NVIDIA's AI Engineers: Agent Inference at Planetary Scale and "Speed of Light" — Nader Khalil (Brev), Kyle Kranen (Dynamo)

0:00 Agents can do three things. They can access your files, they can access the internet, and then now they can write custom code and execute it. You should really only let an agent do two of those three things. If you can access your files and you can write custom code, you don't want internet access because that's one, a safety vulnerability, right?

0:15 If you have access to internet and your file system, you should know the full scope of what that agent's capable of doing, otherwise malware can get injected or s- something bad can happen. And so that's a lot of what we've been thinking about is like, you know, how do we both enable this because it's clearly the future, but then also, you know, what, what are these enforcement points that we can start to, like, protect?

0:33 [upbeat music] All right. Welcome to the Latency podcast in the Chroma Studio. Welcome to all the [laughs] guests here. Uh, we're back with our guest host, Vibhu. Welcome. Good to have you back. And our friends, uh, Nader f- and Kyle from NVIDIA. Welcome. Yeah, thanks for having us. [laughs] Yeah, thank you. Actually, I don't even know your titles. Uh, I know- [laughs]...

0:55 you're like architects something of Dynamo. Yeah. I, I'm one of the engineering leaders f- and ar- uh, architects of Dynamo. And you're director of something developers- Of developer tech. Yeah... you're the developers' developers'- [laughs]... developers guy at NVIDIA. Open source, agent marketing, Brev- Yeah, yeah... and like dev tools and stuff- Yeah... kind of in the focus.

1:12 And we're, we're kind of recording this ahead of NVIDIA GTC, which is coming to town, uh, again, uh, or taking over town [laughs], uh, which, uh, which we'll all be at. Um, and we'll talk a little bit about your sessions and stuff. Yeah. Yeah. Yeah. We're super excited for it. One of my favorite memories for Nader, like, you always do like marketing stunts.

1:28 [laughs] And like while you were at Brev, you like had this surfboard that you like went down to GTC with. [laughs] Yeah. [laughs] And like Na- uh, NVIDIA apparently liked it so much that they bought you. [laughs] Like, what, what, what was that like? What was that? Yeah. Yeah, we, we, um... Our logo was a shaka. We, we, uh, we were always just kind of like trying to keep true to who we were.

1:47 I think, you know, so in the stuff startups, you're like trying to pretend that you're a bigger, more mature company than you are, and it was actually Evan Conrad, SFConfU, who was just like, "You guys are-" Yeah, previous guest. Yeah. Yeah. Amazing. Oh, really? Yeah. Amazing. Yeah. He was just like, "Guys, you're two dudes in a room. Why are you pretending that you're not?"

2:00 [laughs] Uh, and so then we were like, "Okay, let's make the logo a shaka." We brought surfboards to our booth to GTC and the energy was great. Yeah. Um, some palm trees too. [laughs] They, they actually poked out over like the, the walls so you could, you could see the Brev booth- Oh, that's so funny... and no one else- Okay... just from very far away. Oh, so you remember it back then?

2:18 Yeah, I, I remember it. Pre-acquisition I was like, "Oh, those guys are cool." Dude, that makes sense 'cause, uh, we-- So we signed up really last minute and so we had the last booth. It was all the way in the corner. And so I was, I was worried that no one was gonna come, so that's why we had like the palm trees. We really came in with the surfboards. We even had one of our investors bring her dog- Wow...

2:35 and then she was just like walking the dog around to try to like bring energy [laughs] towards our booth. Yeah, Steph. But yeah. Yeah, she's the best. You know, as a conference organizer, I love that, right? Like, it's like, uh, everyone who sponsors a conference comes, does their booth. They're like, "We are changing the future of AI," or something, some generic bullshit.

2:51 And like, no, like actually try to stand out, make it fun, right? They-- And people still remember it after three years. Yeah. Yeah. [laughs] Yeah. You know what's so funny? I'll s- I'll send- I'll give you this clip if you wanna, if you wanna add it in, but, uh, my wife was, at the time fiancée, she was in medical school and she came to help us 'cause it was like a big moment for us and so we, we bought this Cricut.

