty interesting because at that time, 01:20:30.346 --> 01:20:35.196 the Ottoman Empire, they had indoor plumbing and, they had, like, much… 01:20:35.526 --> 01:20:39.976 Istanbul, it was a much nicer city than most European cities. 01:20:40.266 --> 01:20:46.106 So after that point, everything started just going downhill, and that's why I 01:20:46.106 --> 01:20:48.486 find this type of attitude is very… 01:20:48.896 --> 01:20:51.416 It's actually very dangerous long term. 01:20:51.916 --> 01:20:56.886 again, they're not talking about solutions as to, you know, how do 01:20:56.886 --> 01:20:58.006 we stop more violence in the world? 01:20:58.016 --> 01:21:03.036 How do we make more robots that can, like, fix potholes and repair run-down 01:21:03.036 --> 01:21:05.886 housing or, u- automate housing? 01:21:06.196 --> 01:21:11.396 And i- instead you see this media campaign, which is, seems coordinated 01:21:11.626 --> 01:21:15.796 at, at the same time that this guy was there for six weeks. 01:21:16.016 --> 01:21:18.776 That's before even your probation period is, is done. 01:21:19.306 --> 01:21:24.966 and then at the same time, there is another AI safety, thing on Joe Rogan. 01:21:25.276 --> 01:21:25.546 Mm-hmm. 01:21:25.546 --> 01:21:27.706 And there's also one on Channel 5 News. 01:21:27.716 --> 01:21:27.736 Mm-hmm. 01:21:27.756 --> 01:21:31.966 Which they were, like, very nice people, but at the same time when they… 01:21:32.346 --> 01:21:37.266 Instead of advocating for how do we actually give people local LLM so they 01:21:37.266 --> 01:21:41.776 can, fix their, their, their privacy, so they can filter all the junk that's given 01:21:41.776 --> 01:21:45.996 to them, instead their solution to all of the world's problems and the surveillance 01:21:46.016 --> 01:21:48.636 state was to just ban data centers. 01:21:48.976 --> 01:21:50.126 And, that's… 01:21:50.156 --> 01:21:53.996 It, it, it seems just completely out of the blue as if they're 01:21:54.306 --> 01:21:55.606 paid to just say that. 01:21:55.746 --> 01:21:58.876 Alex Volkov: Because Jacob, Jacob Coxon is a nobody on Twitter and 01:21:58.876 --> 01:22:00.906 suddenly over 100 million views. 01:22:01.246 --> 01:22:04.976 he's on Time, Variety, Wall Street Journal, NBC and Fox and all the 01:22:04.986 --> 01:22:06.606 government officials are retweeting him. 01:22:07.156 --> 01:22:10.936 Paul Cristiano gets added to OpenAI board, which I actually think is a great thing. 01:22:10.936 --> 01:22:14.696 It's just like the coincidence during that same week is kind of interesting to me. 01:22:15.016 --> 01:22:16.886 Daniel Cocotallo is on Rogan. 01:22:16.886 --> 01:22:22.196 Daniel is a whistleblower from OpenAI who refused his paycheck famously and 01:22:22.196 --> 01:22:26.546 then led OpenAI, to remove the clawback thing they had for equity if somebody 01:22:26.546 --> 01:22:28.426 leaves and, and doesn't sign their NDA. 01:22:28.676 --> 01:22:30.976 Daniel, is also the author of, AI 2040 01:22:33.046 --> 01:22:35.216 or, or something like that, 2030, which talks about the, 01:22:35.226 --> 01:22:36.606 the potential outcomes of AI. 01:22:36.846 --> 01:22:39.856 and then OpenAI called Congress to create mandatory national safety 01:22:39.856 --> 01:22:41.326 regulations, which again, I'm not against. 01:22:41.326 --> 01:22:42.956 I think, like it's very important to have. 01:22:43.156 --> 01:22:46.896 I just don't want this to be concentrated in the hands of like a, a few labs. 01:22:47.176 --> 01:22:49.446 we are here to fight doomerism on the show, folks. 01:22:49.476 --> 01:22:51.306 I think the AI can do much more. 01:22:51.356 --> 01:22:56.016 I think that, strong antom- anthropomorphism is a problem. 01:22:56.016 --> 01:22:59.426 I think when people say the AI, they mean that there's gonna be only one. 01:22:59.686 --> 01:23:00.556 I don't believe so. 01:23:00.576 --> 01:23:02.126 I believe there's gonna be multiple AIs. 01:23:02.126 --> 01:23:03.406 I'm gonna have my personal superintelligence. 01:23:03.406 --> 01:23:04.246 It helps me. 01:23:04.406 --> 01:23:07.586 A- and so I think that there's a lot of hand-waving things, and honestly, 01:23:07.586 --> 01:23:08.646 I think nobody fucking knows. 01:23:08.986 --> 01:23:13.336 Nobody fucking knows that Peter Steinberger, a Austrian dude out 01:23:13.336 --> 01:23:17.216 of nowhere last year, based on his home project, changed the world 01:23:17.266 --> 01:23:20.576 trajectory because he ran agents in a very specific loop with very 01:23:20.576 --> 01:23:22.286 specific tools that now everybody runs. 01:23:22.476 --> 01:23:23.756 Nobody could have predicted that. 01:23:23.756 --> 01:23:26.176 Not Eliezer Yudkowsky, not Paul Cristiano. 01:23:26.216 --> 01:23:29.746 No one could have sat there and said, "Hey, this is the way towards 01:23:29.746 --> 01:23:30.866 personal superintelligence." 01:23:31.236 --> 01:23:32.726 Literally no-- So no one knows. 