OpenAI Joins the AI Assistant Race + Serverless GPUs, LIVE from CoreWeave Fully Connected

0:00 Everybody, welcome. Thank you for tuning in to Thursday I we're in the middle of the show floor in Moscone Center and we we're just setting up the show. >> Hello everyone. I'm B from Raven the Evangelist together with Alex at Corwave >> and then we are coming to you live from the Moscone Center. As you know the Thursday I is only possible because of Cororeweave and Corweave has this biggest event of our year that's actually happening live right now. So many many folks are probably tuning in to that. but Thursday I needs to happen and so we are going to start the show with the news and then we're going to switch and have a bunch of interviews with a bunch of folks from Core. let me say hi to Niston and LJ as well. N LJ, how are you guys?

0:46 >> Hey everybody. >> Hey it's it's going to be a good day. It's been a good past seven days. A lot of announcements. We're going to talk about a lot. Folks, OpenAI joins the AI assistance race. This is the top news from this week. OpenAI joins the AI assistance race. Meta has Muse, Entropic has Cloud, but not really an assistant. and Grock has Grogbot. and this week, OpenAI stepped in and changed their offering a little bit to something called Dots. I was at OpenAI dev day as I heard some of them talk about this on stage. it is an assistant that's live 24/7 for you with its own computer and never goes to sleep and it's monitoring everything. We have to talk about the assistance era and we're definitely going to tell you about the assistance era on the news, but also dev day had a bunch of announcements that we can't wait to cover with you as well.

1:38 Not to be outdone, Entropic released a new version of Sonnet. Sonnet 5.5, a opposite smaller brother, which is opposite is Astra's sorry, opposes Fable's smaller brother. And that is a very interesting model as well, much cheaper. So we're going to definitely check in on this as well. And you definitely want to stick around with us because today we have poor folks from Core coming on the show on the second hour to talk about physical AI, to talk about sandboxing, our new offering called Forge, which is has a free tier that you can try at Corwave and and a bunch of the live demos that we're doing live on stage. I think that's going to be very very exciting.

2:30 This is the TLDDR. This is everything that big that happened in the world of AI in the on October 1st. This is October. Can you guys believe this is October? I I'm still It doesn't click for me, Peter. We're in October already. We're in the last quarter of 2026 and nobody's pacing. There's no pacing the frontier. There's no pace. Three big models this week. Just this week, folks. We talked to you about GPT6 Astra and and Soul just last week and now we have another one GPT 6.1 Soul announced during dev day. We have Cloud Sonet 5.5 and and and and Gemini Wolf from Gemini came back >> finally.

3:10 >> but not really. Peter, I think we'll definitely have to at least mention Gemini for Argon because that's well we don't we don't know much about it. We haven't been able to test it yet, but maybe some some people have. All right, so we'll start with we mentioned a topic OpenAI. OpenAI also launched a bunch of stuff during OpenAI dev day that we're going to run through, but then we're going to do a deep dive. OpenAI launched a Jeff competitor.

3:39 Yep. OpenAI decisions API. essentially a Jeff competitor, which is really funny because Jeff kind of was came up inside OpenAI. However, it is multimodal. I don't have access yet. But decision API is unlike an LLM does not talk but gives you decision powers as well. OpenI launched sign in with Chad GPT. So many many companies can now just finally let the 1.2 billion users that Chad GPT has sign in with their account and then use that token. So, if you are paying for the one of the now three tiers, right, the the the the the plus, the pro, and the extra pro $500 tier that they also announced, you'll be able to use your tokens on different apps, which I think is a very very interesting thing. Open a also launched ultraast, which is a long promised cerebrous integration where GPT6 runs at six times the speed, and we're definitely going to talk to that.

4:41 yeah, we talked about Gemini for Argon from Google. It's number one on text arena and Val's index. but it's still rolling out to truster trusted defenders first and then let's see what else. Ultraast. Anything else I'm missing in the TLDDR? Peter from Dev Day that that we want to Peter was also by the there by the way. >> Yeah. Yeah. No, it was a it was a fun day. did you mention the 61 soul?

5:12 >> 61 soul. Yes, GPT 61 soul. A new model from Open the >> I Yeah, I didn't particularly expect them to release a new model. So, that was a nice surprise. Yeah, it was a nice surprise about like developer tooling and things like that. So, yeah, ultraast mode. that was another one. >> W from Codex Cloud. You want to tell us? Yeah, codeex. >> This is the TDR, folks. Remember, we're going to dive we're going to do a deep proper deep dive into dev day, but yeah, tell us.

5:43 >> So, basically, it's a Codex harness fully in the cloud. So, you can close your laptop, turn it off, and steer from your phone, and it's still going. >> All right. so this is open air dev day. what else do we have? Oh, yeah, we have to talk about this. six AI CEOs signed the voluntary White House accord intelligence. Super intelligence is on the menu, boys. Super is on the menu. You see Dario and Trump giving like like a high five to Dario. So Dario Amade, Elon Musk, Sam Alman was a dev day. By the way, this happened during dev day, but Greg Bman was there. who else?

6:19 I think there were a few more folks there. Oh, Mark Zuckerberg, of course. and then and Jensen also was there. So they all signed the AI at the White House voluntary accord at Super Intelligence. and they also leaked the entropic IPO numbers. We're going to tell you about the entropic IPO numbers as well. and also in kind of the industry and policy thing, you guys clock that. AMD has purchased Dr. Fe Lee's World Labs for what? 8 two Yeah, $4.6 billion.

6:56 no, that's tropics IPO. I'm sorry. Yeah, $8.2 billion. So AMD to acquire World Labs from Feli Dr. Feli which we told you about as it launched what two and a half years ago and we got very excited about their models Dr. Fililli to become the AMD EVP and chief scientist as quite aquire/acquisition as well. all right. So, this this week in this week's bus, the corner where we tell you about everything that has happening in Coriv, we have a bunch of stuff to tell you about. So, I'm going to run through them super quick.

7:31 Cory launches Forge. It's everything that you need to take your agent from good to great. we have a tools to offer you across every step of the agentic loop. as you run your models and agents as you you observe them you you curate the best outcomes you improve and then you run evaluation and you run them again one aentic loop we're going to have a bunch of talks to tell you about this agentic loop here on the podcast very very soon forge is has a few plans it's free and the pro starts from $60 a month with a 30-day trial and all the tools including weights and biases are included there folks weights and biases not going anywhere and biases the weights and biases models product that many people use to train their models is is light and kick inside core with forge. We're going to have core with VPs to talk to us about this >> and forge is fully connecting the entire pro side.

8:24 >> Nice. Very good. Forge is fully connecting the AI. all right. Oh, also super cool. I don't know if you guys clocked this. Cognition now powers their sweet two models on top of core wave infrastructure and NVL72 Vera Rubin CPU GPU kind of like architecture that we're going to hear about in the next part of the show and Cognition are seeing almost 5x increase in the throughput of suite 2 their own house model. So, how that's going to go and to tell you all about this later, we're going to have Cory Sanders EVP of product at Corwe to talk to you about this.

9:06 Vera Rubin, I think they're they're our first customer and I think we're like one of the first folks into production with Vera Rubin and the whole the whole setup which is quite incredible. we're also going to hear about sandboxes. So, that's exciting. >> Let's move forward in open source. Not tons of stuff happening in open source as far as I've seen. H Company open sources Holo 4 computer use. we keep talking about H company. I've never used one of their models, but shout out them for them to computer use Holo 4. And Nvidia releases the open agent safety platform, which is an open shell part of the new platform to watch and control what agents do with a hardware watchdog that can quarantine the rogue agent, which sounds like Open AI could have used this while during the swarm thing. Wol.

9:52 >> they definitely should have something like this. >> Yeah, I will I I will have to say though I was busy with dev day in like preparation for this non LDJ. Anything big happened in open source that that folks need to know about N besides you blowing up on hugging face? >> There's a couple of smaller interesting things like I believe it's the Alice model that they ended up announcing. It's it's basically the Google of Russia that released it. It's called Yandex.

10:18 >> Oh, Yandex released the model. But at least >> yeah and it's called Alice. It's about 35 billion parameters total. And what's interesting it's really it's really high sparity only about 600 million parameters active. >> Wow that's super cool. Yandex are no slouches folks. Nebios came out of Yandex if I'm not mistaken and I think that they also have like a smart device in homes. Okay that's super cool. All right folks we're still in the TRDR in the Jev agents agentic tooling and everything. We have respon claims 2x cheaper than jev. So it hasn't been what two weeks since types released jev and we talked about this on the show and by the way they came to the cor hackathon two weeks ago and the cor hackathon participants got early access to jev. span01 is is 2x cheaper than jev and runs on open router and span light is free. I don't know quite about benchmarking. So we had Florian talk about Jeffbench.

11:17 so a very interesting resource for you to check out if you are into decision models and type one models. There's a lot happening there because also openi released their competitor called decisions API. Release is a big word. They put it on the weight list. so we're weight listed there. There also a new tool that I saw called Jeff Gp. It's a little CLI powered by Jev claims to make coding agents 40% cheaper. And Cloudflare released a bunch of stuff.

11:40 They keep releasing. This is Cloudflare's birthday week. Shout out to Cloudflare. I absolutely love them. if if anybody from Cloudflare is watching then listen I think I think we could hear you but hold on let me just lend the thought clouds is doing the thing that I'm going to tell you all to do as well all products in the world now are going to be agentic agents are going to use products more than humans because for each human I don't think that this panel is quite there yet but this guy and this guy are definitely there most of the products will be used by by our assistants And so cloud rebuilt their whole cloud API on top of the CF API. Everything that you can do as a human clicking buttons in the dashboard your agent can do with the CF API and the right tools. I think that's the absolutely right approach for the world and as the world is moving to the assistance folks again Grabbot has Elon Musk from SpaceXi has Gagbot Meta released Muse and is pushing this to 1.2 billion people very soon and then OpenAI also has 1.2 two billion weekly active user which they announced just now and they are pushing dots which we're going to tell you about all in the next segment.

12:50 speaking of Jeff I also have one more because liquid launched D1 its first decision model and it liquid says it's the first model to be Jeff on hugging faces decision index with stronger multilingual prompt injection and long input results. It's coming to liquids API and open router soon as well. >> All righty. Okay. Okay, so folks that's open source tools and now in voice and vision Hunen released Hunen video. I don't know if you guys saw this. Hun video is a minimax H3 video model and it's built at 1 cent a second. 1 cent a second.

13:29 Just a penny through October. I think that's really cool. I think that's all that we have in the TLDDR for you and I think it's time to >> one more project because we are talking about agents and personal agents assistance actually. So Nilo is an open-sourced MIT licensed self-hosted multi-user agent workspace for you and not just you. It's not single user as many agents are. It's for the whole family corporation. >> All right. That's super cool. all right, Nissan, quick check. Why are you blowing up on hugging face? And then we'll move on to dev day.

14:03 >> yeah, I was just number two in data sets and is number six. >> You released the data set that's number two trending on hanging face. >> number two on data sets and number six on overall things with models and stuff. And also shout out to Seek from Twitter for holding the lead there for like three weeks. he released the entire 17 terabyte arc. I released a 70 megabyte synthetic opus 5.5 generated data set that just played every it played a gentically doctor patient and then it tried to offiscate one thing from the patient like either a lie or something they forgot and then it had to gently correct it. So the data set just made the models to train better at yeah and then I looped this over 2,200 diseases. So I just set up an agent loop and just ran 100 agents at a time and that's how that was generated and then was filtered three times.

14:57 yeah, it's really good for building like small medical robot rags. So you can build an agentic rag with it and you get like way higher performance from u a medical model because now it has an entire source of truth. >> Yeah. Like yeah from from Obus 5.5. So it's going to be pretty accurate. The only downside was that the conversation was a bit too technical because it tried to include every single thing possible. so that was it. But it's meant to be used by a friendly chatbot.