3:07 It's like a vi- like a vinyl, uh, printer 'cause like how else are we gonna label the surfboard? So we got a surfboard. Luckily was able to purchase that on the company card. [laughs] We got a Cricut and it was just like fine-tuning for enterprises or something like that that we put on the, on the surfboard. And it's 1:00 a.m. the day before we go to GTC.

3:23 She's helping me put these like vinyl stickers on and she goes, "You son of a..." She's like, "If you pull this off, you son of a bitch." [laughs] And so, uh, right-- pretty much after the acquisition, I stitched that with the news of the acquisition and I sent it to our family group chat. Aw. [laughs] Aw. [laughs] Yeah, yeah. No.

3:38 Well, she, she made a good choice there. Was that like basically the origin story for Launchables? Is that- We- And maybe we should explain what Brev is and- Yeah. Uh, I mean, Brev is just-- it's a developer tool that makes it really easy to get a GPU. So we connect a bunch of different GPU sources. So the basics of it is like how quickly can we SSH you into a G- into a GPU?

3:54 And whenever we would talk to users, they wanted a GPU. They wanted an A100. And if you go to like any cloud provisioning page, usually it's like three pages of forms, or in the form somewhere there's a dropdown and in the dropdown there's some weird code that you know to translate to an A100. And I remember just thinking like every time someone says they want an A100, like the piece of text that they're telling me that they want is like stuffed away in the corner. Yeah.

4:16 And so we were like, "What if the biggest piece of text was what the user's asking for?" And so when you go to Brev it's just big GPU chips with the type that you want. With beautiful animations that you worked on. [laughs] Pre-- like pre you can d- like now you can just prompt it. Yeah. But back in the day- Yeah, yeah. Those were handcrafted... handcrafted artisanal code. Yeah.

4:31 I was actually really proud of that because, uh, it was in-- I, I made it in Figma. Yeah. And then I fou- I was like really struggling to figure out how to turn it from like Figma to React. So what it actually is is just an SVG and I, I have all the styles and so when you change the chip, whether it's like active or not, it changes the SVG code and that somehow like renders, like looks like it's animating but it-- we, we just have the transitions slow.

4:51 But it's just like the-- a JavaScript function to change the like underlying SVG. Yeah. And that was how I ended up like figuring out how to move it from, from Figma. But yeah, that's ar-artisan. [laughs] Speaking of marketing stunts though, he actually used those SVGs or kind of used those SVGs to make these cards. Oh, yeah. Like a GPU gift card- Yes... that he handed out everywhere.

5:13 That was actually my first impression of Nader. Yeah, yeah, yeah. Yeah, I think I still have one of them. Yeah. [laughs] They look great. Yeah. I have a ton of them still actually in our garage but just they don't have labels. We should honestly like bring, bring them back. But, um, I found this old printing press here actually just around the corner on Van Ness and it's a third-generation San Francisco shop and so I come in an excited startup founder trying to like...

5:32 And they just have this crazy old machinery and I'm in awe 'cause the, the whole building is so physical. Like you're seeing these machines. They have like pedals to like move these saws and whatever. I don't know what this machinery is. But I saw all three generations. Like there's like the grandpa, the father and the son and the son was like around my age. Oh, it's like a holy, holy trinity. Yeah, yeah.

5:49 [laughs] It, it's funny 'cause we, we-- so I just took the same SVG and we just like printed it and it's foil printing so they make a, a mold that's like an inverse of like the A100 and then they put the foil on it and then they press it into the paper. And I remember once we got them he was like, "Hey, don't forget about us."

6:04 You know, I guess like early Apple and Cisco's first business cards were all made there and so he was like, "Yeah, we, we get like the startup businesses but then as they mature they kind of go somewhere else." And so I actually, I think we were talking with marketing about like using them for some local- We should go back and make some cards. Yeah, 100%. Yeah, yeah. But yeah.

6:18 Yeah, you know, I, I remember, you know, as a very, very small Brev investor I was like- [laughs]... what-- why are we spending time like doing these like stunts for GPUs? Like, you know, I think like- [laughs]... as a, you know, typical like cloud har-hardware person you go into an AWS, you pick like T5 XXL whatever and then- Yeah... you just like confirm a list and you look at the specs. Like why animate this GPU?