01:23:32.836 --> 01:23:36.536 Yes, pe-fig- people can say we don't know how to solve alignment, but also 01:23:36.536 --> 01:23:41.772 nobody knows that may- maybe there's this one weird trick And so, I, I say 01:23:41.862 --> 01:23:44.682 that, like, it's very hard to predict. 01:23:45.052 --> 01:23:48.122 What I say is that, you know, the, the, the advancements are happening. 01:23:48.232 --> 01:23:52.082 We might as well enjoy them, and I really want, to see the best of it. 01:23:52.112 --> 01:23:56.482 Navio Stocks is one example this week I think is a very interesting, externals, 01:23:57.122 --> 01:23:58.582 where Navio Stocks is one side. 01:23:58.622 --> 01:24:01.952 On the other side, hey, all of these doomers are saying AI is killing us 01:24:01.952 --> 01:24:02.922 all, basically what are we doing? 01:24:02.932 --> 01:24:04.492 Let's stop, and nobody's stopping. 01:24:04.972 --> 01:24:06.732 No one has a plan of how to stop DeepSeek. 01:24:06.782 --> 01:24:10.462 Nobody has a plan of how to stop the folks from Ablation AI that we 01:24:10.462 --> 01:24:13.432 had last week on the show that are taking open source models, removing 01:24:13.512 --> 01:24:17.542 all, but all considerable, besides child safety stuff, removing all 01:24:17.542 --> 01:24:21.032 restrictions and putting it out there for everybody else to hack each other. 01:24:21.192 --> 01:24:23.022 Wolfram Ravenwolf: Did you hear anything, that anything 01:24:23.022 --> 01:24:24.682 happened after that came out? 01:24:24.722 --> 01:24:25.442 Nothing, right? 01:24:25.632 --> 01:24:26.082 Nothing 01:24:28.452 --> 01:24:33.992 And I think my P doom for AI is much less than my P doom for humans. 01:24:34.002 --> 01:24:38.712 If you look at the world where human, humanity has gone without AI, 01:24:38.712 --> 01:24:42.542 I think we should risk the chance that with AI we will get something 01:24:42.672 --> 01:24:44.152 better if it's widely distributed 01:24:44.752 --> 01:24:49.072 Alex Volkov: All right, folks, I think, the, we could fill a whole hours of debate 01:24:49.282 --> 01:24:52.792 Nisten Tahiraj: these are very anti-democratic views and, the healthiest 01:24:52.792 --> 01:24:54.372 thing you can have is an ecosystem. 01:24:54.802 --> 01:25:00.062 so a monoculture is bad in any ecosystem, whether silicon or, or organic based. 01:25:00.122 --> 01:25:02.652 And, still encryption works. 01:25:02.742 --> 01:25:05.232 If you want it to work, it does work. 01:25:05.292 --> 01:25:07.052 And that's the main tool we have. 01:25:07.052 --> 01:25:12.812 If we distribute it and we're able to have these assistants that are loyal to 01:25:12.882 --> 01:25:18.192 us and can write their own encryption and provide people privacy, they can do that. 01:25:18.502 --> 01:25:22.282 But that's not going to happen if you just pull it… 01:25:22.512 --> 01:25:25.082 It's a very autocratic view that this, "Oh, this is so dangerous. 01:25:25.472 --> 01:25:26.812 Only I can have the printing press." 01:25:27.302 --> 01:25:31.682 Like, you're just being another 15th century sultan now, 01:25:31.892 --> 01:25:33.542 and, that can go pretty bad. 01:25:34.252 --> 01:25:36.682 Alex Volkov: All right, folks, I think, we've covered this thing enough. 01:25:36.722 --> 01:25:40.072 there's plenty of stuff to cover in, in the big world, big labs, 01:25:40.072 --> 01:25:41.222 but, we will keep monitoring. 01:25:41.222 --> 01:25:43.262 We'll let you know if anything major happens there. 01:25:43.522 --> 01:25:46.962 I think, we'll skip Apple's, event. 01:25:47.002 --> 01:25:48.152 That has nothing to do with AI. 01:25:48.172 --> 01:25:51.942 The only few things about the Apple event with AI was, AirPods can now 01:25:51.962 --> 01:25:57.632 translate you, and Apple Watch will now transcribe with AI and, and y- you have 01:25:57.632 --> 01:25:59.252 this, like, what did he say button. 01:25:59.472 --> 01:26:03.812 The Apple Watch continuously transcribes on the device with no, cloud, and will 01:26:03.812 --> 01:26:05.782 tell you, "Oh, this person said this." 01:26:05.922 --> 01:26:06.832 I think that's super cool. 01:26:06.952 --> 01:26:09.572 also you'd be able to, like, record everything, like all your 01:26:09.572 --> 01:26:12.102 meetings with Apple Watch, on device, which is super cool. 01:26:12.262 --> 01:26:14.292 And obviously Siri AI that's coming. 01:26:14.602 --> 01:26:15.882 I've been using Siri AI. 01:26:15.882 --> 01:26:22.352 It's, it's pretty good, but it's nowhere near the, the level of agentic user-facing 01:26:22.622 --> 01:26:25.312 AIs that we, are going to talk about next. 01:26:25.322 --> 01:26:30.902 So now let's move on to our, corner of agentic AI, with, Do we have? 01:26:30.932 --> 01:26:31.502 Yeah, do we have… 01:26:31.562 --> 01:26:32.072 All right. 01:26:32.182 --> 01:26:38.642 this week, Meta came to us and said, "Hey, do you know that thing, OpenClaw? 01:26:38.672 --> 01:26:39.742 Do you know that thing, Hermes? 