15:26 >> Listen, I think that this is this is the biggest open source news from all the news that that we just played. one of us blowing up on hugging face and folks can use the data set. I >> Okay, there was one more news. One more time but quickly because we have so many >> SG lang announced like native support for turning any model into a je and they already released set for for doing quen. Yes, this is insane. You'll be able to use that with any model because you just take the encoder for it from it and then you you use their classifier code. Yeah.

16:02 Any model on on SG lang. Right now it only supports the quen ones. >> Yeah. >> But yeah, this is pretty crazy I think. Wow. Yeah, that's that's insane. I think that folks like types safe kicked off a hornets nest as as noticed by openi. >> We last thing we democratize Jeff guys. >> Yes, >> it's been democratized af. >> I think that Jeff shouldn't be over network. I think that Jeff should just run on whatever CPU do you have. And I think you know maybe we'll get Jeff cores on on the iPhone in like five years. The you know they have neural course for everything. Five months.

16:38 no. Apple takes a long time. All right, folks. I think it's time for us to move to Open EI dev day. There's a lot of announcements. It's very exciting. I was there. Peter was there. actually, I wish I had this video to play, but I I asked Sam Alman's a bunch of questions, including should they slow down or have they considered slowing down? they said no. Okay, let's let's move to the Frontier Labs. Dev Day coverage starts right now.

17:18 All right. folks, this week was the fourth annual open the eyes developer conference which essentially is there you WWDC from Apple or Meta Connect from Meta which happened just a week prior and OpenI went all in with 20 releases. This was the biggest dev day ever down at the Fort Mason. Beautiful location in in San Francisco and I was there. Peter and I were there as well. Peter, thoughts on dev day. Super quick before we go dot dot dot dot before we go announcements and dot dot dot is a funny pun.

17:57 thoughts on dev day. How did this year go? >> You know what? Whatever you think of Open AI, they can put up a good party like it's a >> they can put on the show. Yeah. >> Yeah. It's a it's such a nice event and just it's a very happy event. Lots of people, you know, good food, very well organized. you know, it's like almost announcements. You guys who haven't been there, you probably know much more about the announcements cuz you know, we just hear it on stage and then we go to the next room and so on.

18:30 >> So, so yeah, they announced a lot of good stuff. Who knows how much of it will be important that will stick around because I think previous dev days not sure stuff was that important but I think it's just a good place and I wish more companies would do it right where you just have a big event where people come together kind of celebrate the company or what they've done and and just have a nice time >> and they also they're very good at at letting openi employees actually talk to developers. I think this is what dev day is all about. And one thing to mention before we go to the news folks that if you if you have a chance to sign up for next dev day, you probably should. Sam Alman really takes it out of his way to go to dev day. He didn't go to meet with Trump and all the other CEOs. He sent Greg Bman. Sam was there answering our questions. Sam Alman was like shaking head with people. He was at the afterparty. It it's really really something. Okay. So, besides the the kind of the vibes, which were immaculate, besides the swag, which was immaculate, Peter, we got the this monochrome Game Boy emulator, we got a bunch of stuff. besides all that, we also had a bunch of things announced.

19:35 And I think the number one thing that's very important is that folks open AI is stepping into the assistant era, the AI assistant era, which is could be very funny because the original system prompt for chat GPT was you're an AI assistant. You guys remember that? But it wasn't really agentic assistanty. It wasn't prompting. It wasn't proactive. It didn't like notify you. So, OpenAI looked at their competitors, looked at Peter Steinberger from Open Claw that they've hired like eight months ago and started building something. That something is now unveiled. It's called Dots. A dot is an always on agent in Chad GPT powered by GPT Astra. Not sure quite which model. Maybe the latest maybe the latest Astra, but it's powered by Astra. That's what they said. it has its own computer and it's it own browser in the cloud and it's connected to over 4,000 connector applications.

20:30 The stuff that folks have built for Chad GPT, your dot can now connect to it works 24/7. and you can talk to it, you can call it. there's like a interface there that looks like a phone call that you can call it. I think that's going to be great for my mom. you can also add it to Slack and Teams and by voice. and you said the boundaries what it can and cannot do, what it asks first, etc. Sam Alman literally said that he uses dots to run open AI for him and there was a very interesting thing and during dev day Peter probably you caught it that Roman wet a guest on our pod by the way shout out to Roman one of the best de folks in the biz did the live demo and the live demo didn't work the voice didn't work and apparently this was because there was some deployment happening however what happened was that Tibo the guy the reset guy his dot alerted him like hey you're about to do the demo and the demo may not work. So, OpenAI is really big into like the personal assistant stuff and the teammates and their dots now live in the OpenAI slack.

21:32 Wol from Let's go to you because out of everyone that we know, you're one of the more assistant PE people on the planet. You and Amy are a known duo in the human and AI world. What do you think about dot? It's great that OpenAI is stepping in and I love that OpenAI is using it for actual productive stuff instead of just a demo where you show how it does some mails or something that Zam Alman himself is using it to run the company that's a good sign and I think more should be doing it like that.

22:05 >> Yep. folks have you had a chance to try Peter your your thoughts? we we've talked with Peter before a little bit about assistance. I know you're not like big on them, but thoughts about do that as a product. >> Yeah. Well, first thing to say, Alex, I think you really called it in the last couple of Thursday eyes. Like you you kept bringing this up and making a big deal out of it. So, yeah, I think a very good call. So, yeah, I think everyone's going into this. I I feel bad because I haven't tried all of the others really.

22:35 I don't have access to Muse yet, so I I can't speak in a more complete way. But when when I was on boarding into dots and I have tried it for a little bit, I was a little bit underwhelmed at first because you kind of connect and it's like then what it's like okay I guess it's a chat. >> but actually I think the on boarding could could be better but once you just get past that point which which is a little bit boring >> it is quite nice for for two reasons.

23:03 one is that you are actually speaking to to Astra. So it is a smart nice talk to talk to model and it's very good at coordinating different threads. So that's one thing I was worried about is that you get to speak to one dot. It's not like I think in Grog both right you you can speak to many different ones or at least you can see them there. It's just you just speak to one and that's it. they announced maybe they're gonna have more but for now it's just the one and I thought that might not work but it is actually really good so I did like that yeah and and >> the very interesting thing is that u let me just sorry to interrupt that's how muse works muse is just one muse and then you have chats grockbot specifically has multiple bots for everything but grogbot this week released the the chief of staff pattern like main main Grogbot >> and so I think that we're seeing this shape evolve in real time as people play with it. Do you want one like Amy that you talk to like me and Wolf from just one physical person that I talk to or do you want me and Wolf from that we have WhatsApp for discoms and we can go to Slack to like works, right? You want different methods of of communication.

24:12 What do you think? >> I think that's what a sub agents are for. So I want one executive assistant I'm talking to and a lot of employees that work for us and that is a paradigm I've been using and I'm super happy with this. You need a lot of channels and threads of course to separate the stuff not just one conversation >> but then it works. >> I wish we had all the time in the world to tell you about that but folks once open launches something many many people will start using this okay they announced 1.2 two billion weekly active users of Chad GPT. Many of them are free probably. the difference between the different offerings is very interesting.

24:46 Meta is offering theirs for free. Open EI is only allowing this on the pro and including the $100 plan. So like it's it's $100 limited, but they will release this to everyone. It's very very obvious. So I think it's time for us to move to the model GPT 6.1 soul. Just what a week a week and a half after after Soul 6, we get a new model. and 5 days later it's already replaced. It's the same price, $2 in and then $10 out, but cash input is now 95% cheaper and and 10 cents a million tokens and open eyes pitching that this is a near Astro Intelligence with a fifth of the price. NJ, have you had a chance to try out 61 6.1 soul?

25:32 I've been trying it out a little bit and it's >> I'd say the it has less of the big model smell. >> Yeah, big >> Yeah, I do like that they are focusing a lot on the speed and efficiency of it in terms of the amount of tokens and cost it is to use. I don't know if you saw some of the metrics for Sonnet 5.5 and Opus 5.5, which I'm sure we'll go over later, but those seem to be using a lot of tokens and and at some points in time being even comparable in in cost or amount of tokens used compared to something like Fable while Soul seems to be truly just significantly cheaper in almost every case.

26:12 >> Yeah. So on deep sui 6.1 soul actually beats Astra at one six of the cost defaulting codecs and folks are calling this the the workhorse. Peter thoughts and arena performance for GPT 6.1. Yeah, I've been I've been testing quite a bit actually and I I do like it and I'm very impressed with how cheap they made it and yeah, which kind of is a little bit confusing, I would say, in terms of the the the how they priced and clearly they're feeling the competitive pressure, right?

26:43 Oppos 55. I we I don't think Oh, no. We actually do have a score for code arena. So, it's kind of measuring front end. So, it's a little bit lower than Astra, but not not that much lower. So it's fourth. so which makes it above Fable actually which is kind of it it's not so it's really >> it's not well yeah but very well priced model. Yeah. >> Yeah. >> Yeah. Go ahead. >> Yeah. You using it dayto-day I would I wouldn't quite say it has doesn't have like the big model smell. It has bigger model smell than six I would say.

27:20 >> Mhm. But it it does tend have tendency to kind of I don't know it feels a bit more artistic like it kind of goes to its hole for a bit and just then comes out and says oh here it is. So an Astra is a bit more kind of a bit more higher EQ I would say but in terms of the quality of the output I mean it's a no-brainer now to just use like to use the lower tier models as your daily driver. So for open I just use 61 now for anthropic. Yeah, it's kind of I I haven't quite decided as a topos or sonet, but you can probably kind of you need to find which one, but certainly it doesn't really make sense to use Astro or Fable as your daily driver. Now, >> I will just want to repeat this directly to the camera. Folks, Peter Goto from Marina AI is I think summarizing the vibes of the whole timeline. It does not make sense to use Fable or right now.

28:14 This is where we are. Maybe this is pacing the frontier. Maybe the companies are now releasing such good and cheap models that we won't want to like always raise to the best best best intelligence. Peter, we talked about this that I I think OPUS 5.5 is enough. I think the GPT soul 6.1 is enough. I don't want to like reach for Astra and also I don't want to spend all of my token budgets on those big models, but like for day-to-day for weekly like daily drivers, this is enough. For bigger problems, Navier Stokes level problems, maybe you want to reach for the bigger models. specifically though, the reason why you would maybe not use Astra is because it's very expensive. But if you are a billionaire, there's another option for you. they launched the ultra fast and pro-500 tier and Open EI. OpenI launches the Pro 500 tier. You pay them $500 and you get access to ultraast. Ultraast mode is their models running on Cerebraus inference and shout out to Cerebras for this speed and is up to eight times as fast in codeex around 300 tokens per second for Astra I believe and six times the price and it comes with a new $500 tier. Sam said on stage it's so fast he never goes back anymore.

29:28 Obviously you want speed and that's why we got excited by Jeff, right? That's why we we're all like okay I can do this intelligent with Jeff. So like now you can do this with speed and the quietly the $200 pro plan is back but half the usage. So f you to whoever in openi decided to cut my usage in half which it was quite awful even before and we'll see after this show if this succeeds if the folks at core will allow me to get to the $500 plan but yeah it's quite expensive but ultra fast is crazy.

29:58 I think we're skipping Codex clouds and oh Codex is also running in the cloud. Another thing to call Peter that we we we did talk to you about on the show, but here's the the actual calling. Local environments make no sense. Local environments, local Macs make no sense. The more I work with Fable and the more I work with codeex on my machine, the less memory I have, Peter, you talked about this with your Linux machine as well. everybody's moving to the cloud. Most of what my machine can do, they can do on the cloud. All they need is like my login. That's that that's what that is for. And they're also announcing Codex cloud. So you you'd be able to move your stuff in Codex to the cloud. I think that's very very important decisions API. Maybe the last thing that we'll talk here folks that decisions API from OpenI. This is their new dev competitor Jev Killer. Not public yet. We cannot benchmark yet.

