WEBVTT 00:00:00.055 --> 00:00:01.014 Alex Volkov: Anthropic committed 00:00:01.655 --> 00:00:05.534 to 5% of its shares to a new Anthropic public benefit fund. 00:00:05.910 --> 00:00:10.190 Josh Long: if you are running a PHP server, you need to make sure that it is patched. 00:00:10.310 --> 00:00:11.829 Like, hopefully you patched yesterday 00:00:12.000 --> 00:00:17.060 Alex Volkov: OpenClaw and Hermes work the best with Anthropic models, and then Anthropic just 00:00:17.120 --> 00:00:20.480 shut them down. And now we finally see why. 00:00:20.920 --> 00:00:22.780 Hello and welcome everyone 00:00:24.160 --> 00:00:25.000 to ThursdAI 00:00:25.800 --> 00:00:26.120 for 00:00:27.320 --> 00:00:28.319 October 22nd. 00:00:31.120 --> 00:00:32.680 Cold open today, scary. 00:00:33.120 --> 00:00:36.820 The internet feels like it's failing, and a lot of it is 00:00:37.960 --> 00:00:38.840 potentially because 00:00:39.640 --> 00:00:41.639 of AI. So we're gonna talk about this. 00:00:41.800 --> 00:00:45.820 Uh, obviously we're not selling doom and gloom here, uh, but we do have, 00:00:46.506 --> 00:00:51.066 Josh Long, AKA the Joshmeister, joining us at the end of the show to talk about, to 00:00:51.306 --> 00:00:52.705 break down some of the security 00:00:53.586 --> 00:00:56.105 incidents, uh, many of them powered by AI 00:00:57.026 --> 00:00:58.006 that just happened 00:00:58.906 --> 00:01:00.786 this week, many this month. 00:01:01.186 --> 00:01:04.765 If you heard anything about the PHP core vulnerabilities, if you heard about the 00:01:04.866 --> 00:01:07.376 GitLab RCEs, uh, and the, 00:01:08.091 --> 00:01:11.205 there's one not related to AI, which is the eMudra, certificate. 00:01:11.926 --> 00:01:16.026 We're gonna break it down to you and also give you practical things, what you can do 00:01:16.126 --> 00:01:20.606 today, right now, to protect yourself, uh, with your agents as well. 00:01:21.725 --> 00:01:24.965 But also we're going to talk about Anthropic stepping in 00:01:25.566 --> 00:01:30.571 to the personal agent era with Claude Duet, which I think to me is one of the, 00:01:30.832 --> 00:01:32.611 the highest highs. I just got access this morning. 00:01:32.631 --> 00:01:34.312 I'm so excited. I haven't even tried this. 00:01:34.752 --> 00:01:37.091 Um, we're gonna talk about a lot of stuff. 00:01:37.152 --> 00:01:41.752 Anthropic becoming a PBC, public benefit company, with giving away 5% of its value, 00:01:42.392 --> 00:01:44.111 um, a lot more. 00:01:44.552 --> 00:01:46.732 Joining me directly from 00:01:48.032 --> 00:01:53.632 GTC Berlin is Wolfram Ravenwolf from our booth at GTC Berlin at Corwe. 00:01:53.672 --> 00:01:54.432 Wolfram, how are you? 00:01:55.152 --> 00:01:57.192 Wolfram Ravenwolf: Hi, Alex. I hope the audio is fine. 00:01:57.352 --> 00:01:59.432 I'm in the middle of the GTC. 00:02:00.032 --> 00:02:00.451 Alex Volkov: That is awesome. 00:02:00.472 --> 00:02:01.551 Wolfram Ravenwolf: At our booth right now. 00:02:01.632 --> 00:02:02.572 I see still 00:02:03.632 --> 00:02:05.231 stuff going on over here. 00:02:05.792 --> 00:02:09.912 Alex Volkov: You're joining, from GTC with Jensen Huang standing by you, ready to do an interview, 00:02:09.932 --> 00:02:10.131 yes? 00:02:11.572 --> 00:02:14.171 Wolfram Ravenwolf: Here he is. Oh, that's another colleague of mine, but 00:02:14.682 --> 00:02:14.941 Alex Volkov: that is 00:02:15.762 --> 00:02:15.981 Jensen. 00:02:16.041 --> 00:02:16.721 Wolfram Ravenwolf: Maybe later. 00:02:17.462 --> 00:02:17.762 Alex Volkov: All right. 00:02:17.922 --> 00:02:19.281 Wolfram Ravenwolf: Still waiting for him to come along. 00:02:19.722 --> 00:02:23.311 Alex Volkov: We w- will definitely chat about some of the stuff that CoreWeave launched, in our 00:02:23.332 --> 00:02:25.332 new segment called From the Forge. 00:02:27.092 --> 00:02:27.251 And 00:02:28.852 --> 00:02:30.031 Wolfram, what are you gonna show us? 00:02:30.492 --> 00:02:31.591 Wolfram Ravenwolf: Some of this stuff here. 00:02:31.692 --> 00:02:32.131 Alex Volkov: Oh. 00:02:32.332 --> 00:02:33.731 Wolfram Ravenwolf: This is a new one. 00:02:34.772 --> 00:02:34.892 The 00:02:35.692 --> 00:02:40.912 NVIDIA HGX Rubin NVL8, and this thing right here on the desk is 00:02:41.332 --> 00:02:43.972 1.8, 1.2 million dollars. 00:02:45.012 --> 00:02:45.192 Alex Volkov: Oh. 00:02:46.012 --> 00:02:47.051 Wolfram Ravenwolf: Just for perspective. 00:02:47.372 --> 00:02:49.292 Alex Volkov: Can you see if it attached to the table? 00:02:50.012 --> 00:02:52.492 Wolfram Ravenwolf: Uh, I can't take it, sorry. 00:02:54.212 --> 00:02:54.892 Maybe later. 00:02:55.732 --> 00:02:58.722 Alex Volkov: All right, uh, we we have a big show, Wolfram, hopefully you can join us from 00:02:59.002 --> 00:03:02.181 the hotel, but it's great, that you're showing us the floor of GTC. 00:03:02.472 --> 00:03:07.027 the highlight for me, I tune into Jensen's, keynote, famous keynote, 00:03:07.827 --> 00:03:12.877 and yeah, the highlight for me is how strongly NVIDIA is leaning into the claws 00:03:13.317 --> 00:03:16.837 and that, previously, and like obviously you can see Jensen over there with claws 00:03:17.197 --> 00:03:17.316 on 00:03:17.437 --> 00:03:17.656 Wolfram Ravenwolf: Yeah, yes. 00:03:17.677 --> 00:03:18.977 Alex Volkov: on their hands. Yes. 00:03:19.487 --> 00:03:22.367 do the employees also wear claws, all of those folks? 00:03:22.847 --> 00:03:25.327 It looks like the employees have claws, like red claws, some of them. 00:03:25.447 --> 00:03:30.006 Wolfram Ravenwolf: Yeah, yeah, there's a whole area dedicated to claws, but also Hermes agents or 00:03:30.407 --> 00:03:32.047 personal agents in general. 00:03:32.327 --> 00:03:36.506 Alex Volkov: Yeah. So, uh, personal agents, I think, is a huge thing, in GTC as well. 00:03:36.827 --> 00:03:40.597 Obviously, Jensen is kind of like, focusing on that, and he ended the show 00:03:40.737 --> 00:03:42.016 floor with the Hermes agent. 00:03:42.317 --> 00:03:44.156 We're also gonna chat about that as well. 00:03:45.077 --> 00:03:48.817 A big week. LDJ, welcome to the show as well. 00:03:50.057 --> 00:03:55.757 Um, I think, eh, we have to talk about the security stuff just after the TLDR, 00:03:56.237 --> 00:03:59.707 but, I think at times like this, you're, maybe you're happy that you're a non on 00:03:59.747 --> 00:04:00.486 the internet, no? 00:04:01.767 --> 00:04:03.086 LDJ: Good point, good point. 00:04:05.087 --> 00:04:06.387 Alex Volkov: Because stuff is... 00:04:08.087 --> 00:04:10.696 I don't want to doom and gloom, but there's a lot of vulnerabilities that 00:04:10.786 --> 00:04:12.247 we're, starting to see. again, we're gonna cover 00:04:12.907 --> 00:04:14.626 a lot of them at the end of the show. 00:04:14.747 --> 00:04:15.827 Today's show is 00:04:16.627 --> 00:04:18.066 full of guests, folks. 00:04:18.416 --> 00:04:18.797 this week 00:04:19.416 --> 00:04:20.516 also we saw a release 00:04:21.457 --> 00:04:24.547 from a company called Recursive AI, or Recursive SI, I guess. 00:04:24.687 --> 00:04:27.807 I think they rebranded. Uh, we're gonna have to keep up LDJ tabs on who did 00:04:27.907 --> 00:04:30.266 rebrand the to SI and who didn't. 00:04:30.667 --> 00:04:34.927 And, we're gonna have Tim Shi and Prince Kanuma talk about to us about Flint, which 00:04:34.967 --> 00:04:36.247 is a model that they, 00:04:37.063 --> 00:04:40.642 trained almost from scratch using a recursive process from, 00:04:41.988 --> 00:04:44.663 the Eureka machine, something they called the Eureka machine. 00:04:44.818 --> 00:04:45.982 they just announced it yesterday. 00:04:46.083 --> 00:04:47.403 So we're gonna chat about Flint as well. 00:04:47.803 --> 00:04:51.682 OpenAI has had quite a week since Dev Day. 00:04:51.763 --> 00:04:55.752 They said, hey, we're gonna either ship something or give you a reset, or for the 00:04:55.793 --> 00:04:57.092 next 30 days straight. 00:04:57.313 --> 00:04:59.652 They always kind of do that, and so they've shipped some 00:05:00.293 --> 00:05:03.012 caching updates, which are very, very, very interesting. 00:05:03.413 --> 00:05:05.932 and I think I'm really hoping everybody else will, will catch up. 00:05:06.783 --> 00:05:07.782 specifically that cache 00:05:08.423 --> 00:05:11.742 almost never, never expires, which is honestly quite crazy. 00:05:12.183 --> 00:05:15.739 Uh, and yeah, a lot of, a lot of big news, a lot of big news. 00:05:15.779 --> 00:05:18.309 I have a whole folder here that we need to get through. 00:05:18.979 --> 00:05:21.149 Aldija, what was, what was one highlight for you this week? 00:05:23.149 --> 00:05:26.308 LDJ: Uh, yeah, there's a few. There is a Reflection Beam is now an artificial 00:05:26.389 --> 00:05:28.789 analysis, and and so we have some results there. 00:05:28.909 --> 00:05:31.469 There's a new DeepSeek Flash that was announced, 00:05:32.349 --> 00:05:32.988 and uh, 00:05:33.629 --> 00:05:34.588 besides that, 00:05:35.789 --> 00:05:36.308 there is 00:05:37.829 --> 00:05:38.088 just 00:05:39.829 --> 00:05:43.288 extra benchmark news of people auditing benchmarks, which I think is really good. 00:05:43.429 --> 00:05:47.949 Some, some flaws found in Terminal Bench 4, and uh, yeah, I think a lot of cool 00:05:47.989 --> 00:05:48.549 things to cover. 00:05:49.379 --> 00:05:55.168 Alex Volkov: And we have Nisten Tahiray, who's now at DevRel at PrismML, and finally added the 00:05:55.249 --> 00:05:57.449 title, Nisten. Nisten, we are on the 00:05:58.209 --> 00:06:01.678 one highlight of the week that's the most important thing that, that you're, that 00:06:01.699 --> 00:06:04.248 you want to cover. if security is that for you, 00:06:04.969 --> 00:06:07.968 a reminder, we have a full segment about that, but please let us know. 00:06:10.045 --> 00:06:14.465 Nisten Tahiraj: You know you're locked in when you just, like, forget to check Twitter and uh, 00:06:15.565 --> 00:06:17.405 and do all of this stuff. 00:06:17.765 --> 00:06:22.284 The only reason I came back to check was because, uh, my Muse agent finally 00:06:22.885 --> 00:06:25.705 just ran out of, uh, out of quota, and, uh, 00:06:26.485 --> 00:06:27.885 I had to resort to using 00:06:28.525 --> 00:06:32.244 freaking DOD, which is just absolute garbage of a product. 00:06:33.111 --> 00:06:36.991 Uh, other than I really liked this new model, by the way. 00:06:37.151 --> 00:06:38.630 I started using it, uh, 00:06:39.231 --> 00:06:39.711 yesterday, 00:06:40.391 --> 00:06:42.411 the 42B coding model, uh, 00:06:42.551 --> 00:06:42.810 Alex Volkov: Flint? 00:06:43.431 --> 00:06:44.871 Nisten Tahiraj: Yeah, yeah, yeah. 00:06:44.911 --> 00:06:49.950 I started running a lot of agents with Flint because it's, it's, it's extremely 00:06:50.151 --> 00:06:50.511 fast. 00:06:51.151 --> 00:06:56.431 Uh, it took a bit, look, it's not that good, it is still pretty tiny, remember, 00:06:56.471 --> 00:06:57.631 but it is extremely fast, 00:06:58.391 --> 00:07:03.090 and uh, I just put Haiku on a loop to just find out what was the best way to prompt 00:07:03.151 --> 00:07:03.311 it. 00:07:03.991 --> 00:07:06.811 And then once you figured that out, it was actually 00:07:07.511 --> 00:07:12.591 pretty close. I got it pretty close to doing the job of Haiku, maybe like 96% 00:07:13.391 --> 00:07:19.150 of the score. So, uh, I actually like that, that model, and uh, I think it's 00:07:19.231 --> 00:07:22.311 a, it's a little bit underhyped, if anything. 00:07:22.791 --> 00:07:22.970 Alex Volkov: Yeah. 00:07:22.991 --> 00:07:23.391 Nisten Tahiraj: Right now. 00:07:24.231 --> 00:07:27.790 Uh, yeah, I would say that is, uh, that is the highlight. 00:07:29.431 --> 00:07:34.460 Alex Volkov: I, we're gonna talk with Prince Kanuma, our friend of the pod, MLX King, who had 00:07:34.781 --> 00:07:40.241 day one support for Flint, and the Tim Shi from the folks at, Recursive SI 00:07:40.881 --> 00:07:44.520 is, hopefully gonna join us. We're trying to get Tim, so hopefully monitoring the, 00:07:45.106 --> 00:07:49.035 the channels, to talk to us about the best way that, that they know how to prompt. 00:07:49.066 --> 00:07:52.306 the cool thing about Flint is that they trained it in like 4 months using, 00:07:52.475 --> 00:07:56.236 recursive processes, so I'm really excited for that conversation. 00:07:57.086 --> 00:07:59.165 the highlight for me came in today. 00:07:59.301 --> 00:08:03.575 so shout out to Edwin Arbus and shout out to Adi Osmani for my invite for Claude 00:08:03.776 --> 00:08:05.896 Duet, which is, this is the highlight for me. 00:08:06.176 --> 00:08:10.435 It came literally yesterday, Anthropic announced their entry into the personal 00:08:10.616 --> 00:08:10.955 agent 00:08:11.776 --> 00:08:12.096 era. 00:08:12.776 --> 00:08:17.516 Uh, personal agent, as we talk to you on the show now for almost 00:08:18.176 --> 00:08:20.776 a year, personal agents are AI 00:08:21.416 --> 00:08:23.535 agents that run in 24/7. 00:08:23.656 --> 00:08:27.436 They have their own computer, their whole- their own browser, and they are proactive. 00:08:27.496 --> 00:08:29.896 They're doing tasks for you because they have context for you. 00:08:30.211 --> 00:08:34.436 so Nisten just mentioned Muse and mentioned Dots, which is OpenAI's one. 00:08:34.706 --> 00:08:38.806 Claude is stepping into this era with Claude Duet, and I just got access, just 00:08:38.826 --> 00:08:41.185 set it up this morning, so I didn't have tons of time to play with it. 00:08:41.596 --> 00:08:43.036 but it already knew a bunch. 00:08:43.196 --> 00:08:44.995 So that, one of the biggest things that 00:08:46.136 --> 00:08:51.696 now is a problem in this era is if there's a new one, those products are very sticky. 00:08:52.135 --> 00:08:55.236 And Nisten, maybe this is why one of the reasons why you're saying, Dots is not 00:08:55.296 --> 00:08:57.416 Dots for you, because the products are very, very sticky. 00:08:58.256 --> 00:09:01.275 It's only as good as the information and the context you give it about your 00:09:01.336 --> 00:09:05.435 personal stuff. And moving that and moving those connections, moving that data feels 00:09:05.616 --> 00:09:06.846 really, difficult right now. 00:09:07.086 --> 00:09:11.696 Not to mention many folks in this era, report on AI fatigue, AI agent fatigue. 00:09:12.056 --> 00:09:13.375 So, I don't have fatigue. 00:09:13.656 --> 00:09:15.556 I'm excited to check out Claude Duet. 00:09:15.616 --> 00:09:19.506 I think, I've been telling you guys on the show for the longest time what I wanted is 00:09:19.586 --> 00:09:23.026 exactly this format with Opus 5.5 intelligence, and that's what we got. 00:09:23.386 --> 00:09:24.665 So this is the highlight for me. 00:09:24.946 --> 00:09:26.866 Uh, I think with that, let's jump to TLDR. 00:09:26.906 --> 00:09:29.505 Nissa, unless you have anything else to, you want to add, 00:09:30.226 --> 00:09:31.405 we can jump into TLDR. 00:09:31.466 --> 00:09:32.025 Or LDJ. 00:09:32.186 --> 00:09:36.425 Nisten Tahiraj: No, I, I, I have an entire rant about just how bad dot is, 00:09:37.186 --> 00:09:37.485 so. 00:09:39.021 --> 00:09:43.065 Alex Volkov: all right, we, there's a whole metric ton of, of news and models, this week. 00:09:43.386 --> 00:09:47.421 And again, because we have 3 interviews, let's just run through, the TLDR, uh, 00:09:47.446 --> 00:09:50.006 super quick. Let me just make sure that we have all of it. 00:09:50.966 --> 00:09:51.246 Okay, 00:09:52.582 --> 00:09:54.181 let's jump into the TLDR. 00:09:58.502 --> 00:09:59.461 Which is, let's go. 00:10:11.182 --> 00:10:11.801 Right, here's 00:10:12.862 --> 00:10:17.022 most of what we're gonna talk about, most of everything that happened in the week, 00:10:17.382 --> 00:10:21.142 of ThursdAI for October 22nd in about 5 minutes. 00:10:21.502 --> 00:10:25.061 Let's try to get into this because there's a lot to talk about. 00:10:25.262 --> 00:10:30.381 First of all, Anthropic decided to put 5% of its shares into a public benefit fund, 00:10:30.782 --> 00:10:34.941 and there is a lot of criticism against Dario, which Dario, in a very weird way, 00:10:35.101 --> 00:10:37.422 answered on Twitter. that conversation is 00:10:38.282 --> 00:10:42.111 very deep. Dario answers with very long threads, and very thoughtful as well, so 00:10:42.292 --> 00:10:43.851 we thought that we'd at least mention this. 00:10:44.262 --> 00:10:47.722 also, Anthropic, as we said in the beginning, opened the waitlist for Claude 00:10:47.822 --> 00:10:53.122 Duet. This is a Claude that that has a 24/7 cycle, it's on computer, and it works 00:10:53.262 --> 00:10:55.101 without being prompted, and it's proactive. 00:10:55.222 --> 00:10:56.001 This is Claude's 00:10:56.822 --> 00:10:57.502 answer to 00:10:58.622 --> 00:10:59.301 Grokbots, 00:11:00.582 --> 00:11:01.301 Meta Muse, 00:11:01.942 --> 00:11:06.581 obviously the Hermises and the OpenClaws of the world, and the most recently Dots 00:11:06.622 --> 00:11:07.461 from OpenAI. 00:11:08.102 --> 00:11:11.641 We've been telling you that we want that format with Opus 5.5, and that's exactly 00:11:11.662 --> 00:11:13.022 what Anthropic is stepping into. 00:11:13.262 --> 00:11:16.522 In a very interesting thing, it seems like Anthropic is following and not leading the 00:11:16.622 --> 00:11:18.701 category, but we'll see, we'll see how good it is. 00:11:18.782 --> 00:11:20.102 I just got access this morning. 00:11:20.581 --> 00:11:25.702 Also, OpenAI added 12-hour prompt caching built for agents that never sleep. 00:11:25.942 --> 00:11:28.541 This is like a big, big, big thing everybody's very, very excited about. 00:11:29.128 --> 00:11:33.328 and uh, GPT Astra Pro runs now parallel agents in ChatGPT. 00:11:33.698 --> 00:11:38.237 OpenAI also acquired Orchestra Bio, and there's a lot more stuff. 00:11:39.198 --> 00:11:44.318 Specifically, OpenAI models, Astra, Sol, and Luna are now in Google Cloud, in 00:11:44.478 --> 00:11:46.298 public preview. So if you are 00:11:46.918 --> 00:11:50.298 part of the Google Cloud, maybe with your startup, now you have access to the best 00:11:50.598 --> 00:11:54.838 OpenAI, models, and Google now becomes kind of the cloud that has all the the 00:11:54.938 --> 00:11:56.417 frontier models on there. 00:11:57.148 --> 00:11:59.588 Gemini 4 Argon, it's been a week since Argon. 00:11:59.708 --> 00:12:04.358 It did 25 trillion tokens in 3 days, and 4 Argon models beat Minecraft. 00:12:04.678 --> 00:12:06.358 We should mention a week with Argon. 00:12:06.478 --> 00:12:08.437 Uh, we did see multiple folks, 00:12:09.444 --> 00:12:11.883 chat about Argon on our timelines as well. 00:12:12.204 --> 00:12:17.343 And then, uh, Inception Mercury 3, a diffusion LLM claimed a thousand plus 00:12:17.444 --> 00:12:20.544 tokens with 1 million context, by by their own vendor. 00:12:21.944 --> 00:12:22.224 So, 00:12:22.984 --> 00:12:26.493 we have a new segment here, uh, Security Week, and I d- I hope we don't have to 00:12:26.614 --> 00:12:28.994 keep the segment up, but it does seem like 00:12:30.054 --> 00:12:30.853 the Defender 00:12:31.534 --> 00:12:33.133 window is kind of closing. 00:12:33.254 --> 00:12:37.234 Defender's window is what Greg Brockman called the very short time before 00:12:37.334 --> 00:12:39.494 attackers have access to very capable AI 00:12:40.574 --> 00:12:41.613 models and swarms. 00:12:41.949 --> 00:12:44.923 th- this week OpenAI's Daybreak and some other folks 00:12:45.824 --> 00:12:51.103 helped a, the PHP Foundation to find a very, like, two critical holes in the PHP 00:12:51.264 --> 00:12:51.984 engine itself. 00:12:53.179 --> 00:12:58.109 they patched the PHP engine, and PHP runs 70% of the internet, many folks scrambled 00:12:58.209 --> 00:13:00.418 really hard, to try and patch their systems. 00:13:01.009 --> 00:13:04.569 however, this is a big deal that will continue to to be talked about. 00:13:04.729 --> 00:13:06.688 Many, many websites went offline. 00:13:07.119 --> 00:13:09.119 folks like the NSA and the FBI 00:13:10.119 --> 00:13:13.518 posted about, hey, you should take your website offline if you're not patched. 00:13:13.559 --> 00:13:17.969 It's that bad. and also some trusted root certificates for HTTPS turned out to have 00:13:18.229 --> 00:13:21.979 a private key. It was literally the number 4, which everybody's making a big deal of, 00:13:21.999 --> 00:13:24.768 but that's not the big deal. That's just the level of incompetence there. 00:13:24.929 --> 00:13:27.628 Uh, eModra certificates have been compromised. 00:13:27.929 --> 00:13:32.669 this is a very scary week in security in the world, in addition to multiple 00:13:32.769 --> 00:13:35.809 companies like Railway being hacked and GitLab finding vulnerabilities. 00:13:36.129 --> 00:13:38.049 I'm talking about this really fast, and it's very scary. 00:13:38.359 --> 00:13:42.779 we will talk about how to try and and maintain sanity in this world, very soon 00:13:42.824 --> 00:13:44.218 with with Josh Long. 00:13:44.319 --> 00:13:45.678 He's gonna join us at the end of the show. 00:13:45.839 --> 00:13:50.019 And OpenAI's Daybreak program is quietly behind a lot of this year's biggest bug 00:13:50.158 --> 00:13:54.428 finds, uh, and I think that this is very important, but again, the defender's 00:13:54.489 --> 00:13:59.449 window is closing, uh, so a lot of companies were hacked, including OpenCode 00:13:59.729 --> 00:14:00.449 and Railway. 00:14:01.569 --> 00:14:03.988 Also, Microsoft said that China's state- aligned 00:14:04.609 --> 00:14:09.499 Storm used agentic AI, pen test framework RText against hundreds of organizations. 00:14:10.614 --> 00:14:14.088 the good news there is both NVIDIA and CoreWeave building fences around the 00:14:14.169 --> 00:14:18.064 agents. OpenShell and Zero Trust AI framework from CoreWeave, both announced, 00:14:18.064 --> 00:14:20.929 this week. OpenShell obviously been talked about from NVIDIA for a while. 00:14:21.649 --> 00:14:27.309 In open source and LLM news, DeepSeek shipped a new MIT licensed Flash with a 00:14:27.409 --> 00:14:28.569 big jump on Terminal Bench. 00:14:28.689 --> 00:14:33.968 We are always very, very happy to see the whale, resurface, and holy crap did the 00:14:34.049 --> 00:14:38.908 whale resurface. DeepSeek also open source Deep Infer, the engine, the inference 00:14:38.969 --> 00:14:40.689 engine that DeepSeek runs on. 00:14:41.149 --> 00:14:43.448 there's also a bunch of speculation about a new 00:14:44.789 --> 00:14:49.638 unnamed model called Space Bunny, and Minimax unveiled that Space Bunny was M3.1 00:14:50.079 --> 00:14:50.438 Flash. 00:14:51.119 --> 00:14:54.779 And as Nisten mentioned, Recursive Flint came out of nowhere. 00:14:54.879 --> 00:14:59.659 Richard Soter's company, Recursive, Richard Socher's company, Recursive, came 00:14:59.699 --> 00:15:04.718 out with Flint, a local coding model, and it's small and local, and we have both 00:15:04.839 --> 00:15:10.479 Prince Kanuma and the Tim Shi hopefully joining us very, very soon at 11, at 10 AM 00:15:10.639 --> 00:15:12.239 Pacific to talk about, 00:15:13.359 --> 00:15:13.798 uh, 00:15:15.399 --> 00:15:16.199 to talk about that. 00:15:17.195 --> 00:15:21.315 very excited about that. In tools and agentic engineering, or the personal agent 00:15:21.635 --> 00:15:26.934 corner, if you want to call it, Jensen Huang, CEO NVIDIA, closed his GTC keynote 00:15:27.235 --> 00:15:29.865 with a Hermes agent building and training a robot. 