6:37 And, and I, I do think like it just shows the level of care that goes throughout Brev and- Yeah... and now-- and also Dynamo. And, and NVIDIA. I think that's what the, the thing that struck me most when we first came in was like the amount of passion that everyone has. Like I think, um, you know, you talk to, you talk to Kyle, you talk to like every VP that I've met at NVIDIA goes so close to the metal.

6:56 Like I remember it was almost a year ago and like my VP asked me, he's like, "Hey, what's Cursor? And like are you using it and if so why?" And I'm just like surprised at this and he downloaded Cursor and he was asking me to help him like use it and I thought that was-- or, or like just show him what he, you know, why we were using it.

7:09 And so the amount of care that I think everyone has and the- Passion... appreci- passion and appreciation for the moment, right? This is a very unique time so it's really cool to see everyone really like, uh, appreciate that. Yeah. One thing I wanted to do before we move over to sort of like research topics and the, the, the stuff that Kyle's working on is just tell the story of the acquisition, right? Like not many people have been- [laughs]...

7:29 been through an acquisition with NVIDIA. What's it like? Uh, what-- Yeah, just an- anything you'd like to say. It's a crazy experience. I think, uh- You know, we we're... The, the, the thing that was the most exciting for us was our goal was just to make it easier for developers. We wanted to find access to GPUs, make it easier to do that, and then all...

7:47 Oh, a- actually, your question about launchables. So launchables was just make one-click expe- like, one-click deploys for any software on top of the GPU. Mm-hmm. And so what we really liked about NVIDIA was that it felt like we just got a lot more resources to do all of that. I think, uh, you know, NVIDIA's goal is to make things as easy for developers as possible, so there was a really nice, like, synergy there. I think the...

8:05 You know, when it comes to, like, an acquisition, I think the amount that the soul of the products align, I think is gonna be... is gonna speak to the success of the acquisition. Yeah. And so it in many ways feels like we're home. This is a really great outcome for us. Like, we... You know, I love brev.nvidia.com. Like, you should, you should use it. It's- It's the front page for GPUs. Yeah. Yeah.

8:23 If you want GPUs- Yeah... you go there, you get 'em. Yeah. And it's like internally it's growing very quickly. I- Yeah... I don't remember, you said some stats there. Yeah. Yeah, yeah. It's, uh... I, I wish I had the exact numbers, but, like, internally, externally, it's been growing really quickly. We've been working with a bunch of partners, with a bunch of different customers and ISVs.

8:36 If you have a solution that you want someone... that runs on a GPU and you want people to use it quickly, we can bundle it up, uh, in a launchable and make it a one-click run. If you're doing things and you want just, like, a sandbox or something to run on, right? Like, OpenCL, huge moment, super exciting. Our, uh...

8:51 And we'll talk into it more, but, you know, internally, people wanna run this, and y- you know, we have to be really careful from the security implications. Do we let this run on the corporate network? Security's guidance was, "Hey, run this on brev." It's in... You know, it's, it's, it a, it's a VM. It's sitting in the cloud, it's off the corporate network, it's isolated.

9:06 And so that's been our stance internally and externally about how to even run something like OpenCL while we figure out how to run these things securely. But- Yeah... yeah. I think there's also... Like, you almost, like, were the right team at the right time when NVIDIA is starting to invest a lot more in developer experience or what- whatever you call- Yeah. [laughs]... uh, UX or I, I don't know what you call it.

9:27 Like, software. Yeah. Like, obviously NVIDIA's always investing in software, but, like, there's like... this is, like, a different audience. Yeah. It's a wider developer base. Yeah. Yeah. Right? Yeah. Yeah. And, you know, it's funny. It's like, it's not, uh- So, like, what, what is it called internally? What, what is this that people should be aware that is going on there? Uh, what, like developer experience, or?