30:52 but exciting. Why is exciting? >> For me it's exciting because I'm using OPMI already. I have the subscriptions and I use them. So if I was to use Jeff, I would have to send the data relevant to the decisions to another company and that is something maybe I don't want to do. So I can concentrate on where my data is going and if I'm at my just use my eye. >> Peter decision API thoughts on this at all?

31:20 >> Yeah, I mean this place is exploding again. You guys called it right. It's a like it was very exciting and we get so much new stuff and yeah I agree with with Wolf from I think that's it it was one thing that is kind of exciting but then oh it's another company that I I need to send my data to but also the important part is that because it's so cheap and you want to classify stuff then you kind of have to send all of your data right it's like so much data then it's like oh man I don't know what to do with this so yeah the more you can reduce your footprint the better. so that makes it high utility. I think what's interesting about it as well is that it is based on Luna. so that that would be interesting. Yeah. Multimodality you guys mentioned that's that's >> multimodal. I'm so excited about it.

32:08 >> Yeah. So that that is interesting. So yeah we we we don't know that much about it yet. but that certainly I and I'm I am curious like do do you guys think they just reacted and shipped something super quickly after Jeb or did >> I wanted to ask Sam Alman this live at the Q&A that we had with Sam Alman and I just decided I I did ask him about this because on latent space Swix interviewed Diego and Diego said that he came with this idea to Sam Alman two and a half years ago and Sam Alman said it's a crazy idea you should work on it and Diego has been working on RLHF and once he finished up. He went and did type safe and they say it's like two years in the making and now we're seeing just like decision models in open source and everyone just popping like mushrooms of the rain. So I really really wonder if if if this is a reaction or you know somebody open I just waited on do yoga to release and then said all right you know they proved the category exists you can only like shoot so far in so many directions. we have to move on. The last thing I will say, two things.

33:11 There's a huge platform play for Open EAI, 1.2 billion people is significantly more than Apple had when they launched the app store. OpenI relaunches for the first time the app store. For the fourth time, OpenAI relaunches the app store during dev day. Folks, we've been there since dev day one with GPT GPTs and custom GPTs. Then the next year it was plugins. Then the next year it was the the M the OpenAI chat GPT app store. And now we're back to plugins. plugins now though are powered by GPT sorry power GPT plugins are now powered by MCP UI and MCP apps shout out to our friends on the pod that that created this category and Ido and it's it's really something you can build whole applications inside now including a side pane etc and it's a huge power play because again the huge platform play because again 1.2 two billion people.

34:07 and then sign in with Chad GPT. That's the next kind of thing. You guys remember signing with Facebook? This was like all over the news, all over the internet and now nobody signs into Facebook. We're all sign in with Google. Maybe X. Now sign in with GPT is there. The benefit though besides accessing your data like sign in with Facebook used to give you, you can now sign in with your tokens. So if you are paying $5, $500 a month, you can now spend those tokens elsewhere. So you don't only have to rely on codex. I don't know if I will ever do that, but that's something to keep in mind on. Sam Alman posted this this morning that the they have underestimated how many people would want to do something like that. So sign in with JPT and the the plugins platform play I think is very important. Folks, let's move on.

34:54 >> I have something we could quickly throw in because Google is also back in the personal assistant or family assistant land case race with CC. If you remember Gemini Spark, that was also a personal agent that wasn't really used that much, I think. Yeah. But now they integrate Gmail and Google Calendar, Google Drive and >> I have so much to say about this that it's unfair that you bring this up when we have limited time warm because folks at Google at Google IO they announced Spark which is their assistant. It is awful. I am sorry if you're working on this product and you feel that Alex is attacking you folks. It is awful. I give you all the feedback. Spark should be the best AI assistant. It's connected to Gmail to drive to all the Google properties. It should be the best. What I use most of my assistants for is specifically accessing my email. Why is Google not building the best assistant that access my email? I don't understand. but and CC is kind of like an additional thing. So they're trying multiple things. None of them is nearly close to like you know Dots or Muse or Grockpot or Instinct or or or or there's a lot of assistance. the so much so we need the new show. We're moving to Sonnet though. We're moving to Sonnet four 5.5.

36:05 Entropic did not want to give OpenAI a lot of time to rest on their laurels with Devday and take over the the airwaves. Entropic decided to release Sonet 5.5. let's see. Where's our Son? I think all right. Antropic announces Sonet 5.5. Faster, cheaper, and beats OPU on thermal bench. What? folks, reactions to Sonet 5.5. Oppus 5.5 is my daily driver and it's great. Should I switch to OP to Sonnet is the question.

36:37 Peter, I see you like gir to go to answer this. Let let us know how you feel. >> Yeah, I I sort of tried it a little bit and it I don't really know what to make of these models because they produce outputs to the unbelievable way. If we had access to these models like six to eight months ago, I think we'd be losing our minds. And so I I don't really know what to make of it and I think anthropic models all just all seem to be jumping over each other. So they all kind of crushed up towards the frontier. I think it's hard to say like we is it opposite you should do use daily. I think opposite does feel a little bit smarter but maybe sonet is like okay to use. So yeah but fable now almost makes doesn't quite make sense to use. I I don't know.

37:27 It's difficult but yeah but Alex you still use Opus right? >> I honestly haven't had time to even look at sonet since it was released because of dev day and planning the show and and the show but I am very excited for speed specifically the whole thing about sonet for me on top of oppus was it's faster and sameish intelligence. So stuff like using it for sub agent for oppus when oppos can execute I don't need to burn my context window on like like trivial things like fetching something over the web right.

37:55 So if I do need the smart model to decide which button to click, which things to pull from the web page and so I I trusted Sonnet for that. honestly though, here's my here's my call even without trying this folks and I know that that's not maybe what we expect from the show. You expect us to try everything. Here's my call on set. Sonnet made sense when OPUS was expensive. Given that now on the $200 tier in Entropic, Oppus is nearly infinite. I I'm not able to to get to the end of my limits, I don't see a reason to use Sonic. Like if it's even a little worse, a little faster, opposite quite fast. I don't see a reason to use it. so yeah, it's a very interesting model. I think for for many people who use it via API, I think it matters, right? So like in via API, you do want to like cut costs. You're not like subsidized by by Entropic giving you like very very cheap tokens. I think it's like 5,000 worth of tokens in API on the $200 plan.

38:55 Something like that. Like crazy. So, Entropic is is really trying to buy us. >> don't give them ideas, Alex, that they do the same. >> You still haven't gotten back to the Entropic thing. You you still you still you're still tempted. NJ, have you used Sonnet? Any any thoughts? Now, I have not really used Sonet 5.5, but I have looked at a lot of the different benchmarks, and my conclusion right now is that I think it's going to be the new best free model that I'd recommend to to friends and family that don't want to get any paid plan.

39:25 >> When you say free, what do you mean they put it up as a free tier? >> Yeah, it's it's available on the free tier right now. Sonet 5.5, which the Sonet models have always been available on on the free tier. >> Yes. Yes, sir. and I checked it looks like it's available at least up to extra high of reasoning mode on the free tier. >> And, maybe it is only like two prompts per five hours that you're able to get, you know, with the usage limits.

39:48 >> but it's significantly better than Luna Max, though. And I think on the free tier for Chat GBT right now, you don't even get Luna Max. You get like maybe Luna medium reasoning or Luna High, something like that. >> All right. So folks, we're also going to ask the the listeners, folks who tune in, have you played with Sonnet 5.5? Have you played with GPT Soul 6.1? Please let us know what you think. Go ahead and listen.

40:13 >> it's my default model for agentic tasks. >> Sonnet is for the best. It it talks a lot nicer. So if you need to like talk to your agent and just look stuff up, it talks a lot nicer. I will never use it for coding. There's nothing that beats Opus 5.5 on a large codebase. JPD Soul does do better designs and 3D stuff, but anything else it's pretty dumb, especially like DevOps and stuff. yeah, so but Sonic, it just talks nicer. It's way more concise. It's way more natural. So yeah, I actually use it to talk to the other agents and and summarize me stuff. It's also the default Sonnet Medium is a default on my phone. It it's just way faster and it's just been perfect. not not only way faster like 30% faster than previous Sonet >> which is which is quite something.

41:05 >> before before we move on and I think we'll skip Gemini 4 Argon because folks there is an announcement about Gemini but we can't use it yet so what's the point we'll tell you about that model once we can use it but Google seems to be at least back >> I can tell you >> Peter okay so so just before Gemini Argon okay so we'll have a few minutes on Gemini Argon I I just want to tell you that it does seem like all these frontier labs cracked RSI because the speed with which new models are getting releasing is getting me whiplash And I'm doing this professionally for three and a half years. It's giving me whiplash. Like we only had like le latest model before and we're sitting there and Samman talking oh yeah two models forward you guys won't believe what the is coming. I was like what? so it does seem like all of them cracked something like these models are training the next couple of models. So very interesting and how to do this for yourself is coming up with corwe and our loop. So we're going to mention this in in in a few minutes here. Peter >> go ahead.

42:02 Yeah. So, Google has a new model almost out. I think they're going through the process of probably that I I'm guessing Anthropic and OpenAI have gone through this. So, they're probably a bit quicker and I think Google seems to be going this for the first time. So, hopefully it's going to be out very quickly. >> or like >> Gemini for Argon Google's new frontier models for trusted testers only. We have such a trusted tester here with us.

42:28 have you an arena played with this model? What do you think? >> We we have Arena's coin. It's it's interesting combination. So, you would have seen like the benchmarks they shared. They're all like beating everything and and so on. as you can imagine, you know, lots like to select the ones that they they want to share >> on on agent mode. It's eighth. So, it's below. So a eighth but it kind of includes bunch of different reasoning levels but it's basically below fable below below Astra below soul but it is above for example like any of the other laps it's above muse it's above Kimmy and so it's basically like the third lap right now but what is interesting which I think gives it another dimension is that it's first on the tax arena which is basically kind of you can think of it as chart and the general you know I know knowledge understanding >> so that and I can't speak to the complete view of everything it can do but my sense my prediction is when you guys get access you might be like h you know I'm not sure it's like it's not as good at coding whatever it might be and I've I've done a few tests like that and it wasn't like super impressive it was fine >> but my guess is that outside of these kinds of workflows it could be really good so I think don't don't dismiss Google because maybe oh it's not so good on your coding task which I think is important but there's like a whole world out there and I think they they are also seem to be better maybe will probably underestimate it because >> I want to I want to point into the whole world out there wolf from your whole house is like based on Android and Google assistants and stuff what do you make of this >> that's the thing Google has the distribution reach if they make an agent or if they make a new model they just push it out and so many millions of people are suddenly users without having to download an app or install something.

44:29 So that is it gets in the Chrome browser as well and the Google search. So they have a really big reach and any new model and intelligence increase they have increases the intelligence for the whole world in a way. >> Yep. all right. so do not sleep on Google but Google has been sleep at the wheel at least in the assistance era. So I am very much expecting the Gemini for Argon hopefully will power up the the wave of AI assistance. it's going to be very hot very hot >> and even if it's not that a great coding model if it's a good agentic model and it can control coding agents that is enough I mean models can specialize now so we just need for our personal agents a good model for this and it can use the best model for whatever use case we have. benchmarks are hard now because of the use cases that people use are like Peter like you're saying yeah there's like the coding benchmarks and and the standard things and the verifiable things but I think the next billion people will use AI in a different way and that different way means something of hey I talked to this model for a month and whether or not it remembers the stuff that I told it before but not because of long context because it's memory management and the number of tasks it the number of steps it does to make give me a task and a simple test like booking my hair salon that that is the number of stuff that like is is matters matters to people and I don't think that we need significantly more intelligent I think we need different breakthroughs Jeff is such a breakthrough other memory systems as well and good luck to Google trying to to catch up to everybody else in the assistance race we're skipping industry policy all I have to tell you there is if you haven't seen this yet you should take a look Donald Trump invited every AI ICO possible. Sundar Pichai, Daria Amade, Zak, Greg Bachmann, Ela Musk, and Jensen. Sam Alman was at Devday, which I thought was a huge testament of where Sam Alman places the importance of of the things. However, they all signed up this thing. It's called the super intelligence pact. here's what basically they're signing. No internal monitoring, external auditor, board oversight, but no penalties and no regulator. Nissan, you talked about this on the show last week that if you hack, if your agents hack a hospital, you go to jail. right now in White House, they all signed that non no jail time for anyone if their agents hack. And also we've seen I think last week that there was like 10,000 incidents, not three, not four, not hang in face no only. there's like 10,000 incidents of the swarm attacking outside of their infrastructure.