00:15:30.625 --> 00:15:31.734 Uh, Hermes, is 00:15:32.415 --> 00:15:34.455 Jensen is getting Hermes filled, and it's great to see. 00:15:34.775 --> 00:15:35.874 Uh, Nvidia. 00:15:37.055 --> 00:15:37.434 Uh, 00:15:39.255 --> 00:15:40.374 let's see what else is in there. 00:15:40.695 --> 00:15:45.415 Uh, From the Forge, which is a new corner, folks, no longer This Week's Buzz after 3, 00:15:45.575 --> 00:15:48.104 3 something years. This Week's Buzz is not the corner anymore. 00:15:48.165 --> 00:15:52.325 We, we're, we now have a new corner called From the Forge, powered by CoreWeave. 00:15:52.625 --> 00:15:57.304 Reflections Beam, which we mentioned last week, is now, on Weights and Biases, and 00:15:57.385 --> 00:16:01.054 it was trained on Weights and Biases, with Weights and Biases, curves behind it. 00:16:01.425 --> 00:16:02.605 we also have RL's 00:16:03.265 --> 00:16:04.305 rollouts in Forge, 00:16:05.251 --> 00:16:07.970 and NVIDIA Rubin's new NVL8 00:16:08.571 --> 00:16:11.010 lands on CoreWeave this quarter, so very exciting things. 00:16:11.321 --> 00:16:15.491 NVIDIA also announced a bunch of stuff at GTX, including the NVIDIA DSX. 00:16:16.251 --> 00:16:20.491 It's a pre-tested AI factory, deployable in as little as 14 days, so we should 00:16:20.611 --> 00:16:21.971 mention that a little bit as well. 00:16:22.531 --> 00:16:25.890 In vision and video, there's a lot of stuff happening, just an incredible amount 00:16:25.911 --> 00:16:27.400 of stuff happening. we, we use some of them. 00:16:27.641 --> 00:16:28.161 Kling 4 00:16:28.801 --> 00:16:32.060 does 30 seconds of video in a single pass, and Kling 4 is really, really good with a 00:16:32.121 --> 00:16:34.161 bunch of input variables. 00:16:34.721 --> 00:16:35.240 Flux 3 00:16:36.280 --> 00:16:39.720 Max now renders 10 seconds of video in 12 seconds. 00:16:39.881 --> 00:16:40.300 this is 00:16:41.161 --> 00:16:41.241 a 00:16:42.281 --> 00:16:46.350 another feat by our friends at Fal, releasing a Max version, optimized Max 00:16:46.391 --> 00:16:47.410 version of a video. 00:16:47.871 --> 00:16:51.751 And, Odyssey, I don't know, Nisten, LDJ, if you guys saw, but Odyssey ML released a 00:16:51.911 --> 00:16:56.060 world model, and then they put a fruit fly's brain inside that world model. 00:16:56.341 --> 00:17:00.800 So now we literally have a world model that's a simulation and a simulated entity 00:17:00.901 --> 00:17:04.081 there, like the fruit fly brain that everybody was getting excited about a few 00:17:04.101 --> 00:17:06.141 weeks ago. Uh, they put that in the world model. 00:17:06.176 --> 00:17:08.371 Uh, we should at least play that video because that's super cool. 00:17:08.691 --> 00:17:13.541 Our friends from Crea released a, the top non-OpenAI image editor, in their, in 00:17:13.681 --> 00:17:16.171 Arena. voice and audio is also blowing up. 00:17:16.851 --> 00:17:20.291 I really, really hope we have some time to play some of these clips because Eleven 00:17:20.331 --> 00:17:25.471 Lab launched Eleven Real Time, taking about half a second, and also Eleven Real 00:17:25.491 --> 00:17:29.511 Time Translate, where if I talk, you can hear me in different language in real 00:17:29.571 --> 00:17:30.730 time. It's really cool. 00:17:31.351 --> 00:17:35.121 Mistral also shipped Voxroll 3, which Voxroll is their kind of, transcription 00:17:35.261 --> 00:17:37.991 service, but this is now a whole voice agent stack. 00:17:38.550 --> 00:17:41.241 And, our friend, from Underdog, 00:17:42.371 --> 00:17:44.651 Sigil Wen released Bark 3. 00:17:44.931 --> 00:17:50.090 It's a 163 megabyte speech model that ties to NVIDIA's ParKit and runs completely on 00:17:50.171 --> 00:17:52.610 device, fully transcribing on device. 00:17:53.081 --> 00:17:57.161 And Josh Long from Intego closes the show with this week's security mess. 00:17:57.521 --> 00:17:59.280 And right, let's go, let's get into it. 00:17:59.361 --> 00:18:01.241 This is the TLDR. This is everything that... 00:18:02.401 --> 00:18:04.920 the we have to talk about, which is a lot. 00:18:05.241 --> 00:18:05.721 Uh, 00:18:06.961 --> 00:18:10.270 we should start, I think, with Anthropic, but we'll definitely, we'll definitely 00:18:10.331 --> 00:18:10.891 move forward. 00:18:11.611 --> 00:18:17.051 Uh, anything else, Nisten, LDJ, anything huge from this week that we must mention? 00:18:17.551 --> 00:18:20.611 Anybody in the comments, anything else that we should at least mention that 00:18:20.670 --> 00:18:22.790 happened this week before we move forward? 00:18:24.511 --> 00:18:26.230 LDJ: Yeah, maybe uh, Mercury decide. 00:18:26.590 --> 00:18:31.351 Also, I know you, you mentioned Mercury's, uh, Mercury 3 model, but they also have 00:18:31.431 --> 00:18:35.870 released a competitor to OpenAI's Decision API and of the JEV model, so 00:18:35.951 --> 00:18:36.331 Alex Volkov: Yeah. 00:18:36.351 --> 00:18:37.910 LDJ: Maybe that's something worth covering. 00:18:38.351 --> 00:18:38.790 Nisten Tahiraj: By the way, 00:18:39.391 --> 00:18:43.710 uh, if you're listening on the show, go on VLLM's Hugging Face. 00:18:44.391 --> 00:18:49.170 You have a free endpoint for D3 Turbo 27-bit decider. 00:18:49.431 --> 00:18:50.731 It scores better than JEV. 00:18:51.031 --> 00:18:52.230 It accepts videos, 00:18:53.071 --> 00:18:54.570 and you can run up to like- 00:18:54.611 --> 00:18:55.770 Alex Volkov: accepts videos, Nisten? 00:18:56.351 --> 00:19:01.830 Nisten Tahiraj: Yeah, and yeah, and you can run up to 500 million tokens, uh, decisions, or 00:19:01.911 --> 00:19:04.570 500,000 decisions a day for free. 00:19:04.831 --> 00:19:05.870 You don't even need an account. 00:19:06.231 --> 00:19:09.230 Just go on their Hugging Face, search D3 Turbo VLLM, 00:19:09.831 --> 00:19:11.150 grab a free key, and just use it. 00:19:11.751 --> 00:19:13.171 Alex Volkov: This is D3 from Liquid, right? 00:19:13.871 --> 00:19:17.011 Nisten Tahiraj: No, from VLLM trained their own decider models. 00:19:17.071 --> 00:19:18.271 Alex Volkov: Oh, they they also have a D, 00:19:18.491 --> 00:19:18.651 Nisten Tahiraj: Yeah. 00:19:19.011 --> 00:19:19.550 Alex Volkov: Wow, okay. 00:19:19.930 --> 00:19:22.951 Nisten Tahiraj: And uh, they're the the best ones right now. 00:19:23.171 --> 00:19:23.631 Uh, I would 00:19:23.691 --> 00:19:24.290 Alex Volkov: Are you serious? 00:19:24.311 --> 00:19:27.350 Nisten Tahiraj: say I I tested them, they're better than the Perplexity decider, 00:19:28.051 --> 00:19:32.510 and uh, they have a free API, which is so generous. 00:19:32.891 --> 00:19:34.670 I'll just send, I'll just send the link. 00:19:34.691 --> 00:19:37.160 Alex Volkov: Wow, okay, let's at least show this. I think this is like a cool thing to show, 00:19:37.221 --> 00:19:37.520 Nisten. 00:19:37.541 --> 00:19:37.920 Nisten Tahiraj: Yeah. 00:19:37.941 --> 00:19:39.600 Alex Volkov: So this one, D3, decision 3. 00:19:39.621 --> 00:19:43.501 Nisten Tahiraj: Yeah, yeah. So you're going to see, uh, it just says try it for free, 00:19:44.381 --> 00:19:46.641 and you can just get an account and a key right now for free. 00:19:46.661 --> 00:19:48.240 Alex Volkov: Very cool, very cool. 00:19:48.261 --> 00:19:49.700 Nisten Tahiraj: I, uh, I have like 5. 00:19:50.741 --> 00:19:51.041 Uh, 00:19:51.181 --> 00:19:51.840 Alex Volkov: You have like 5? 00:19:51.901 --> 00:19:54.240 Nisten Tahiraj: Just, yeah, yeah, just get your API key for free right now. 00:19:54.341 --> 00:19:55.760 Alex Volkov: Are you serious they're judging video? 00:19:56.701 --> 00:19:58.080 Nisten Tahiraj: Uh, yeah, yeah. 00:19:58.101 --> 00:19:58.280 Alex Volkov: Wow. 00:19:58.301 --> 00:19:59.840 Nisten Tahiraj: I- I- I- it's extremely good. 00:19:59.901 --> 00:20:02.541 I've run it through all the benchmarks I've had. 00:20:02.981 --> 00:20:05.300 I have just maxed out like, uh, 5 00:20:06.060 --> 00:20:07.241 of the API limits 00:20:07.901 --> 00:20:10.101 on it. It's, it's the best one. 00:20:10.261 --> 00:20:12.460 V3 Turbo, it's the best one. 00:20:12.581 --> 00:20:14.020 V3 Flash, also 00:20:14.701 --> 00:20:15.300 insane. 00:20:16.061 --> 00:20:16.361 Uh. 00:20:16.641 --> 00:20:17.440 Alex Volkov: That's very cool, Nisten. 00:20:17.581 --> 00:20:19.460 Nisten Tahiraj: Yeah, for, for running locally, yeah. 00:20:20.621 --> 00:20:23.491 Alex Volkov: I, dude, I have all these things I prepared for the show. 00:20:23.511 --> 00:20:27.271 Look, this is a book of news, and this could be one of the coolest things, and it 00:20:27.311 --> 00:20:28.670 just like came. is, this is great. 00:20:28.711 --> 00:20:29.550 So thank you, Nisten. 00:20:30.187 --> 00:20:33.906 Let us move to Frontier because Frontier is very, very interesting. 00:20:44.986 --> 00:20:47.146 Hi, folks. On Saturday, Anthropic committed 00:20:47.787 --> 00:20:51.666 to 5% of its shares to a new Anthropic public benefit fund. 00:20:52.147 --> 00:20:56.747 The board can add up to 2% more every 50% rise in the company's value, 00:20:57.702 --> 00:21:02.621 and the founders who pledged 80% of their equity plus employee matching program, and 00:21:02.702 --> 00:21:06.901 Anthropic says roughly 20% of the company is now earmarked for public good. 00:21:07.782 --> 00:21:09.541 This sounds like a huge, huge thing. 00:21:10.262 --> 00:21:13.572 it's, this is an LLC, it's not a charity, it's a public benefit, charity. 00:21:13.652 --> 00:21:18.031 It can switch between ways of giving, and the equity can't be used for lobbying or 00:21:18.252 --> 00:21:22.172 politics. It still needs stockholder approval, and they also published their 00:21:22.212 --> 00:21:26.611 first annual public benefit report, and the trustees wrote it largely describes 00:21:27.092 --> 00:21:29.171 inputs, so next year they want outcomes. 00:21:30.012 --> 00:21:33.291 So largely describes the money going in and how much they're gonna put into this, 00:21:33.332 --> 00:21:35.972 but no outcomes and no, uh, news. 00:21:36.092 --> 00:21:37.372 So, but I respect that line. 00:21:38.157 --> 00:21:43.081 meanwhile, Dario is fighting the critics, which is, you know, many, many folks, 00:21:43.342 --> 00:21:45.181 given his positions and stances. 00:21:46.132 --> 00:21:50.351 Nathan Calvin posted a long critique of Anthropic, very public debate. 00:21:50.492 --> 00:21:52.331 I don't know if you guys have have seen this at all. 00:21:52.632 --> 00:21:56.192 if you don't know, Nathan Calvin is the general counsel of ENCODE, a small AI 00:21:56.312 --> 00:21:59.511 policy nonprofit that pushed for California's SB 53 00:22:00.232 --> 00:22:04.131 bill, and last year he went viral when he said OpenAI subpoenaed him and his group 00:22:04.212 --> 00:22:05.511 in the middle of that fight. 00:22:05.872 --> 00:22:09.191 So this isn't just an Anthropic hater, it's a safety policy guy. 00:22:09.512 --> 00:22:13.872 His arguments were Anthropic's record now looks like a race to super intelligence. 00:22:14.552 --> 00:22:18.432 More than half a trillion dollars in compute deals, Claude leading a quarter of 00:22:18.512 --> 00:22:22.191 Anthropic internal AI development, which we talked about on last week's show, 00:22:23.032 --> 00:22:27.751 which is basically the path towards RSI, and a draft federal bill that would 00:22:27.851 --> 00:22:29.011 preempt state laws. 00:22:29.572 --> 00:22:29.771 His 00:22:30.692 --> 00:22:35.591 basic position and argument is Anthropic, instead of being like a safety shop, is 00:22:35.932 --> 00:22:39.742 becoming a let's race, accelerate everything, lab. 00:22:40.662 --> 00:22:41.162 And this 00:22:41.942 --> 00:22:46.061 triggered Dario, who answered him directly on Twitter by via quote tweet, a very long 00:22:46.141 --> 00:22:47.562 quote tweet, and saying, 00:22:48.628 --> 00:22:52.187 9 commercially costly safety calls were done on Anthropic's side. 00:22:53.593 --> 00:22:54.593 specifically he mentioned 00:22:55.393 --> 00:23:00.273 the we there's things we do that hurt us financially, and we still do them because 00:23:00.393 --> 00:23:02.443 we we do want that safety, built in. 00:23:02.883 --> 00:23:07.143 And the the main call that stuck with me is that Anthropic had Mythos class models 00:23:07.443 --> 00:23:13.003 back in February, and they held them from general availability until June, 00:23:13.603 --> 00:23:16.203 when they claimed the cyber safeguards were ready. 00:23:16.548 --> 00:23:19.512 Uh, I don't know if the world was ready for Mythos in February. 00:23:20.233 --> 00:23:24.373 Obviously, they announced Mythos in around, was that April back when we were 00:23:24.513 --> 00:23:26.872 AI Engineer London, and then, uh, 00:23:27.553 --> 00:23:32.432 Dario also admitted 3 of Calvin's 9 points were communication mistakes, including the 00:23:32.553 --> 00:23:36.773 letter to the Congress they walked back, and he he admittedly said, uh, there was a 00:23:36.853 --> 00:23:40.282 mistake. and Dario said he will have more to say soon. 00:23:41.723 --> 00:23:45.263 So, folks, on the panel, Nisten, I'm very interested in hearing your thoughts on 00:23:45.283 --> 00:23:47.123 this. Is this pre-IPO PR, 00:23:47.753 --> 00:23:49.092 or is this real accountability? 00:23:49.193 --> 00:23:51.512 Is somebody holding Anthropic to real accountability? 00:23:51.673 --> 00:23:53.832 And what about this public benefit thing? 00:23:53.993 --> 00:23:55.592 You talked about communism. 00:23:56.393 --> 00:24:00.352 Uh, is this, uh, are they reinventing, uh, uh, 00:24:01.153 --> 00:24:02.072 what is it called, uh, 00:24:02.873 --> 00:24:03.252 uh, 00:24:04.153 --> 00:24:05.573 yeah, are they reinventing... 00:24:07.193 --> 00:24:09.632 The word escapes me, something from first principles. 00:24:11.073 --> 00:24:11.592 Nisten Tahiraj: Uh, 00:24:12.713 --> 00:24:16.113 yeah, I don't know, political opinion aside, I think just like any other 00:24:16.193 --> 00:24:20.152 company, there's a lot of conflict there on the, on the inside. 00:24:21.033 --> 00:24:24.192 So I kind of, I more like Dario than dislike him. 00:24:24.273 --> 00:24:26.913 I think he's just very heavily influenced by, 00:24:27.793 --> 00:24:28.773 uh, the whole 00:24:29.513 --> 00:24:33.993 cultish thing that that was in there before and is trying to to play both 00:24:34.073 --> 00:24:36.513 sides. And, uh, at the end of the day, 00:24:37.193 --> 00:24:40.413 Opus Low is still the best model you can run, so... 00:24:41.833 --> 00:24:46.352 I think they do have a point that you need the job done and just fire up a whole 00:24:46.433 --> 00:24:46.773 bunch of 00:24:47.873 --> 00:24:49.232 Opus Low at the end. 00:24:50.153 --> 00:24:51.913 So, are they gonna get surpassed? 00:24:52.313 --> 00:24:55.913 Maybe. They have plenty of tricks up their, up their sleeves, so I, I don't 00:24:55.933 --> 00:24:58.572 know, kind of don't really have that much of an 00:24:59.353 --> 00:25:00.553 opinion on this, uh. 00:25:00.632 --> 00:25:01.752 Alex Volkov: What about the PBC thing? 00:25:01.992 --> 00:25:04.112 specifically about giving back, 00:25:04.833 --> 00:25:08.553 given that the IPO is coming up and a lot of people will invest a lot of money in 00:25:08.593 --> 00:25:12.512 this IPO, based on the streets, based on the stuff that I hear, like regular folks 00:25:12.552 --> 00:25:14.242 are very excited about that, coming. 00:25:14.262 --> 00:25:17.903 Nisten Tahiraj: That one, even what they were saying, it just went from one ear into the other. 00:25:17.983 --> 00:25:19.102 That just sounded like. 00:25:19.663 --> 00:25:19.882 Alex Volkov: Yeah. 00:25:20.103 --> 00:25:21.323 Nisten Tahiraj: Yeah. Just. 00:25:21.383 --> 00:25:23.023 Alex Volkov: Alexi, thoughts on the, on the PBC thing? 00:25:23.063 --> 00:25:27.103 Obviously, Elon Musk and, and, and OpenAI have a big thing about, you know, OpenAI 00:25:27.183 --> 00:25:32.232 being a charity, being, you know, building AI in, in the open for that, and now, you 00:25:32.273 --> 00:25:34.232 know, Anthropic is stepping in there in a different way. 00:25:34.313 --> 00:25:36.452 So Anthropic first started as a pro, 00:25:37.273 --> 00:25:39.673 you know, capital company, and now it's turning into a PBC. 00:25:40.033 --> 00:25:40.252 LDJ: Mm-hmm. 00:25:40.433 --> 00:25:40.692 Alex Volkov: Um. 00:25:40.793 --> 00:25:44.832 LDJ: Well, yeah, to clarify, Anthropic has actually been a PBC since its founding. 00:25:45.673 --> 00:25:47.332 back before ChatGPT was created. 00:25:47.393 --> 00:25:53.112 It's just now they are announcing this specific fund, uh, that's like dedicated 00:25:53.233 --> 00:25:57.313 to the, basically the nonprofit portion of their business that's like exclusively 00:25:57.393 --> 00:25:59.832 won't be used for, uh, commercial actions. 00:26:00.433 --> 00:26:03.073 Um, but yeah, they they've been established as a PBC for a while. 00:26:05.193 --> 00:26:07.532 Alex Volkov: All right. Uh, so this is now the commitment, right? 00:26:07.593 --> 00:26:10.513 So in terms of funds and and uh and and and things. 00:26:10.833 --> 00:26:11.113 Um, 00:26:12.313 --> 00:26:16.713 oh, okay, folks, the other thing we have to talk about from Anthropic, besides the 00:26:16.773 --> 00:26:19.213 IPO stuff, is Claude Duet. 00:26:19.853 --> 00:26:22.172 Here's how it looks. This hopefully this won't 00:26:22.773 --> 00:26:24.592 show you guys my personal details. 00:26:24.693 --> 00:26:27.152 Uh, Uh, personal agents are really hard to, to demo. 00:26:28.743 --> 00:26:29.942 welcome to Claude Duet. 00:26:30.823 --> 00:26:33.952 This is obviously, you're live on the show. 00:26:34.272 --> 00:26:39.092 Please don't expose any of my private details. 00:26:39.433 --> 00:26:43.232 Tell folks how you're different from just, 00:26:44.713 --> 00:26:45.752 from just Claude. 00:26:48.153 --> 00:26:50.033 Claude announces an always on, 00:26:50.713 --> 00:26:53.862 and open the wait list, which now you're seeing on your screen, always on 00:26:54.903 --> 00:26:57.303 personal agent assistant powered by Opus 5.5. 00:26:57.463 --> 00:27:01.692 It- it's a Claude that does not wait for you, and works proactively. 00:27:02.073 --> 00:27:06.863 you don't have to prompt it. It has 24/7, computer, brings over your memory from 00:27:06.963 --> 00:27:11.362 Claude, your skills and connected apps, and it reaches, into more than 6,000 00:27:11.443 --> 00:27:13.243 connectors and plugins they have on their system. 00:27:13.643 --> 00:27:16.382 every few hours checks your accounts, checks your emails, surfaces what needs 00:27:16.403 --> 00:27:20.203 your attention, and apparently it's really good at at that surfacing specifically. 00:27:20.483 --> 00:27:23.703 It has its own cloud computer, and it keeps working even when your laptop is 00:27:23.843 --> 00:27:24.202 closed. 00:27:26.708 --> 00:27:29.408 So, obviously, they thought about security a lot. 00:27:29.508 --> 00:27:31.797 Anthropic does that, and they have launched 00:27:32.558 --> 00:27:34.838 multiple things that wait for your approval. 00:27:35.198 --> 00:27:38.637 here we have a first glimpse of the web interface of Claude. 00:27:38.998 --> 00:27:41.137 As you can see, this is not the regular Claude. 00:27:41.318 --> 00:27:44.098 This is, now you can get there to claude.ai/messages. 00:27:44.518 --> 00:27:46.397 here's a response from from my duet. 00:27:46.478 --> 00:27:47.317 You can't rename it, 00:27:47.918 --> 00:27:50.278 unlike, you know, Muses and other places. 00:27:50.358 --> 00:27:54.618 I I I haven't seen a place where I can just rename duets, so I think they aren't 00:27:55.118 --> 00:27:57.877 anthro- anthropomorphizing this. 00:27:58.168 --> 00:28:01.408 Claude says, or I'm Claude, the same model you can use at Claude AI. 00:28:01.723 --> 00:28:03.267 here I work as Alex's assistant. 00:28:03.348 --> 00:28:07.117 I keep track of tasks, read the apps he connects and work in threads. 00:28:07.278 --> 00:28:09.198 I will not show you his private details on the show. 00:28:09.358 --> 00:28:10.717 Hopefully that's thank you. 00:28:11.278 --> 00:28:11.618 Uh, 00:28:12.238 --> 00:28:15.037 do you work proactively? 00:28:15.438 --> 00:28:17.057 So this is the same thing 00:28:17.878 --> 00:28:19.817 that, uh, let me actually go like this, yeah. 00:28:20.038 --> 00:28:23.707 Uh, this is Cloud Duet. This is the same thing that, we've talked to you about 00:28:24.348 --> 00:28:24.707 where, 00:28:25.797 --> 00:28:29.527 OpenAI released Dots, and, you know, obviously Meta Muse is changing the 00:28:29.608 --> 00:28:33.847 category with, uh, with with regular folks and just announced Meta Muse Wonder last 00:28:33.967 --> 00:28:34.207 week. 00:28:35.488 --> 00:28:36.527 This is a huge thing. 00:28:36.798 --> 00:28:37.678 as far as I 00:28:39.278 --> 00:28:44.298 want to kind of highlight, we b- we on the show have been, or at least some of us on 00:28:44.318 --> 00:28:47.518 the show have been assistant pilled since early, early 00:28:48.358 --> 00:28:49.597 Clawdbot days. 00:28:50.078 --> 00:28:54.518 You guys remember rolling back to Clawdbot, which was the, the, the, the 00:28:54.678 --> 00:28:57.748 pre, before renamed name for OpenClaw. 00:28:58.348 --> 00:29:01.308 There is a reason why Anthropic reached out to Peter Steinberger and asked him to 00:29:01.388 --> 00:29:02.668 change the name for Clawdbot. 00:29:03.308 --> 00:29:09.288 Also, a reminder that OpenClaw and Hermes work the best with Anthropic 00:29:09.348 --> 00:29:11.788 models, and then Anthropic just shut them down. 00:29:12.348 --> 00:29:14.548 And now we finally see why. 00:29:14.778 --> 00:29:17.418 this is Anthropic entry into the world of personal agents, 00:29:18.618 --> 00:29:21.507 and there's a few things I want to talk about with, with the team here, uh, 00:29:21.628 --> 00:29:23.727 without like delaying. We have a lot of, a lot of news. 00:29:23.788 --> 00:29:26.387 I can spend a whole hour talking about this thing. 00:29:26.748 --> 00:29:30.608 Specifically, I want to talk about the fact that all of the companies now, all 00:29:30.668 --> 00:29:33.427 the major frontier labs now have a personal assistant. 00:29:34.788 --> 00:29:38.767 Anthropic has Duet, which is new and waitlisted, and they're releasing it very, 00:29:38.788 --> 00:29:40.288 very slowly. Uh, 00:29:40.988 --> 00:29:41.348 uh, 00:29:41.988 --> 00:29:42.608 OpenAI has 00:29:44.108 --> 00:29:48.207 Dots, which Nisten absolutely hates with passion, uh, and wants to go on the rant, 00:29:48.228 --> 00:29:51.928 but we don't have time, but at least one sentence from Nisten on Dots will be 00:29:51.988 --> 00:29:52.348 interesting. 00:29:53.148 --> 00:29:57.466 Meta has Muse, which is now leading the category as far as I'm concerned. 00:29:57.828 --> 00:30:01.668 Not only that, their model is trained for Muse specifically, and that flywheel is 00:30:01.748 --> 00:30:03.147 crazy, and Meta has the surface. 00:30:04.168 --> 00:30:05.847 Google and Gemini, they have Spark. 00:30:06.288 --> 00:30:09.967 Many people don't know about Spark, and which is a failure on on Google's folks to 00:30:10.048 --> 00:30:11.408 promote it and also make it good. 00:30:11.658 --> 00:30:13.918 and the Spar- Spark is quite bad. 00:30:14.018 --> 00:30:16.857 I've posted about this publicly, and I'm willing to give them feedback. 