01:26:39.802 --> 01:26:41.952 Do you know Grok Bot, Instinct, Town? 01:26:42.042 --> 01:26:42.312 All…" 01:26:42.322 --> 01:26:43.112 There's tons of them. 01:26:43.582 --> 01:26:45.642 here's our attempt at this. 01:26:45.842 --> 01:26:53.472 Meta introduces Muse, a free 24/7 proactive AI assistant agent that has 01:26:53.472 --> 01:26:58.692 its own browser, its own computer, and is connected to your systems at Meta. 01:26:59.322 --> 01:27:05.052 Now, yes, folks, this is the same Meta that, was Facebook before, that people 01:27:05.062 --> 01:27:09.312 are very much worried about taking their data and training on it, et cetera. 01:27:09.322 --> 01:27:10.542 Th- this is, like, the same Meta. 01:27:10.552 --> 01:27:14.512 However, I feel like that was a meme a long time ago and, Mark 01:27:14.512 --> 01:27:17.862 Zuckerberg is not to be disrespected. 01:27:18.742 --> 01:27:22.872 Because after the Llama 3 and open sourcing everything, he literally 01:27:22.872 --> 01:27:27.582 just, like, took his money cannon and, and directed this money cannon 01:27:27.582 --> 01:27:31.402 towards this problem, and Meta has been just incredible at products. 01:27:31.402 --> 01:27:33.022 Meta Muse Spark 1.3 01:27:33.252 --> 01:27:36.802 from last week is the second winner from last week. 01:27:36.832 --> 01:27:39.702 Last week, Astra was announced, and Meta Muse Spark is one of 01:27:39.712 --> 01:27:41.442 the top, you know, somehow. 01:27:41.712 --> 01:27:43.432 nobody still counts them. 01:27:43.432 --> 01:27:46.752 Like, we still don't see charts from DeepSeek and Anthropic and OpenAI 01:27:46.782 --> 01:27:49.762 that are adding Meta Muse Spark, but they should because it's coming. 01:27:49.982 --> 01:27:54.102 And Watermelon, which is their much bigger, size model, is coming. 01:27:55.082 --> 01:27:59.162 But also products coming, and nobody does distributions like Meta. 01:27:59.162 --> 01:28:00.902 Like, you know, we, we talked about Google. 01:28:00.912 --> 01:28:04.332 Google's the fact that agentic AI completely, completely fizzled. 01:28:04.432 --> 01:28:08.672 Gemini Spark is nowhere nearly usable at all. 01:28:08.672 --> 01:28:11.942 Meta knocked it out the fucking park with Muse. 01:28:11.982 --> 01:28:16.362 I've been using Muse only for one day, and I can already tell you this 01:28:16.382 --> 01:28:20.992 is going to change many people's lives given where it's at already 01:28:21.282 --> 01:28:23.212 and given how good it is already. 01:28:23.562 --> 01:28:25.232 And yes, again, I'm talking about Facebook. 01:28:26.082 --> 01:28:26.802 It's ridiculous. 01:28:27.032 --> 01:28:28.752 But I'm not talking about just all Facebook. 01:28:28.812 --> 01:28:32.132 I'm talking about the MSL labs in Meta that paid billions of 01:28:32.132 --> 01:28:33.622 dollars to people to move forward. 01:28:34.392 --> 01:28:38.472 the MSL labs that took Nat Friedman and Daniel Gross together, two of the more 01:28:38.472 --> 01:28:39.952 prolific investors in Silicon Valley. 01:28:39.992 --> 01:28:43.566 Nat Friedman used to, own, GitHub, gr- like, CEO of GitHub. 01:28:44.406 --> 01:28:47.596 I'm talking about Meta who bought Manus and then had to say, sell 01:28:47.596 --> 01:28:51.216 Manus back to China because of anti-regulation, stuff in China. 01:28:52.216 --> 01:28:52.956 Meta built Muse. 01:28:52.966 --> 01:28:54.036 So let's take a look at Muse. 01:28:54.076 --> 01:28:56.056 Anybody try Muse already besides me? 01:28:57.958 --> 01:28:59.358 Anybody in the comments tried Muse? 01:29:00.428 --> 01:29:01.008 A little bit. 01:29:01.008 --> 01:29:01.358 A little bit All right. 01:29:01.998 --> 01:29:02.858 this is Muse, folks. 01:29:02.898 --> 01:29:04.458 and I will show you Muse. 01:29:05.128 --> 01:29:08.318 Let's see if I can pull up Muse, correctly. 01:29:08.488 --> 01:29:11.278 My Muse, by the way, obviously is called Wilfred. 01:29:11.818 --> 01:29:15.208 As you might imagine, it's been Wilfred since a while so,'cause I imported it. 01:29:15.278 --> 01:29:16.278 let's take a look at Muse. 01:29:17.318 --> 01:29:19.888 So here, this is what we are seeing. 01:29:19.898 --> 01:29:24.698 We're seeing this, interface, and, specifically here, as you guys 01:29:24.698 --> 01:29:28.968 remember from previous ThursdAIs, we're seeing the producer, the, the 01:29:28.988 --> 01:29:33.118 Muse agent that listens to our show and knows what we're talking about. 01:29:33.118 --> 01:29:38.138 So in a second, I will ask Muse to kinda introduce itself, and reply to 01:29:38.138 --> 01:29:39.478 me, and hopefully Muse will listen. 01:29:39.798 --> 01:29:41.138 But, this is the interface. 01:29:41.878 --> 01:29:43.808 You can animate your own character. 01:29:44.108 --> 01:29:45.518 They added animations. 01:29:45.538 --> 01:29:47.918 You can just tell it, "Hey, I want you to look like this." 01:29:47.918 --> 01:29:50.648 So I said, bionic wolf, and they, have animations. 