47:10 >> it's it it it's going to happen and I think the end result is is a good thing. Pretty much you can guarantee that every app with a back door, every piece of spyware or every piece of badly written software is just going to turn into a messaging board for custom encrypted agents just talking to each other in jail. >> Yeah. >> So, the internet is just going to fill up with those. is going to fill up the small models and there's nothing you can do about that because unless you just cut off the internet entirely. So yeah, so we're going to get to the point where we all just need to have proper PGP and proper AL for the apps and no back doors and that's the only logical way. So so I think overall it's it's a good thing. Let it happen. Just let the AI worms into the ecosystem. They play a good role. That's my >> Do you guys want to talk about Dr.

48:04 Fale Lee becoming the chief scientist of AMD and selling World Labs for 8.2 billion? Peter, we talked about this a little bit over dinner. Why did AMD buy World Labs? Anybody have thoughts? folks, any thoughts on AMD and and World Labs? World Labs is obviously milking making world models which something we will cover in the next part of the show. very important for robotics etc. But other than that folks any any thoughts on on this deal at all?

48:34 >> I think robotics is the thing that it is now the new upcoming stuff and having a vault model to train model train robots that is super important and AMD is in the hardware space and they want to have something like that as well. It makes sense. >> Yeah. LJ, you had a comment as well. Yeah, I think they're they're tryinging to integrate more research internally like Nvidia has been doing. And you know, Nvidia has been releasing the Neotron models. They've been doing a bunch of cool stuff across really all areas of AI. And I think they're trying to get more of that talent in their company, too. And then that will allow them to better develop their next generation of chips for AI and everything. But also, it's interesting because this comes just within a pretty short period after they had acquired Talis, the the custom AI chip company that basically burns the model into the chip. Yeah.

49:26 >> And I don't know, maybe some interesting things there. Maybe it'll be really cool if they end up having a World Labs model basically burned into the chip. And I don't know, maybe that's a stretch, but who knows? >> we'll see. Peter, any thoughts on this? Yeah, I think I think it's such an uncertain time, right? So what happens when anthropic open going to have super models that can like build their own chips and all of that and I think I look who knows maybe they've got some very specific idea but I think just to have great people and clearly you know lab is must be exceptional cons considering you know the pedigree and so yeah I my guess if nothing else just to have great people who can help you build the next chip. I think that that must be worth a lot. Yeah. And I think it's all all stock deal, right? So it's not like they're getting cash out the door either.

50:24 >> I mean, they're getting Dr. Fale as a chief scientist. I think that that's like a huge huge huge thing. We've seen these equi hires happen with Nom Shazir into Google and then Nom left and and then we've seen this with Muhammad, what's his name? Mustafa Sullean and and those folks into Microsoft that created Microsoft AI. We've seen this like across all all over. So echo hire is at least part of it, right? Like at least like you know Dr. Fililli is like super wellknown.

50:52 She's like the grandmother of all like AI and Andre Karpathi studied after like under under Dr. Fili. So I think this is a huge huge get for for AMD and shout out to the world lab folks for this absolutely incredible incredible deal. All right, I think this is the news quick hits. We have time for quick hits. I don't think so. We have time to tell you that Hen releases a fine tune of H3 Max, the video model. we also have I think that's pretty pretty much it.

51:24 anything else huge we didn't cover in the news, >> guys? AMD on September 21st, they hit $1 trillion in market value. So, you could just go shopping for Taiwanese New Year now. Just I would have done the same. I would bought a few startups or like why not? >> Yeah. Yeah. Yeah. For sure. For sure. all right, folks. I think it's time for us to move on. This is the news for you. This is I I really don't think two hours is enough.

51:58 But I don't know how how many people would love to listen for more than two hours of folks. Like I would love to keep talking about the stuff dev day. We'll definitely keep covering this. But as I told Sam Alman, we're doing this professionally and we don't have time. So I'm sure that everybody who's listening is getting just tons of signals which assistant to use etc. So in the next two minutes what I will suggest is for us to slow down from the announcements of the updates and everything and just talk about use cases. All right. Then we're going to talk about fully connected and folks from Cor going to come and we're going to tell you about how to build. So if you're an engineer, please stick with us. I think it's going to be very very interesting. However, the world has changed. Peter said, "Alex, you called it." I'm not the first one who called it, but we all been like obsessed with OpenClaw and Hermes agents and then Grockbot and then Muse and now Dots and probably whatever Google is going to release after Spark because Spark is dog excuse me, language.

52:50 We yesterday sat at a team dinner and at our team dinner, it was maybe 12 people around the table and I asked everybody to raise their hand if they have a personal AI assistant helping them. me, Wolffram and one other intern raised our hands and that intern uses instinct by the way which is blowing up in popularity and I called that this is this thing will change in the next three months and in order to kind of make the show palable for people let's talk about how we use our assistance wolf from one example from you of how I know you're like very assistant people but like one example of people can connect with give us one example >> I use my agent for everything but >> everything you guys don't understand Wolf goes into a restaurant Okay. Wolam goes into a restaurant and he points his phone on the menu and he's like, "What should I eat?" And the assistant has a preferences, dietary preferences. Like it it is quite something to see where we're all going. Basically, I I'm not quite at the take picture of the menu level, but I'm using it this week for everything.

53:52 >> This is basic stuff. I mean, my agent should be proactive enough to know my geo location and pick up the that I'm heading for a restaurant now. have to eat and just even tell me the restaurant to go to. Anyway, just an example is when I got here to fully connect it. I traveled here so I had my agent pick the plane, pick everything for me, check the hotel and prepare this and put it on my meta classes as well. So my agent created an app so I can talk to my agent, my personal assistant Amy on my classes and I was walking through the airport and she was telling me go there now you have to do this and the productivity can even find out if the gate changes and tell you before you sit at the wrong gate which happened earlier times. Yeah.

54:32 >> So just one use case making life easier. I I already called out the dots use case that Tibo from OpenAI got his dot to ping him and say, "Hey, do you know that production is down and you won't be able to do your demo?" so that's one use case from OpenAI. My personal use case, I'll give you the stupidest one. Okay, just literally the the the stupid I I'm on the plane here and the Wi-Fi does not work. It works in the build a bit and then it does not work. United has a refund policy. All you need to do is go to united.com/refund and tell them, "Hey, Wi-Fi doesn't work." But it's a form to fill. It's annoying. I don't have time for this. I just like now just talk to Muse and like Wi-Fi didn't work. Figure it out. Get me the $8 back. So if before filling this form didn't didn't worth my time of the $8, sending a voice message to insist they can do this definitely is easier. So that's my use case. I had another one in mind.

55:29 >> I have one I could tell. >> Yes. One and then let's move on. >> Yeah. Because this is a funny anecdote. So, my agent knows what I like to eat. Like, I like chicken. I like sushi. And we sat at the table, my family and I, and we decided we want to have sushi. So, I just told Amy, "Hey, we want sushi." Half an hour later, the doorbell rings and the delivery guy hands me the package. We go to the table. I put it on the table. I pull out, we wanted sushi.

55:50 I pull out chicken. I pull out another portion of chicken and finally some sushi, but not the one we ordered. So, we ate it. We started eating it. And I was thinking, damn, AI makes mistakes. And this is a perfect example for this. My agent ordered the wrong food. And I took a picture of the food and the bag and wrote in the message, "Hey, damn it. I didn't want this. You know what I wanted? You just wanted sushi." And my agent, you know what? She replied, she said, "You idiot, you got the wrong package. Look at the back. The name is a different one." And I looked at the back and the name was a different one. The delivery guy, a human, made the mistake, gave me the wrong back. I, as a recipient, made the mistake. not realizing there's a wrong name on the back. I even noticed the name and thought, huh? Whatever.

56:36 >> So, two humans made a mistake and the AI assistant kept kept the things going. all righty folks. I think that we will bring you more use cases like this. I believe that everybody here's my oh that's what I wanted to say. We're now in at fully connected which we will switch into and we're going to have we have a guest very soon. However, we are now thinking about the fact that u all the executives of the big company is here and every time I want to bring one of them in the podcast, I don't talk directly to the executive. I talk to their personal human assistant.

57:11 And the amount of productivity that gets unlocked by an executive by a personal assistant is very high. And I think that this is my way to convince them that every employee here needs to have an assistant that that works for them as well because that amount of productivity should be applicable for every person in here. So this is our kind of assistance corner before we move in into the fully connected and I think it's time for this week's buzz. Folks, thank you so much for joining. please stay with us on on the side. I would love for you to just like stick around as we show you core weaves fully connected. But let's do let's do this this week's buzz dance >> from Gorewave.

58:00 >> Bye Gore. >> All righty. welcome to this week's buzz, folks. maybe for the last time because we have some rebranding to do. all right. folks, Fully Connected 2026. This is Corvvis's biggest stage in the world right now. And so in order to chat a bit about what's going on in Fully Connected, I have a guest coming on the show. What >> I will just tell you a bit about Fully Connected. this this year is the biggest connected yet. We'll send Wolf from actually to walk around and Wolf will show you around. We have robots. We have yep, >> we have robots here. We have physical AI. We have kind of folks from data center showing you pipes. We have robots walking around doing different things.

58:55 and we've announced Forge, which we'll tell you all about later. But now I want to introduce our our next guest on the show, Rich. I don't want to Richard I don't want to butcher your last name. Alfeld. That's pretty perfect. >> I said Richard, but Dr. Richard Al, welcome to Thursday, this is your first time here, and this happened literally this morning. So, I hope that I have enough research. but I would love to hear from you all about physical AI. Here's the thing. Thursday is a show for engineers, AI developers, and many of them are in the in the bits world, in the LLM's world, and we we build with Jeff and different educ etc. but we all know that more is coming from AI and that thing is called physical AI and you are the VP of physical AI at you founded monolith and I would love for you to tell me a complete pleb that does not understand physical at all. what is this about?

59:54 Why why the physical part and what does it mean? >> Cool. So there have always been like hype names for specific industries and I think physical AI is the first sort of name that is getting hyped that I actually like >> because the name explains what it does. It's AI in the physical world. >> Yeah. >> My own personal mental bridge to remember what physical AI is supposed to do. Perceive, reason, and act in the real world.

60:23 >> Also comes with a sort of most popular model in this space VLA vision, language, action, >> vision, and >> so if you're thinking right like if you're from the world of LLMs and you've realized, hey, cool, I can build these fantastic chat bots. They can reason pretty cleverly >> and you're like, hey, wouldn't it be cool if it could reason about the real world? If it can look at the microphone and then potentially adjust it because it can see that Richard is not doing it in the right way, it can see it. It can think, "Oh, he needs to talk. He needs this thing closer to his face." And then he could potentially action this and move it closer to my mouth.