00:30:16.898 --> 00:30:19.997 So, folks from Google, if you're listening, I'm very willing to give you 00:30:20.018 --> 00:30:22.398 feedback. also there's entries to the category. 00:30:22.608 --> 00:30:27.287 Instinct is a big entry to the category, and Hark from, uh, um, uh, 00:30:28.368 --> 00:30:31.607 Hark Pro, whatever that's called, and a bunch of others, a lot. 00:30:31.848 --> 00:30:33.607 Obviously, we have the, the standalone 00:30:34.488 --> 00:30:40.138 self-assistance Hermes and OpenClaw, which is now what Jensen pushes to people, that 00:30:40.178 --> 00:30:43.788 people can build upon, and Pi Durable, which makes, easy for you to build one. 00:30:45.748 --> 00:30:50.687 So now every, every company has a personal assistant, and one of the reasons is 00:30:51.067 --> 00:30:52.627 those, they are very sticky, 00:30:53.548 --> 00:30:56.888 very sticky. Nisten, you want to talk about your experience with, with Muse and 00:30:56.908 --> 00:30:59.187 then moving away from Muse and how you feel about that? 00:31:01.188 --> 00:31:03.048 Nisten Tahiraj: Yeah, it look, it it's not that good. 00:31:03.108 --> 00:31:06.648 It's kind of dumb sometimes, but it just goes. 00:31:06.828 --> 00:31:08.307 It is so energetic. 00:31:08.788 --> 00:31:13.468 I give it tasks that involve, I don't know, categorizing videos, going into like 00:31:13.548 --> 00:31:17.307 a, a simulation and stuff, and it just does it. 00:31:17.988 --> 00:31:21.548 Uh, whereas, uh, Dots just feels like 00:31:22.228 --> 00:31:28.047 that government employee that tries to avoid all responsibility while still being 00:31:28.148 --> 00:31:33.388 right and handing off things and just giving you like 3-line, uh, updates. 00:31:34.028 --> 00:31:36.987 I- it's just com- so un-agentic. 00:31:37.508 --> 00:31:41.627 Yeah, uh, uh, it- it's the most useful thing. 00:31:42.028 --> 00:31:46.307 Meanwhile, Dots just struggles and stops the job because it couldn't upload a 20 00:31:46.428 --> 00:31:49.948 megabyte file through the interface, so it- it just stopped working. 00:31:50.428 --> 00:31:56.388 Whereas Muse is really opening entire VMs, backing up 80 gigs of data, is- is 00:31:56.428 --> 00:31:59.187 just going completely, uh, completely nuts. 00:31:59.588 --> 00:32:00.708 Su- super useful. 00:32:01.308 --> 00:32:07.307 Uh, most, most useful agent they have, uh, they have deployed to the masses, I think. 00:32:07.588 --> 00:32:09.798 Alex Volkov: Yeah, the, the agentic part is very interesting there. 00:32:09.878 --> 00:32:12.517 Like, you want your agent to be able to go and do stuff. 00:32:12.598 --> 00:32:14.557 Obviously, you want safety, right? 00:32:14.598 --> 00:32:18.217 So, like, we're, we're talking in, in Black October, essentially, one, one of 00:32:18.238 --> 00:32:23.167 the, one of the worst cybersecurity months, in, in the history of technology. 00:32:23.548 --> 00:32:28.308 So you don't want your agent taking too many things and going, like, willy-nilly 00:32:28.348 --> 00:32:30.428 with your data, but, definitely 00:32:31.288 --> 00:32:34.587 you also want it to be agentic and not stop and wait for you, otherwise it's 00:32:34.648 --> 00:32:37.988 wasting your time. Claude Duet specifically, they're talking about they 00:32:38.048 --> 00:32:42.368 will highlight the most important things, but we'll decide what other f- what other 00:32:42.408 --> 00:32:47.418 things are important. So, as far as I've seen back in my, you know, Hermes days 00:32:47.478 --> 00:32:53.068 when, I could use my subscription for Claude with, with Hermes, with OpenClaw, 00:32:53.388 --> 00:32:56.508 Opus has the best decisions towards, like, what to surface to you. 00:32:56.828 --> 00:32:58.727 And the proactivity part is, like, very interesting. 00:32:58.828 --> 00:32:59.907 I'm really excited to see. 00:33:00.228 --> 00:33:03.168 Uh, it's getting a little bit, um, 00:33:03.788 --> 00:33:09.427 cramped in my AI assistant zoo, where there's no instinct and there's no, uh... 00:33:10.868 --> 00:33:12.668 Dots doesn't send me that many notifications. 00:33:13.028 --> 00:33:16.807 and here's the biggest part, and LDJ would love to chat with you about this, is that 00:33:17.028 --> 00:33:19.787 all these companies now realize that this is the real stickiness. 00:33:20.388 --> 00:33:25.807 It's way more st- so if you're Claw, Claw, if you're Claude coding all day, you can 00:33:25.868 --> 00:33:27.288 just switch to a different 00:33:27.948 --> 00:33:29.708 harness with Opus behind it. 00:33:30.028 --> 00:33:30.427 It's not 00:33:31.188 --> 00:33:32.988 gonna be the same, but it's gonna be pretty much the same. 00:33:33.708 --> 00:33:37.787 If your assistant is building workflows with you and knows everything and has the 00:33:37.868 --> 00:33:41.128 password stored in a secure way, et cetera, switching to a new assistant is 00:33:41.268 --> 00:33:42.027 significantly 00:33:42.668 --> 00:33:43.868 more of an effort to you. 00:33:44.468 --> 00:33:45.748 And so it's very sticky. 00:33:45.908 --> 00:33:49.227 Not only that, for regular folks, and I think that this is the biggest part, LDJ 00:33:49.328 --> 00:33:53.467 would love you for you to chime in, for regular non-techy folks who don't Claude 00:33:53.548 --> 00:33:56.187 code and get to the limit very, very fast. 00:33:57.388 --> 00:34:01.847 This is a significant boost in the number of tokens they will use because these 00:34:01.948 --> 00:34:04.387 assistants run 24/7 and do things proactively. 00:34:05.108 --> 00:34:09.027 Thoughts on that? Did, did you already experience some of these, like, personal 00:34:09.148 --> 00:34:11.887 agents? Which ones, and what do you think about Duet specifically? 00:34:12.588 --> 00:34:15.327 LDJ: Yeah, so I've been using Dots lately, and 00:34:16.227 --> 00:34:19.827 I, I think there's a, a lot more room there where it could get a lot better. 00:34:20.428 --> 00:34:23.988 And overall, just, I think this, this goes for really all the bots overall. 00:34:24.028 --> 00:34:28.408 I think that as the models get more capable, get better trained to interact 00:34:28.588 --> 00:34:28.808 and- 00:34:29.428 --> 00:34:33.308 and kind of, uh, predict what the user would want in the future, I think they'll 00:34:33.328 --> 00:34:35.347 get a lot better over the- the next couple years. 00:34:35.948 --> 00:34:39.327 But at the moment, it still has been definitely like a net positive usefulness 00:34:39.388 --> 00:34:39.667 for me, 00:34:40.468 --> 00:34:45.427 and it's been just suggesting like, oh, hey, here's that, uh, here's that specific 00:34:45.508 --> 00:34:49.187 model that you were waiting for artificial analysis scores on, like, here's- and 00:34:49.227 --> 00:34:51.727 here's the cost efficiency curve that you're curious about on that that's 00:34:51.788 --> 00:34:55.608 finally published, and, and, and just the things like that, which are also relevant 00:34:55.708 --> 00:34:55.828 to 00:34:56.548 --> 00:34:58.108 our information we cover on the show. 00:34:58.428 --> 00:34:58.707 Yeah. 00:34:59.268 --> 00:35:04.527 Alex Volkov: Oh, 100%. Uh, all of these agents now on Wednesday report to me and say, hey, he- 00:35:04.587 --> 00:35:06.417 here's some stuff. specifically Grok, 00:35:07.068 --> 00:35:08.108 Bots is really good at that. 00:35:08.268 --> 00:35:10.787 specifically because, like, a lot of the information that we talk about on the show 00:35:10.948 --> 00:35:12.087 exists on the timelines. 00:35:12.868 --> 00:35:15.868 Last week, I asked my Grok Bots to just scroll the timeline for me and tell me, 00:35:15.948 --> 00:35:17.348 like, I don't have time to be on Twitter. 00:35:17.428 --> 00:35:18.588 Nisten, you can do that as well. 00:35:18.868 --> 00:35:21.668 I don't have time to be on Twitter, just scroll the timeline for me and tell me, 00:35:21.748 --> 00:35:23.608 like, what's important, and and it did. 00:35:23.708 --> 00:35:24.768 It was very interesting. I was like, 00:35:25.548 --> 00:35:27.818 can I just, like, stop, using Twitter? 00:35:28.137 --> 00:35:29.797 But no, not really. right, so 00:35:30.758 --> 00:35:32.437 back to Claude Duet. 00:35:32.798 --> 00:35:35.838 I want to show you guys a cool thing that it just told me. 00:35:36.107 --> 00:35:39.454 I asked, how do you tell what other assistants are doing? 00:35:39.774 --> 00:35:43.052 And it says, your Claude Code sessions can send tasks to me directly. 00:35:43.174 --> 00:35:48.613 This is a significant difference from Dots from OpenAI, where Dots sessions are not 00:35:48.694 --> 00:35:49.874 aware of Dots itself. 00:35:50.254 --> 00:35:51.414 I- I've tried to ask it. 00:35:51.724 --> 00:35:54.704 Claude in your other channels can pass notes to me, and I can read those 00:35:54.804 --> 00:35:55.683 channels, and 00:35:56.284 --> 00:35:59.353 if another session writes to an app you're connected, I can read work there. 00:35:59.434 --> 00:36:03.513 So it already knows that I have Git- GitHub and Linear and email as well. 00:36:04.794 --> 00:36:04.953 And 00:36:06.314 --> 00:36:07.394 I'm very excited about this. 00:36:07.514 --> 00:36:12.344 Like, literally got this this morning, and right now, uh, still the regular Claude is 00:36:12.463 --> 00:36:16.104 is helping produce the show, but I'm really excited about the fact that, w- we 00:36:16.224 --> 00:36:21.664 now have an entry from Anthropic, because Anthropic is really known at at taste, and 00:36:21.743 --> 00:36:24.344 we've been chatting about different, offerings from different 00:36:25.104 --> 00:36:28.904 products and launches. Anthropic is really known at taste, and taste is something I 00:36:28.944 --> 00:36:32.804 expect my personal assistant to do, not only design taste, but the stuff that I 00:36:32.824 --> 00:36:35.163 can already say about, Duet that are very, very good. 00:36:35.644 --> 00:36:37.564 They have a standalone interface. 00:36:38.044 --> 00:36:40.604 This is very important. I don't know how OpenAI messed this up. 00:36:41.004 --> 00:36:44.643 OpenAI's Dots is sitting inside the OpenAI stuff, and it's like over there on the 00:36:44.664 --> 00:36:45.763 top. There's no standalone. 00:36:46.524 --> 00:36:49.224 Muse, Grokbots, whatever, they all have standalone with integrations. 00:36:49.324 --> 00:36:52.234 This is very important. so not only, does Claude have standalone, they have a 00:36:52.314 --> 00:36:54.094 standalone app as well. 00:36:54.654 --> 00:36:57.713 Also very important. I want to open the Claude app separately. 00:36:58.014 --> 00:37:02.123 OpenAI shoved them all together, and it's really, I think it hurt them, uh, because 00:37:02.384 --> 00:37:05.563 I, I want that separation between my personal assistant and between the work 00:37:05.624 --> 00:37:09.183 that I do. Maybe my personal assistant is for my personal stuff, and OpenAI's work 00:37:09.244 --> 00:37:13.513 is for work stuff, right? so that is already a big, big thing for, for, for 00:37:13.614 --> 00:37:15.193 Claude, besides the intelligence, right? 00:37:15.494 --> 00:37:20.684 And if Fable 5.5 comes, which was rumored to come this week, and my assistant is not 00:37:20.983 --> 00:37:21.684 that, that, 00:37:22.464 --> 00:37:23.903 that better. That's crazy. 00:37:24.304 --> 00:37:24.784 But also, 00:37:25.784 --> 00:37:27.702 folks, Opus 5.5. 00:37:28.544 --> 00:37:29.583 Opus 5.5. 00:37:29.664 --> 00:37:31.744 This is exactly what I want for my assistance, right? 00:37:31.824 --> 00:37:32.964 Nisten, I see you shaking your head. 00:37:33.024 --> 00:37:35.864 This is like, this is the level of intelligence that I want for my personal 00:37:35.904 --> 00:37:37.063 assistant, Opus 5.5. 00:37:37.104 --> 00:37:39.084 And like, the only way to get it right now 00:37:40.064 --> 00:37:45.584 is beg your Grokbots for Opus 5.5, because apparently last week Elon Musk said that 00:37:45.624 --> 00:37:49.723 Grokbots, for the difficult task, they will route you to Opus 5.5, or 00:37:50.604 --> 00:37:51.403 getting Claude to it. 00:37:52.324 --> 00:37:55.424 So I'm very excited. If it's not clear, very, very excited. 00:37:55.524 --> 00:37:56.963 Let's go Anthropic. Thank you for this. 00:37:57.044 --> 00:37:59.974 I've been waiting. last comments before we move on to OpenAI stuff. 00:38:00.454 --> 00:38:05.974 Nisten Tahiraj: Uh, just, just give Muse a, uh, sandbox with Tailscale and Opus on it, and just 00:38:06.014 --> 00:38:08.473 give it an MCP to talk to Opus for everything it needs. 00:38:08.774 --> 00:38:10.013 It works amazingly well. 00:38:11.173 --> 00:38:12.133 That's how I use it. 00:38:12.534 --> 00:38:13.874 Alex Volkov: Yeah, but see, Nisten, here's the problem. 00:38:13.894 --> 00:38:15.473 Nisten Tahiraj: It's not the same. It's not the same. 00:38:15.854 --> 00:38:16.294 I agree. 00:38:16.774 --> 00:38:20.313 Alex Volkov: Like, like when a notification comes, I want that intelligence and that taste to, 00:38:20.354 --> 00:38:23.494 like, judge it and not, like, have to decide, oh, I need to go to him today. 00:38:23.614 --> 00:38:25.524 Yeah, I agree. Muse Wonder is great. 00:38:25.684 --> 00:38:28.944 Muse Wonder, I upgraded my Muse for Wonder, and it it's performed 00:38:29.004 --> 00:38:32.634 significantly better. Muse does a lot of stuff for me, including, including 00:38:32.714 --> 00:38:36.913 calling. Okay, one use case for a personal assistance, no matter which ones, but Muse 00:38:36.954 --> 00:38:38.074 is currently the best at this. 00:38:38.494 --> 00:38:44.254 Muse called, let me see, my Muse called my kid's school to see when I need to come in 00:38:44.294 --> 00:38:45.934 tomorrow for a parent-teacher conference. 00:38:46.694 --> 00:38:49.513 I didn't know about this, it wasn't on my calendar, they didn't send an email about 00:38:49.534 --> 00:38:54.454 this. Muse called the school, and they patiently walked through which kid when, 00:38:54.934 --> 00:38:56.294 and then it put it on my calendar. 00:38:57.014 --> 00:38:58.073 Just this thing 00:38:58.934 --> 00:39:00.653 that was done before I asked. 00:39:00.853 --> 00:39:03.533 Well, no, I actually had to improve the call, but it was beautiful. 00:39:03.774 --> 00:39:06.413 Uh, and I think that, like, a lot of, a lot of the... 00:39:07.814 --> 00:39:10.493 How should I say? A lot of the use cases we don't need, like, more intelligence 00:39:10.534 --> 00:39:14.334 for, but a lot of the decision making and the, the taste making in assistance, we do 00:39:14.414 --> 00:39:15.434 need, something like Opus. 00:39:15.454 --> 00:39:16.953 So I'm very, very excited. 00:39:17.314 --> 00:39:18.353 Uh, what? 00:39:19.114 --> 00:39:21.093 Yes, Muse called the school, which is great. 00:39:21.253 --> 00:39:24.614 All right, folks, let's move on to our next segment for OpenAI. 00:39:24.844 --> 00:39:26.799 very, very excited to talk to you guys about this. 00:39:27.600 --> 00:39:28.720 from Anthropic to OpenAI. 00:39:29.950 --> 00:39:34.269 they shipped the most boring features this week, but the best ones. 00:39:35.310 --> 00:39:37.389 12-hour prompt caching in the API. 00:39:37.830 --> 00:39:43.099 Writing to the cache costs 2 and a half times the input price, versus the regular 00:39:43.160 --> 00:39:47.080 1, 1.25 for the default 30-minute cache. 00:39:48.055 --> 00:39:51.735 basically, OpenAI is changing the way caching happens, and they're pitching it 00:39:51.815 --> 00:39:55.054 for always-on assistants like Dots, which reuse the same instructions and tools all 00:39:55.094 --> 00:39:56.254 day, hours apart. 00:39:56.695 --> 00:40:00.575 Thibault from the Codex team says the Codex and work the cache for logged in 00:40:00.615 --> 00:40:02.614 users now never expires. 00:40:03.895 --> 00:40:04.974 This is the big piece. 00:40:06.175 --> 00:40:09.335 You can come back to a conversation from last week and you still hit the cache. 00:40:09.455 --> 00:40:13.615 There's no penalty. This is a big thing that's known as a no-no in the world of 00:40:13.695 --> 00:40:17.254 Claude Code, for example, that you have this long ass thread and you come back to 00:40:17.294 --> 00:40:19.375 it and you pay all the tokens from scratch. 00:40:21.000 --> 00:40:22.919 I think that this is like huge, huge news. 00:40:23.919 --> 00:40:25.740 Thoughts on the caching updates, 00:40:26.720 --> 00:40:27.159 folks? 00:40:29.600 --> 00:40:30.680 LDJ: Things will be snappier, 00:40:32.000 --> 00:40:32.699 and we like that. 00:40:33.400 --> 00:40:36.020 Alex Volkov: But also cheaper. And also this, this is the way 00:40:36.800 --> 00:40:38.380 they fit the agentic world, right? 00:40:38.440 --> 00:40:41.919 Like, this is why, like, probably OpenAI is rebuilding this because you come back 00:40:41.960 --> 00:40:44.579 to your dot after 5 hours, 00:40:45.280 --> 00:40:49.000 and all of the instructions and all of the system prompts and all of the tools, 00:40:49.120 --> 00:40:50.599 whatever, they're already cached. 00:40:51.100 --> 00:40:53.740 and I gotta wonder if they found some breakthrough to do it. 00:40:53.900 --> 00:40:55.379 Always on cache is crazy. 00:40:56.100 --> 00:41:00.360 Like, obviously, obviously they need to do something where to store it, and storage 00:41:00.420 --> 00:41:01.179 now is not cheap. 00:41:03.260 --> 00:41:04.759 Nisten Tahiraj: Yeah, some crazy DevOps 00:41:05.540 --> 00:41:09.639 their team had to, has to pull, continue pulling for that for years in a row, yeah. 00:41:10.019 --> 00:41:10.239 Alex Volkov: Yep. 00:41:10.459 --> 00:41:12.699 LDJ: Yeah. Hey, Astra, design a new SSD. 00:41:13.940 --> 00:41:15.860 Alex Volkov: Maybe, maybe Astra design a new SSD. 00:41:15.980 --> 00:41:18.300 I am very disappointed in Astra. 00:41:18.540 --> 00:41:20.780 I've been talking about this, I'm not the only one. 00:41:21.180 --> 00:41:24.830 Some folks, like Swix and some folks like Tio are saying that Astra is the most 00:41:24.910 --> 00:41:26.470 disappointing release from OpenAI. 00:41:26.820 --> 00:41:28.740 I j- it just stops all the time. 00:41:29.100 --> 00:41:33.399 And so when they launched Dots, and I was like, yay, Dots, and I tried it out, and 00:41:33.440 --> 00:41:36.980 it was like powered by Astra. I was like, uh, I'd rather this be powered by, by 00:41:37.780 --> 00:41:39.899 s- GPT 5.6 Soul, for example. 00:41:41.180 --> 00:41:46.230 However, there's Astra and there's Astra Pro, and OpenAI also, has an update about 00:41:46.490 --> 00:41:52.269 GPT-6 Astra Pro, in ChatGPT runs multiple agents on different parts of the 00:41:52.370 --> 00:41:55.089 task in parallel, and you can follow each of the tasks. 00:41:55.620 --> 00:41:56.660 Astra Pro is 00:41:57.380 --> 00:41:58.440 ridiculous. 00:41:58.550 --> 00:41:59.430 Th- this is like a, 00:42:00.116 --> 00:42:00.876 this is not a 00:42:01.636 --> 00:42:05.715 just a huge model. This is a huge model, with many big models in a trench coat 00:42:05.756 --> 00:42:06.786 together. and 00:42:07.802 --> 00:42:11.211 they run, a lot of agents right now to for your task, and I think that's super cool. 00:42:11.892 --> 00:42:16.551 And OpenAI, I don't know if you guys saw this, OpenAI bought Orchestra Bio, a drug 00:42:16.612 --> 00:42:21.051 development coordination platform, and now is folding it into OpenAI's Rosalind, 00:42:21.092 --> 00:42:23.412 which is their life sciences stack. 00:42:23.972 --> 00:42:27.721 and they're offering Rosalind as a, free for university labs as well. 00:42:28.362 --> 00:42:30.602 So shout out to OpenAI for some of this work. 00:42:31.102 --> 00:42:34.502 They also have, Alexander Madry coming back to OpenAI to work on safety and 00:42:34.582 --> 00:42:34.981 security, 00:42:36.548 --> 00:42:41.047 and they are a big part of this week's cybersecurity stuff with Daybreak. 00:42:41.148 --> 00:42:42.268 Daybreak is OpenAI's, 00:42:42.953 --> 00:42:44.554 responsible access to 00:42:45.274 --> 00:42:47.714 models that can do vulnerability analysis, 00:42:48.474 --> 00:42:51.384 and many companies are getting access to Daybreak after the 00:42:52.024 --> 00:42:56.814 very famous, Hugging Face not being able to use any OpenAI models to check on the 00:42:56.933 --> 00:42:57.694 agent swarms. 00:42:59.614 --> 00:43:00.843 Righty. Argon? 00:43:01.483 --> 00:43:02.963 Let's talk about Argon. It's been a week. 00:43:03.244 --> 00:43:08.903 Argon is Gemini 4, folks from Gemini said that they have seen 25 00:43:09.184 --> 00:43:13.983 trillion tokens in 72 hours since Gemini 4 Argon was released. 00:43:14.494 --> 00:43:18.173 Logan, a friend of the show, says Argon processed about 25 trillion tokens in the 00:43:18.214 --> 00:43:19.013 first 3 days. 00:43:21.054 --> 00:43:24.133 96 million tokens every second, and Google's open, 00:43:24.894 --> 00:43:27.214 openly compute, compute constrained. 00:43:27.654 --> 00:43:30.894 Sundar said on Big Technology Podcast that Google is compute constrained, 00:43:32.014 --> 00:43:32.414 which is 00:43:33.094 --> 00:43:33.974 just insane. 00:43:34.654 --> 00:43:38.430 And, my favorite Argon story from this week, which Claude suggests that this is 00:43:38.470 --> 00:43:39.049 my favorite. 00:43:39.670 --> 00:43:41.790 It's not really Alex's favorite, it's Claude's favorite. 00:43:42.070 --> 00:43:43.230 Let's call things like they are. 00:43:43.530 --> 00:43:47.249 4 Argon agents beat Minecraft from scratch in 27 hours. 00:43:48.130 --> 00:43:53.730 No human help, no pre-scouted seeds, and they wrote 2,700 of their own 00:43:53.810 --> 00:43:56.500 program, also died 605 times on easy mode. 00:43:56.980 --> 00:43:58.840 Argon is now in anti-gravity, and 00:44:00.395 --> 00:44:01.704 Argon, was used 00:44:03.085 --> 00:44:07.834 to help find PHP 2 shell from Wiz, and Daybreak, like we mentioned, was used as 00:44:07.875 --> 00:44:13.355 well. So this is Argon a week after it came out, LDJ or Nisten, have you used the 00:44:13.475 --> 00:44:14.674 Gemini offering at all? 00:44:16.915 --> 00:44:20.114 The one thing, we're gonna wait for Wolfram to chime in here, but Wolfram is 00:44:20.155 --> 00:44:24.165 our kind of like, main residential Google and Gemini guy because his whole house 00:44:24.265 --> 00:44:28.505 runs on on Android. He does not use Argon because it rejected his prompts, 00:44:29.285 --> 00:44:29.625 which is 00:44:30.245 --> 00:44:31.004 kind of crazy. 00:44:32.205 --> 00:44:34.445 Nisten Tahiraj: All right, it it looked pretty good for 3D. 00:44:34.645 --> 00:44:35.125 I just, 00:44:35.765 --> 00:44:37.290 just got kind of, lost. 00:44:37.410 --> 00:44:40.050 Are the API prices good on it, or? 00:44:41.020 --> 00:44:43.619 Oh, $2 input, $10 output. 00:44:43.900 --> 00:44:44.080 Alex Volkov: Yeah. 00:44:44.500 --> 00:44:45.799 Nisten Tahiraj: Through November 30. 00:44:46.020 --> 00:44:47.099 Yeah, it's not, it's not cheap. 00:44:47.420 --> 00:44:49.079 Alex Volkov: It's, but it's Sonnet prices, it's not- 00:44:49.100 --> 00:44:50.320 Nisten Tahiraj: It's not that bad, actually. 00:44:50.420 --> 00:44:54.099 Alex Volkov: Yeah, it's not bad. It's Sonnet prices, it's not Astra or, or Fable prices. 00:44:54.225 --> 00:44:56.109 and open source has been very big this week, folks. 00:44:56.130 --> 00:44:58.229 So let's go to open source because I think it's exciting. 00:44:58.430 --> 00:45:01.250 The whale has resurfaced, and there's a bunch of open source stuff that we need to 00:45:01.290 --> 00:45:01.710 talk about? 