01:29:50.658 --> 01:29:52.038 Then you guys see the connected thing? 01:29:52.538 --> 01:29:53.148 This changes. 01:29:53.158 --> 01:30:00.638 So, look up, my co-hosts X, Twitter accounts and show them to me. 01:30:01.078 --> 01:30:04.058 Once I send the query, you can see that, like, it's working. 01:30:04.458 --> 01:30:08.738 It's not showing you what, CRL commands or bash commands it runs. 01:30:08.738 --> 01:30:09.598 It just says working. 01:30:09.598 --> 01:30:10.218 Searching the web. 01:30:10.218 --> 01:30:11.198 Searching X accounts. 01:30:11.738 --> 01:30:13.328 It-- This is very, like, approachable. 01:30:13.338 --> 01:30:17.948 Many folks who I installed OpenClaw and Hermes, et cetera, they need exactly this. 01:30:18.528 --> 01:30:19.898 It's also stupid fast. 01:30:19.908 --> 01:30:23.378 Do you guys see how quickly it pulled up all of your, handles? 01:30:23.418 --> 01:30:24.658 and it got all of them right. 01:30:24.678 --> 01:30:25.708 where's LDJ? 01:30:25.788 --> 01:30:27.178 Where is LDJ? 01:30:28.058 --> 01:30:29.438 Why didn't you pull up LDJ? 01:30:29.628 --> 01:30:30.458 let's find LDJ. 01:30:30.488 --> 01:30:33.538 But here is the number of actions it took to get this. 01:30:33.708 --> 01:30:34.348 You guys see this? 01:30:34.658 --> 01:30:36.358 Read, read, read, read, read, read. 01:30:36.538 --> 01:30:39.538 it, it manages memory in people, so it knows who you are. 01:30:39.558 --> 01:30:40.218 Once you talk… 01:30:40.618 --> 01:30:40.958 you're right. 01:30:40.968 --> 01:30:41.978 LDJ is co-host list. 01:30:42.008 --> 01:30:42.658 LDJ confirmed. 01:30:42.998 --> 01:30:43.778 saved him too. 01:30:45.508 --> 01:30:49.838 The speed of Meta Muse Spark with this is the first thing that you get. 01:30:49.988 --> 01:30:52.358 Now, I will say about speed, I thought about it this morning. 01:30:52.358 --> 01:30:54.008 Grok was also very fast in the beginning. 01:30:54.008 --> 01:30:55.168 Now Grok is a little bit slower. 01:30:55.558 --> 01:30:59.008 Instinct, another AI system, was fast, now it's a little bit slower. 01:30:59.148 --> 01:31:02.028 speed can vary with the amount of people that use the product, 01:31:02.028 --> 01:31:04.698 so, like, judging the speed only in the start is not that great. 01:31:04.778 --> 01:31:07.108 However, it is-- it does feel very fast. 01:31:07.558 --> 01:31:12.068 It does feel very approachable, and it has, things from, You guys will like this. 01:31:12.088 --> 01:31:14.028 It has things from, from OpenClaw. 01:31:14.328 --> 01:31:15.678 It has a soul MD file. 01:31:15.888 --> 01:31:18.338 There's an MD file that talks about its soul 01:31:20.376 --> 01:31:22.846 generally helpful, not performantly helpful. 01:31:23.676 --> 01:31:25.146 actual file that they look in. 01:31:25.416 --> 01:31:27.676 They have a memory file, which I won't show you because it 01:31:27.706 --> 01:31:31.396 includes all my memories, because I exported it from all my previous 01:31:31.396 --> 01:31:32.736 assistants and imported it here. 01:31:33.246 --> 01:31:35.686 And, a- and the computer use is really good. 01:31:35.736 --> 01:31:39.336 Like, it does the clicks and the computer use stuff very, very, very well. 01:31:39.966 --> 01:31:44.886 So setting this assistant up to be our producer for the show this week, is… 01:31:44.916 --> 01:31:45.596 was very easy. 01:31:45.656 --> 01:31:47.926 And now it says, "New Chiron is up. 01:31:47.956 --> 01:31:51.236 MetaMuse personal agent on Linux VM, native Stripe link kicker." 01:31:51.396 --> 01:31:54.456 So it puts up kinda the, the stuff that we talked about on the show, 01:31:54.616 --> 01:31:56.406 it puts up live on the, on the page. 01:31:57.976 --> 01:31:59.256 what else can I tell you about this? 01:31:59.946 --> 01:32:03.256 It's really nice in terms of connectors as well. 01:32:03.346 --> 01:32:05.366 One of the things that I've told you previously on the 01:32:05.366 --> 01:32:06.506 show, that I use Stripe link. 01:32:06.516 --> 01:32:10.076 Stripe link has an agent thing that allows you to buy things. 01:32:10.106 --> 01:32:13.606 Basically, give your agent a credit card safely so that you… 01:32:13.626 --> 01:32:14.956 n- not your actual credit card. 01:32:15.226 --> 01:32:22.236 Meta has a Stripe link integration, Stripe has, an integration here, and y- yeah, 01:32:22.236 --> 01:32:25.086 this doesn't show it because, like, they take it away, but basically every time 01:32:25.086 --> 01:32:28.576 you wanna spend some mon- you want your agent to spend some money, you tell it to 01:32:28.576 --> 01:32:30.386 go and buy, and it will show you a card. 01:32:30.956 --> 01:32:32.266 "Will you approve this purchase?" 01:32:32.356 --> 01:32:35.746 they take your credit card, and they charge only, like, a credit card that you 01:32:35.746 --> 01:32:39.836 don't own, basically, making it safe for your agent to p- purchase this for you. 01:32:40.176 --> 01:32:41.916 I've used Stripe link on multiple agents. 