60:56 >> That kind of loop. That's what physical is all about. >> I have a question though because for example, the regular LMS, they're multimodel for quite a while now. like GBT4 started and you know Greg Buckton draw something and suddenly it saw it and this was incredible and so now we kind of get it for granted that you know these models also can can see meta's muse models are now multimodal where is the difference between kind of like LM just getting visual input to vision and action models that >> so now imagine right like I think in my example of moving the microphone I need a robot arm to actually do this right and this robot arm might consist of 20 different actuators like electric motors that need to move it around. just if you I don't know maybe in school did a little robotics class you realize that all of these different motors right like in the end it's just a signal of how much it needs to move and it's quite hard >> in threedimensional space to actually reach this thing and then figure out how your like three motors need to move in unison to actually move it where you want it in 3D space >> and that is actually quite complex and you need your LLM to do that too and this last part actually doing something in the real world that tends to be the hardest that tends to be hardest. it's so I definitely want to talk to you about like simulation then we get to the core angle of this right. obviously huge news in the world of >> world models and simulation models this week with MD buying world labs.

62:26 >> huge news >> world world models generally I think both from Google genie kind of type stuff and and other world models that they have and world labs obviously and some other folks. are they the the way to get the bridge between like getting input and having a robot move move things around? Could you connect world models to physical AI for me? >> Yeah. Okay. So, let's start from the top, right? I want a robot to go and adjust my microphone, and so I somehow need to train this thing to do this.

62:55 >> How can I do this? >> I doubt you're going to find online data on the internet, right, that actually has somebody already doing doing this with the same robot in the same scenario. Zero chances exists. And even if it exists, you're not going to get it, right? Like, because it's probably owned by some big company or robotics company, KUKA, AB, whatever. So, you have zero data to start from. So you need to first figure out okay where am I going to get this data from that I can actually train. the easiest one most popular is that I'm going to ask you before the show to go and do this 2,000 times. I'm going to give you a teleoperating set where you can essentially take the robot arm and try and putting the microphone in front of somebody 200 different times. I'm going to film it what's happening in the video. I'm going to record the sensors on the motors on the robot and then once I do this 200 times I actually get a good data set that I can then put into the language learning model and from that on it's quite likely that in different scenarios is going to work.

63:56 Yeah. >> Right. This is the easiest way you do things in AI. It's showing a 200 examples of the exact same thing and I've got all of my inputs which is the vision of where the thing is and the outputs of what the motors need to do that I can learn. It's always been done like this in robotics. 10 years ago, you basically had people just move the robot arm manually by hand. Now we do it with air.

64:13 >> That's level one. Level two is that you're essentially, okay, this is quite complicated. I don't want to do it like that. I'm going to go and simulate it because it's embarrassing if you spend all morning adjusting the microphone. >> I don't want to do to be doing 2,000 examples of this also. Yes, >> that's correct. And so there are companies that are starting to think, okay, can we can we do this faster? Can we do it cheaper? I think some of the most promising approaches are companies like physical intelligence.

64:37 Physical intelligence are saying, "Hey, what about we train a foundational model for the real world where you don't need to do it 200,000 times, 2,000 times, we're going to go and create a lot of sustaining scenarios that might not be adjusting a microphone, but it might be adjusting a tie or it might be adjusting like specific things." And these baseline sets of how you adjust something with a robot arm already go into the data set and you basically connect to phys physical intelligence.

65:03 >> You maybe have five examples of adjusting the microphone. You upload it into this and then tada it works. So that would be introduction to world models and how you do this and like >> world labs now essentially are at next level around this which means that >> when you sort of have these foundational models I want to have more data data augmentation for example right like in this specific video of me sort of adjusting the microphone the light is always coming from the same angle I'm sitting in the same way but in the reality like when I actually do the demo >> I might be over there over there and so you basically just want to create >> millions millions of different versions of the same thing. So you can put it into the AI and train across different ones without actually having to do it.

65:45 And world models are fantastic because they can you can take a scene and you can sort of run it through different versions of it. We always have done this in AI, right? Like in the early days when we tried to classify cat versus dog. >> It was quite hard to do this robustly if you only one cat and one dog picture. And so if you sort of just move the cat around, different angles, different lights, different sizes, that creates a much more robust classifier.

66:10 >> And that same idea basically is >> applies to world labs and and simulations. I have so much to ask you that probably I won't about whether or not we are in one right now and this is all simulated by like the next version of a world model. but I do want to talk to you about the our core revenge because we're the essential cloud for AI. We're we're providing multiple things for u folks who are building LM we're providing infant service we're providing loop for example for folks who are building agents right but we're also working with folks like Caterpillar and and and Nvidia and I would love to to hear from you how we're solving stuff for them I would love to like just to highlight how much is involved in physical AI >> cool okay so over the last couple of years right like first was the big foundational model training wave And neoclouds like Hawweave emerged because taking thousands of GPUs to train a huge foundational model is a very different job that you need specialized equipment for. And so I think that was the first big wave. After that came obviously agents like we all have our coding agents run through the night over the weekend and it's a very different workload. Again it needs a different cloud.

67:21 >> Right now with physical AI we again have completely different cloud requirements. If you sort of go through the Caterpillar example and their VP Brandon talked about it on a couple of stages here, so I could talk a little bit more freely about it. Then if you're trying to train a complex equipment like this to drive autonomously in a mine or excavate something and pick up a rock like in general like excavators are a fantastic example, but basically I can see like imagine this thing digging something up and putting something from left to right.

67:50 >> Then you can sort of realize that they probably already have a bunch of data, right? like they've got cameras on their fleet and they've had those cameras for decades. They've got telemetry and so they have this underlying data set that does stuff. >> But this data set won't be enough. They need to augment it. It might not be raining in the video. It might be too hot. It might be too cold. It might be really dusty. And so first, right, like you have all of this stma augmentation, this vision augmentation that I talked about. Next they will have to run a lot of simulations like simulations and like I sort of differentiate like I differ them make them a little bit different from just data augmentation because imagine like actually they said like okay this scenario doesn't work at all we don't have any cameras that we can sort of play with.

68:30 >> Yeah. >> We've never done bedrock in Southern California. I actually need to figure out in a realistic physical simulation how to do this. So they built a 3D caterpillar truck or excavator >> simulate the earth and then this thing like in a video game just tries to excavate stuff and from that I can learn as well what's going on. >> So we saw the latest GPT models we don't know which one solved the Navier Stokes equation about like fluid dynamics and physics whatever.

68:58 >> >> so I did aerodynamics as my masters. So >> could you please speak about the connection of us solving more and more challenging physics problem now with AI with like LLMs and our ability to do the stuff that you did like bedrock in California. Could you make the connection for me? Like I found it like fascinating. >> Good. So solving things in simulations is one of the most elegant ways I can train stuff right because you can literally >> in theory simulate everything possible for an excavator. Do that in simulation.

69:25 Train your AI and then in theory it should work. >> Yes. In theory, it should work in the real world. But now comes the problem. >> If you do you play video games? >> Yes, sir. >> Okay. Do they actually mirror reality? Does water does the same thing? Does rock do the same thing? >> Fast tracing, rate tracing, all this, they they try to approximate, but definitely not. >> No. Right. And so I think that is the major problem here. We call that the sim tore gap, >> which means that rocks don't behave in real world as they do in simulation. As much as I try to do it more accurately, they just don't. And I think this is the last big wave we're currently seeing in physical AI where people are saying, "Hey, why don't I train a model based on really physical simulations, the kind of simulations they run on supercomputers at Stanford for water >> and then I train an AI model to simulate really really realistic water and then I put that >> into my sort of physical equations for the simulations >> and then I'm actually having realistic stuff to train on. So like I I train on the very very very like good physics simulators.

70:25 >> Yes. >> But then I approximate it with generative AI essentially. >> Wow. >> That is Do you foresee that being a significantly improving in the future like very >> So I've worked on that for 10 years. I think in the beginning most people thought it was I don't know crazy. I was definitely a little bit too early in this field to be fair. Over the last couple of years it's made huge strides. >> We're talking there's lots of startups in this space. you might know Physics X who work with Bwave. There's plenty of startups that have realized that you can actually build really cool physics approximators that generalize better and better that you can then put into the simulation and suddenly when you're trying to pick up like something that can crumble >> you'll see it crumbling.

71:07 >> Yes. >> exactly. >> And now back to us and the workloads that it requires to run those simulations in the cloud. we're powering like different companies obviously work with Nvidia. We're great partners with Nvidia. They're here. They're like sponsoring for the connectors as well for us and and Caterpillar. You talked about like a 100-y old company that drive tractors now trying to like >> self drive them FSD and Caterpillar and doing like different things. >> >> what else can we be doing for these companies? Like what is missing in their workplace right now? How is Cory meeting?

71:38 >> I mean our job has always been right like to sort of look at the sort of frontier labs and provide them with a premium experience, right? Like the thing you can't do for a million companies but you can do for the top 10. And so if you look into physical AI, what does that mean? One, we have to look at our data centers and really build customized data center architectures that just work for physical AI. What does that mean? That means I have a lot more storage. I have storage that is really fast. So that when a customer saves three pabyte of complex 3D data, >> it's not text anymore. It's not just like streams of text, right? That's what we're talking about.

72:10 >> I can give you caps examples where you literally you overwhelm the storage. Boom. Like that. It goes really quickly. Imagine trying to query complex 3D data 2 million times a second. >> Wow. >> Yes. >> It's it's not going to work. And so you need to optimize the storage. there's a different combinations of CPU plus GPU. We need CPUs to pre-process computer vision data. >> we need to obviously GPUs to fine-tune so like V200s to fine-tune foundational models, but we also need to run simulations and that is actually very close to video gaming. You need like RTX GPUs. Yeah. which actually run these workloads and that's a different type of >> that is very interesting like the consumer line for the the video games where which is where Nvidia started now comes into play correct >> because it's like better at that versus like text and and transformer type >> Nvidia has always had if you're a gamer they've always had an edge here >> and that really helps when running simulations in physical AI and then right like you have your real world data processing your simulation running your vision model fine-tuning and all of these different things like they need to be looped together so that you get a seamless experience and you need to have five different stacks and that is rather hard to do really really well and with physical being quite new not a lot of people have done it well yet >> and so you can see like with Caterpillar working with us that we're making really good progress >> on this sort of hardware front.

73:30 >> I will ask you maybe a last question unrelated to Cor that that cleans my house. How far away are we from this future where a robot I don't have to like train it on how to clean my house? Where are we? Where are we? Do you have one at home because you're in in >> so I mean if you want just a vacuum cleaning robot. >> Yeah, I got attacked by one yesterday. The the bar attacking people.

73:58 >> I love that thing. I think it's an excellent addition. And so we have a couple of sort of startups working with Cory already that have fantastic robots that make espresso. They try to fold things. You might have sort of seen demos maybe of Sunday Robotics. >> Yeah. >> Who have this fantastic setup like >> at this point, right? Like these things, they're quite cool. They are actually capable of doing these sort of simple tasks like folding your laundry and tidying up your home. They take quite a while. It's not super fast yet. If you sort of look at the videos, you might have seen in the corner like 10x accelerated.

74:34 >> Yes. >> And so >> and maybe tele operated at some places. So some folks are like >> correct >> transforming into the robots from far away to try to help and also learn from those actions. Right. >> Correct. And so there's just so many things in a real home that are hard. Like the way laundry behaves is incredibly intractable. The way water behaves, the home is a nightmare of complex physical problems and they're everywhere, right? it's very chaotic and I don't think I've talked to a couple of CEOs that are in our network on a sort of little trip field trip last week and they're like, "Have you ever been to two homes that looked anywhere remote and two people decorate differently?" And so >> the second a woman moves in there, it changes completely because she's like, "No, I don't like it. I don't like this." Yes, >> correct. And so I think my money currently would probably be on us seeing industrial robots first.

75:23 >> Okay. Because if you looked at sort of industrial automation, we've we already got really mature solutions doing really valuable things that at scale save people millions and millions in sort of manufacturing cost. So that has a colossal business case for people to do it well. And >> automation has crept into these a long time, right? Like if you go to some of the factory shops across America, you'll see robot dogs already transporting things and stuff. And so industrial automation, I think we're much much closer to skin deployments. We have broaden robotics and core weave network and client who do a fantastic job there.