00:45:15.030 --> 00:45:18.069 Open source AI, let's get it started. 00:45:21.590 --> 00:45:24.709 Alrighty, let's switch to open source, 00:45:25.470 --> 00:45:30.550 and I think the start, the start is DeepSeek V41 Flash 00:45:31.109 --> 00:45:32.229 1017. 00:45:32.669 --> 00:45:33.329 The whale 00:45:34.270 --> 00:45:36.010 has resurfaced with a new checkpoint. 00:45:36.310 --> 00:45:39.250 They've been doing this. This is the third DeepSeek checkpoint, if I'm not mistaken, 00:45:39.270 --> 00:45:42.149 right? Like, this is back-to-back checkpoints on the DeepSeek 00:45:42.830 --> 00:45:47.149 V4.1 Flash, which is now still MIT licensed, which is incredible. 00:45:47.590 --> 00:45:50.670 and DeepSeek is now showing up in a big, big way. 00:45:51.110 --> 00:45:54.029 Terminal Bench 4 goes from 31 to 42. 00:45:54.830 --> 00:45:58.209 It's a huge jump on on DeepSeek, and outputs get shorter. 00:45:58.390 --> 00:45:59.790 It's both better and cheaper to run. 00:45:59.950 --> 00:46:02.870 DeepSeek says it come from reinforcement learning, where the model could take its 00:46:02.950 --> 00:46:07.129 own work to end-to-end projects, which sounds like a little bit like s- s- smells 00:46:07.150 --> 00:46:09.290 a little bit like RSI. on DeepSWE, 00:46:10.210 --> 00:46:15.770 they claim 77% ahead of GPT 5.6 solo at 73 and Opus 5 at 74, 00:46:16.049 --> 00:46:17.569 not 5.5, just Opus 5. 00:46:18.050 --> 00:46:22.009 that, they op- they uh, DeepSeek also opened their search and fetch API to 00:46:22.050 --> 00:46:26.710 everyone, 75 cents per thousand calls and free through, Claude Code and Codex 00:46:27.590 --> 00:46:29.230 until October 31st. 00:46:29.990 --> 00:46:32.020 That is very interesting, as an announcement. 00:46:33.096 --> 00:46:34.216 thoughts on DeepSeek, folks? 00:46:34.376 --> 00:46:36.795 Any, LDJ, I would love to hear from you or your like, 00:46:37.542 --> 00:46:40.671 you know, your timeline. Is DeepSeek getting, excitement there? 00:46:41.772 --> 00:46:44.571 Is this new checkpoint better than the previous checkpoints? 00:46:44.712 --> 00:46:48.632 LDJ: Yeah, I mean, I would definitely say it seems like this checkpoint is certainly 00:46:48.712 --> 00:46:49.951 better than the previous one. 00:46:50.032 --> 00:46:52.171 Um, I, I really feel like they just should have called it 00:46:52.872 --> 00:46:58.392 V4.1 or 4.2 or, or, or whatever next increment instead of this, you know, 00:46:58.912 --> 00:47:03.071 th- they're doing the whole date scheme that OpenAI used to do with the GPT-4 00:47:03.272 --> 00:47:04.791 Turbo and 4o models. 00:47:05.312 --> 00:47:05.492 Alex Volkov: Yeah. 00:47:05.672 --> 00:47:09.632 LDJ: But besides that naming scheme rant, though, yeah, it's it's impressive. 00:47:09.752 --> 00:47:12.911 Uh, they're just continuing on the pace, but it seems like 00:47:13.992 --> 00:47:19.971 I wouldn't really say this is threatening, you know, the the Fables and Astra 00:47:20.012 --> 00:47:25.232 isn't certainly not like, you know, Fable 5.2 or Astra 6.1 that would be bound to 00:47:25.272 --> 00:47:25.992 come any day now. 00:47:26.432 --> 00:47:26.631 Alex Volkov: Yeah, 00:47:27.352 --> 00:47:27.572 b- 00:47:28.352 --> 00:47:30.632 but remember, this is an MIT licensed model. 00:47:31.072 --> 00:47:32.911 it is a big model, right? It's not a tiny one. 00:47:32.992 --> 00:47:34.071 You can't run it locally. 00:47:34.992 --> 00:47:38.811 So before we talk about the next model that you can run locally, this is not 00:47:38.872 --> 00:47:40.212 something that you can run locally, but 00:47:41.032 --> 00:47:44.112 so, so we see some of the evals here that are interesting. 00:47:44.552 --> 00:47:45.991 DeepSWE, I think, is the highest one. 00:47:47.032 --> 00:47:52.271 The, this interesting Agents Less Exam also 35.4, and Apex Bench at 72. 00:47:52.612 --> 00:47:55.452 so shout out to DeepSeek for MIT license, as always. 00:47:56.292 --> 00:47:59.772 You know, we have, we have to applaud them as well. 00:47:59.972 --> 00:48:03.552 Let's see here, the Pareto frontier looks insane on this model. 00:48:03.652 --> 00:48:08.811 The Pareto frontier for cost versus execution is just absolutely mogging. 00:48:09.332 --> 00:48:14.522 Look at this. This is like just DeepSeek on this top right quarter, which is, what 00:48:14.582 --> 00:48:14.901 is this? 00:48:16.002 --> 00:48:20.412 This is SWE, yeah, on DeepSWE, you can see other models on kind of 00:48:21.272 --> 00:48:22.652 across the map for pricing. 00:48:22.912 --> 00:48:26.951 So DeepSWE on pricing is just, just absolute, absolutely crazy. 00:48:28.048 --> 00:48:29.447 but, let's see what else. 00:48:30.888 --> 00:48:34.607 They also launched DeepSeek Search, and artificial analysis posted they have a 00:48:34.728 --> 00:48:36.368 search index, for search APIs. 00:48:36.548 --> 00:48:37.227 I think Exa 00:48:37.828 --> 00:48:39.787 is the, the, the top one on there. 00:48:39.868 --> 00:48:42.548 Yeah, Exa. Shout out to Exa for also releasing some stuff. 00:48:42.848 --> 00:48:47.247 LDJ: About running on a local model, you could actually run it on the 512 gigabyte Mac 00:48:47.328 --> 00:48:50.027 Studio, if anybody still has that before they stop selling them. 00:48:50.728 --> 00:48:53.647 Alex Volkov: Yeah, I think those are in hard, in, in hard supply. 00:48:53.998 --> 00:48:56.278 Yam Peleg, any thoughts on, on the latest DeepSeek? 00:48:56.877 --> 00:48:57.218 Welcome to 00:48:57.318 --> 00:48:59.947 Yam Peleg: Bro, the, destroy, they broke the internet. 00:49:01.198 --> 00:49:02.158 Alex Volkov: DeepSeek announces DeepInfer, 00:49:03.998 --> 00:49:07.978 which is the open source inference stack for DeepSeek, inference engine that serves 00:49:08.038 --> 00:49:09.838 the V4.1 flash in production. 00:49:09.918 --> 00:49:11.038 It's Apache 2 license. 00:49:11.278 --> 00:49:17.137 It runs on NVIDIA Blackwells and Huawei's Ascend 950, and it is, 00:49:17.358 --> 00:49:19.217 their own speculative decoding there and batch 00:49:19.838 --> 00:49:20.718 inference as well. 00:49:21.398 --> 00:49:24.267 Tim Dettmers, the QLoRA guy said it's a very, very big deal. 00:49:24.348 --> 00:49:27.548 He says DeepSeek inference system is on par with OpenAI stack, 00:49:28.667 --> 00:49:30.267 and it's now Apache 2 licensed. 00:49:30.628 --> 00:49:34.158 That's Tim Dettmers, uh, the QLoRA guy, the guy who who invented the the 00:49:34.238 --> 00:49:38.227 quantization method. and he says that it's better what Anthropic is running. 00:49:40.788 --> 00:49:43.664 Again, I I will repeat this slowly for the folks in the back. 00:49:44.224 --> 00:49:48.063 Tim Dettmers is one of the top folks at inference in the world, says what DeepSeek 00:49:48.504 --> 00:49:49.203 released with DeepInfer 00:49:50.064 --> 00:49:51.603 in Apache 2 license fully 00:49:52.224 --> 00:49:52.384 is 00:49:53.104 --> 00:49:58.524 on par with the OpenAI's inference engine and production, and is better than what 00:49:58.623 --> 00:49:59.463 Anthropic is running. 00:50:00.704 --> 00:50:02.103 at least for the user size. 00:50:02.464 --> 00:50:07.254 It doesn't ship KV cache system, or prefill decode disaggregation, yet, so you 00:50:07.334 --> 00:50:08.774 have to plug in yourself. 00:50:09.024 --> 00:50:13.883 semi-analysis measured VLLM within 6% of its decode throughput, so open source 00:50:14.164 --> 00:50:16.504 isn't that far behind from, from that as well. 00:50:17.004 --> 00:50:17.223 Like, 00:50:17.844 --> 00:50:19.583 Yam, okay, now broke the internet, yes. 00:50:19.644 --> 00:50:21.764 Yam Peleg: Bro, that's the secret sauce. 00:50:22.324 --> 00:50:23.984 that, that's the thing. 00:50:24.644 --> 00:50:29.104 Uh, look, I don't know if, everybody got the, how huge that is. 00:50:29.624 --> 00:50:32.614 look, inference providers, that's their bread and butter. 00:50:32.864 --> 00:50:38.624 the better you do this, the more you can serve more users with the same 00:50:38.904 --> 00:50:40.103 equipment, meaning that 00:50:40.984 --> 00:50:44.383 you just get to keep the rest of the margin for yourself. 00:50:44.424 --> 00:50:48.264 Like, you make this better, you, you literally pr- print money. 00:50:48.823 --> 00:50:54.584 And it, it, we all anticipated that, you know, the, the big labs 00:50:54.744 --> 00:50:56.443 probably has some, some tricks. 00:50:56.504 --> 00:50:56.844 Alex Volkov: Mhm. 00:50:56.864 --> 00:50:58.184 Yam Peleg: They probably work on this. 00:50:58.474 --> 00:50:58.973 that's their 00:50:59.594 --> 00:51:01.914 profit line at the end, so obviously. 00:51:02.634 --> 00:51:04.414 But we never got to see 00:51:05.390 --> 00:51:06.349 just how 00:51:07.189 --> 00:51:10.270 insanely efficient these things can be. 00:51:10.590 --> 00:51:16.270 And we also, the prices of DeepSeek models inference on 00:51:16.390 --> 00:51:17.410 DeepSeek versus 00:51:18.070 --> 00:51:19.250 other providers, and 00:51:20.610 --> 00:51:22.010 DeepSeek themselves were, 00:51:23.020 --> 00:51:24.299 competitive, to say the least. 00:51:24.420 --> 00:51:26.880 They were, phenomenal, phenomenally cheap. 00:51:26.960 --> 00:51:32.539 So we all, uh, we all, uh, we all guess that probably they have, you know, they 00:51:32.600 --> 00:51:33.340 have a very good 00:51:34.160 --> 00:51:34.839 stack for 00:51:34.920 --> 00:51:37.019 Alex Volkov: Yeah, ju- just to show the pricing again, 00:51:37.080 --> 00:51:37.180 Yam Peleg: Yeah. 00:51:37.200 --> 00:51:39.640 Alex Volkov: This is, this is just on the flash model. 00:51:39.880 --> 00:51:43.599 They're just like mogging the Pareto frontier on cost, and they they have been 00:51:43.640 --> 00:51:47.279 always doing this, and they only can do this if they have very, very, very good 00:51:47.400 --> 00:51:47.859 inference 00:51:48.720 --> 00:51:51.280 stack. And now this inference stack is open sourced. 00:51:52.356 --> 00:51:52.775 it's very 00:51:52.876 --> 00:51:54.515 Yam Peleg: Even more, look, even more. 00:51:54.836 --> 00:51:57.535 DeepSeek are known to release models 00:51:58.242 --> 00:52:03.232 you and we all thought, okay, they serve their models better than everybody else. 00:52:04.152 --> 00:52:08.382 They get to charge for it, through the API, so the model, gets open source. 00:52:09.102 --> 00:52:12.181 But now you also have the inference stack. 00:52:12.982 --> 00:52:14.062 It's just gonna be a bomb. 00:52:14.182 --> 00:52:15.847 It's just gonna be a bomb everywhere, seriously. 00:52:15.932 --> 00:52:20.152 Alex Volkov: I'm super excited about, prices everywhere going cheaper if we can run stuff on this. 00:52:20.177 --> 00:52:23.202 Like, our, I know our folks, already are looking at, like, the differences between, 00:52:23.537 --> 00:52:25.101 VLLM and, and 00:52:25.782 --> 00:52:25.981 DeepInfer, 00:52:26.621 --> 00:52:28.621 et cetera, and that's very, very interesting. 00:52:28.782 --> 00:52:31.821 Also, obviously, DeepSeek has been releasing other stuff, like their 00:52:32.102 --> 00:52:35.321 attention mechanism, DSAs, were released, like, earlier this year, also, blew 00:52:35.462 --> 00:52:36.561 everything else. the, 00:52:37.857 --> 00:52:41.457 the algorithm that they released for other stuff, this now includes speculative 00:52:41.497 --> 00:52:43.197 decoding. It's very, very big for open source. 00:52:43.257 --> 00:52:44.137 I'm very, very excited. 00:52:44.833 --> 00:52:46.473 speaking of open source, 00:52:47.153 --> 00:52:48.272 let's also talk about, 00:52:49.563 --> 00:52:50.682 Minimax, right? 00:52:50.843 --> 00:52:52.003 We have to talk about Minimax. 00:52:52.038 --> 00:52:54.953 I don't know, Peter, if you, if you're already excited to chime in, there was a 00:52:55.153 --> 00:52:57.883 model that was trained under the name, was that 00:52:58.503 --> 00:53:01.592 pony something? Hold Space Bunny was, the model was trained on a Space Bunny. 00:53:01.633 --> 00:53:06.132 I think it was open in open code as and, it's now Minimax M3.1 Flash, 00:53:06.883 --> 00:53:12.063 with 15 cents per 1 million tokens, and the weights are coming on October 23rd, so 00:53:12.183 --> 00:53:15.522 tomorrow. Peter, any any thoughts on Minimax, how that landed on Arena at all? 00:53:15.563 --> 00:53:17.723 And also, welcome to the show. Uh, good to see you. 00:53:18.523 --> 00:53:22.162 Peter Gostev: Yeah, actually, uh, I don't have too much to say, unfortunately, about Minimax. 00:53:22.283 --> 00:53:26.302 I just, but just to add on DeepSeek, I think, uh, first of all, yeah, that's a, 00:53:26.363 --> 00:53:28.182 that's an incredible thing that they're releasing 00:53:28.283 --> 00:53:28.502 Alex Volkov: Mhm. 00:53:28.523 --> 00:53:32.603 Peter Gostev: that much, but I think it's interesting that, um, I was looking at the Pareto 00:53:32.803 --> 00:53:34.882 Frontier on, on Arena, and it did 00:53:35.483 --> 00:53:38.483 land on it, but it's not like it crushed it. 00:53:38.768 --> 00:53:39.292 it. So I think 00:53:40.053 --> 00:53:44.133 it's kind of interesting. Yeah, maybe if they actually just had a bit more compute, 00:53:44.253 --> 00:53:47.273 maybe they would have reduced the prices even more, because I guess there's no 00:53:47.333 --> 00:53:49.212 point in them reducing prices 00:53:50.573 --> 00:53:50.693 by 00:53:51.453 --> 00:53:53.893 50% if, if they can't serve it. 00:53:54.013 --> 00:53:56.573 So I guess maybe that's why they're open sourcing as well. 00:53:56.653 --> 00:54:00.212 But yeah, it's kind of, um, yeah, that that's interesting. 00:54:00.253 --> 00:54:02.292 I kind of considering all of the, 00:54:03.173 --> 00:54:07.712 all of the, I know, messaging around it and all of the reception, you'd think 00:54:07.773 --> 00:54:10.213 that, oh, they're gonna crush it completely. 00:54:10.733 --> 00:54:14.253 I would say, in my mind, they they kind of got to the 00:54:15.013 --> 00:54:17.013 category of models that are very competent, 00:54:17.653 --> 00:54:20.832 uh, and they're they're pushing up from the lower end. 00:54:21.213 --> 00:54:21.432 Alex Volkov: Mhm. 00:54:21.453 --> 00:54:24.453 Peter Gostev: They're not, um, they're not like super cheap. 00:54:24.573 --> 00:54:26.392 So I, I want to dig into this a bit more. 00:54:26.453 --> 00:54:27.672 Why? I don't really understand why. 00:54:27.733 --> 00:54:29.392 Maybe it's just using a lot of tokens or something. 00:54:29.453 --> 00:54:30.492 Not, not totally sure. 00:54:31.013 --> 00:54:36.133 Um, but yeah, I think it's another thing is that for, for things like for Minimax, 00:54:36.253 --> 00:54:37.573 for Kimi, for 00:54:38.573 --> 00:54:41.793 Mimo and others, like, I think it's really difficult, right? 00:54:41.853 --> 00:54:43.453 Because how do you compete with DeepSeek? 00:54:43.493 --> 00:54:48.233 I think, like, Opus has an answer for it, and like GPT maybe Astra and others have 00:54:48.273 --> 00:54:48.753 an answer for it. 00:54:48.773 --> 00:54:48.872 Alex Volkov: Yeah. 00:54:48.893 --> 00:54:50.132 Peter Gostev: But like, what about the others? 00:54:50.213 --> 00:54:51.532 I think it's, it's hard for them. 00:54:53.053 --> 00:54:56.773 Alex Volkov: But also there's a lot of people who use AI, and I, I agree with you, but it's 00:54:56.812 --> 00:54:57.152 just... 00:54:58.573 --> 00:55:00.472 somehow they are, um, uh... 00:55:02.293 --> 00:55:02.753 I think... 00:55:04.133 --> 00:55:07.532 the highlight for me between those different models, not only DeepSeek is the 00:55:07.573 --> 00:55:11.512 best one, they're also like fully, fully Apache 2 MIT licensed open source, like 00:55:11.613 --> 00:55:13.592 fully, and they don't require any stuff. 00:55:13.653 --> 00:55:14.412 You can just run them. 00:55:15.013 --> 00:55:16.932 And also this is the smallest DeepSeek. 00:55:17.333 --> 00:55:19.832 DeepSeek V4, the big one, is going to land at some point. 00:55:19.933 --> 00:55:23.532 That's going to absolutely mog the frontier, especially with the pricing. 00:55:23.883 --> 00:55:28.162 and notable as well, NVIDIA is kind of like ownership of everything. 00:55:28.603 --> 00:55:33.463 The Deep Infer stack runs both on NVIDIA stack and on Huawei chips, which is also 00:55:33.643 --> 00:55:36.922 very, very notable because that's where like most of the Chinese folks, because 00:55:36.963 --> 00:55:38.283 the export controls can run. 00:55:38.703 --> 00:55:42.252 right, so skipping Minimax, it's pretty good, 74 and DeepSui, and weights are due 00:55:42.313 --> 00:55:44.073 tomorrow, so we're gonna m- mention them. 00:55:44.338 --> 00:55:48.193 Uh, one more open model this week is Flint, right? 00:55:48.213 --> 00:55:52.173 So we always talk about kind of the, the, the, the, the, the kind of the Chinese 00:55:52.253 --> 00:55:54.233 models and then also, open source models. 00:55:54.353 --> 00:55:59.429 We're going to chat with, Prince Kanuma about Flint In huge open source news, but 00:55:59.529 --> 00:56:02.049 open source/agentic news, Jensen Huang 00:56:03.229 --> 00:56:06.008 closed his GTC keynote 00:56:06.829 --> 00:56:12.628 by showing how he uses the Hermes agent to build something to train 00:56:12.749 --> 00:56:17.319 robots. Wolfram, you were there, and also this is a ping from your personal 00:56:17.399 --> 00:56:17.798 assistant 00:56:18.639 --> 00:56:19.898 after watching my news. 00:56:20.319 --> 00:56:22.038 Can you tell us about that a little bit? 00:56:22.119 --> 00:56:25.119 So shout out to News Research and the Hermes team for that incredible. 00:56:25.759 --> 00:56:26.619 Let's take Jensen. 00:56:26.639 --> 00:56:27.939 Yam Peleg: Let's go, man. Let's go. 00:56:27.959 --> 00:56:32.478 Alex Volkov: Open claw, yes, very strong, and then pull him towards Hermes, which is like, like 00:56:32.559 --> 00:56:34.479 huge. although they're still wearing claws. 00:56:34.518 --> 00:56:37.738 There's still the images of Jensen with like the red claws, like all over GTC, 00:56:37.799 --> 00:56:39.438 Wolfram showed this, this before. 00:56:39.689 --> 00:56:43.488 tell us how it felt to see Hermes like on the big stage, and what did they actually 00:56:43.609 --> 00:56:44.209 show and say? 00:56:45.129 --> 00:56:49.208 Wolfram Ravenwolf: Yeah, just, just to clarify, so they are still calling these kinds of agents claws 00:56:49.249 --> 00:56:54.349 apparently, but he always mentions Hermes in the same spot as, uh, OpenClaw and even 00:56:54.409 --> 00:56:59.589 some others he mentioned. And the thing is now that, um, at the end of the keynote, 00:56:59.609 --> 00:57:05.269 he was demonstrating how to use an agent and local AI and cloud AI together to 00:57:05.449 --> 00:57:10.009 create, uh, or to train the little, uh, micro ducks, the little robots that 00:57:10.089 --> 00:57:15.089 Hugging Face or Pollen Robotics created, which now belong to NVIDIA, after all. 00:57:15.529 --> 00:57:19.408 And, um, yeah, he was using Hermes agent locally on his system, and apparently 00:57:19.489 --> 00:57:23.888 there are skills for this, and it was training the robot on the local system, 00:57:24.009 --> 00:57:29.569 but also using online inference to, yeah, to basically, uh, get the 00:57:30.089 --> 00:57:31.648 horsepower that he didn't have. 00:57:31.729 --> 00:57:32.908 But the agent was doing it all. 00:57:32.969 --> 00:57:37.509 It's, it's like we are using our Wolfred and Amy's, you know, you have an idea, the 00:57:37.609 --> 00:57:42.108 agent does everything that is required and uses all the resources that are necessary, 00:57:42.129 --> 00:57:45.709 and then in the end you have a walking micro duck, and he even demonstrated it on 00:57:45.809 --> 00:57:49.988 stage with 4 little ducklings walking around and doing some movement that have 00:57:50.009 --> 00:57:50.888 been trained that way. 00:57:51.449 --> 00:57:53.368 Alex Volkov: It's very interesting how these things work. 00:57:53.849 --> 00:57:56.668 Micro duck is Poland Robotics, which was purchased by Hugging Face. 00:57:56.729 --> 00:57:58.809 Now Hugging Face is part of NVIDIA, and 00:57:59.569 --> 00:58:02.529 Jensen is standing on stage saying, hey, go to Pollen Robotics, buy these 00:58:02.609 --> 00:58:04.548 MicroDucks. It's really, really, really funny. 00:58:04.569 --> 00:58:04.989 Uh. 00:58:05.009 --> 00:58:07.908 Wolfram Ravenwolf: He said they are sold out already, but buy one anyway. 00:58:07.929 --> 00:58:11.888 And I got mine in August already pre- ordered, so I hope I could get it at the 00:58:11.929 --> 00:58:12.789 end of the year because 00:58:13.409 --> 00:58:17.309 it's so interesting how robotics is now taking off, and I felt that at the GTC as 00:58:17.369 --> 00:58:21.449 well, that robotics is, yeah, it's getting ever more important, the world models. 00:58:21.769 --> 00:58:25.628 Other labs like, uh, Black Forest Lab, they are also building these world models 00:58:25.729 --> 00:58:28.588 and cooperating with NVIDIA to get robots 00:58:29.209 --> 00:58:30.449 going farther away. 00:58:30.489 --> 00:58:33.728 So I think we are seeing an inflection point with robotics soon as well. 00:58:34.369 --> 00:58:38.878 Alex Volkov: All right, folks, this is a good reminder to tell you that the show is sponsored by 00:58:38.959 --> 00:58:43.338 none other than CoreWeave, and for the last 3 and a half years we had This Week's 00:58:43.399 --> 00:58:46.359 Buzz, because, you know, Weights & Biases is part of CoreWeave and Buzz, 00:58:47.199 --> 00:58:48.318 this was our joke, whatever. 00:58:48.549 --> 00:58:51.749 it's time to, it's time to announce a new segment on the show. 00:58:52.089 --> 00:58:55.928 this replaces This Week's Buzz, and this is called From the Forge. 00:58:56.489 --> 00:58:56.494 So, 00:59:05.025 --> 00:59:07.625 Segment sting: From the Forge powered by CoreWeave. 00:59:12.025 --> 00:59:12.984 Wolfram Ravenwolf: Cool new intro. 00:59:13.545 --> 00:59:14.745 Alex Volkov: Yeah, cool new intro. 00:59:14.995 --> 00:59:19.505 Wolfram, you're at GTC Berlin, uh, representing us, at GTC Berlin from 00:59:19.605 --> 00:59:22.415 CoreWeave, and, we, last week we talked about Reflections Beam. 00:59:22.745 --> 00:59:23.385 Reflections Beam 00:59:24.185 --> 00:59:27.924 shipped their open weights, and we're supporting them immediately on CoreWeave. 00:59:27.945 --> 00:59:32.185 And CoreWeave, also said that Reflection used weights and biases inside Forge while 00:59:32.285 --> 00:59:33.681 training, Reflections Beam. 00:59:33.731 --> 00:59:36.570 I need to get used to this. we told you about Reflection Beam 2 weeks ago. 00:59:36.971 --> 00:59:41.811 On the 16th, the full weights landed, 501 billion parameters, Apache 2 license, and 00:59:41.931 --> 00:59:44.971 we're running this on CoreWeave Serverless Inference from day 0. 00:59:45.381 --> 00:59:50.101 they used weights and biases inside CoreWeave Forge during Beam's training, 00:59:50.501 --> 00:59:51.160 and we, 00:59:51.981 --> 00:59:57.201 we knew Beam back when it was just a loss curve, said our social channels, which is, 00:59:57.241 --> 00:59:58.981 I think, is really absolutely crazy. 00:59:59.271 --> 01:00:02.871 Reflection says that Beam served a trillion tokens on its first day, which is 01:00:02.951 --> 01:00:06.350 a huge, huge news for Western open source. 01:00:06.621 --> 01:00:09.260 and it's also landed on Arena, and artificial analysis. 