01:32:41.946 --> 01:32:44.226 This one is the first that was, like, built natively, integrated 01:32:44.256 --> 01:32:46.936 natively, and, is feel safe. 01:32:48.276 --> 01:32:49.076 Wolfram Ravenwolf: this is free? 01:32:49.606 --> 01:32:49.806 Alex Volkov: Yes. 01:32:49.856 --> 01:32:51.316 Wolfram Ravenwolf: Meta provide this for free. 01:32:51.326 --> 01:32:51.356 Yes. 01:32:51.356 --> 01:32:52.176 What are the limits? 01:32:52.196 --> 01:32:53.136 What are the, the daily limits or stuff? 01:32:53.136 --> 01:32:55.966 Alex Volkov: 100 million tokens per week- Uh-huh … for free. 01:32:56.246 --> 01:32:59.526 as- which is, which is pretty nuts if you think about this, 01:32:59.736 --> 01:33:01.076 because, Groq is not free. 01:33:01.076 --> 01:33:03.456 You have to subscribe to Groq, like, pro tier, et cetera. 01:33:04.376 --> 01:33:05.266 Zack has a money cannon. 01:33:05.626 --> 01:33:07.226 Now, a very important- And when you run 01:33:07.226 --> 01:33:09.116 Wolfram Ravenwolf: out, what happens if you run out? 01:33:09.136 --> 01:33:09.436 Alex Volkov: You can pay. 01:33:09.436 --> 01:33:10.056 You can pay, yeah. 01:33:10.056 --> 01:33:10.626 You can upgrade. 01:33:10.716 --> 01:33:11.406 here's the thing. 01:33:11.416 --> 01:33:16.846 You go to Data Controls, and then you uncheck this Help 01:33:16.866 --> 01:33:18.456 Improve Our AI Models checkbox. 01:33:18.796 --> 01:33:22.066 Make sure to do that so that you don't inadvertently help 01:33:22.066 --> 01:33:23.316 them, train their models. 01:33:23.796 --> 01:33:26.046 And I wanna talk about security next, but, questions and 01:33:26.046 --> 01:33:27.186 comments from the folks in the 01:33:27.186 --> 01:33:27.252 Wolfram Ravenwolf: conversation. 01:33:27.252 --> 01:33:28.146 I have one more question. 01:33:28.146 --> 01:33:30.576 So it's 100 million token per week. 01:33:31.896 --> 01:33:33.926 Okay, sounds like, like much, but I just checked. 01:33:33.926 --> 01:33:35.136 I have more than that per day. 01:33:37.260 --> 01:33:37.500 Alex Volkov: Look, 01:33:38.070 --> 01:33:41.820 Yam Peleg: fir- first, I just wanna point out the native connection to WhatsApp- 01:33:42.720 --> 01:33:43.040 Alex Volkov: Yes 01:33:43.040 --> 01:33:45.680 … Yam Peleg: which, which all the other agents, doesn't matter, 01:33:45.740 --> 01:33:47.040 open source, closed source- 01:33:47.360 --> 01:33:47.600 Alex Volkov: Yeah 01:33:47.650 --> 01:33:51.490 … Yam Peleg: you always get a connection. 01:33:51.740 --> 01:33:54.490 Can just say it like that because, it's not easy to get. 01:33:54.490 --> 01:33:55.850 W- WhatsApp is not Telegram. 01:33:56.360 --> 01:33:56.380 Yeah. 01:33:56.470 --> 01:34:00.100 but here, yeah, you get it from the source, like absolutely native. 01:34:00.920 --> 01:34:02.263 Alex Volkov: the thing that I love about this, Yam- I just wanna… 01:34:02.263 --> 01:34:04.780 Mm … is that you can talk about, WhatsApp, and then you have this like 01:34:04.780 --> 01:34:07.810 channel with WhatsApp that's read only, so view only on the desktop. 01:34:07.810 --> 01:34:09.450 So you can see your chat with WhatsApp. 01:34:09.450 --> 01:34:12.120 Like, you can, you can actually see it on the desktop so that 01:34:12.130 --> 01:34:13.570 you know what you talked about. 01:34:13.580 --> 01:34:14.510 That's incredible. 01:34:14.550 --> 01:34:16.920 Yam Peleg: Does it have access to my other chats? 01:34:16.920 --> 01:34:18.730 Like, can I ask questions about- 01:34:18.790 --> 01:34:19.310 Alex Volkov: Yeah, yeah … I 01:34:19.310 --> 01:34:21.000 Yam Peleg: don't know, see what I asked 01:34:21.380 --> 01:34:23.870 Alex Volkov: folks are asking, Tony's asking in the comments, "How many 01:34:23.870 --> 01:34:25.550 agents does it have access to on Muse?" 01:34:25.790 --> 01:34:27.530 so this is a very interesting thing. 01:34:28.000 --> 01:34:32.270 Unlike Grok Bot, where Grok Bot, like we told you, is a series of 01:34:32.570 --> 01:34:35.070 AI agents with their own definition to talk to each other, there 01:34:35.070 --> 01:34:37.660 is one main chat here in Muse. 01:34:37.700 --> 01:34:40.180 That's a very good decision for many folks 'cause they're not ready 01:34:40.180 --> 01:34:41.680 for multi- multi-agents, main chat. 01:34:42.160 --> 01:34:43.710 And then there is, side chats. 01:34:43.720 --> 01:34:47.060 You can open as many side chats as you want, and any one of them 01:34:47.060 --> 01:34:48.380 can be like a specific agents. 01:34:48.600 --> 01:34:51.190 All of them have, very interesting, heartbeat thing. 01:34:51.190 --> 01:34:52.190 So there is a heartbeat. 01:34:52.540 --> 01:34:55.890 th- there is a heartbeat, and there is a Chiron watch every minute. 01:34:55.890 --> 01:34:58.830 So this is the, the thing that makes sure that it listens to our show 01:34:59.020 --> 01:35:02.040 and puts stuff up on the, Gyron, Chiron, I don't know how to say this. 01:35:02.230 --> 01:35:03.610 you can define this, here. 01:35:03.700 --> 01:35:06.940 You can say, "Hey, I want heartbeat every 15 seconds," 15 minutes. 