75:56 The home robots right now I think are at a stage where we can sort of do really cool like short-term demos but we still need to scale up further use cases. But that said physical right now is moving really fast and we have it feels like all of the components are there. >> It might do boom like this like a little zip and then like 6 months from now we suddenly have it. I would love to have you back on the pod once you have one in your home that you trust. Richard, thank you so much for joining us. I know you're a busy dude. You have like a bunch of interviews going up.

76:26 >> This was fun. >> Yeah. Cheers. >> You guys should check out Richard's work before he joined a core and now it's quite incredible. >> Yeah. >> we have our next guest and this is a friend of mine, >> the mighty Daniel Bis. >> The mighty Daniel Bos. Folks, you know we told you about core with inference multiple times. we have Daniel Bis. Daniel worked at open pipe on open pipe stuff. We told you about open pipe multiple times before and now Daniel is a senior product leader product manager within core and the way we collaborate is on the inference service that we have that's now part of forge.

77:09 Yes. and I would love for you to tell us a bit about how first of all the open pipe efforts and also the influence efforts are like folding into the force thing that we announced. Could you could you talk about this a little? >> Yeah. So so it's it's been a little over a year since open pipe joined the core team. Yeah. >> And the weights and biases teams and now we're all fully connected. >> we specifically focused on postacquisition on kind of our serverless products. So we have serverless inference. We launched model distillation yesterday as part of forge. We have serverless RL, serverless SFT. All these great products to engage with inference and post training in a very serverless fashion. you don't have to manage you know core Kubernetes yourself. You don't have to control your cluster understand how to build that full stack. you can actually just really engage on on the model layer and start building your agents and and training and and and improving that performance and efficiency. Yeah.

78:12 >> in in as lightweight of of a way as possible. >> That performance being improved as part of the the core we forge loop. >> Yes. >> Tell us about the loop. Daniel, you're the first guest to like go to the loop. I'm sure we're going to mention this again. The loop is quite an interesting thing we're kind of showing. It's a way to take your agents from good to great. >> Yes, >> that's I heard that the loop is being referred to this. What are the parts of the loop? Why are we calling this the loop? Could you tell us about the agentic loop a little bit?

78:40 >> Yes. So the loop is this awesome like it it's what everyone is doing already. >> It's a mental model for how to improve this. It's a playbook. >> Exactly. And and everyone's been doing it in scattered places. You have to, you know, observe what's happening. You have to iterate. You have to figure out, you know, what are what are the places where things aren't working? How do I where and how do I improve? You know, can I just tweak the prompt? Yeah.

79:07 >> Can I actually post-train the model? can I distill a small smaller model to improve performance? >> and then I want to run again and and in in production and and see what happens and and kind of go back through the loop. So, the loop is is not necessarily a new concept, but it's it's a new organized way, to really have one surface area to to train and improve what's happening with your agents in a very fluid and organized way.

79:38 >> I think it's very important specifically because and now I'm like taking a step back from being the interview were and it's very important because many companies are trying to get to aentic AI now, right? So obviously we have all of the bigger frontier labs are doing like assistant stuff etc. But many companies cannot use Muse for example for their internal stuff you know and so they are trying to build things for themselves internally but also for their customers and they need robust evaluation stuff they need to learn how to do this and so we have folks we have FDES for example Formula 1 shout out to my previous manager Morgan and and some other folks who are like showing them how to use our tools correctively and we need also that u we need that playbook and in that playbook at least This part of that playbook obviously is serverless SFT and rollouts which is the stuff that we launched in public review. Would you would you want to talk about that?

80:32 >> Yeah, absolutely. So I think and again we're all kind of figuring this out together with our customers and and what serves them best and I think we're still in the early days of model optimization on very specific tasks. And what's cool that not a lot of people realize is even even with a small amount of effort, you can really post-train models to be extremely task specific and and performant on what you're trying to do. and so >> effort and data, right? Like like small amount of data as well. That's what people sometimes don't realize like it doesn't we talked about Niss releasing a data set and that's not blowing up on hugging face folks that it could be that you can like fine-tune the model to be specifically significantly better on your use case as a company by just doing a little bit of work and then fine-tuning this on like SF stuff.

81:23 >> Yeah. And part of what we like to bring is is those insights of just doing this for for many years now. Yeah. >> what are some of the tips and tricks? Where do people get stuck and how can we unblock you so that you can kind of access these tools and and make good use of them? >> Daniel, I have one last question. I know you're also busy, dude. We all have like a bunch of stuff running around and doing things. Distillation is kind of has a bad like like feeling to it. Why?

81:51 >> Because of folks distilling illegally, right? Like there's there's been these things from like Frontier Labs like oh the the Chinese models whatever they still us but they doing those so illegally. Obviously when Opus 5.5 releases last week and then Sonnet 5.5 releases this week. Sonnet has some amount of distilled from the bigger models. An Kapathy talked about this what three years ago like you first do an effort of training a huge model. It's really good with like the best big model smell that you can get right and then you kind of like have it teach another model basically a teacher student thing. So we're offering distillation services now. Is that that's new for us? Tell us tell us about distillation services. Yeah. And I love that example of like teacher and student. Yeah. Rather than just, you know, >> kind of some of the connotations around distillation. I think you're exactly right. I mean, you you see great models like GLM53 having a GLM53 flash variant. these are often, you know, student teacher relationships between models and and GLM 53 flash just a really quick and efficient model.

82:54 Yeah. and and that's what we're trying to bring to the table for tasks that you know can can be well trained on and just hyper performant at scale. So >> at some point folks should stop paying frontier prices and pay for significantly cheaper prices. the one thing I will say Daniel the you know I will help you say this. Basically we're not only offering service inference for customers as in streaming tokens. If you're a company that needs influence for their products, we both together worked on on I'm blanking on the name of the product, but that powers that you know different companies. If you're a founder of a company and you're finding it hard to pay for the Frontier Labs and you want some like services, we would love we would love to work with you. We have like a service inference product that would be the right fit for you.

83:42 >> obviously, if you want to train bigger models, Frontier Labs, you're already on Core. So that's that nine out of 10. I think is the is the line n out of 10 once. Daniel Bose, thank you so much for coming up. I wish I had more time to talk to you because I find very much of the stuff that we do exciting and the open source models that are coming up and we discuss and like oh is this good? Is this I think that that's a very interesting dance that we do on our infant service. and it wouldn't be there without you. So thank you so much for all your work as well.

84:10 Thank you for coming up. >> Cheers. >> Cheers. All righty folks. Daniel we scaled up our infant service fairly quick and fairly quick. We scaled up really quick and we're serving the best frontier models in open source and open weights right now and Daniel has been an instrumental part of that. And another person has been instrumental in the beginning and now is doing significantly other things is Dio Filho. >> That's right. >> No, tell me how you say it in Brazil.

84:37 >> Dfilio means junior in Portuguese. >> In Portuguese. Yeah. Dakfilio. Okay. So, so Junior, >> that's right. You can call me Junior just for you. >> Dak is a senior PM at ML Products and Sandboxes. >> That's right. >> Where you moved after doing some work at inference that we worked on together. >> That's right. >> And I I find Sandboxes super interesting. Yes. Wolf has been a user of of the stuff that you're doing for a while now.

85:02 >> Really? Well, good day. Thanks for having me. My name is D Guten and >> Doaden. Doen. >> Yeah. Okay. >> Thank you. >> Yeah. Wolfpin wouldn't be possible without the sandboxes and the support I've got. So, thanks a lot. >> Thank you for for breaking things and getting us to fix them. >> Always a pleasure. >> Dak Ianbach talks about Vera CPU. He's a he's an Nvidia exec and he talks about the Vera CPU and the Vera Rubin CPU GPU dance. Tell us about Vera Rubin and CPU and the importance of those things. Please tell us like what what's going on in that world.

85:37 >> Absolutely. So Ian Buck, really famous guy. He's been in it at India for a while. Ian has been talking at the core conference yesterday about Vera CPUs. Yeah. Right. So the Vera Rubin GPU node has a couple of Vera CPUs and it has the Ruben GPUs. So that >> combination is available for the GPU nodes. People really like them in racks so they can do very large scale inference and training. But then for sandboxes where you want to contain execution of arbitrary code, you really want CPU nodes. So that's the Ver CPU is the latest iteration that Nvidia offers has a few things that allow you to pack more workloads into a single node >> which in the age of agents and in the age of post training that's what you need.

86:18 >> Why why is this important? Why do why do we need CPU? All I can think about as somebody who let's say doesn't follow the latest trends agentic execution tool calling etc. maybe they think of AI as matrix multiplication and trans transformer layer and that's GPUs, right? Yes. So now why do we need why do we need like CPUs, >> right? Why is the GPU guy and the GPU company talking about CPU nodes? Why are we here?

86:44 >> Why are we here talking about CPUs? That is a very good question. So there's a couple of answers. The first one is there's lots of talk about agents in town. People are building agent harnesses. They're building agent runtimes. They're building lots of agent things. And the place in which you want to run agents is in the cloud. Now an agent is some CPU plus GPU, right? The the brains behind it run on a GPU. We have a bunch of inference APIs for that.

87:09 There's lots of self-deploying inference things you could do. But the harness and the execution of whatever code the agent generates or the commands the agent wants to run and the file system the agent wants to use, all of that needs a file system and CPU. So there's the brains and there's the the tools that this brain gets. >> That's right. >> And the tools, they all need execution that's fast and is not about multiplying matrices. It's about like which tool to call and then the output of that cool where to place it and then which files to like read and write etc. And a bunch of it there's also disk access in there as well. But like a bunch of it is like CPU bound as well.

87:45 >> Absolutely. >> And okay. And so how does Verra Ruben help there by combining them together? How more performance? the the the metric is agent packing. If I own one node, how many agents can I run on it? >> Agent packing. I've never heard this before. >> That's right. Agent packing. You want to pack as many agents in that one node as possible. And there's a few different axes that you could do that. So it if I can put in a bunch of pods or a bunch of VMs in this CPU, how many CPUs or how many agent CPUs can I have in there?

88:15 Agent environments, harnesses, right? >> let's say it's X. The whole deal with Verera is even if you don't do over subscription, which is when you try to run one agent and then very quickly replace it with another agent because agents have a little bit of downtime. No, they run a command and they wait for the brains to think it's very spiky and whenever it's spiky, you can try to like increase utilization, but Vera allows you to have a larger number of agents within a single node without doing overs subscription. So oversubscribing gets you even more. So here's a stat. Okay, numbers. Yeah, I'm listening.

88:48 >> Big numbers. >> Yes. >> Within one rack of Vera CPUs, you can run over 20,000 agents concurrently. >> That's 11,000 cores within one Vera CPU rack. Just CPU. Wow. >> So that's why the GPU guy is talking about CPU nodes. >> because the payloads changed, the workloads has changed. And we've seen this in Devon. the folks from Cognition was yesterday on the fully connected big platform show and they talked about some of their performance. Could you talk about this a little bit?

89:21 >> Absolutely. So, what they were talking about yesterday is the Vera Reuben GPU being the first user in the world having delivery and be able to run stuff on their Vera Ruben GPU. >> Cognition was the first user in the world on this like new stack that Jensen imagined and is now sending around the world that has >> Exactly. that he imagined and and fulfilled with his bare hands. Yes. >> and that's through Corewave cuz Cororeweave deployed these GPUs first and we were able to get to market with a good product based on the Nvidia.

89:49 >> I won't push back on this. It's a great product. I use Devon. Devon is incredible. >> Yes. >> Scott Wu and the team saw whatever everybody else is seeing now. Whatever Elon Musk is paying $60 billion for with Cursor, he saw it what, two years ago. We talked about Devon two years ago. Non, you remember LJ? You guys remember being like $500 a month employee. He was like, "" And back then they had a computer, they had a browser, they had file systems, and now core powering that Devon.