01:00:09.321 --> 01:00:11.001 There's a bunch of stuff as well. 01:00:11.371 --> 01:00:15.331 also on Forge, Nvidia and You.com used RL rollouts, which is our new, 01:00:16.007 --> 01:00:19.026 offering for, to post-train Nemotron 3.5 Lightning. 01:00:19.457 --> 01:00:20.017 BrowseComp 01:00:20.937 --> 01:00:26.356 bumped up from 37 to 45% on that, rollout using those RL rollouts, with 30% fewer 01:00:26.437 --> 01:00:29.896 tool calls. So if you are building any type of AI, CoreWeave Forge is the place 01:00:29.917 --> 01:00:33.957 for you to go. And robotics folks, physical AI launchables, Wolfram, you were 01:00:34.017 --> 01:00:37.277 there while, while we announced this, one command and you're running Isaac Sim 01:00:37.337 --> 01:00:41.916 template. Wolfram, you also said that in GTC Berlin, we talk, like, they announced 01:00:41.956 --> 01:00:43.567 a bunch of stuff, with Rubin stuff. 01:00:43.647 --> 01:00:45.426 Could you tell us about the NVL 8? 01:00:45.567 --> 01:00:47.526 What's that and like how, how that's gonna plug in? 01:00:48.407 --> 01:00:51.186 Wolfram Ravenwolf: Yeah, in addition to what you already mentioned, there were two announcement 01:00:51.247 --> 01:00:57.227 call we've made. So the first one is about NVIDIA HGX Rubin NVL 8, which is 01:00:57.287 --> 01:00:58.627 coming to CoreWeave this quarter. 01:00:59.207 --> 01:01:05.106 It's 8 Rubin GPUs, 2 terabytes of HBM4, and it's built for large KV cache 01:01:05.167 --> 01:01:07.567 and concurrent long context, uh, agents. 01:01:08.167 --> 01:01:12.766 And it's, uh, very efficient, so NVIDIA reported 10 times more tokens per megawatt 01:01:13.127 --> 01:01:15.926 than the H- HGX B200, 01:01:16.567 --> 01:01:18.686 um, tested with Kimi K2 Thinking. 01:01:19.007 --> 01:01:19.127 So 01:01:19.247 --> 01:01:21.146 Alex Volkov: Tokens per megawatt is an interesting one. 01:01:21.327 --> 01:01:21.787 That is. 01:01:21.807 --> 01:01:26.306 Wolfram Ravenwolf: We can run more agents for more users, and hopefully that will also affect the price, 01:01:26.367 --> 01:01:26.707 of course. 01:01:26.767 --> 01:01:26.986 Alex Volkov: Yeah. 01:01:27.687 --> 01:01:32.406 Wolfram Ravenwolf: And uh, the second announcement, uh, is very important for Europeans, especially 01:01:32.487 --> 01:01:36.346 bec- I mean, it's, it is important for everyone, and this is not a plausible 01:01:36.407 --> 01:01:39.026 thought I see, but I'm in a hotel room and the lighting is terrible. 01:01:39.127 --> 01:01:39.386 Something. 01:01:39.407 --> 01:01:40.767 Alex Volkov: Yeah, it's okay, it's okay. Keep going. 01:01:41.167 --> 01:01:46.626 Wolfram Ravenwolf: Security. Uh, zero trust AI frameworks that move security from promises, like we 01:01:46.647 --> 01:01:50.347 are not doing this, we are not trading, into customer-controlled architecture. 01:01:50.487 --> 01:01:55.507 So your keys, explicit support access when needed, hardware attestation, confidential 01:01:55.607 --> 01:01:59.626 execution, and no customer data being used to train provider models because 01:01:59.967 --> 01:02:04.527 everything is encrypted, it's end to end, and uh, this is coming, and this is 01:02:04.607 --> 01:02:09.066 something like in Europe, a lot of people, they can't use or are not allowed to use 01:02:09.247 --> 01:02:10.347 uh US services. 01:02:10.367 --> 01:02:14.987 Say in Germany, many companies require German hosting, but if this makes it 01:02:15.047 --> 01:02:19.066 impossible for a US company to look into what is going on here, that makes it more, 01:02:19.527 --> 01:02:21.446 uh, possible to use such services. 01:02:21.767 --> 01:02:23.126 And for everybody, security 01:02:23.807 --> 01:02:25.967 and privacy being enhanced is a good thing. 01:02:26.727 --> 01:02:29.527 Alex Volkov: Yep, and we offer a bunch of those services on Forge. 01:02:30.073 --> 01:02:31.592 Check out forge.coreweave.com. 01:02:32.353 --> 01:02:34.753 All right, folks, this was, uh, uh, from the Forge. 01:02:34.773 --> 01:02:36.213 This is the first from the Forge. I'm excited. 01:02:36.293 --> 01:02:38.593 There's a lot of news. There's a lot of news to to continue. 01:02:38.653 --> 01:02:40.492 Obviously, GCC Berlin, Wolfram, you were there. 01:02:40.833 --> 01:02:43.272 let's move on because we have tons of stuff to talk about. 01:02:43.733 --> 01:02:45.892 there's a lot of stuff in multimodal and video. 01:02:46.013 --> 01:02:49.572 Before by the way, the, the transition that you guys saw from The Forge, this is, 01:02:49.713 --> 01:02:55.072 uh, the, the, the Mini Max M3 Max, which is, uh, we've talked about a couple of, 01:02:55.293 --> 01:02:57.732 times on the show, and the music is from Eleven Labs. 01:02:59.068 --> 01:03:01.867 Let's talk about AI art and diffusion and video models, etcetera. 01:03:11.668 --> 01:03:14.828 All right, welcome back, Peter. folks, we have quite a few, 01:03:15.983 --> 01:03:20.223 I don't know how to call this category, AI art, video, generative diffusion AI. 01:03:20.343 --> 01:03:23.303 Now there's no diffusion models in LLM, so it's very interesting, but let's call 01:03:23.343 --> 01:03:25.269 this, taste category or whatever. 01:03:25.439 --> 01:03:29.889 Kling 4, it generates up to 30 seconds of video in a single pass up to 4K. 01:03:30.529 --> 01:03:32.968 We're now at the era where we're generating 30 01:03:33.649 --> 01:03:37.609 seconds in 4K with stereo audio and better lip sync. 01:03:37.809 --> 01:03:43.509 I think it's a good thing to go and try, and and find a good video with Kling 4. 01:03:43.889 --> 01:03:49.158 videos, voices, and images as references, and and it's like 3 times as cheap as 01:03:49.719 --> 01:03:50.399 Seedance. 01:03:51.085 --> 01:03:53.165 There's a lot of capabilities in there. 01:03:53.365 --> 01:03:55.404 4K specifically is the most exciting one. 01:03:55.845 --> 01:03:59.005 So let's take a look at Kling 4 and then maybe play. 01:03:59.205 --> 01:04:04.874 So video duration from 3 to 30 seconds, 1080p and 4K, dynamic range 4K, aspect 01:04:04.935 --> 01:04:06.654 ratio 21 by 9 is a new one. 01:04:06.975 --> 01:04:11.385 they they have new languages, Wolfram noticed the German and the and other 01:04:11.505 --> 01:04:14.145 languages as well. Dialects and accents is pretty cool. 01:04:14.505 --> 01:04:16.405 Uh, let's take a look at... 01:04:17.224 --> 01:04:18.645 there's so much to take a look at. 01:04:18.705 --> 01:04:19.545 Let's take a look at the 01:04:20.465 --> 01:04:21.464 start frames, for example. 01:04:24.465 --> 01:04:26.125 Demo audio: You never listen to a word I say. 01:04:26.145 --> 01:04:27.464 Because I've had enough! 01:04:29.390 --> 01:04:32.785 Alex Volkov: for me, this was getting stuck, but like, here's an example of the input frame of 01:04:32.825 --> 01:04:35.305 the first frame, and this was the video that they showed. 01:04:35.585 --> 01:04:37.665 let's take a look at another one. This, I think, is a full- 01:04:38.144 --> 01:04:39.925 Demo audio: Hi everyone, I'm so happy you're here. 01:04:39.945 --> 01:04:40.825 Alex Volkov: Wow, the lip sync is great. 01:04:41.185 --> 01:04:42.184 Demo audio: Welcome to my channel. 01:04:43.945 --> 01:04:45.880 Alex Volkov: The lip sync is absolutely, absolutely perfect. 01:04:45.934 --> 01:04:49.575 let's take a look at this one. This is the start frame, and let's look at this guy. 01:04:52.095 --> 01:04:54.574 The internet is a little stucky for me, so let me try again. 01:04:56.175 --> 01:04:58.934 Demo audio: Mic check check, light slow, I'm stepping in clean, no cut, one take, keep a shot 01:04:58.975 --> 01:05:01.134 rock tight, words run fast, but the flow scene bass from the quiet. 01:05:01.155 --> 01:05:02.754 Alex Volkov: All right, this is like, the lip sync is great, 01:05:03.055 --> 01:05:03.155 Unidentified speaker: Nice. 01:05:03.175 --> 01:05:03.695 Alex Volkov: but uh, 01:05:04.335 --> 01:05:06.634 this was, this was like, very, very clearly AI 01:05:07.475 --> 01:05:08.195 generated. 01:05:08.885 --> 01:05:10.445 does this lead the video category? 01:05:10.525 --> 01:05:12.925 I'm not sure, because there's a lot of competitors there. 01:05:13.065 --> 01:05:15.904 Flux 3 Max from Black Forest Labs, and Thal, 01:05:16.865 --> 01:05:22.215 which we've seen the Fal folks optimize the Mini Max M3, wi- with the Max version 01:05:22.255 --> 01:05:26.105 that runs super fast. Now Black Forest Labs, and, Fal released Flux 3 Max. 01:05:26.485 --> 01:05:30.745 Headline is speed, 10 second clip in about 12 seconds, so not quite real time, but 01:05:30.804 --> 01:05:33.804 very, very close. The folks at Fal, shout out, huge shout out to them. 01:05:34.165 --> 01:05:36.565 Uh, we can take a look at one of those as well. 01:05:36.925 --> 01:05:40.744 Uh. Peter, how, how, how is the arena in video generation these days? 01:05:40.845 --> 01:05:42.544 What what's, what's the highlight there? 01:05:42.605 --> 01:05:46.484 Who, who are these models breaking through the top, or are we still like in the Omni 01:05:46.565 --> 01:05:47.724 category from Google? 01:05:47.765 --> 01:05:53.744 Peter Gostev: Y- yeah, there's definitely a competition from, uh, Minimax, um, yeah, 01:05:53.845 --> 01:05:57.024 H3. I don't think, we don't have Kling yet, but you know what? 01:05:57.405 --> 01:05:59.745 They, actually, I'm just looking at the leaderboard. 01:05:59.845 --> 01:06:02.485 We haven't had one for a while, right, I think. 01:06:03.005 --> 01:06:05.644 So it'd be nice to see where they land. 01:06:06.205 --> 01:06:06.345 Alex Volkov: Yeah. 01:06:06.365 --> 01:06:09.085 Peter Gostev: Yeah, Kling 3 was like 19th now. 01:06:09.205 --> 01:06:14.144 So that's really the ga- you, I, I kind of forgot that, yeah, actually we haven't had 01:06:14.245 --> 01:06:16.885 a, a response from, um, from them in a while. 01:06:16.925 --> 01:06:17.205 So, 01:06:17.885 --> 01:06:21.825 yeah, if they go towards the top, that would basically add another competitor. 01:06:22.645 --> 01:06:25.805 Yeah, we probably had like 3 or 4, now it's gonna be 4 or 5. 01:06:26.125 --> 01:06:26.265 Alex Volkov: Yeah. 01:06:26.525 --> 01:06:29.044 Peter Gostev: Top tier competitors, so that that's really cool to see. 01:06:29.125 --> 01:06:33.525 Alex Volkov: And, and I think for folks who haven't, like, tracked this category for a while 01:06:33.645 --> 01:06:36.365 because they said, okay, you know, this is all like AI slop, et cetera. 01:06:36.805 --> 01:06:40.064 The video models is improving in a crazy rate. 01:06:40.125 --> 01:06:43.665 When we talked to you about Seedance before, Seedance was the only one that 01:06:43.725 --> 01:06:47.404 look at take up to like 5 references and do text to video and image to video. 01:06:47.685 --> 01:06:51.425 All of these now can do video to video and references for faces and characters, and 01:06:51.465 --> 01:06:54.324 the character consistency is insane, is insane. 01:06:54.365 --> 01:06:58.574 So now we're looking at realism, and uh, and and speed, basically, because Seedance 01:06:58.655 --> 01:07:01.794 took like, I don't know, 2, 2 and a half, 3 minutes to generate 30 seconds, and now, 01:07:02.105 --> 01:07:03.405 you know, this one is faster. 01:07:03.505 --> 01:07:06.344 Wolfram, go ahead, and then we'll show some examples of the, the PAL one. 01:07:07.104 --> 01:07:12.744 Wolfram Ravenwolf: And H3 is also taking off in the local AI scene because I'm running this on my 3090, 01:07:13.105 --> 01:07:14.164 and it works very well. 01:07:14.425 --> 01:07:15.724 Alex Volkov: Minimax H3, yeah. 01:07:16.345 --> 01:07:20.944 Wolfram Ravenwolf: At Minimax H3, yeah, you can now run a, a front, yeah, it's, it's almost frontier. 01:07:20.985 --> 01:07:23.924 I would call it frontier. I mean, you have Omni and Reference and all of this, and 01:07:23.945 --> 01:07:28.244 you can do this on your own GPU, and the community is fine tuning and optimizing 01:07:28.305 --> 01:07:30.124 it, and so it's really fast now. 01:07:30.225 --> 01:07:32.104 It is amazing what is happening here. 01:07:33.385 --> 01:07:37.124 Alex Volkov: All right, folks, I think it's time for us to, uh, to move. 01:07:38.025 --> 01:07:39.184 Oh, we're 5 minutes late. 01:07:39.425 --> 01:07:42.904 I'm so sorry. Okay, folks, it's time for us to to get to our next guest on the 01:07:42.945 --> 01:07:44.320 show. let me just run through this. 01:07:44.320 --> 01:07:48.704 so Cling 4, Flux 3 Max, and Crea 3 also launched a a model. 01:07:49.025 --> 01:07:52.324 It's number 4 on Arena multi-image edit leaderboard, so shout out to the folks at 01:07:52.465 --> 01:07:55.564 Crea. and yeah, Odyssey ML we don't have time to go to. 01:07:55.660 --> 01:07:59.805 All right, folks, time for us to move on because this week there's a lot of great 01:08:00.025 --> 01:08:02.214 open source, news from Western Labs. 01:08:02.335 --> 01:08:08.024 I am very, very happy to, invite, Tim Shi from, from Recursive AI and Prince 01:08:08.185 --> 01:08:11.395 Canuma, our friend, uh, second time, I think, or third time from the pod from 01:08:11.435 --> 01:08:12.740 like the local AI MLX community. 01:08:12.740 --> 01:08:15.435 think. Uh, Tim, how about this is your first time on the pod, 01:08:16.235 --> 01:08:19.195 and I know this this came together super quickly, so really appreciate you coming 01:08:19.235 --> 01:08:21.164 on, because you guys just launched Flint. 01:08:21.525 --> 01:08:25.164 Would love to just have you introduce Recursive AI and then, and then yourself, 01:08:25.205 --> 01:08:27.804 and then we'll, we'll get to Prince, and then we're gonna talk about this most, 01:08:27.885 --> 01:08:30.044 most, most, most exciting release from you guys. 01:08:31.685 --> 01:08:33.604 Tim Shi: Yeah, super excited to be here, and thanks for having me. 01:08:33.925 --> 01:08:35.464 Um, Recursive is building, 01:08:36.125 --> 01:08:40.784 um, yeah, self self-improving AI for us to build super intelligence, and we want to 01:08:40.885 --> 01:08:42.184 automate knowledge discovery. 01:08:42.245 --> 01:08:46.284 So the idea, uh, the process of ideation, implementation, validation, the whole 01:08:46.365 --> 01:08:49.684 scientific method, and we're applying to the science of AI first. 01:08:50.205 --> 01:08:54.404 So Flint was like the first demonstration of this technology, or we call it Eureka 01:08:54.484 --> 01:08:58.825 Machine, where you can basically apply it to the discovery of the training of the 01:08:58.885 --> 01:08:59.684 mod- training of 01:09:00.305 --> 01:09:05.744 a small language model, and then we, uh, basically optimize model for, uh, Mac. 01:09:05.885 --> 01:09:10.004 So it will run really fast on Mac, and it has like a lot of capabilities that can 01:09:10.125 --> 01:09:12.204 sort of do for your daily workflows. 01:09:12.405 --> 01:09:18.184 Um, yeah, so my background is I started my career actually with, uh, working with 01:09:18.325 --> 01:09:22.364 Richard and then NLP Labs in, uh, at Stanford and then OpenAI, and then I 01:09:22.405 --> 01:09:26.524 started a company called Cresta, basically one of the early attempts to apply, um, AI 01:09:26.564 --> 01:09:31.444 to enterprise. And then, yeah, super excited to be here and also, um, yeah, to 01:09:31.525 --> 01:09:32.364 talk about a release. 01:09:33.325 --> 01:09:34.934 Alex Volkov: That is awesome. absolutely thank you for coming. 01:09:34.955 --> 01:09:36.274 We're gonna get to Flint in a second. 01:09:36.355 --> 01:09:39.615 I also like a bunch of questions about the Eureka Machine, whatever you guys call the 01:09:39.655 --> 01:09:41.135 Eureka Machine, RSI, et cetera. 01:09:41.265 --> 01:09:43.025 Prince Canuma, second time on the show. 01:09:43.145 --> 01:09:44.764 Some folks know you as the MLX King. 01:09:44.785 --> 01:09:47.525 You want to briefly introduce yourself so we can get to the, the, the meat and 01:09:47.624 --> 01:09:51.645 potatoes of the most important thing, in, in open, in Western open, source AI this 01:09:51.705 --> 01:09:51.905 week? 01:09:53.025 --> 01:09:54.825 Prince Canuma: Absolutely. Thank you for having me, Alex. 01:09:54.945 --> 01:09:56.425 Hi, Tim, uh, once again. 01:09:56.785 --> 01:09:58.025 Uh, Prince Canuma here. 01:09:58.265 --> 01:10:00.784 I am the creator of MLX VLM. 01:10:01.145 --> 01:10:05.185 Over the past 3 years, I've been building infrastructure to run AI on Apple devices 01:10:05.625 --> 01:10:11.144 efficiently. We just crossed 7.5 million downloads in our engine, and, uh, happy to 01:10:11.625 --> 01:10:16.384 have brought uh, day zero support for Apple Silicon on MLX for Flint. 01:10:17.935 --> 01:10:21.864 Alex Volkov: folks, I'm looking at the Flint website, which I think I should like just the show 01:10:21.995 --> 01:10:23.554 while team we have the questions for you. 01:10:23.685 --> 01:10:25.124 let's pull this up here. 01:10:25.565 --> 01:10:25.965 this is 01:10:26.685 --> 01:10:29.804 quite the model, and a very small one at that. 01:10:30.045 --> 01:10:32.895 So could you introduce us to Flint before we talk about the Eureka machine 01:10:32.975 --> 01:10:37.264 sizes-wise, what are you guys aiming for, and why is it so cut and fast? 01:10:38.545 --> 01:10:38.804 Or how? 01:10:39.025 --> 01:10:39.224 Tim Shi: Yeah, 01:10:40.156 --> 01:10:43.811 we, we think about Flint as a first demonstration of our Eureka machine, sort 01:10:43.851 --> 01:10:46.090 of our, system to automate scientific discovery. 01:10:46.696 --> 01:10:51.677 and the Flint has, our goal is to optimize it for, like, your daily coding use cases. 01:10:51.956 --> 01:10:57.607 So developers can run Flint on a Mac, M4 and M5, and get really fast speed, and you 01:10:57.647 --> 01:10:58.167 can see, 01:10:58.873 --> 01:11:01.803 and be able to, like, sort of drive a lot of their daily coding use cases. 01:11:03.803 --> 01:11:08.242 Alex Volkov: Yeah, Prince, you saw this as well, like, there's like over 125 tokens per second on 01:11:08.383 --> 01:11:08.883 M4. 01:11:09.023 --> 01:11:09.402 Prince Canuma: Uh, 01:11:09.823 --> 01:11:10.242 Alex Volkov: Pro and 01:11:10.343 --> 01:11:10.942 Prince Canuma: 100, 01:11:11.743 --> 01:11:16.423 189 as of yesterday, but this is more of like a branch that I'm working on. 01:11:16.903 --> 01:11:17.122 Alex Volkov: Yeah. 01:11:18.183 --> 01:11:18.643 And, um, 01:11:18.663 --> 01:11:18.963 Tim Shi: Yeah. 01:11:19.503 --> 01:11:19.803 Alex Volkov: is, 01:11:20.783 --> 01:11:23.802 tell us about, so, okay, so I do want to like, talk about sizes a little bit, 01:11:23.843 --> 01:11:28.143 right? It's a 40 billion, parameter model, but like a very, very sparse MoE, right? 01:11:28.343 --> 01:11:31.563 Uh, uh, Tim, could you talk to us about like how... 01:11:32.783 --> 01:11:34.462 Okay, so the Eureka machine, once we get to, 01:11:35.063 --> 01:11:36.083 did it make the decisions? 01:11:36.103 --> 01:11:39.303 Like, how do you guys arrive at that, like, very, very, very sparse kind of 01:11:39.343 --> 01:11:40.842 architecture, 2 billion parameter active? 01:11:40.943 --> 01:11:41.102 Is 01:11:41.743 --> 01:11:43.303 that what, like, enables the speed? 01:11:43.573 --> 01:11:46.672 and how is it compared to, like, other models in open source that are running at 01:11:46.733 --> 01:11:47.653 the same-ish size? 01:11:48.853 --> 01:11:52.033 Tim Shi: Yeah, we actually first look at the sort of hardware substrate, right? 01:11:52.053 --> 01:11:56.373 So we work with closely with partners like, um, you know, N- NVIDIA and AMD, and 01:11:56.413 --> 01:12:01.453 we look at, you know, for a model to be running locally, what are the, basically 01:12:01.653 --> 01:12:02.853 the shape of the model to be 01:12:03.533 --> 01:12:07.752 eff- efficient, right? So for MacBook Pro, um, we basically asked the model to search 01:12:07.853 --> 01:12:12.213 what is like the most, um, efficient architecture, and it come up with 01:12:12.253 --> 01:12:16.773 something that's like actually way sparser than, um, uh, sparser than like 01:12:16.853 --> 01:12:17.692 comparative models. 01:12:18.293 --> 01:12:22.333 Um, so we ended up with a final architecture of 40 billion total parameter 01:12:22.413 --> 01:12:25.232 with about 2.5 billion active per token, 01:12:25.853 --> 01:12:29.232 and it's, um, basically what the Eureka Machine think will project it to offer the 01:12:29.253 --> 01:12:32.793 best combination of speed and coding performance, sort of among like all the 01:12:32.893 --> 01:12:37.513 configurations you can have on the, uh, of on the on the Mac, on the Mac, um, yeah. 01:12:38.373 --> 01:12:38.453 Alex Volkov: I 01:12:39.173 --> 01:12:43.692 find your last sentence incredible enough so that I will ask you to repeat it. 01:12:44.333 --> 01:12:47.773 This is what the Eureka Machine thinks that is the most optimal. 01:12:49.493 --> 01:12:49.793 Let- let's 01:12:49.813 --> 01:12:50.752 Tim Shi: That's right, yeah, yeah. 01:12:50.858 --> 01:12:52.303 Alex Volkov: the- let's talk about it. It's- it's incredible. 01:12:52.343 --> 01:12:54.763 You guys are like, literally the name is Recursive. 01:12:55.153 --> 01:12:58.042 Recursive, was that rebranded to SI or not yet? 01:12:58.138 --> 01:13:00.208 but Recursive Superintelligence, I think, is the concept. 01:13:00.929 --> 01:13:03.718 and then now you're referring to the fact that, like, the thing that built the 01:13:03.759 --> 01:13:06.459 thing, that's how you guys arrived at the architecture. 01:13:06.819 --> 01:13:11.438 How involved were you guys at the building of this, the decision making versus the, 01:13:11.674 --> 01:13:13.168 the Eureka Machine decision? 01:13:13.449 --> 01:13:14.688 And what is the Eureka Machine? 01:13:14.709 --> 01:13:16.268 I think it's time for us to actually talk about 01:13:17.249 --> 01:13:20.608 what is that, because I think it's maybe more important than the actual model. 01:13:22.449 --> 01:13:25.748 Tim Shi: It's funny that you mentioned it. We actually got the domain recursive.si, um, 01:13:26.089 --> 01:13:26.249 and 01:13:26.889 --> 01:13:29.348 it's also stands for recursive super intelligence. 01:13:29.489 --> 01:13:34.009 recursive self-improvement is also recursive super intelligence, um, and we 01:13:34.569 --> 01:13:37.929 are thinking about basically the Eureka machine as a basically basically a machine 01:13:37.988 --> 01:13:39.649 to automate scientific discovery. 01:13:39.869 --> 01:13:44.388 So it will basically have, um, think of it as like a community of scientists and 01:13:44.469 --> 01:13:48.579 agents that will come up with ideas and, you validate ideas and implement those 01:13:48.639 --> 01:13:53.878 ideas. So it runs in a loop, um, keep, you know, trying different things that will 01:13:53.919 --> 01:13:57.238 potentially improve the, um, your objective. 01:13:57.339 --> 01:14:02.539 So you define a certain objective for this particular example, it's going to be speed 01:14:02.599 --> 01:14:06.139 and coding performance. It's a good combination of what we think is going to 01:14:06.179 --> 01:14:08.199 be good for developer on the machine. 01:14:08.519 --> 01:14:13.498 And then the Eureka machine basically search for different combinations of, uh, 01:14:13.559 --> 01:14:16.278 different architecture. And of- of- obviously, like, you know, 01:14:17.119 --> 01:14:20.438 it's not inventing, like, completely new, like, say, transformer architecture, 01:14:20.479 --> 01:14:23.598 right? Obviously, there's like a lot of unknown unknowns in scientific discovery, 01:14:23.999 --> 01:14:27.599 but within the known boundaries of what you can, um, search for, I think it's sort 01:14:27.639 --> 01:14:29.959 of find some interesting discoveries. 