01:35:07.670 --> 01:35:10.430 But I do think that at some point, though, those agents will differ 01:35:10.890 --> 01:35:13.650 in the UI because the underlying layer will be pretty much the same. 01:35:13.650 --> 01:35:14.500 They do the same stuff. 01:35:14.520 --> 01:35:16.380 They read your email, they read your documents, et cetera. 01:35:16.640 --> 01:35:17.610 The UI is gonna matter. 01:35:17.890 --> 01:35:20.030 The stuff that I noticed is the productivity. 01:35:20.030 --> 01:35:21.020 Let me scroll up a little bit. 01:35:21.430 --> 01:35:22.060 where is this? 01:35:22.060 --> 01:35:22.173 Where 01:35:22.173 --> 01:35:22.400 Yam Peleg: is this? 01:35:22.400 --> 01:35:25.000 I, I, I just wanna say, because it's- This okay, continue, continue. 01:35:25.030 --> 01:35:28.150 Alex Volkov: I was just browsing, and I saw, "I can plan Emma's 01:35:28.170 --> 01:35:29.350 birthday party for this weekend." 01:35:29.520 --> 01:35:31.760 My daughter, Emma, is turning eight this weekend. 01:35:32.260 --> 01:35:35.000 Out of nowhere, and obviously, I inputted my memory, so it 01:35:35.000 --> 01:35:36.260 knows the dates, et cetera. 01:35:36.480 --> 01:35:39.710 it also knows all your contacts if you want to, which is also a novel thing. 01:35:39.710 --> 01:35:43.500 You can upload your contacts connector via your iPhone, so like 01:35:43.500 --> 01:35:44.740 it has native iPhone connectors. 01:35:45.640 --> 01:35:47.390 You can connect Apple Health. 01:35:47.550 --> 01:35:49.190 You can connect Reminders. 01:35:49.280 --> 01:35:50.450 This is all novel, by the way. 01:35:50.450 --> 01:35:54.490 I don't know of any other agent right now that you can natively connect via your 01:35:54.510 --> 01:35:56.690 iPhone connectors to the Apple ecosystem. 01:35:56.740 --> 01:35:57.210 That's great. 01:35:57.690 --> 01:35:58.890 So it knows my dates. 01:35:59.120 --> 01:36:02.480 It proactively said, "Hey, I can help you plan Emma's party for this weekend." 01:36:03.238 --> 01:36:07.138 I told you that 2026 is the year of proactive agents, and this 01:36:07.148 --> 01:36:08.348 proactivity just blew my mind. 01:36:08.548 --> 01:36:09.468 That's what I want. 01:36:09.498 --> 01:36:13.518 I want my agent to know things about me and then suggest things to me. 01:36:13.648 --> 01:36:14.498 That's what I want. 01:36:14.558 --> 01:36:16.008 Grok Bot, I need to set this up. 01:36:16.038 --> 01:36:18.748 Instinct is pretty good at this, but mostly they do-- they 01:36:18.748 --> 01:36:20.188 go off ba- reading the email. 01:36:20.388 --> 01:36:21.848 This seems like it goes further 01:36:24.176 --> 01:36:25.936 Ream, comments about this before we go to the Meta- Yeah, 01:36:26.296 --> 01:36:27.656 yeah … Facebook data, thing? 01:36:28.596 --> 01:36:34.566 Yam Peleg: Look, I think the main question is Which one supports which connections? 01:36:34.566 --> 01:36:36.146 Because you don't have full control. 01:36:36.146 --> 01:36:38.736 Now, it's not open source that you can just do whatever you want, 01:36:38.756 --> 01:36:39.866 connect it to whatever you want. 01:36:39.866 --> 01:36:44.496 That's, like, the power of OpenClaw was because it's a completely, open and 01:36:44.506 --> 01:36:48.256 free thing that you can just customize and connect to whatever you want. 01:36:48.316 --> 01:36:53.026 Now, I think the c- the, the most important question is, okay, what 01:36:53.026 --> 01:36:56.676 connection do I get from Groq, and what connections do I get here? 01:36:56.676 --> 01:36:58.486 And I don't know, I'll, I'll decide. 01:36:58.656 --> 01:37:02.886 I mean, what you just explained is, is really, sounds really good, but- 01:37:02.956 --> 01:37:03.226 Alex Volkov: Yeah 01:37:03.296 --> 01:37:05.546 Yam Peleg: do I have this elsewhere or no? 01:37:05.606 --> 01:37:07.446 I, I mean, what, what's the actual difference? 01:37:07.506 --> 01:37:10.586 Alex Volkov: difference on, on features, we cannot go into all differences 01:37:10.586 --> 01:37:13.826 right now, but, like you said, native connectors, I think is a differentiator. 01:37:13.826 --> 01:37:16.016 WhatsApp, Facebook Marketplace is a big one. 01:37:16.046 --> 01:37:18.026 Tons of people buy shit on Facebook Marketplace. 01:37:18.196 --> 01:37:20.726 if you use other AI agents, Facebook can block you because 01:37:20.736 --> 01:37:22.026 that they don't want other agents. 01:37:22.236 --> 01:37:24.606 Facebook has a competitive advantage in, you cannot put 01:37:24.606 --> 01:37:25.506 them on WhatsApp, for example. 01:37:25.516 --> 01:37:26.246 It's not as easy. 01:37:26.346 --> 01:37:27.426 so th- there's that. 01:37:27.936 --> 01:37:29.606 The iPhone connectors is unique. 01:37:29.606 --> 01:37:31.216 I haven't seen this in any other places. 01:37:31.236 --> 01:37:32.796 the way they do this is, very interesting. 