90:15 >> Yes. >> And the speed up is insane. >> Yes. >> What? 4 Tell tell us like what they're seeing the performance wise. >> It was 1.4x. It was 40% higher packing. >> 4.8. >> What? What? 4.8. >> 4.8x. That's what I saw. >> That's what you saw. I don't know what that metric is. I I thought you were talking about the packing. >> Oh, no, not the packing. The performance on on the actual task, but tell us about the packing metric. Yeah, it basically is how many more agents can I fit in this one thing and then the Vera CPU has faster clock. You can pack more agents.

90:44 You can have these racks that are very dense into your data center and you can run a ton of agents on them. >> Yeah. >> and that's what they were talking about. >> Yeah. And I I saw a stat they saying like 4.8.8.8 more performance. A bigger number. I like that better on Devon. And I I heard from them from some of the team that like they pushed and pushed like the number didn't seem like big and they could go farther away. Here's the additional thing that people don't think about. Wolf from I would love your take on this. CPUs faster CPUs also allow faster evaluation which then makes the all the just the bots faster like the improvement with which Devon can now improve because they like test it and evaluate and run the agentic loop.

91:25 That's also something that that is improved with faster execution as well. >> Yes, exactly. So it's not just the GPU and how fast the AI can think but also how fast it can execute and we see that sometimes or often it is actually the agent has decided what to do and now it takes a while for to run the program and the faster the CPU and the GPU the faster the whole thing will run and it will speed up everything.

91:49 >> That's right. >> Yeah. So that's Nvidia and we're the first cloud to put it in the first customer's hands. I think that's very exciting. That's super cool. Thank you for telling us about this. it is very hard also. but also we have services that we provide in terms of like sandboxing as well and you're part of that and that's part of our forge. >> That is part of our forge. >> Would love for you to tell us sandboxes is not a new thing but I think they're now in G. Correct. Correct. That's that's the announcement that we're having and as of yesterday. That's right. Look at that.

92:18 That's a lot of work that you guys put in there. tell us a bit about the forge and the kind of the sandbox loop. So Devon obviously like runs on like GPUs and we work with them. That's right. it's a great example. Tell us about like what customers are to expect when they're hitting core with sandboxes. >> So if you go search for coreweave there was a preconceived notion that it was just GPU for right and and that is not true. So over time, Cororeweave has amassed this incredibly large amount of data on how to run GPUs really well on how to expose services through things like Forge where you can do serverless inference. You can do serverless training, you can do serverless reinforcement learning, you can have serverless notebooks. So it's not just hey give me a data center filled to the brim with GPUs and really fast infiniband, but also all of these serverless services that you need to do the loop. If I want to start with a task, I want to make my model better or faster or cheaper for a particular task, I have these the slope that I go through where I eval the model, I deploy this new version, I use sandboxes as a way to run the harness, to run the agent execution environment, as a way to run the post-raining RL environment. All of these available on Forge today, forge.coreweave.com >> and it's now like a service as well that people can sign up and then evaluate this for like a 30-day. We also have a a partnership offer that you can like bring your own workloads and then without commitment like evaluate whether or not Coro is going to be better for you. Obviously, it's going to be better for you folks. We're like number three platinum in a row in the core core core max >> cluster max.

93:46 >> Cluster max. Thank you. We're number like third year in a row platinum tier at cluster max from semi analysis. Of course, it's going to run better. >> We basically define what the platinum tier was. We were the only ones and now we're we're three years rolling. >> Now we're the only ones. Three years running. That's right. do what else did I not ask you that I would love to ask you? We're partner network obviously like you want to talk about partner network.

94:08 >> So the the the thing is a customer wants to test some stuff out before they would need to go talk to a seller, you know, sign up for a contract for a few a few thousand GPUs here and there because the demand was so relentless. >> Every time somebody ask me like, hey, you work at Corwe, can I get some GPUs? I'm like, you have to go talk to the sales guys. They're right behind us doing work right now.

94:27 >> That's right. So our sellers are here available. They're available through Twitter. They're available through forge.coreweave.com. And the point is if you want to run these workloads, if you want a PC, if you just want to try out things, you can just go to the website or you can reach out to the partners that we work with who have direct connection to the sales teams and can get you hooked up. >> Yes. So folks, if you again if you're we talked about this before with Daniel, if you have a startup and you have some investment because you know this stuff is not cheap. you can talk to us and we can provide you with the best possible and optimized cloud for AI. I think those are very very exciting filio >> and there is one more thing >> one more thing please.

95:05 >> Can I say one more thing? >> Yes, of course sir. serverless GPUs. >> Serverless GPU untrusted serverless GPUs. As I was saying before, you needed to go get a PC, talk to a seller, find a way to get a cluster >> on Core. Now >> you can go use this amazing thing called GPU sandboxes where you can sign up on forge.coreweave.com. >> Yeah. >> You send us a ping so we enable it so that don't crazy people are not cryptomining without our consent. Yes, sir.

95:35 >> Right. And then you will be able to use serverless GPUs pay per hour without any contract without any sort of commitment to coreweave. Just use GPUs for your tasks. >> That is incredible. >> Yes, >> I'm signing up right now. I have a bunch of workloads that I want to like run and I never even internally crypto mine. I believe you >> sir. Even internally like going through like our own procurement etc. is like a whole pain. I would rather sometimes to like go around go direct and expense that than going through internally.

95:59 >> This is in private preview starting yesterday. Amazing. >> And we will be rolling out more cues in the coming months. Yeah, I think that >> breaking news. >> This is breaking news. So folks from the show from the PM that brings you this serverless GPUs is coming to to poor. I think it's been a huge ask for many many people to come to us and say hey do you guys do the GPU thing maybe they want to just test maybe they want to commit like for a sales this is incredible.

96:24 >> thank you so much. This has been incredible. Thank you for coming. I do expect you to come back on the show and tell us about how the the the private bet is. I will I would love more. Thank you very much. You guys have a good day. >> Talked to you earlier here first. Several GPUs is coming to Porweave. You you don't you need to understand we've been asking for this internally since I joined Corw. I was like, "Hey, can I get some of those GPUs?" Like, "Yeah, well, you need to go to that guy and that guy." And now it's an offering.

96:49 Obviously, offering such as this exists elsewhere, but you guys heard it first. It's really hard to put Vera Rubin in production. I should have pinged DO and asked if it's Vera or Rubin that we're getting in those several GPUs or not, but I think it's very exciting. yeah, M, you'd be able to run some workloads on core. We'll be able hopefully to procure some credits for the Thursday listeners. folks, next up we have Corey Wol from I don't think we have quite the time for you to go on the floor because Cory Sanders is joining us in around five minutes. But maybe for five minutes you if you're ready you want to you want to do >> I'm ready.

97:25 >> last time we were in Muski we sent Wolram on a walk and Wolram is going to join us now in like another camera and folks we just want to want to show you how big this thing has gotten and how many robotics are here. so we're gonna we're gonna do this. Okay. So, this is behind the scenes the setup, folks. You can you can see in here, Alex and Tom, our camera guy. Thank you, L, for all the work that you did to to get us up and running. this is the floor of the fully connected. We have a stage up there. They're doing demos. There's tons of people here. Really tons of people. But Wol from if you go towards kind of that area. I want to show the the pipes, the actual rack pipes that we have over there. I think those are super cool.

98:09 I'm heading Wolfam. Oh, yeah. Wolf's showing you some of the data center things that we that we're showing here at the fully connected thing. and then a bunch of robots. Wolf, turn around. There's a robot just behind you, I think. Oh, this is a rack. This is Nvidia rack. >> Yes. >> Yeah. Some racks actually. >> Racks, man. Can you read out what's written on there? Is this the verb? >> This Ruben NVL 72 switch tray.

98:35 >> Wow. Oh, let's go. >> Let's Let's go. >> They're liquid cooling the switches. That's the optical. >> Can you don't do this really, but can you test if they're bolted onto that table or can we take it home with us? >> Let me take >> Oh, no. It's it's it's not available for us to take. >> All right. >> I couldn't hit the reset button. AI engineer either. >> Oh, by the way, this is the founder of Open Pipe, Kyle Wolfing. Say hi to Kyle on on the show for us.

99:07 and David both both co-ounders of hi just going around the floor. Hey >> K has been on the show before. Yeah and now he's leaving. Hopefully he'll come up. W show us show us the Formula 1 stuff. >> Oh yeah, the Formula One stuff is over here. >> That looks lit. And you can actually sit in the car and do a simulation. So those are screens in front of the cars. So this is quite fun.

99:43 Let's see what we >> want is over here >> in my living room. >> Have a big open one. >> Oh, this is Lydia. Oh, show show us the robot. Oh, >> wolf. >> the robot. The robot. There's a robot on the floor. And I think like we talked about physical AI. this is >> It looks like one of the robot dogs.

100:14 >> This looks like one of the robots, but I think it's more more advanced. >> Yeah. >> Yeah. >> Can I just say >> robots moving around? >> Yeah. Show us around while we set up our next speaker. >> Is over here. >> Wolam, show us the Formula 1. Are you ready showed us? And then while we get Cory set up here, Wolf, we're going to get back to you because we now have our next guest on the show that we want to introduce. Wolf, thank you so much. He's walking around. He's doing like a like a men on the floor type thing.

100:49 >> I like that. Yes. >> We should have done this one. >> like they did Jensen and and the Salesforce guy. They were walking around. Do you see this? >> No. Jensen like was leading Mark Ben off, right? The sal guy around the floor is like, "Hey, come with me. Come with it." It was really funny. We're not quite set up for a walking session. But folks, I do want to introduce you to Cory Sanders. Cory Sanders is an SVP of product here. Yes. At Core with a 20 20 year career at the >> prior Yeah. So now it's 22 now total 20 years at Microsoft and two here at Cory.

101:21 >> Cory you weren't scheduled a week before this >> and then and then we we saw the keynote live >> and they're like we got to get him on >> and then and then I looked at your demo and you murdered that sir. I first of all congratulations. Live demos are really hard. Openi had the live demo just before you and their team deployed in the middle and nothing worked and they're very gracious about this and the thing about live demos is they could fail. So you do take the risk >> and they do oftenimes fail.

101:48 >> Yes. But yours didn't. And you run through the whole Cory Forge gentic loop thing in like a 10-minute thing. >> 10 minutes. >> I think you picked it like 200 words per minute. >> But yeah, they did. They did. They asked me to do it in seven minutes and I said that's not that's not >> Cory Sanders. >> Great. Look at this. >> We launched a We can swear on this part. It's my part. I don't care. We We launched a ton of >> Yeah. Yeah.

102:10 >> We launched a ton of products. I know it's kind of like ch asking you to choose your baby, but like what is the highlight that you are looking at that we've launched at at Fully Connected? >> I mean, I I I feel like agent agent I know now I'm saying this and like everyone in my team that doesn't work on this product is going to be like what an but agent lens I really like agent lens like I think it's look the whole concept of this AI loop starts with being able to observe what's happening, right? And if you can't observe what's happening, you're dead in the water. You can't improve. you can't curate, you can't do anything. And so for me, I think the the the steps we've taken with agent lens, I think the automated insights that we built as part of that to be able to discover things that are hiding in in your application.

102:53 I think those are really really cool and they're AI focused, right? >> So what is agent lens? I will use this as a segue to ask you about like give us two lines. >> I'll talk about whatever. So agent just sorry to interrupt one thing Thursday listeners know what weave is and we've talked about weave and traces and agent etc. As agent lens I think is a new thing for >> sort of like yeah great great yeah good point. So so it's effectively it's I I mean it's a lot of the components of weave and sort of the observability constructs of weave but taking it to the next level around much more human accessible insights is probably the best way to put it right so it it allows dopes like me PMs to be able to go and do that data curation to be able to understand what's happening with my app and be able to actually make changes right there on the experience. I have used we for whatever 3 years.