01:14:31.879 --> 01:14:33.088 Alex Volkov: That is, very interesting. 01:14:33.169 --> 01:14:34.709 How how recursive is it as well? 01:14:34.809 --> 01:14:40.018 Is the Eureka Machine, like, update the Eureka Machine, inference or code or, 01:14:40.659 --> 01:14:43.779 directions? Can you talk about, like, the recursive part as well? 01:14:43.799 --> 01:14:47.729 Because, when you say Eureka Machine came up with this model, this model is a small 01:14:47.849 --> 01:14:50.359 model. h- how recursive is the recursive part? 01:14:52.319 --> 01:14:57.038 Tim Shi: Yeah, so the Eureka Machine, um, is a larger system and model with combination 01:14:57.059 --> 01:15:02.019 of different models and, um, harnesses, but, uh, uh, we're sort of moving towards 01:15:02.079 --> 01:15:04.058 eventually Flint, we'll build the next version of Flint. 01:15:04.439 --> 01:15:08.959 Um, currently, um, it's basically a larger set of systems, models looking at the 01:15:09.199 --> 01:15:11.559 small model and say, what are things to improve? 01:15:11.839 --> 01:15:16.839 So, for example, um, um, I think Thomas Wolf actually, um, posted about this as 01:15:16.879 --> 01:15:19.919 well. It's like an AI researcher looking at the LM head would say, you should 01:15:19.959 --> 01:15:23.519 always keep it in high precision and never decide to quantize it to 4 bits. 01:15:23.839 --> 01:15:27.638 But our Eureka machine actually tried that idea and validated this actually a good 01:15:27.719 --> 01:15:30.998 idea for, uh, for, for Mac and local coding use cases. 01:15:31.479 --> 01:15:34.718 So some, some of the, some things like that that maybe human scientists will 01:15:34.799 --> 01:15:39.739 overlook, but, um, a, a Eureka machine, like an automated AI scientist, would 01:15:39.799 --> 01:15:40.118 just, 01:15:40.879 --> 01:15:44.998 you know, combining all the existing knowledge, propose a hypothesis that might 01:15:45.079 --> 01:15:48.958 eventually be true. Those are the discoveries that we think that our machine 01:15:48.979 --> 01:15:50.519 will be able to make, um, 01:15:51.559 --> 01:15:52.939 down the road for many different domains. 01:15:53.039 --> 01:15:55.799 I think it'll be really valuable for, for the, for the science itself. 01:15:57.439 --> 01:16:00.398 Alex Volkov: I, I think that's kind of incredible, especially like some of the decisions 01:16:00.439 --> 01:16:02.559 that, you know, maybe humans didn't go through. 01:16:02.639 --> 01:16:05.839 Uh, let- let me go to Prince for a bit, for- for the model itself. 01:16:06.179 --> 01:16:07.608 Uh, Prince, you've been testing out like, maybe 01:16:07.749 --> 01:16:07.969 Prince Canuma: Yes. 01:16:08.029 --> 01:16:09.348 Alex Volkov: Day- Day 0 support as well. 01:16:09.669 --> 01:16:13.209 Tell us, is this a mo- is this a good model for a daily driver for coding right 01:16:13.249 --> 01:16:13.388 now? 01:16:14.388 --> 01:16:18.429 Prince Canuma: Yes. So actually, I rank it around one of the top 3. 01:16:18.549 --> 01:16:21.669 I have a- a list that I keep at all times because people ask me all the time. 01:16:21.709 --> 01:16:22.209 Alex Volkov: Yeah. 01:16:22.229 --> 01:16:27.789 Prince Canuma: Right now it's- it's Flint, it's Cohere, and it is Qwen 27B, the, the most 01:16:27.869 --> 01:16:32.149 famous one. Those are really good models, and actually Flint was surprising to me 01:16:32.309 --> 01:16:36.228 because a lot of models you have to, like, ask a question and then ask it again and 01:16:36.349 --> 01:16:37.429 continuously do that. 01:16:37.869 --> 01:16:39.868 Flint, tool calling, great, 01:16:40.549 --> 01:16:44.848 as well as, like, building artifacts, like, hey, build me a game, take documents 01:16:44.909 --> 01:16:47.348 from the internet, store it somewhere in my computer. 01:16:47.589 --> 01:16:48.748 It can do all of that. 01:16:49.309 --> 01:16:53.708 And uh, right now it's my main daily driver for, for a couple weeks on the 01:16:53.829 --> 01:16:54.149 native. 01:16:55.749 --> 01:16:57.229 Alex Volkov: Wow, that's quite incredible. 01:16:57.349 --> 01:16:58.148 And uh, Tim, 01:16:59.109 --> 01:17:02.489 kind of like I'm, I'm always assuming when I bring folks from different labs for 01:17:02.549 --> 01:17:04.989 training models, like I'm always assuming there's something else that, that they 01:17:05.069 --> 01:17:05.789 still don't like 01:17:06.549 --> 01:17:08.668 can talk to us about, about like the, the new iterations. 01:17:09.069 --> 01:17:12.648 But given the name itself, it's like very, very obvious where you guys are going to, 01:17:12.709 --> 01:17:13.789 to, to improve these models. 01:17:14.109 --> 01:17:17.149 I guess my question to you guys is like, why start from 01:17:18.149 --> 01:17:19.429 local and why 01:17:20.349 --> 01:17:21.189 bet on open? 01:17:23.549 --> 01:17:25.368 Tim Shi: I think the answer to open is very clear. 01:17:25.549 --> 01:17:30.049 I think that, you know, historically, uh, models being like in closed, uh, source, 01:17:30.109 --> 01:17:33.989 at least in the Western Hemisphere, and I think that, uh, a lot of Chinese open 01:17:34.029 --> 01:17:38.569 source models, open way models actually gained a lot of adoption in, um, uh, large 01:17:38.629 --> 01:17:41.669 enterprises, and we see the curve of, like, companies shifting. 01:17:41.689 --> 01:17:45.729 You know, maybe they start with large, you know, closed models, and they realize that 01:17:45.949 --> 01:17:50.668 the cost is going to be very high, and then they start shifting towards, uh, open 01:17:50.709 --> 01:17:54.868 way models. And we actually did that before at Cresta, like we with like OGPT 01:17:54.988 --> 01:17:58.829 models, and we shifted to fine tune our small, we actually have like tens of th- 01:17:58.869 --> 01:18:03.149 uh, tens of like different, uh, open way models, small models running production, 01:18:03.229 --> 01:18:07.309 running different things like summarization or, you know, some, um, auto 01:18:07.549 --> 01:18:09.008 triaging tickets and all these like 01:18:09.749 --> 01:18:10.908 specialized ta- tasks. 01:18:11.269 --> 01:18:15.248 And it's obvious like enterprises would eventually own their intelligence in a 01:18:15.309 --> 01:18:19.588 sense like they would prefer to have control over their intelligence instead of 01:18:19.629 --> 01:18:25.249 like relying on a external partner who, you know, may n- may may still be very 01:18:25.349 --> 01:18:29.549 costly at the time. So we think that this could be the future where people, some 01:18:29.589 --> 01:18:32.728 people would choose open way models, some people would choose like closed source 01:18:32.789 --> 01:18:36.908 models, but at the end of the day, it's the customer's choice, and I see the usage 01:18:37.029 --> 01:18:40.248 of a model is going to be, um, 01:18:40.909 --> 01:18:42.389 really high in the next year or so. 01:18:43.229 --> 01:18:46.328 And, and to answer your question about small, I think that we're just getting 01:18:46.389 --> 01:18:48.469 started. This is our first model release, and, um, 01:18:48.509 --> 01:18:48.648 Alex Volkov: Mm. 01:18:48.669 --> 01:18:51.589 Tim Shi: yeah, we'll see how, what the Eureka Machine can do in the future. 01:18:52.229 --> 01:18:54.968 Alex Volkov: Well, Tim, we, we, I don't know how familiar you are with Thursday AI, but 01:18:54.989 --> 01:18:58.969 like we, we've been doing, a segment on open source AI and modeling for the past 3 01:18:59.048 --> 01:19:02.429 and a half years, and every time that we have somebody from the labs come in and 01:19:02.449 --> 01:19:06.429 then we see Apache 2, we have, we have applause because I think it's very, very 01:19:06.449 --> 01:19:07.608 important for the ecosystem. 01:19:07.689 --> 01:19:11.069 So thank you guys for releasing this in like a very open licensing as well. 01:19:11.479 --> 01:19:14.639 I also want to mention the speed with which you train this. 01:19:14.719 --> 01:19:16.814 I think this is like one of the highlights, just before we, we drop off, 01:19:16.889 --> 01:19:19.028 this is one of the highlights in the blog post for me. 01:19:19.339 --> 01:19:20.018 could you talk about 01:19:20.939 --> 01:19:22.219 how fast this happened and, 01:19:23.085 --> 01:19:24.484 how? Just how? 01:19:26.525 --> 01:19:30.004 Tim Shi: Yeah, actually, a lot of kudos to the technical team, and I think, um, 01:19:30.685 --> 01:19:34.945 there's a lot of effort went into this, but I'm also very impressed by how fast 01:19:34.985 --> 01:19:39.585 we're able to put together, the, um, whole model factory and be able to, like, train 01:19:39.604 --> 01:19:42.644 the first model, and in about, like, less than 4 months. 01:19:42.985 --> 01:19:44.144 And I think a lot of this actually- 01:19:44.185 --> 01:19:44.845 Alex Volkov: 4 months 01:19:45.464 --> 01:19:46.685 to release a model from scratch. 01:19:48.105 --> 01:19:51.824 Tim Shi: We actually didn't have any pre-training, post-training effort until like 1 month 01:19:51.865 --> 01:19:52.845 ago, and we were 01:19:53.625 --> 01:19:56.944 debating what are the use cases, product use cases of like, you know, Eureka 01:19:56.985 --> 01:19:57.724 Machine. We thought 01:19:58.345 --> 01:20:03.185 we should be using it to train models, and I think that we just decided to put 01:20:03.225 --> 01:20:04.524 together teams, but also 01:20:05.625 --> 01:20:09.964 rethinking how, you know, agents can be used in the whole process and how Eureka 01:20:09.985 --> 01:20:10.905 Machine can be used to 01:20:11.545 --> 01:20:12.584 accelerate that process. 01:20:12.744 --> 01:20:13.225 So we 01:20:14.225 --> 01:20:17.764 apply auto research to almost every stage of the pipeline, from pre-training to 01:20:17.865 --> 01:20:21.544 post-training to RL to inference, and we think that 01:20:22.185 --> 01:20:27.225 that's going to be, um, uh, a way for a lot of labs to accelerate, to catch up, to 01:20:27.305 --> 01:20:31.345 like, you know, build models in a, in a, in a way that wasn't imagined before. 01:20:32.464 --> 01:20:34.315 Alex Volkov: I think, before, so that is incredible. 01:20:34.375 --> 01:20:37.695 Four, four months from nothing to no pre- training to a model that runs and one of 01:20:37.715 --> 01:20:40.394 the fastest to, to run on local hardware, I think it's incredible. 01:20:40.685 --> 01:20:44.565 maybe for folks who are listening who are like, I can't wait to try the Flint model. 01:20:44.804 --> 01:20:47.305 Uh, excited about the recursive stuff, excited about Eureka Machine, but like 01:20:47.365 --> 01:20:48.604 can't wait to try the Flint model. 01:20:48.820 --> 01:20:52.004 Uh, maybe from both of you, any, comments about how best to use this? 01:20:52.085 --> 01:20:55.064 Like all models come in with their tricks, with their, with their like specific 01:20:55.085 --> 01:20:56.435 stuff. Prince, you already mentioned something. 01:20:56.635 --> 01:20:57.635 Uh, team would love to hear from you. 01:20:57.955 --> 01:21:01.604 What, what do you find from, from using Flint, like, locally, that folks need to 01:21:01.665 --> 01:21:03.065 know about, like, how to prompt? 01:21:03.145 --> 01:21:06.365 Is there, like, any specific things about reasoning or quirks like that that are 01:21:06.405 --> 01:21:10.045 different from different models? Uh, would love to hear i- like each one of you give 01:21:10.085 --> 01:21:13.385 us a little bit of a tip to folks who are listening who would like to use those 01:21:13.545 --> 01:21:13.864 locally. 01:21:15.185 --> 01:21:15.564 Oh, and also, 01:21:15.585 --> 01:21:16.244 Tim Shi: Yeah, we'll be curious. 01:21:16.265 --> 01:21:20.124 Alex Volkov: I'll bring in Nisten, the co-host, who has been using this all day and wants to to 01:21:20.225 --> 01:21:22.385 tell you all about this, but let's start with with use cases as well. 01:21:24.665 --> 01:21:24.865 Tim Shi: Yeah. 01:21:25.105 --> 01:21:25.245 Alex Volkov: Yeah. 01:21:25.305 --> 01:21:27.304 Tim Shi: I would say this model is more agentic. 01:21:28.065 --> 01:21:31.284 Yeah, I would say this model is more agentic than than, um, feels like you can 01:21:31.345 --> 01:21:32.785 run it locally, do a lot of things. 01:21:33.185 --> 01:21:37.984 Um, but remember, this is the small model, so, you know, and compared to, um, like 01:21:38.105 --> 01:21:41.125 Opus and all these larger models, you typically would need to give it more 01:21:41.304 --> 01:21:44.785 instructions and guardrails and, you know, tips on what to do. 01:21:45.385 --> 01:21:45.604 And 01:21:46.385 --> 01:21:50.445 it's just the nature of the model, but also you can, um, yeah, because also I'm 01:21:50.645 --> 01:21:53.224 just curious, like, what, what other things you guys were able to build with 01:21:53.265 --> 01:21:56.545 this? So, yeah, feel free to share on social media about things you build and 01:21:56.625 --> 01:21:57.544 how you use the model. 01:21:59.505 --> 01:21:59.625 Nisten Tahiraj: So 01:22:00.505 --> 01:22:05.825 I, I have it doing both stuff for, uh, smaller models, just doing testing, but 01:22:05.865 --> 01:22:11.345 the most fun thing I have with it is I made a big city-sized GTA mod 01:22:11.744 --> 01:22:12.304 simulator, 01:22:13.145 --> 01:22:17.584 and, uh, this was probably the smallest, fastest thing that I, I could actually run 01:22:17.785 --> 01:22:18.625 as a character, 01:22:19.304 --> 01:22:22.104 uh, going around, uh, to make decisions and stuff. 01:22:22.345 --> 01:22:24.184 I just want to know what, what went into, 01:22:24.785 --> 01:22:27.345 uh, RLing it, uh, so well. 01:22:27.385 --> 01:22:32.565 A- again, it's not perfect, but it's pretty easy to, to prompt. 01:22:33.145 --> 01:22:35.505 It is actually pretty, pretty obie- obedient. 01:22:35.545 --> 01:22:40.344 So I want to, what went into RLing this, and do you guys plan on turning it into a 01:22:40.625 --> 01:22:42.984 JEV as well? Because it's kind of ideal size. 01:22:45.265 --> 01:22:46.924 Tim Shi: Yeah, I think, uh, that's, that's a good question. 01:22:46.985 --> 01:22:50.984 I think the RL, I think our, I think put a lot of effort to making sure the model is 01:22:51.025 --> 01:22:53.804 very agentic. And the JEV question is qu- um, interesting. 01:22:53.825 --> 01:22:56.825 I'll be curious to see if someone could make a demo around, like, you know, just 01:22:56.985 --> 01:23:00.205 model modest by itself to train itself to be a JEV. 01:23:00.545 --> 01:23:02.004 I think that would be a cool demo to show. 01:23:02.705 --> 01:23:02.964 Alex Volkov: No. 01:23:03.425 --> 01:23:03.605 Prince Canuma: Yep. 01:23:03.705 --> 01:23:03.964 Alex Volkov: Uh. 01:23:03.985 --> 01:23:08.285 Prince Canuma: The good, the good thing is that you can run up to like 8 agents in parallel and, 01:23:08.465 --> 01:23:09.924 and have really good performance. 01:23:09.985 --> 01:23:10.864 So you're talking about 01:23:11.465 --> 01:23:15.765 each agent, like 32 tokens per second and averaging like nearly 500 tokens per 01:23:15.825 --> 01:23:16.104 second 01:23:16.705 --> 01:23:22.224 on a, on a M5 Max, and this could work on a 48 gigabyte, even up to like 36 01:23:22.865 --> 01:23:27.685 gigabytes. And if you have more memory, great, but you can run a parallel agent 01:23:27.785 --> 01:23:28.104 with it. 01:23:29.345 --> 01:23:33.864 Alex Volkov: I think th- this is a great place to end our interview on because specifically, 01:23:34.225 --> 01:23:37.785 given the next segment about security and other stuff, and given the personal agents 01:23:38.025 --> 01:23:41.905 stuff, many people would like to run agentic models locally completely so that 01:23:41.985 --> 01:23:43.744 no information leaves their systems. 01:23:44.025 --> 01:23:47.435 And, you know, that that's one of the big upsides for local agents completely, like 01:23:47.495 --> 01:23:48.734 everything's running on your system. 01:23:48.960 --> 01:23:52.125 Tim Shi and Prince Canuma, thank you so much for coming and telling us all about 01:23:52.305 --> 01:23:52.605 Flint. 01:23:52.785 --> 01:23:53.244 Tim Shi: My pleasure. 01:23:53.425 --> 01:23:53.725 Alex Volkov: Running, 01:23:54.960 --> 01:23:58.859 running this locally. We're very, very excited about the new lab as well as like 01:23:58.900 --> 01:24:01.799 a new leader in in the category, and the speed is really, really impressive. 01:24:01.820 --> 01:24:03.599 So, so definitely give Flint a try. 01:24:03.840 --> 01:24:07.909 I will be monitoring the other stuff, so team feel free, and welcome to come back 01:24:07.970 --> 01:24:10.710 once we, once the Eureka Machine decides to release a bigger model. 01:24:10.810 --> 01:24:13.669 I I'll I'll we'll we'll be definitely down to to host that as well. 01:24:13.850 --> 01:24:16.150 Thank you so much for joining as we move to our next segment. 01:24:16.249 --> 01:24:17.349 Thank you, Alex. Bye-bye. 01:24:17.490 --> 01:24:18.570 Cheers. Thank you, Alex. 01:24:19.050 --> 01:24:20.809 Uh, thank you, team. Thank you, Prince Canuma. 01:24:20.970 --> 01:24:21.730 And now, 01:24:22.730 --> 01:24:26.370 I think I'll bring everybody back to the stage for this as we introduce Josh Long. 01:24:26.480 --> 01:24:29.580 Josh, I don't think you've been in video form on the pod before. 01:24:29.640 --> 01:24:30.759 We definitely chatted. 01:24:32.416 --> 01:24:35.795 I should have, I should have chatted with my Claude to rearrange the segments 01:24:35.856 --> 01:24:39.185 because, uh, we just finished a very high note, local, open source, Apache, like, 01:24:39.226 --> 01:24:43.446 works amazing, runs, and now we have to talk about the Black October, the very 01:24:43.626 --> 01:24:47.385 scary stuff that happens in the world of cybersecurity. 01:24:47.756 --> 01:24:52.256 Josh is a cybersecurity evangelist, Josh, some folks are calling whatever's 01:24:52.276 --> 01:24:54.795 happening right now in cybersecurity is the Black October. 01:24:55.492 --> 01:24:56.232 there's a lot of 01:24:57.012 --> 01:25:01.152 vulnerabilities and hacks are happening all the same time, so much so that folks 01:25:01.212 --> 01:25:02.611 are saying, is the internet done? 01:25:03.612 --> 01:25:07.191 Which is very scary. Just to name a few of them, the bigger ones, and the reason why 01:25:07.252 --> 01:25:11.932 we're talking about them to you in the AI show, AI news show, because a lot of them 01:25:12.052 --> 01:25:15.811 are AI related. Some of them are maybe swarm stacks, some of them were discovered 01:25:15.852 --> 01:25:20.491 by AI. we'll start with PHP2Shell, which is a huge vulnerability in the most, 01:25:21.592 --> 01:25:21.991 popular 01:25:22.632 --> 01:25:24.912 language in the world, the most websites run on. 01:25:24.952 --> 01:25:28.371 I think, like, the the latest stats are talking about 70% of the world's websites 01:25:28.432 --> 01:25:29.611 are still running on PHP. 01:25:29.912 --> 01:25:34.551 PHP2Shell basically gives an ability to take over that machine without any, like, 01:25:34.632 --> 01:25:35.991 credentials. There's also a, 01:25:36.667 --> 01:25:38.546 an updated CVE in, 01:25:39.992 --> 01:25:41.311 eMudhra, was that eMudhra, right? 01:25:41.352 --> 01:25:45.802 Like, the certificate, update that lets everybody say that they have a, valid 01:25:45.842 --> 01:25:51.111 HTTPS. There's just some of the stuff, including LiteLLM getting a supply chain 01:25:51.212 --> 01:25:51.532 attack. 01:25:53.012 --> 01:25:53.932 Josh, you're smiling. 01:25:54.572 --> 01:25:57.981 H- how are you still able to to smile, given that everything is going on right 01:25:58.042 --> 01:25:59.231 now and folks are f- freaking out? 01:25:59.322 --> 01:25:59.741 I want, 01:26:00.401 --> 01:26:01.461 I want some chats about that. 01:26:01.792 --> 01:26:05.701 Josh Long: okay, so for for one thing, it gives me an opportunity to get to talk about fun stuff 01:26:05.762 --> 01:26:11.032 like this on shows like, uh, ThursdAI, so, uh, that part of it I don't mind. 01:26:11.511 --> 01:26:16.811 Um, but yeah, I do think that, so there there tends to be a lot of like Y2K like 01:26:16.912 --> 01:26:18.532 hype around things like this, right? 01:26:18.612 --> 01:26:21.772 Everyone Yeah. starts freaking out, and it's like, oh my gosh, it's the end of the 01:26:21.832 --> 01:26:27.576 world. But, it, so a lot of the standard kind of disclaimers apply here, 01:26:27.657 --> 01:26:30.712 right? talking about uh, the PHP2 shell vulnerabilities first. 01:26:30.993 --> 01:26:36.123 there were, a few, independent groups who discovered these, two vulnerabilities. 01:26:36.558 --> 01:26:41.842 that we had Calif and Wiz and even OpenAI, and they were all kind of like finding 01:26:41.903 --> 01:26:43.482 these independently and reporting them, 01:26:43.823 --> 01:26:43.963 Alex Volkov: Yeah. 01:26:43.983 --> 01:26:47.903 Josh Long: um, which is really interesting because that doesn't seem like that happens very 01:26:48.023 --> 01:26:51.882 often. The last time that I remember that this being like a big thing 01:26:52.543 --> 01:26:55.762 where a whole bunch of people independently found the vulnerabilities 01:26:55.822 --> 01:27:00.363 and reported them was back with, uh, I think it was in 2017, toward the end of 01:27:00.383 --> 01:27:04.222 the year, um, with Spectre and Meltdown, and those kind of like came to light all 01:27:04.263 --> 01:27:07.962 of a sudden in January 2018, and everyone was freaking out about those too. 01:27:08.223 --> 01:27:08.402 Alex Volkov: Yeah. 01:27:08.463 --> 01:27:12.962 Josh Long: Um, so kind of unusual to have a situation like that where multiple groups are 01:27:13.023 --> 01:27:15.923 independently discovering the same vulnerabilities and reporting them at the 01:27:15.983 --> 01:27:16.503 same time. 01:27:16.863 --> 01:27:21.903 Alex Volkov: So, I said that, Daybreak, the, the kind of responsible security program that they 01:27:22.003 --> 01:27:26.083 allow slowly people into to, to be able to extend the defender's window, whatever 01:27:26.162 --> 01:27:27.662 Greg Bachman called the defender's window, right? 01:27:27.683 --> 01:27:31.783 Like, we have a limited time until all these, like, other intelligence catches up 01:27:31.823 --> 01:27:33.243 to the frontier of cybersecurity. 01:27:33.463 --> 01:27:37.212 we have limited time, so Daybreak is their system to let people into, and with GPT 01:27:37.333 --> 01:27:43.293 5.6, oh, sorry, 6.1 Sol, th- this is like w- where they discovered the PHP 01:27:43.393 --> 01:27:45.492 vulnerabilities and disclosed it to PHP Foundation, right? 01:27:45.733 --> 01:27:50.922 And also Calif and and Wiz and folks, th- they apparently use Gemini Argon, f- for 01:27:51.003 --> 01:27:52.942 that as well. both arrived at the same thing. 01:27:53.023 --> 01:27:56.863 What is that thing, Josh? Could you tell us, like, what w- what folks need to be n- 01:27:57.223 --> 01:28:01.712 what folks need to know about the the PHP 2 shell, and what they need to do as well? 01:28:02.593 --> 01:28:07.712 Josh Long: Yeah, so, uh, the the most important thing is, first of all, if you are running a PHP 01:28:07.913 --> 01:28:10.073 server, you need to make sure that it is patched. 