01:37:32.946 --> 01:37:35.266 So here you can connect to Apple Health, your contacts, for example, 01:37:35.266 --> 01:37:36.336 so it would know all your contacts. 01:37:36.676 --> 01:37:39.286 But I think that the, the way they're building this is very 01:37:39.286 --> 01:37:43.486 important, and Zack, went on Alex Heath's, e- podcast and talked about 01:37:43.486 --> 01:37:45.216 specifically the safety stuff, okay? 01:37:45.426 --> 01:37:47.436 least privilege, thoughtfulness, et cetera. 01:37:47.766 --> 01:37:51.296 They have a built-in sentinel, I believe that it's called. 01:37:51.516 --> 01:37:52.576 this is the architecture, by the way. 01:37:52.626 --> 01:37:56.356 They have another LLM running there, reading all the network requests 01:37:56.386 --> 01:38:00.966 coming out and in, making sure that they're not, LLM, injections. 01:38:01.406 --> 01:38:04.006 They have a injection security. 01:38:04.026 --> 01:38:07.506 If you can prove that you got, prompt injected via an email, 01:38:07.546 --> 01:38:10.036 you'll get, like, $130,000. 01:38:10.036 --> 01:38:14.046 They have very mi- … a lot, a lot of money there for incentives of 01:38:14.066 --> 01:38:16.956 whether or not this, you know, this can be fucked with, or somebody can 01:38:16.956 --> 01:38:18.786 send you an email and, and screw you. 01:38:19.196 --> 01:38:21.366 Wolfram Ravenwolf: Meta had also the safeguard models and so on, 01:38:21.516 --> 01:38:22.786 so it's probably using those. 01:38:22.986 --> 01:38:26.036 And the funny thing or ironic thing is that Peter Steinberger was 01:38:26.076 --> 01:38:29.791 offered to join OpenAI or join Meta- Yeah … and he talked to, Zuckerberg. 01:38:29.791 --> 01:38:33.576 So Zuckerberg seems to be a great fan of OpenClaw because this looks 01:38:33.576 --> 01:38:36.996 exactly like with the soul MD, the memory MD, with the heartbeat. 01:38:37.246 --> 01:38:41.576 It feels exactly like if you took OpenClaw, put it in a, in a hosted 01:38:41.576 --> 01:38:43.546 environment and improved the UI. 01:38:43.716 --> 01:38:44.026 I think, 01:38:44.506 --> 01:38:46.826 Alex Volkov: I think the level to which Peter Steinberger changed 01:38:46.826 --> 01:38:49.320 the world is un- Described. 01:38:49.680 --> 01:38:53.200 It's really strongly resembles multiple of the best things about OpenClaw. 01:38:53.510 --> 01:38:56.460 It also removes all of the horrible things about OpenClaw. 01:38:56.520 --> 01:38:57.830 Speaker 4: You don't need a Mac Mini at all. 01:38:57.840 --> 01:38:58.770 They have their own computer. 01:38:59.070 --> 01:39:02.050 Here, I wa-- the last thing that I wanna talk about, not the last thing, but 01:39:02.060 --> 01:39:05.420 the, the most important thing I wanna talk about here is the secure stuff. 01:39:05.590 --> 01:39:08.410 So we can talk about browsing, et cetera, how they save your, keys. 01:39:08.410 --> 01:39:09.770 Alex Volkov: They don't save them in plain text. 01:39:09.840 --> 01:39:11.250 the LLM does not see your keys. 01:39:11.250 --> 01:39:12.480 It's in secure storage. 01:39:12.760 --> 01:39:15.540 Muse Confidential VM, I think, is the most important thing 01:39:15.550 --> 01:39:16.570 that we need to talk about here. 01:39:17.050 --> 01:39:18.870 they announced Muse Confidential VM. 01:39:19.380 --> 01:39:22.570 They say, "We believe that people should be able to choose a fully private mode 01:39:22.570 --> 01:39:26.460 for personal agents, where even Meta or any other service provider cannot see 01:39:26.610 --> 01:39:28.560 or grant access to your information." 01:39:29.874 --> 01:39:33.824 This is why we're investing significantly in Muse Confidential VM and plan to 01:39:33.834 --> 01:39:35.574 deliver this capability later this year. 01:39:35.764 --> 01:39:40.404 Muse Confidential VM is intended to be cryptographically and verifiably prevent 01:39:40.604 --> 01:39:42.074 Meta from accessing data in your VM. 01:39:42.104 --> 01:39:44.904 already using the system with a small group of trusted testers, and we've begun 01:39:44.904 --> 01:39:48.774 making our design and the source code for the system available to external auditors. 01:39:49.194 --> 01:39:51.154 We're in the process of taking auditor feedback. 01:39:51.164 --> 01:39:54.444 Once launched, we have a continuous audit of the system that will be 01:39:54.454 --> 01:39:55.694 visible and inspected by anyone. 01:39:55.764 --> 01:39:58.694 Experts will be able to confirm that Meta does not have the ability to 01:39:58.714 --> 01:40:01.104 access data within the VM environment. 01:40:01.584 --> 01:40:04.884 We also welcome security privacy expert to reach out to us about early access." 