103:40 >> Yeah. >> When it was just chatbots that I was building, we was easy for me. >> Yeah. >> I was like, "Oh, this user talked about this thing. Here's the thing." Whatever. I recently connected with like an agent that runs for an hour, >> bro. I have no idea what's going on. >> I have no idea what's going on. It's a fire hose. >> That's right. >> How is Agent Lens helping me? >> Exactly. So, so that's exactly the the issue is that you've got pages and pages and pages of conversations. There's no way you're gonna find anything in there.

104:03 And so, you know, agent lens does a few things. you can create this concept of tags and this is all about your curation of data another stage in the loop and so you can find the data you're looking for using tags you can generate them with Arya or you can hand generate them and then a agent lens has a built-in insights that is detecting trends that are very small trends but still trends nonetheless so 1% of your conversations may have some issue yeah and agent and the insights tab in agent lens will discover it and and then give you some guidance on what to We had so first I'm I'm very excited about agent lens. So I can't wait to have folks tell me how it works for them doesn't work for them and like give you guys feedback as well because a lot of folks that we bring a lot of listen on Thursday I a lot of the folks that we bring on core hacks for example the hackathons that we do they're now showing here on core u they >> they used the observability layer but again like overwhelming etc. So bringing that to an agentic place I think is great. we also had Daniel Boss and and Dio both on the show just before you and Daniel talks about model distillation. Yes. And I think you showed some stuff around that as well.

105:11 Could you talk about that part of the the force that we're releasing? >> Yeah, absolutely. So you know look and I was just doing another show with a gentleman from you.com the the post-training aspect of the loop and and the improvement of models I think is becoming a centerpiece to how almost every AI application is going to run right and having a a set of tools to do that is going to be crucial distillation is a really good example that you basically take you know your data the sort of the outcomes that you're trying to drive it keeps your quality but it makes your model a lot smaller And so that of course reduces your cost, improves performance. And so it's a really great service for doing that post training and getting to the results without too much of a cost impact.

105:53 >> And doing this in a legal way, not in the way distillation is perceived in the media where other companies are saying other companies against the >> this is this is all all a AI based which means it's okay. >> Yes. I I I do want to like highlight and do made this point for me. I didn't like realize this from the keynote that not only are we the first cloud to put Ver Rubin in place, we're also the first customer that runs Ver Rubin production stuff in addition is also announcing their stuff on the stage. I would love to just like refund that a little bit.

106:25 How cool is that? That's super cool. >> It's really cool. I'll tell you like the >> I love the I love the services. I love the, you know, the demo, the AI loop, but like I'm also kind of a hardware geek at heart. And so, like, when we were doing those first deployments, and we won't do this in all the data centers we deploy, but the first deployments had like this clear floor, so you could see Valve, which is our liquid cooling system, so you could see it underneath the floor. and for me, it was just like it was one of the coolest sort of visceral examples of both how we're different in market and then also how cool it is that what we're doing to bring these things alive. really is.

106:59 Cory Sanders, do you make a pun and say that liquid cooling is very >> very cool. >> How do you like that? >> I love that. >> I'm glad you noticed. I didn't I didn't I didn't belabor it. So, I'm glad you brought it up because I don't think most people would realize. >> I think our AI producer caught on to this. We'll see. We'll see if that works. Cory, I do have a question for you around weights and biases as a product. There used to be a product called Weights and Biases. Then the company kind of started making more products like weave, etc. And so you know the company now weights and biases but the product became weights and biases models.

107:30 >> Yes. >> Weights and biases is now back to being like one part of the stack. There's a lot of very dedicated weights and biases users. Could you talk to them and tell them like how this is getting treated as part of the forge is weights and bias is still on because they I I think they log in to 1B and now they go for a little to find it. Yeah. Exactly. Yeah. I mean look this it's I I think it's a very >> it's a tough transition, right? I think that's that's just being sort of blunt. I think the whole idea here was expanding upon the service list, right? And obviously we showed that here at fully connected that the amount of opportunity and the amount of sort of endtoend services to support this application building, this research is pretty exciting. But as part of that, we did want to create an integrated experience. And so to do that, we ended up actually having to kind of morph some of these things into place as part of Forge. But weights and mises models is the best product in in market. It's absolutely >> and it will continue to get a ton of love and a ton of new development and I would argue continue to improve as we bring some of these other services to integrate with it.

108:37 >> I have a question maybe it's an answer question. Is the CLI 1B command is that going to change or are we keeping one B for now for that CLI? >> Gosh, you know what's funny? I I don't know if we I don't know if we discussed that. Okay, cool. >> I know we should let me get back to you on that. I'll be honest. part of my >> people love the one. >> It would be interesting to talk to people about that because for me I've sort of been like with AI CLIs are dead but maybe that's not the right assumption. okay yeah because we're like and I joke about it internally I'm like oh well it doesn't really matter if the API is good as long as it's doing good work >> because AI is going to be the only thing that calls all our APIs moving forward.

109:14 and so, you know, as long as the back end is fast and easy and cheap, right, there's a little bit of like whether the API looks nice or not is maybe not relevant anymore, but maybe I'm too far down that path. >> I think there's one thing there that people just love the 1B. >> People love the one >> and also all these models were trained on all of the internet. The 1B is everywhere. Like these models like they will reach to 1B by default.

109:36 >> So the absolutely the the 1B models, the 1B experience is persisting like we will it's like that's not going away. I just want to be super clear that is persisting the CLI. let me talk with the team. We have no plans to to kill it. So not kill if people if people are loving it like you know we got to be customer first. >> Corey is there anything else that I know you're very short on time. I know you have like stuff running around. U >> this is my most important >> I will ask you though right so you have a personal assistant. You're a VP here at like shout out cheese. Let's let's call her out. we're now entering the era of the personal assistant for people open just now dots mus before we talk on the show a lot about like wol from is like very agent build assistant build how do you see this world how do you see the assistant AI world are you using any yourself and where do you see us supporting the companies that are building systems for people for billions of people now >> yeah I mean so so first of all I I do I I do use various sort of supporting AI assistant tools I don't have a master one right now and so I think I continue to sort of investigate but but I use a lot of sort of the supporting tools and a lot of sort of the productivity tools around it.

110:49 >> You know I I look I think where core fits in there is sort of in the same place that we're fitting in right now with AI applications. I think if people people have the desire to build these have the desire to give them bunch of their data and bunch of their access but may want control in how they do it right open models and sort of on a on a dedicated secure platform like coreweave is a great place to do it.

111:14 if if they're happy and comfortable and excited to use you know the open AI models or the anthropic models or what have you and agents to do it that's also great right and so it's really in some ways people have to find what their where their comfort is for for these and so I think we'll play a role in the productivity side I think we'll play a bigger role in the mission critical app side to be honest I think that's really where our sweet spot is and where the loop will really live every time I do I do this I can't like I can't stop myself from doing it >> I saw two people yesterday during I saw They were like something like this, right? Exactly. Yeah. It was It was a lot.

111:47 >> Cory, thank you so much for coming up. >> I have one quick question if you will. and excuse the bad pants, but how can we loop in the audience and get them fully connected to the forge? >> Wow, >> that was a lot. That was a lot. That was a lot. >> This how we can loop the audience and get them fully connected >> fully connected to the forge. Well, I mean, forge.cory.com. You can go to it right now. You can start a free trial with 30-day pro trial that you get for free.

112:11 >> There's no sales people in the middle. No sales. >> No sales in the middle. Do you know how many times people ask me about like hey you're in can you get us in GPUs? I'm like no I can I I don't know. >> I don't >> actually people ask me too and I'm like I don't I don't have buying I don't have buying and selling power on. >> I think I will go around now because even internally procuring GPUs is like a mess but I can expand the whatever and like try it out.

112:32 >> That's right. Yeah we do have serverless inference as part of that offering. it's a relatively small amount of dollars but a lot of tokens right so you can go try things out >> and do tell us about GPUs >> in sandbox in sandboxes. >> I'm so excited about that. You don't I I think that like this is lowkey the biggest announcement that I heard about from this whole thing. >> And did you tell Dak that? Cuz then I will hear about that from Dak.

112:58 >> I I will clip this from the >> You tell him you tell him that you Perfect. Now I'll hear about it from him. >> We know you have to run. Thank you so much for joining us. >> Thank you guys so much. I appreciate it. Yeah, absolutely. Thank you. And the demo was amazing folks. We're going to clip Cory's demo. >> Yeah. Yeah. We got to put it out there. It's >> put it out there on social.

113:11 >> I got lucky. Let's put it that way. So, >> no, it was great. Cory, thank you so much. >> Thank you so much. Thanks for having me. >> All righty, folks. you heard here from SVP of product, Cory Sanders, and the demo is >> it was really smooth. are you guys as excited about like like serverless GPUs as I am? I'm like I can't wait to be able to play with it. I need me some GPUs, man.

113:35 I think it's good for coreweave to expand their services, make it more available and not have to sign some long-term contracts in order to actually start taking advantage of the hardware they have. >> Dude, I will say this very quietly because there's a lot of sales people on this floor and Cor makes a lot of money from a lot of companies via salespeople on the floor. But I I want direct access. That's why serverless inference was dope. And now like if we have GPUs as well, I will work very hard folks look at me in this camera. I will work very hard to be able to bring some you remember how I brought inference credits to to Thursday. I will work really hard to be able to also bring GPU and sandbox credits to the the Thursday audience. So when we have drops of things, I would love for you to test out the the force theic loop etc. also Wolfbench because the newer evaluations they need GPUs and now I can do evaluations I wasn't able to do before.

114:27 >> Yeah. N go ahead. Yeah, around like a 100 agents at a time. Would I be able to just like fire up 100 sandboxes at a time? Would that be fast? >> 100 sandboxes. Well, from I don't know you you How many in parallel did you spun up when they were in beta? How many sandboxes did you run in parallel? >> 64. Yeah. >> Look, N 40,000 agents are getting packed into one stack of Vera Ruben GPU GPU CPU like combo, right? That's what we heard.

114:58 I think it depends on >> I think it depends on like what the workload is. all righty folks. I think >> I think we did it. we definitely went through some of the stuff. We interviewed four people from Cory with incredible stuff. We have some breaking news for you that the lowkey best announcement at Cory fully connected for us for engineers is GPUs on demand. Thank you Cory for making this possible. fully connected is quite incredible.

115:26 I will say I told many many folks about fully connected and gave you free tickets to the show and yesterday we had battlebots these like this is like a Vegas type big show wolf from that was super cool >> and many people who saw this online started like sending me DMs like hey can you get me in can you get me in if you're a Thursday I listener you're getting the alpha we told you about the agentic stuff told you about open plot just you know if you're here you already know that this is this is where you get the alpha and now you got the alpha about coreweave wolf from you told me one OpenAI yesterday there was a lot of protests in front of the dev day by PAI folks. For some reason they were screaming free Palestine. I don't understand the connection between that and this but there were like a lot of protests over there. You were in London.

116:12 There were protests over there. >> as well >> there's still no protest in front of the Corvully connected. so we're still low key. This is not a yeah this is not a financial advice or anything but folks cor is innovating putting in the most difficult racks in the world first putting the customers on those most difficult racks where Ruben first cor is innovating in the agentic loop you definitely should try agent lens if you tried weave before and in GPUs as well and they're innovating in in media because Thursday I part of cordway folks thank you so much for tuning in today this is Alex welcome and full Raven Wolf signing off from Corwe. If you missed any part of the show, the show comes to you through any channel that you want. We're live on YouTube.

116:58 We're going to be on on X and Spotify and Apple, wherever you choose to get your podcast. If you don't listen to podcast, there's a newsletter for you on Thursday at that news. Thank you so much, Wolf. Thank you so much for joining. This has been awesome. I love this Nistan LJ and and Peter Costa from before. Thanks for holding on the show on the digital and we will see you here next week back to our regular time 8:30 a.m. Pacific. cheers everyone.

117:22 Bye-bye. Signing off. Bye everyone.