01:28:10.193 --> 01:28:14.133 Like, hopefully you patched yesterday because, um, you know, these are pretty 01:28:14.233 --> 01:28:15.352 critical vulnerabilities. 01:28:15.513 --> 01:28:20.753 Um, a- as far as the the ratings go, um, a 10 is the highest. 01:28:21.032 --> 01:28:24.112 The CVSS score for these, I think, is 9.2. 01:28:24.513 --> 01:28:28.552 So it's like just about as critical as it gets as far as vulnerabilities go. 01:28:29.032 --> 01:28:33.372 Um, so these are very severe flaws, um, and it's very, very important to make sure 01:28:33.433 --> 01:28:35.012 that you are patched, uh, 01:28:35.193 --> 01:28:41.032 Alex Volkov: Severe enough that both the FBI and an NSA accounts on Twitter and the Prime Minister 01:28:41.073 --> 01:28:41.672 of Japan 01:28:42.433 --> 01:28:46.693 and multiple other, like, state, big state, like, companies all came out and 01:28:46.753 --> 01:28:49.253 said, hey, patch or go offline. 01:28:49.833 --> 01:28:49.973 Josh Long: Yeah. 01:28:49.993 --> 01:28:51.312 Alex Volkov: This is like how big this is. 01:28:51.353 --> 01:28:54.772 Like, you have to patch or completely take your system offline until you are able to. 01:28:54.833 --> 01:28:58.883 Don't keep them online. this goes out to everybody who listens to th- th- does run 01:28:58.903 --> 01:29:00.822 like PHP software, whatever WordPress is. 01:29:01.053 --> 01:29:05.592 we will say, though, uh, many companies and clouds, if you are running things on 01:29:05.613 --> 01:29:09.932 the big clouds, they have protected this, and if you have, like, a web firewall in 01:29:09.993 --> 01:29:12.493 front of your website, like a Cloudflare or Vercel, right? 01:29:12.953 --> 01:29:16.283 those, they have released, they have updated their, their web firewalls and 01:29:16.323 --> 01:29:18.302 have released, like, protections, but you still have to patch. 01:29:18.443 --> 01:29:20.063 Like, it's still, like, very important. 01:29:20.123 --> 01:29:20.523 Yeah, Josh? 01:29:20.883 --> 01:29:22.963 Josh Long: Yeah, 100%. Yeah, I was gonna mention that too. 01:29:23.043 --> 01:29:25.023 So I- I- I've seen a lot of, like, 01:29:25.683 --> 01:29:28.702 yeah, again, hype and kind of misinformation going around online where 01:29:28.723 --> 01:29:32.183 people are saying, you know, you might want to just stay off the web for the next 01:29:32.243 --> 01:29:35.762 couple of days, and and I- I'm like, well, let's think about that. 01:29:35.843 --> 01:29:40.783 Let's break that down. Okay, so first of all, the big Fortune 500 companies, um, 01:29:40.963 --> 01:29:43.322 are are going to have patched immediately. 01:29:43.603 --> 01:29:43.783 Um, 01:29:43.803 --> 01:29:43.942 Alex Volkov: Yeah. 01:29:43.963 --> 01:29:47.502 Josh Long: they already had web application firewalls in place anyway, which would have 01:29:47.603 --> 01:29:49.602 mitigated those kind of, uh, attacks. 01:29:50.123 --> 01:29:55.623 So, um, there there has been an increase in attempted attacks, uh, but, 01:29:55.963 --> 01:30:00.543 um, but because just about everybody's behind something like Cloudflare, um, you 01:30:00.563 --> 01:30:00.942 know, they're 01:30:01.803 --> 01:30:06.223 pretty widely protected, like, even if you haven't patched yet, but hopefully just 01:30:06.263 --> 01:30:07.922 about everybody has patched by now. 01:30:08.283 --> 01:30:11.982 Again, if you are running a PHP server, you need to make sure you're on the latest 01:30:12.123 --> 01:30:13.183 patch. Um, 01:30:13.923 --> 01:30:15.843 uh, it- there's no excuse for that. 01:30:16.083 --> 01:30:19.022 Even if you do have a web application firewall, you need to make sure you're 01:30:19.083 --> 01:30:23.082 patched. Um, but, but it's not the kind of thing that people need to be, like, 01:30:23.403 --> 01:30:25.882 freaking out about, "Oh my gosh, like, can I browse the web? 01:30:26.163 --> 01:30:28.522 Am I unsafe?" Well, think about it this way. 01:30:28.643 --> 01:30:33.542 So, um, if, like, you shouldn't be entering personal information into random 01:30:33.643 --> 01:30:35.083 websites anyway, right? 01:30:35.563 --> 01:30:36.002 So, 01:30:36.963 --> 01:30:40.803 like, it's- it's really- it's really probably not something you need to be 01:30:41.083 --> 01:30:43.722 extra, like, height- on heightened alert about, right? 01:30:43.883 --> 01:30:48.242 Um, like, go about your daily business, don't worry about it too much, but it's, 01:30:48.443 --> 01:30:52.063 it's more on the side of the people who are running websites, especially if you're 01:30:52.123 --> 01:30:55.522 not a Fortune 500 company who's got a big security team and all that. 01:30:56.043 --> 01:30:56.362 Um, 01:30:57.043 --> 01:31:01.023 you probably are more likely to be impacted by this because you need to patch 01:31:01.043 --> 01:31:01.462 yourself. 01:31:01.803 --> 01:31:05.543 Alex Volkov: Yeah, and hopefully folks are not self- hosting those completely and hosting them 01:31:05.603 --> 01:31:10.052 on servers, and cloud providers that, that they do kind of like patch because it's in 01:31:10.093 --> 01:31:12.232 the language itself. So that's one thing that happened. 01:31:12.293 --> 01:31:14.288 PHP2Shell is, it's kind of crazy. 01:31:14.668 --> 01:31:17.197 the first class, but there's a few vulnerabilities that were discussed. 01:31:17.258 --> 01:31:21.977 I, I do want to talk about the, the, the, the eMudhra kind of like signature 01:31:22.058 --> 01:31:23.938 signing. Could you walk us through that one, Josh? 01:31:23.978 --> 01:31:26.258 Like, uh, and Nisten, I see you like want to chime in. 01:31:26.338 --> 01:31:30.887 Feel free to chime in here as well, if, if your experience is like different, because 01:31:30.948 --> 01:31:36.028 we need to talk about the class of attacks and discoveries getting faster and faster 01:31:36.068 --> 01:31:39.427 with AIs that are able to do this now, and whether or not this is now our daily 01:31:39.468 --> 01:31:42.937 lives, like, should we always keep our OS up to date and browser up to date now, 01:31:43.218 --> 01:31:45.477 which is good security practices, which you can get to at the end. 01:31:45.618 --> 01:31:47.477 Nisten, you want to chime in here on about something? 01:31:47.518 --> 01:31:47.997 I see. 01:31:48.018 --> 01:31:53.278 Nisten Tahiraj: Yeah, as as someone that built their own router at home on Alpine Linux and sets 01:31:53.378 --> 01:31:58.038 own SE Linux policies, Josh, what about people that just don't care at all? 01:31:58.578 --> 01:32:00.877 Like, they actually just do not, 01:32:01.538 --> 01:32:03.457 uh, care. Like, what does this go? 01:32:03.498 --> 01:32:08.458 Because I know I could run an LLM on their, on their server or turn it into a 01:32:08.618 --> 01:32:11.277 botnet or something, it could harass people. 01:32:11.298 --> 01:32:14.138 But what about, what about people that they just, they still just do not 01:32:14.778 --> 01:32:15.097 care? 01:32:15.738 --> 01:32:17.297 Alex Volkov: When you say they don't care, what do you specifically mean? 01:32:17.318 --> 01:32:17.518 You mean 01:32:17.538 --> 01:32:19.218 Nisten Tahiraj: I mean, literally do not care. 01:32:19.538 --> 01:32:20.718 Alex Volkov: No, but like about which part? 01:32:20.778 --> 01:32:22.638 About the fact that now there's vulnerabilities outside? 01:32:22.658 --> 01:32:24.918 Nisten Tahiraj: About PHP on websites or their servers. 01:32:25.218 --> 01:32:26.058 Alex Volkov: Ah, yeah. 01:32:27.538 --> 01:32:30.377 Josh Long: So, so you're talking about like server operators who don't care. 01:32:30.697 --> 01:32:34.457 Nisten Tahiraj: No, no, just the people that own the these, these websites and that go to them. 01:32:34.778 --> 01:32:34.917 Josh Long: Ah, 01:32:34.938 --> 01:32:35.337 Alex Volkov: Yeah, server 01:32:35.618 --> 01:32:35.718 Josh Long: yeah. 01:32:35.738 --> 01:32:35.898 Alex Volkov: Yeah. 01:32:36.498 --> 01:32:40.097 Josh Long: So, so I, I guess, um, the thing that people should be aware of, if you are 01:32:40.178 --> 01:32:43.718 running a website and you don't have a patch and you don't feel like you need to 01:32:43.778 --> 01:32:48.337 worry about this, um, you probably don't understand the severity of situation, 01:32:48.378 --> 01:32:52.377 right? So, so anytime that there are major vulnerabilities just flapping in the wind, 01:32:52.458 --> 01:32:56.618 right? Like, uh, somebody else could come along and pretty much compromise your 01:32:56.697 --> 01:33:00.138 site. They could, um, could take your site offline, right? 01:33:00.218 --> 01:33:03.277 If, if you're running an e-commerce site, um, guess what? 01:33:03.298 --> 01:33:06.038 You're not gonna have sales, but it gets worse than that because now it's 01:33:06.138 --> 01:33:07.978 reputational damage to you, right? 01:33:08.418 --> 01:33:12.958 So, um, it- nobody's ever gonna trust you again if they know that the last time they 01:33:12.978 --> 01:33:16.038 went to your site, um, it was compromised, right? 01:33:16.138 --> 01:33:20.658 It- and- and, you know, and- ...there are different things that could happen with 01:33:20.698 --> 01:33:24.778 that compromise. Sometimes it's a very, like, public, obvious thing where somebody 01:33:24.898 --> 01:33:25.977 defaces the website. 01:33:26.138 --> 01:33:31.138 Um, other times it might be, um, sort of a more behind the scenes thing where 01:33:31.498 --> 01:33:37.158 they're, uh, they're in there, but they're not making themselves, like, widely known, 01:33:37.258 --> 01:33:41.557 so they're kind of, like, able to steal, um, information that people are providing 01:33:41.618 --> 01:33:42.738 to the website, right? 01:33:43.218 --> 01:33:46.018 Potentially including sensitive information, might be credit card numbers 01:33:46.058 --> 01:33:48.358 or whatever else people are typing into the site, right? 01:33:48.418 --> 01:33:49.418 So, um, 01:33:50.658 --> 01:33:54.617 i- i- it's- it's a- an important responsibility. 01:33:54.698 --> 01:33:58.478 If you are choosing to run your own server, you need to make sure to keep it 01:33:58.618 --> 01:34:00.057 patched. That's super important. 01:34:00.138 --> 01:34:05.078 Uh, if n- if for no other reason than your own, you know, reputation, um, you need to 01:34:05.138 --> 01:34:06.297 make sure to keep it patched. 01:34:06.778 --> 01:34:09.998 Alex Volkov: So, ThursdAI is focused on engineers and folks who are building in the public, 01:34:10.038 --> 01:34:11.857 right? So many of us are running websites. 01:34:11.898 --> 01:34:15.058 There was also some vulnerabilities in Node.js, which folks need to know about 01:34:15.078 --> 01:34:16.258 and, like, patch their Node.js. 01:34:16.358 --> 01:34:20.098 There was also, supply chain attacks that, you know, hopefully they were patched with 01:34:20.203 --> 01:34:23.698 sorry, LiteLLM. The other class of a big vulnerability that was discovered was not 01:34:23.778 --> 01:34:27.497 related to AI at all. It was like complete negligence from the company called eMudhra 01:34:27.578 --> 01:34:31.338 that they signed HTTPS certificates, and instead of, 01:34:31.978 --> 01:34:34.798 Josh, please correct me if I'm wrong here, but like I'm not a cybersecurity expert at 01:34:34.858 --> 01:34:35.938 all, but like instead of a 01:34:36.978 --> 01:34:42.807 24 or 48 like, like long character signing key, the keys were literally just, 01:34:43.168 --> 01:34:46.128 one of them was just like like the number 4, one of them was 56, whatever. 01:34:46.488 --> 01:34:50.607 Some faulty thing, could be, I don't know, led to the fact that those security, 01:34:51.174 --> 01:34:53.134 signatures were compromised completely. 01:34:53.774 --> 01:34:55.933 That is a different class of things to worry about. 01:34:56.034 --> 01:34:59.553 Could you walk us through, what, what a, how should I say, a compromised HTTPS 01:35:00.153 --> 01:35:03.064 signature, affects, how it affects people in the wild. 01:35:03.783 --> 01:35:07.443 Josh Long: Right, um, and and by the way, this is actually not the first time with eMudhra, 01:35:07.504 --> 01:35:11.344 that's not not the first time we've had a, um, a similar situation. 01:35:11.504 --> 01:35:16.744 So, um, I actually, um, wrote a blog post way back in like 2011 about a company 01:35:16.783 --> 01:35:22.304 called DigiNotar, um, that, uh, had a ha- had a situation where 01:35:22.664 --> 01:35:27.783 their um, uh, root certificate authority was, uh, was compromised. 01:35:28.024 --> 01:35:32.844 Basically, what happens in these kind of situations is that, um, uh, anybody who 01:35:32.984 --> 01:35:38.703 wants to can create their own certificate impersonating any website that they want. 01:35:39.144 --> 01:35:44.903 Um, now that doesn't mean that, um, anybody out on the- the- the web could, 01:35:45.184 --> 01:35:46.684 um, reasonably, 01:35:47.904 --> 01:35:52.043 like, imitate Google. This is more of a, like, a- a- a- um, they call it a 01:35:52.144 --> 01:35:54.043 man-in-the-middle type situation, right? 01:35:54.144 --> 01:35:59.504 So if, for example, like, let's say that, um, you went to your local 01:35:59.784 --> 01:36:05.364 restaurant or cafe or whatever, uh, hopped on their public Wi-Fi, uh, publicly 01:36:05.384 --> 01:36:06.504 available Wi-Fi, right? 01:36:06.704 --> 01:36:12.204 And, uh, and somebody, um, running, uh, or who was, um, 01:36:12.543 --> 01:36:14.464 could insert themselves in the middle, right? 01:36:14.543 --> 01:36:18.224 So- so basically the, um, the person hosting that Wi-Fi, 01:36:19.064 --> 01:36:23.763 um, or de- depending on- on the situation with, uh, you know, if it's like an open 01:36:23.864 --> 01:36:25.464 Wi-Fi network, maybe even somebody 01:36:26.104 --> 01:36:30.743 on the same network, um, could find a way to get in between you and the website that 01:36:30.784 --> 01:36:31.904 you're trying to access. 01:36:32.464 --> 01:36:37.504 Um, if they had one of these vulnerability, uh, vulnerable, um, 01:36:37.824 --> 01:36:43.663 certificates issue or that appeared to be issued by eMudhra, um, if your browser 01:36:44.224 --> 01:36:50.103 trusted that eMudhra root certificate, um, then it would trust that you 01:36:50.584 --> 01:36:53.603 were actually going to, for example, Google.com or wherever 01:36:53.624 --> 01:36:53.723 Alex Volkov: Mhm. 01:36:53.744 --> 01:36:57.924 Josh Long: you were trying to go. Um, so it- th- there's a lot of things that kind of have 01:36:57.984 --> 01:37:01.963 to- to be in place that- that- it's that man-in-the-middle aspect of it that makes 01:37:02.024 --> 01:37:06.423 it, like, probably not as big of a deal as I think a lot of people worry that it is. 01:37:07.024 --> 01:37:12.424 Um, I have seen people, you know, s- suggest that it's, uh, a- again, kind of 01:37:12.464 --> 01:37:13.624 one of these world-ending events. 01:37:13.944 --> 01:37:18.944 It is pr- a problem, no doubt, and, uh, so the major browsers, if you're keeping your 01:37:19.024 --> 01:37:22.564 browser up to date, that's really important because they've already removed, 01:37:22.784 --> 01:37:24.123 um, that certificate. 01:37:24.504 --> 01:37:28.923 Um, Apple is the one that's kind of a little bit problematic right now because, 01:37:29.184 --> 01:37:35.024 um, uh, Apple takes a long time to, uh, i- in- in between patches, 01:37:35.144 --> 01:37:38.804 right? So they haven't put out a new version of Safari, at least, uh, unless 01:37:38.863 --> 01:37:41.124 they did, like, in the last hour or so. 01:37:41.304 --> 01:37:41.443 Alex Volkov: Yeah. 01:37:41.664 --> 01:37:44.543 Josh Long: Um, there's not a new version of Safari out yet. 01:37:44.984 --> 01:37:45.344 Um, 01:37:46.224 --> 01:37:50.923 uh, and and Apple does take a long time typically to release patches, just in 01:37:51.064 --> 01:37:56.923 general. Um, but if you're running, you know, Google Chrome, uh, Brave, you know, 01:37:57.344 --> 01:38:01.603 Edge, whatever, they've got patches for all of the, uh, Firefox, for all the all 01:38:01.624 --> 01:38:07.043 the major, um, browsers have already removed the eMudhra certificate, so nobody 01:38:07.084 --> 01:38:12.624 else can pretend to have a, a, uh, a, a signed 01:38:12.704 --> 01:38:14.123 certificate that, you know, 01:38:14.863 --> 01:38:16.744 can let you impersonate something like Google. 01:38:16.784 --> 01:38:18.744 Alex Volkov: Impersonate and then decode traffic. 01:38:18.784 --> 01:38:20.029 I think this is like the biggest thing, right? 01:38:20.029 --> 01:38:22.813 like the man in the middle part that you're talking about, like HTTPS traffic 01:38:22.934 --> 01:38:25.423 is s- secured, and this is why the signatures those exist. 01:38:25.704 --> 01:38:30.274 If they pretend to be one of those, they can say, hey, I am an legit HTTPS like 01:38:30.294 --> 01:38:33.773 thing, but then decode the traffic that's supposed to secure, like credit card data 01:38:33.894 --> 01:38:38.094 and other stuff. if you're still browsing and not going to HTTPS only sites, that's 01:38:38.153 --> 01:38:40.974 a problem. But if you are, now you update your browser. 01:38:41.034 --> 01:38:42.403 I think that this is like the most important part. 01:38:42.449 --> 01:38:47.113 it's very, very important to update the browsers, update the OS systems as fast as 01:38:47.214 --> 01:38:48.373 possible. Um, 01:38:49.054 --> 01:38:54.754 and I would go and say as far as don't go to websites they're not big and known to 01:38:54.934 --> 01:38:58.153 purchase stuff as well and enter your credit cards is always a good advice, but 01:38:58.214 --> 01:39:01.083 the, like, the best advice right now is just don't go to a random janky website, 01:39:01.364 --> 01:39:02.484 enter credit card form. 01:39:02.999 --> 01:39:06.938 Nisten Tahiraj: By the way, the certificate key, what happened there is that during the key 01:39:06.999 --> 01:39:12.358 signing ceremony, the 384-bit key was just changed to a 4, which is so sketchy, but 01:39:12.559 --> 01:39:13.877 also so elegant. 01:39:14.479 --> 01:39:18.248 And, in my opinion, yeah, it was definitely a human that did that 01:39:18.849 --> 01:39:21.549 because they just found out that, like, the private key was literally just the 01:39:21.629 --> 01:39:22.809 number 4. Like, 01:39:22.829 --> 01:39:22.929 Alex Volkov: Yeah. 01:39:22.949 --> 01:39:25.829 Nisten Tahiraj: if it was an LLM, it would have at least done something better. 01:39:25.989 --> 01:39:26.389 That was, 01:39:27.229 --> 01:39:28.028 that was definitely, 01:39:28.629 --> 01:39:31.148 uh, the humans. It's, it's pretty, it's pretty funny. 01:39:31.189 --> 01:39:34.968 But the update, um, I'm like rebooting a whole bunch of stuff right now, I just 01:39:35.029 --> 01:39:35.549 realized. 01:39:36.749 --> 01:39:39.769 Alex Volkov: Yeah, there's a lot of like, there's a lot of, non, like, again, for, for folks who 01:39:39.789 --> 01:39:42.847 are maintaining servers, not folks who are just browsing the web, there's a lot of 01:39:43.488 --> 01:39:48.267 work to do now to upgrade, know, Linuxes and, and, PHP versions and Node versions 01:39:48.348 --> 01:39:49.618 and et cetera. so 01:39:50.498 --> 01:39:54.063 hug ops to everybody who spent a lot of time lately not going to sleep to like 01:39:54.063 --> 01:39:56.098 like try to mitigate the scope of these attacks. 01:39:56.138 --> 01:39:57.798 The last thing that we do want to talk about 01:39:58.498 --> 01:40:02.297 this week also, and it seems to be picking up, is hacks on 01:40:02.938 --> 01:40:07.477 like major companies. So OpenCode was hacked, what, a week ago or so, and they 01:40:07.538 --> 01:40:11.388 like publicly talked about this. And now Railway, which is a deployment place where 01:40:11.428 --> 01:40:14.137 you like deploy your websites and put the websites, was also hacked. 01:40:14.198 --> 01:40:18.558 And like the h- the thing that seems to be happening there, and we would love to chat 01:40:18.638 --> 01:40:22.217 with all of you guys about this, is that the extraction of API keys, specifically 01:40:22.318 --> 01:40:25.917 LLM API keys, is now what's happening because that's money. 01:40:26.198 --> 01:40:28.917 That's literal money to those hackers, right? 01:40:28.958 --> 01:40:32.517 Whereas previously, you know, in the world of, uh, let's say 01:40:33.718 --> 01:40:37.507 DeFi and and crypto, you would hack something because you can, like, sell 01:40:37.548 --> 01:40:40.387 those tokens immediately and then get valuable money. 01:40:40.928 --> 01:40:44.567 when somebody hacks a website, your personal data, you can sell it on the open 01:40:44.808 --> 01:40:47.907 dark web, but o- dark web has your personal data at this point already. 01:40:48.008 --> 01:40:51.168 I'm I'm not sure how much your SSN is worth right now. 01:40:51.877 --> 01:40:56.298 but your API key to OpenAI, if it hasn't have a limit on it, 01:40:56.918 --> 01:40:59.477 that's essentially like a f- free and direct access to your credit. 01:41:00.398 --> 01:41:04.038 And there's a lot of folks who are buying those keys for distillation purposes on 01:41:04.058 --> 01:41:07.138 the black market. There's like many, many folks who cover the API keys on the black 01:41:07.178 --> 01:41:10.477 market. So that seems to be what's happening right now with Railway and uh, 01:41:10.648 --> 01:41:13.357 and OpenCode as well, which I think OpenCode caught up as well. 01:41:13.627 --> 01:41:13.787 and 01:41:14.793 --> 01:41:15.853 thoughts on this, folks? 01:41:15.914 --> 01:41:18.713 Will this keep happening like, like with agent swarms? 01:41:18.754 --> 01:41:20.054 This is, is this the defender 01:41:20.714 --> 01:41:23.574 window that Greg Bachman talked about coming to a close? 01:41:23.914 --> 01:41:25.314 Are we at the place where, like, 01:41:26.194 --> 01:41:28.194 it's gonna be great, but for a while it's gonna be sucky? 01:41:28.233 --> 01:41:30.093 Like, how how to protect yourselves against it? 01:41:30.154 --> 01:41:32.594 Would love to hear opinions around, like, that specific thing. 01:41:35.233 --> 01:41:36.033 Nisten Tahiraj: Just use 01:41:36.874 --> 01:41:38.833 TailScale everywhere and uh, 01:41:39.874 --> 01:41:40.793 eh, yeah. 01:41:41.514 --> 01:41:43.194 TailScale, Two Factor, 01:41:43.954 --> 01:41:48.414 and uh, just just vibe code your own dependencies and 01:41:49.234 --> 01:41:51.233 vibe code your own home firewall, 01:41:52.114 --> 01:41:58.094 and just, just don't trust any, any type of like sales or marketing material for 01:41:58.274 --> 01:42:01.994 a product. If you can't see the code or your agents can't see it, you, you 01:42:02.154 --> 01:42:03.353 literally just cannot trust it. 01:42:03.794 --> 01:42:06.554 I, I don't, I don't know what else to, to tell people. 01:42:07.740 --> 01:42:13.460 Peter Gostev: Again, my, my question, something I don't, I don't really have a good feel for is, us 01:42:13.780 --> 01:42:18.500 people in security, like everyone, are they taking it like properly seriously? 01:42:18.740 --> 01:42:23.460 Because it's just weird how these things keep happening, and like the OpenCode 01:42:23.500 --> 01:42:26.939 team, they didn't have access to the, to Daybreak, right? 