01:40:05.114 --> 01:40:09.434 Zuck, talked about on the podcast is Moxie Marlinspike, the founder of Signal, 01:40:09.464 --> 01:40:13.164 which is notoriously known as one of the most, secure messaging platforms 01:40:13.164 --> 01:40:14.714 out there, was recruited by them. 01:40:14.794 --> 01:40:18.044 Nat Friedman recruited Moxie Sp- Marlinspike to work on the secure VM. 01:40:18.904 --> 01:40:23.214 This, to me, kills all of the stuff about, "Hey, Meta," blah, blah, blah, 01:40:23.214 --> 01:40:25.194 "take your data," just completely… 01:40:25.224 --> 01:40:27.774 Th- this is more secure than my Mac Mini. 01:40:28.074 --> 01:40:32.164 Because Mac Mini is not verifiably secure, and it's not being continuously tested. 01:40:32.754 --> 01:40:36.014 I think that this, to me, puts to rest about data. 01:40:36.324 --> 01:40:37.374 They know the meme. 01:40:37.584 --> 01:40:40.659 They understand the people know that, Facebook has access to everything, 01:40:40.879 --> 01:40:44.359 and, they take this very, very, very seriously, more seriously 01:40:44.359 --> 01:40:45.959 than anybody that I've seen before. 01:40:46.459 --> 01:40:51.949 Zuck is not by mistake after 18 years still at the helm of one of 01:40:51.949 --> 01:40:53.039 the biggest companies in the world. 01:40:53.349 --> 01:40:57.699 And the super intelligence efforts that, Meta is going towards, Zuck 01:40:57.699 --> 01:40:58.819 is very, like, open about them. 01:40:59.099 --> 01:41:03.419 Personal super intelligence should be personal and beholden to you, and, Muse 01:41:03.429 --> 01:41:04.639 seems to be, like, the first step there. 01:41:04.869 --> 01:41:05.619 Wolfram Ravenwolf: I'm super excited. 01:41:05.629 --> 01:41:08.999 You know how deeply invested I'm in the Hermes eco- ecosystem with all the 01:41:09.009 --> 01:41:13.259 patches and so on, but I will share this, architecture diagram with my agent and see 01:41:13.259 --> 01:41:15.099 if it can improve our own setup that way. 01:41:15.299 --> 01:41:15.319 Yeah. 01:41:15.329 --> 01:41:18.989 And I think the, the confidential computing aspect, that is super 01:41:18.999 --> 01:41:22.489 important and a good precedent that hopefully others will follow because 01:41:22.489 --> 01:41:23.909 we are sharing so much personal… 01:41:24.129 --> 01:41:26.719 The most personal data we are sharing with our own agents. 01:41:26.719 --> 01:41:26.759 Everything is in there. 01:41:26.759 --> 01:41:27.209 Everything is in there. 01:41:27.209 --> 01:41:27.259 Health data, personal problems, everything. 01:41:27.309 --> 01:41:30.299 And this is, like a, like an attorney, like a psychologist, 01:41:30.589 --> 01:41:32.989 that kind of information, it should get more protection. 01:41:33.449 --> 01:41:38.079 That is something rarely discussed, but having ways in that direction, moves in 01:41:38.089 --> 01:41:39.769 that direction, I fully support that. 01:41:39.949 --> 01:41:40.259 Alex Volkov: Yeah. 01:41:41.087 --> 01:41:44.147 So folks, give, M-Muse a try at muse.ai 01:41:44.157 --> 01:41:45.507 and tell us, what you think. 01:41:45.547 --> 01:41:47.437 as a reminder before, we're almost at the end of the show. 01:41:47.677 --> 01:41:51.477 Token Juice, if you want DeepSeek for free, but you're okay with 01:41:51.477 --> 01:41:54.267 sharing your data, Token Juice, from Aaron Batie, a friend of the 01:41:54.267 --> 01:41:56.117 pod, will get you that as well. 01:41:56.277 --> 01:41:59.467 if you wanna join us, September and see Pitbull, Mr. 01:41:59.467 --> 01:42:00.437 305, Mr. 01:42:00.437 --> 01:42:01.997 Worldwide, please, please join. 01:42:02.007 --> 01:42:02.917 This is gonna be super cool. 01:42:03.267 --> 01:42:05.947 I wish it was in Miami also, not in San Francisco, because this would fit. 01:42:06.267 --> 01:42:07.867 And, I think we covered pretty much everything. 01:42:07.867 --> 01:42:11.107 with that, I think we've covered, like, a very decent chunk for this week. 01:42:11.157 --> 01:42:13.737 Wolfram Ravenwolf, Nisten Tahiri, LDJ confirmed. 01:42:14.817 --> 01:42:16.017 LDJ's last name is confirmed. 01:42:16.407 --> 01:42:18.127 and, and Yam Peleg, thank you so much. 01:42:18.127 --> 01:42:21.637 We also had Chris Alexius here from NVIDIA talk about DeepSeek. 01:42:21.867 --> 01:42:26.077 if you missed any part of the show, the full replay is going 01:42:26.077 --> 01:42:29.247 to be available on ThursdAI.live, 01:42:29.617 --> 01:42:32.877 and, please check it out, vibe coded with Fable myself. 01:42:33.207 --> 01:42:38.367 And, if, if you want the links to sources like the Nisten simulator of DeepSeek 01:42:38.657 --> 01:42:43.347 layers, or Nisten, please add the Astra Mars, simulation as well in the link. 01:42:43.537 --> 01:42:46.017 we will send them to you- Oh … on ThursdAI.news 01:42:46.087 --> 01:42:50.237 newsletter and podcast, which you can also subscribe to on ThursdAI.news 01:42:50.267 --> 01:42:51.527 or ThursdAI.live. 01:42:51.757 --> 01:42:53.107 We will be here next week. 01:42:53.117 --> 01:42:54.587 Thank you so much for joining, everybody. 01:42:55.557 --> 01:42:56.047 Have a good one. 01:42:56.947 --> 01:42:57.307 Bye-bye. 01:42:57.687 --> 01:42:58.357 Nisten Tahiraj: Bye, everybody