01:42:27.380 --> 01:42:30.580 It just feels like such a trivial thing, like surely you can just do that. 01:42:30.660 --> 01:42:34.140 So I don't know, like, do you feel like they're taking it seriously or not, not 01:42:34.220 --> 01:42:36.019 quite, Penny hasn't quite up yet? 01:42:38.780 --> 01:42:42.240 Josh Long: I- it's, it's interesting because I, I think, um, so I'll, 01:42:43.020 --> 01:42:44.540 and you mentioned vibe coding, right? 01:42:44.620 --> 01:42:47.340 Obviously, there's a lot of people who are, who are just trying to do their own 01:42:47.420 --> 01:42:51.739 thing, and, and, and, uh, and I think that's where a lot of times these API keys 01:42:51.780 --> 01:42:54.640 will leak is because, you know, they're, they're, they're 01:42:55.620 --> 01:42:58.820 unintentionally putting their API key in a public place, right? 01:42:58.900 --> 01:43:00.359 And then someone can grab it, right? 01:43:00.460 --> 01:43:03.700 That, that is- is one of the problems that we're seeing. 01:43:04.180 --> 01:43:05.080 Uh, that's happened 01:43:05.780 --> 01:43:06.499 multiple times, 01:43:07.100 --> 01:43:11.480 um, where somebody's just, uh, unintentionally exposed their own, uh, 01:43:11.900 --> 01:43:16.279 private API key. That's- that's kind of problematic, and- and I think that, um, 01:43:16.460 --> 01:43:20.479 you know, if- again, if you've never run your own service publicly, and now you're 01:43:20.500 --> 01:43:23.180 trying to do that for the first time, and you don't really understand kind of how 01:43:23.220 --> 01:43:26.700 these things work, because most people, uh, probably really don't. 01:43:26.820 --> 01:43:28.740 Like, the average person really doesn't get that. 01:43:29.260 --> 01:43:33.439 Um, they don't understand what the potential problems are with exposing keys 01:43:33.540 --> 01:43:37.979 like that. Um, and so, you know, they just go with whatever output they get. 01:43:38.060 --> 01:43:41.340 It seems to work when they're testing it, and so they just publish it to- to the 01:43:41.420 --> 01:43:46.220 web. And so sometimes that's how these things get- get exposed, which, you know, 01:43:46.540 --> 01:43:49.700 and- and- and I don't expect that everyone's going to be a security expert, 01:43:49.740 --> 01:43:49.940 right? 01:43:49.980 --> 01:43:50.399 Alex Volkov: Yeah. 01:43:50.420 --> 01:43:54.139 Josh Long: And not everybody knows to, you know, prompt like, hey, is there anything about 01:43:54.260 --> 01:43:59.359 this that, like, maybe I should consider doing to harden this before I publish this 01:43:59.420 --> 01:44:01.540 to the open web and start advertising it? 01:44:02.060 --> 01:44:05.560 Um, it would be nice if people thought to do that, or if, uh, that was like sort of 01:44:05.659 --> 01:44:09.379 baked in whenever you're developing something, uh, vibe coding something, but. 01:44:10.380 --> 01:44:11.079 Alex Volkov: There is a reason 01:44:11.740 --> 01:44:11.860 why 01:44:13.140 --> 01:44:13.620 it took 01:44:14.580 --> 01:44:17.200 over 8, uh, Peter, we we talked about this, I think, on the public. 01:44:17.220 --> 01:44:20.479 There is a reason why it took over 8, 9 months for the bigger companies to release 01:44:20.520 --> 01:44:22.019 something like a personal agent to you. 01:44:22.740 --> 01:44:25.779 Like OpenClaw and Hermes, which you run on your computer and you're kind of in charge 01:44:25.820 --> 01:44:27.439 of the security itself, that's one thing. 01:44:27.500 --> 01:44:31.360 There's a completely another thing to allow people to provide their passwords, 01:44:31.620 --> 01:44:33.960 passwords, and then save them in the, in like a local, 01:44:35.060 --> 01:44:36.100 but secure way. 01:44:36.220 --> 01:44:37.499 that the LM doesn't have access to. 01:44:37.540 --> 01:44:41.760 Like, all that infrastructure has to be in place for people to like, to, to, to go 01:44:41.860 --> 01:44:45.119 trust because like if one of those vulnerabilities goes around and actually 01:44:45.140 --> 01:44:49.030 like steals your data from your personal AI assistant, it's gonna be like awful for 01:44:49.150 --> 01:44:50.269 a reputation of the company. 01:44:50.350 --> 01:44:54.450 There is a reason why, like, only now, almost a year after Peter Steinberg 01:44:54.490 --> 01:44:58.330 innovated with the OpenClaw, we're seeing these companies step in, despite them very 01:44:58.390 --> 01:45:01.849 clearly realizing the benefits of, hey, personal agents is like much more sticky, 01:45:01.950 --> 01:45:05.890 sells more tokens, it's harder to to to switch from, et cetera, because there's 01:45:05.970 --> 01:45:09.939 security there. Wolfram, to your, to your metaphor from before, building your house 01:45:10.040 --> 01:45:13.300 versus renting it, so I would, I would add on that metaphor, given the security 01:45:13.420 --> 01:45:14.619 considerations we're talking about now. 01:45:14.930 --> 01:45:18.650 this is the difference between building your house and not really knowing what a, 01:45:18.680 --> 01:45:23.349 a door bolt, security bolt is, and living in a high-rise apartment with a security 01:45:23.410 --> 01:45:27.430 guard at the door, right? Like Muse, the folks have hired, Moxie, Marlinspike, the 01:45:27.450 --> 01:45:29.370 folks from Signal for the E2E connections. 01:45:29.395 --> 01:45:32.560 cetera. They are known for security significantly more than you just like vibe 01:45:32.620 --> 01:45:36.940 coding your server in Nginx and running this on, on like PHP, which is now fully 01:45:37.100 --> 01:45:38.419 compromised, which is scary. 01:45:38.700 --> 01:45:42.440 The practical advice that we want to give folks is, obviously, if you are having any 01:45:42.540 --> 01:45:47.750 API tokens, and you were on those exposed websites, OpenCode, or Railway, et cetera, 01:45:47.950 --> 01:45:51.009 rotate them. Like, know which agent has which API token. 01:45:51.050 --> 01:45:52.070 It's very, very important. 01:45:52.190 --> 01:45:56.069 I know it's a pain in the ass, but at days like this, a pain in the ass is very well 01:45:56.150 --> 01:45:58.549 worth it. Put caps on your API tokens as well. 01:45:59.390 --> 01:46:00.829 Put caps in email notifications. 01:46:01.110 --> 01:46:04.295 Get email notifications when your API keys are like running out. 01:46:04.350 --> 01:46:04.859 put, uh, you know, 01:46:05.660 --> 01:46:09.880 it's better to to to put auto auto refill because then you at least get a 01:46:09.920 --> 01:46:13.000 notification from your credit card statement that like your key has been auto 01:46:13.060 --> 01:46:16.200 refilled. So if you see like multiple ones happening in succession, you'd be able to 01:46:16.240 --> 01:46:19.899 like run and stop this bleed, basically, if your key was exposed. 01:46:20.060 --> 01:46:21.419 Rotating keys is very important. 01:46:22.020 --> 01:46:24.980 And I think the the last thing is very clear for many people. 01:46:25.060 --> 01:46:26.219 Many people don't have this yet. 01:46:26.740 --> 01:46:28.460 Use a goddamn password manager. 01:46:29.460 --> 01:46:32.760 Use a password manager and store those keys so you don't have to remember them 01:46:32.780 --> 01:46:33.980 and put them in files. 01:46:34.340 --> 01:46:40.059 It is, it is increasingly more important now than it used to be before, given that, 01:46:40.219 --> 01:46:41.940 you know, you're giving your keys to agents, etcetera. 01:46:42.100 --> 01:46:43.179 Wolfram, you want to chime in with some 01:46:43.820 --> 01:46:45.199 security practices simple for folks? 01:46:45.260 --> 01:46:47.959 Wolfram Ravenwolf: Yeah, you, you already gave the practice I wanted to mention, but 01:46:48.740 --> 01:46:52.199 you should treat it like a credit card, but it's even worse than that because the 01:46:52.300 --> 01:46:56.239 credit card, you also have the CVC number, and you often have to do a two-factor to 01:46:56.340 --> 01:46:57.119 authorize a payment. 01:46:57.180 --> 01:46:59.959 Alex Volkov: And credit cards are secure and protected by the credit card companies. 01:47:00.060 --> 01:47:03.679 I don't think your API key is gonna get re- refunded by OpenAI if it was leaked. 01:47:03.940 --> 01:47:05.279 Like a very, very big point. 01:47:05.300 --> 01:47:06.179 Like they can, you know. 01:47:06.580 --> 01:47:09.760 Wolfram Ravenwolf: Yeah. So treat them even more securely than your credit. 01:47:09.820 --> 01:47:10.039 Alex Volkov: Yeah. 01:47:10.060 --> 01:47:14.900 Nisten Tahiraj: I, I, I have the most fun advice, just to install TailScale everywhere and give it 01:47:14.980 --> 01:47:18.579 to your, uh, go on Instagram on your Meta Muse, 01:47:19.219 --> 01:47:24.820 and then, uh, you just talk to, uh, Muse on Instagram and just be like, hey, every 01:47:24.900 --> 01:47:29.880 day just go on every single device I have and run a security check of every NPM and 01:47:30.260 --> 01:47:32.000 and PIP, uh, 01:47:32.740 --> 01:47:34.219 needed thing, and just give me a report. 01:47:34.260 --> 01:47:34.719 I- I- 01:47:35.320 --> 01:47:36.679 I'm joking to do it with Muse. 01:47:36.980 --> 01:47:38.140 I- I do it with Opus. 01:47:38.620 --> 01:47:42.660 It- but it's a very simple prompt, just every day at 6 a.m., 01:47:43.340 --> 01:47:48.160 uh, just check every single pip, NPM, Linux package, whatever I have on my 01:47:48.260 --> 01:47:52.799 system, and give me, and also check up the, the latest, like, uh, SNIC and Socket 01:47:52.860 --> 01:47:57.179 IO news, and give me just an overall security report of what I need to do, or 01:47:57.260 --> 01:47:58.679 send me an alert if it's something bad. 01:47:59.140 --> 01:48:01.040 That's it. It just runs it every day. 01:48:01.100 --> 01:48:05.019 It's pretty simple. That's what's saved my ass almost all the time. 01:48:06.040 --> 01:48:10.220 Alex Volkov: Josh, any any other chimes into, like, what folks can realistically do to to, 01:48:10.240 --> 01:48:14.579 like, at least somewhat protect themselves in this world of increasing 01:48:15.960 --> 01:48:17.999 security vulnerabilities and and agent hacks? 01:48:19.000 --> 01:48:24.979 Josh Long: Yeah, I, uh, so generally speaking, um, the the standard security vice advice just 01:48:25.040 --> 01:48:26.659 kind of applies here, so, um. 01:48:26.680 --> 01:48:28.100 Alex Volkov: Eat your vegetables type stuff, yes. 01:48:28.240 --> 01:48:29.259 Josh Long: Yeah, yeah, yeah, yeah. 01:48:29.400 --> 01:48:33.680 So, like, obviously make sure you're using multi-factor authentication on everything, 01:48:34.200 --> 01:48:37.259 uh, you know, make sure you're keeping your operating system, your browser, and 01:48:37.320 --> 01:48:41.579 everything up to date. Um, when there are new patches issued, make sure you, uh, 01:48:41.680 --> 01:48:47.579 apply them quickly. Um, the- the one kind of caveat here, the- the irony is that, 01:48:47.760 --> 01:48:52.199 uh, you know, with some of these, uh, software supply chain attacks, right, the- 01:48:52.240 --> 01:48:57.660 the LiteLLM thing, um, uh, that's, um, one of those awkward 01:48:57.760 --> 01:49:01.360 situations where if you are following that best practice of, like, updating 01:49:01.440 --> 01:49:03.739 immediately, like, as soon as something comes out, 01:49:04.079 --> 01:49:04.219 Alex Volkov: Yeah. 01:49:04.280 --> 01:49:09.319 Josh Long: um, you could ironically end up getting compromised, um, instead because of that 01:49:09.400 --> 01:49:10.639 su- a supply chain attack. 01:49:11.040 --> 01:49:14.699 It's- it's kind of a- one of those unfortunate catch-22 situations, but 01:49:14.960 --> 01:49:19.320 generally speaking, for- for anyone listening, uh, if- if you 01:49:20.000 --> 01:49:24.939 keep everything up to date and, and update pretty quickly, um, that's the best way 01:49:25.120 --> 01:49:26.119 you can protect yourself. 01:49:26.440 --> 01:49:29.540 Alex Volkov: Let's draw a very clear distinction here between what you keep up to date and what 01:49:29.600 --> 01:49:30.799 you need to, like, wait for. 01:49:30.930 --> 01:49:33.930 supply chain attacks, they apply to package management, right, for software 01:49:33.950 --> 01:49:38.590 engineers. So if you are installing from UV or PIP or NPM or like PNP, whatever 01:49:38.730 --> 01:49:43.130 those packages, those packages can get, exposed by hacking the folks who 01:49:43.570 --> 01:49:47.040 distribute those packages and then put like, key extraction things or other 01:49:47.140 --> 01:49:48.220 things that can hurt your system. 01:49:48.630 --> 01:49:50.349 do not update those just willy-nilly. 01:49:50.430 --> 01:49:53.760 There needs to be a period of 24 hours, at least 01:49:54.600 --> 01:49:58.320 to let the folks know that they've been hacked and like claw back the, the, the 01:49:58.400 --> 01:50:00.160 LiteLLM, like exposed. 01:50:00.270 --> 01:50:01.869 this is what's called the supply chain attacks. 01:50:02.590 --> 01:50:06.729 Browsers, and we haven't seen like any supply chains and browser update, like 01:50:06.830 --> 01:50:10.370 update your browsers immediately, update your operating systems as soon as those 01:50:10.510 --> 01:50:14.089 patches come out. And for PHP and other places, you have to patch the actual 01:50:14.230 --> 01:50:16.089 patches that they released on your system. 01:50:16.150 --> 01:50:17.969 So those are like kind of like different things. 01:50:18.270 --> 01:50:21.029 Package manager supply chain attacks do not update immediately. 01:50:21.150 --> 01:50:24.789 In fact, the best practice that we know to to help you against supply chain attacks 01:50:24.890 --> 01:50:29.669 is to put a blocker to say, hey, I'm not installing any packages that are like at 01:50:29.750 --> 01:50:33.749 least 24 hours old, because hopefully in the 24 hours that passed, the, the folks 01:50:33.910 --> 01:50:38.290 who got hacked and the package was like infected, the, the package manager, you 01:50:38.310 --> 01:50:39.989 know, will take it back, right, claw it back. 01:50:40.270 --> 01:50:43.249 That, so package managers do not install the new ones, don't keep them up to date, 01:50:43.270 --> 01:50:47.829 like Josh was saying. Uh, operating systems, phones, uh, you know, browsers, 01:50:47.950 --> 01:50:48.949 those things need to be, 01:50:49.590 --> 01:50:52.769 when you see the update notification, hit that, like literally 01:50:53.710 --> 01:50:53.969 update. 01:50:54.070 --> 01:50:55.210 Josh Long: Yeah. Well, well said. 01:50:55.270 --> 01:50:59.249 That's a good distinction to make, uh, between stuff that, like, consumers are 01:50:59.310 --> 01:51:02.670 running typically versus stuff that is running behind the scenes, right? 01:51:02.750 --> 01:51:02.889 Alex Volkov: Yeah. 01:51:02.910 --> 01:51:04.989 Josh Long: So, so yeah, really good point about that. 01:51:05.310 --> 01:51:10.949 Um, eh, eh, f- for a lot of software, um, supply chain attacks, um, 01:51:11.310 --> 01:51:13.429 s- some- sometimes they're caught quickly. 01:51:13.590 --> 01:51:17.669 Sometimes they're caught, you know, within a couple of hours, like the LiteLLM thing 01:51:17.870 --> 01:51:22.970 recently. So, uh, i- in other cases, they might kind of linger a little longer than 01:51:23.030 --> 01:51:28.429 that. Um, so, uh, but- but I- I think that's that 24-hour thing, unless you know 01:51:28.530 --> 01:51:31.409 there's some vulnerability that you have to patch right now, 01:51:31.710 --> 01:51:31.849 Alex Volkov: Yeah. 01:51:31.990 --> 01:51:34.449 Josh Long: uh, that was just fixed in the latest version, right? 01:51:34.550 --> 01:51:37.570 And if something like that does happen, they're gonna make a big deal out of it, 01:51:37.630 --> 01:51:40.990 right? They'll let people know it's gonna be a well-known, like, public thing, more 01:51:41.030 --> 01:51:41.530 than likely. 01:51:41.870 --> 01:51:42.009 Alex Volkov: Yeah. 01:51:42.150 --> 01:51:45.229 Josh Long: Um, and in those cases, maybe you need to patch right away. 01:51:45.670 --> 01:51:49.489 Um, if it's not that kind of a situation where you know there's some urgent 01:51:49.630 --> 01:51:53.049 vulnerability that's got to be patched, then w- again, when we're talking about 01:51:53.110 --> 01:51:58.370 the software supply chain kind of things, um, probably a good idea to, to wait a 01:51:58.510 --> 01:52:00.150 little bit of time just to make sure. 01:52:00.670 --> 01:52:02.389 Alex Volkov: Hi, folks, we're coming to the end of the show. 01:52:02.470 --> 01:52:05.440 Josh, thank you so much for joining us and, and very clearly, helping folks and 01:52:05.700 --> 01:52:07.520 with, with, with positivity in mind as well. 01:52:07.560 --> 01:52:11.869 I think it's very important because, my, my, my comment to many folks in the group 01:52:11.890 --> 01:52:13.010 chats who are freaking the fuck out. 01:52:13.290 --> 01:52:15.459 Uh, if you are freaking out on X, yes, it's scary. 01:52:15.520 --> 01:52:18.679 Yes, there's a lot of stuff. The algorithm is poised. 01:52:19.000 --> 01:52:23.179 This is what the algorithm loves when you click on the, on the thing and you hover 01:52:23.240 --> 01:52:26.539 and you're like, oh, and even if you comment and write something in that area, 01:52:26.920 --> 01:52:30.340 the, the, the fear and gloom and FUD and like all of these things the algorithm 01:52:30.360 --> 01:52:34.539 absolutely loves. So you are, will get constantly people who are like first 01:52:34.640 --> 01:52:37.480 joining the algorithm and to get some clicks, but also like people who just 01:52:37.520 --> 01:52:38.600 comment on the same stuff, right? 01:52:38.810 --> 01:52:42.290 but you you'll be fed this fear over and over again in different forms and 01:52:42.330 --> 01:52:45.280 different types, different varieties, uh, because the algorithm loves the 01:52:45.340 --> 01:52:48.240 engagement. It does not mean the sky's falling, go out, touch some grass, and 01:52:48.280 --> 01:52:52.300 then come back and run some security stuff because it is important, but it's like the 01:52:52.380 --> 01:52:56.180 the the internet. I've seen some folks say, hey, go take out, you know, 20,000 01:52:56.220 --> 01:52:59.340 dollars in cash to put it because like the the banks are like all done. 01:53:00.180 --> 01:53:03.880 Scary, but I don't know if we're quite there yet because, yeah, the companies are 01:53:04.020 --> 01:53:05.385 patching the PHP stuff, right? 01:53:05.625 --> 01:53:10.035 but definitely there's a lot of stuff that that the can of worms has been opened, and 01:53:10.195 --> 01:53:12.155 be security mindful as well. 01:53:12.385 --> 01:53:14.244 So yeah, I think may- maybe this is the advice. 01:53:14.585 --> 01:53:19.644 We all now have access to, basically security, like, uh, experts, if we can, 01:53:19.665 --> 01:53:22.435 like, knock on Josh's door, everybody, to run stuff for us. 01:53:22.495 --> 01:53:25.875 So definitely use that. the other thing that you did bring, and it's kind of like 01:53:25.915 --> 01:53:28.375 a breaking news type situation, Peter, thank you for bringing this up. 01:53:28.625 --> 01:53:30.324 we talked about, like, agent security. 01:53:30.425 --> 01:53:34.364 1Password finally comes out with the platform for you and your agents to be 01:53:34.405 --> 01:53:37.265 able to run. They they're mentioning, all of 01:53:38.105 --> 01:53:42.444 literally everyone, bot and Anthropic and Claude and OpenAI and Muse and Browse and 01:53:42.545 --> 01:53:43.545 Town and like, et cetera. 01:53:43.735 --> 01:53:44.735 1Password for agents. 01:53:44.895 --> 01:53:47.094 Give your agent access without giving up your passwords. 01:53:47.295 --> 01:53:50.794 When an agent reaches a sign-in form and requests a credential from your vault, if 01:53:50.815 --> 01:53:53.715 you approve the request, 1Password directly auto-fills the credential in the 01:53:53.815 --> 01:53:55.884 destination I think that's incredibly important. 01:53:55.965 --> 01:53:57.684 I've been waiting for this since, 01:53:58.365 --> 01:54:02.805 you know, Muse a- a- announced this, and it's just like absolutely crazy. 01:54:02.885 --> 01:54:04.824 I said, I signed up, let me in there. 01:54:05.000 --> 01:54:07.485 so shout out to 1Password for like finally like, getting there. 01:54:07.805 --> 01:54:09.645 if you are using 1Password, join this waitlist. 01:54:09.685 --> 01:54:12.365 I think it's very, very important. It's kind of like the Stripe link for credit 01:54:12.405 --> 01:54:15.135 cards, but for passwords, and uh, I think it's very important. 01:54:15.415 --> 01:54:16.535 Again, if you are using 01:54:17.335 --> 01:54:21.795 agents at all, even if you are not using a one like password manager, 1Password I 01:54:21.854 --> 01:54:24.214 found for the past 15 years to be the most reliable for me. 01:54:25.621 --> 01:54:26.760 Josh, thank you so much for joining, folks. 01:54:26.781 --> 01:54:29.500 We're 2 and a half hours on the stream, but I think it was very, very important to 01:54:29.541 --> 01:54:33.301 at least mention some of the stuff because it does feel like the defender's window is 01:54:33.381 --> 01:54:38.121 coming to a close, and a lot of swarms, a lot of, like, bad actors with AIs are now, 01:54:38.171 --> 01:54:42.040 taking over. Like Nissen said, you, you do need to take some steps, and dedicate some 01:54:42.141 --> 01:54:44.861 time in your personal life to kind of protect yourself as well. 01:54:45.301 --> 01:54:48.880 The eat your vegetables stuff are always there freezing your credit, and and uh, 01:54:49.301 --> 01:54:52.861 specifically now I think the the the the more, um, how should I say, the 01:54:54.581 --> 01:54:58.861 the the biggest exposure for folks who are not realizing that this is exposure is API 01:54:58.981 --> 01:55:02.500 keys for LLMs. Those are directly connected to your account, especially if 01:55:02.581 --> 01:55:07.520 auto refill is there. Uh, put caps, put notifications on, uh, and then make sure 01:55:07.541 --> 01:55:12.000 that you rotate them if a website was like like Railway was exposed or or OpenCode. 01:55:13.461 --> 01:55:17.940 With that, I think we'll say thanks to folks on the stage, uh, and then we'll 01:55:18.061 --> 01:55:20.401 sign off this is it for Thursday AI for October 22nd. 01:55:20.871 --> 01:55:24.111 the biggest news that we talked about is this, the availability security stuff. 01:55:24.191 --> 01:55:25.211 We talked about, um, 01:55:26.191 --> 01:55:30.590 uh, Flint from, Recursive AI being built by the Eureka Machine, which is a re- 01:55:30.811 --> 01:55:34.411 recursive improvement machine, and is getting a very fast local model. 01:55:34.801 --> 01:55:39.141 we talked about holes in PHP and, your AI keys and, a lot more. 01:55:39.321 --> 01:55:43.080 Anthropic becoming a PBC and Anthropic stepping into the personal agent with, 01:55:43.401 --> 01:55:46.760 Claude Duet, which we got access to and we'll tell you all about next week. 01:55:46.841 --> 01:55:48.780 Hopefully by then the wait list will open. 01:55:49.081 --> 01:55:52.280 Thank you so much for everybody for tuning in, for listening into ThursdAI. 01:55:52.401 --> 01:55:55.981 Josh, thank you for joining. Peter Gosta of LDJ, Wolf from Nisten was here in Yam 01:55:56.001 --> 01:55:58.546 before, and Tim joined us, as well. 01:55:58.627 --> 01:56:00.246 Thank you all for joining ThursdAI. 01:56:00.367 --> 01:56:01.986 We'll see you here next week. Bye, everyone.