9:42 Yeah, yeah. Is this... it's called just- Yeah... developer experience? Or is there, like, a broader- Yeah, I think-... strategy here? In- NVIDIA, um, NVIDIA always wants to make a good developer experience. The thing is, and a lot of the technology's just really complicated. Like, it's not... It's, uh... You know, I, I think, um, the thing that's been really growing or the, the...

9:58 AI's growing and it is having a huge moment, not because, like, let's say data scientists in 2018 were quiet then and are much louder now. The pie is com- right? Yeah. There's a whole bunch of new audiences. My mom's wondering what she's doing. My sister's learned... like, taught herself how to code. Like, the, um...

10:11 You know, I, I actually think just generally AI's a big equalizer and you're seeing a more, like, technologically literate society, I guess. Like, everyone's- [laughs]... everyone's learning how to code. Uh, there isn't really an excuse for that. And so building a good UX means that you really understand who your end user is, and when your end user becomes such a wide, uh, variety of people, then you have to almost, like, reinvent the practice, right? Yeah.

10:30 You have to- And actually build more developer UX, right? Because the, there are tiers of developer base that were added. You know, the, the hackers that are building on top of OpenCL, right, for example, have never used GPU. They don't know what CUDA is. They, they, they just wanna run something. Yeah.

10:47 You need new UX that is not just, "Hey, you know, how do you program something in CUDA and run it?" And then, and then we built, you know, like, when deep learning was getting big, we built, we built Torch and, and... But so re- recently the amount of, like, layers that are added to that developer stack has just exploded because AI's become ubiquitous. Everyone's using it in different ways.

11:05 Yeah, it's moving fast in every direction, vertical, horizontally. [laughs] Yeah. Yeah. Yeah. No, you guys, you even take it down to hardware, like the DGX Spark, you know? It's, it's- Yeah... basically the same system as just throwing it up on big GPU clusters. Yeah, yeah. Yeah. It's Grace Blackwell. [laughs] [laughs] Yeah.

11:19 Uh, we saw the preview at the last year's GTC and that was one of the better performing, uh, videos so far, NVIDIA coverage so far. Awesome. This'll beat it. Um... [laughs] That was actually fun. We hope, fingers crossed. Yeah. Even when Grace Blackwell or when, um, uh, DGX Spark was first coming out, getting to be involved in that from the beginning of the developer experience, and it just comes back to like- You were involved? Yeah.

11:37 Yeah, yeah. Okay. Very directly. Oh, yeah. Say more, say more. Yeah. Yeah. I mean, from... It was just like I, I got an email. We just got thrown into the loop, and suddenly... Yeah, I, it, it was actually really funny 'cause I'm still pretty fresh from the acquisition and I'm, I'm getting an email from a bunch of the engineering VPs about, like, the new hardware GPU chip, like, we're...

11:53 or not chip, but just GPU system that we're putting out, and I'm like, "Okay, cool. Matter's now involved with this for the UX." I'm like, "What am I gonna do here?" So I remember the first meeting I was just, like, kinda quiet as I was hearing engineering VPs talk about what this box could be, what it could do, how we should use it.

12:07 And I remember, uh, one of the first ideas that people were ideating was like, "Oh, the first thing that..." It was like... I think the quote was like, "The first thing someone's gonna wanna do with this is get two of them and run a Kubernetes cluster on top of them." And I was like, "Oh, I think I know why I'm here."

12:19 [laughs] I was like, "The first thing we're doing is easy SSH into the machine." And then, and, you know, just kind of like scoping it down of like once you can do that, every- you... Like, the person who wants to run a Kubernetes cluster on two Sparks has a higher propensity for pain [laughs] than, than, you know, someone who buys it and wants to run OpenCL right now, right?

12:37 If you can make sure that that's as effortless as possible, then the rest becomes easy. So there's a tool called NVIDIA Sync that just makes the SSH connection really simple. So, you know, if you think about it, like, if you have a Mac, uh, or a PC or whatever, if you have a laptop and you buy this GPU and you wanna use it, you should be able to use it like it's a, a GPU in the cloud, right?

12:56 Um, but there's all this friction of, like, how do you actually get into that? That's part of brev's value proposition is just, you know, there's a CLI that wraps SSH and makes it simple. And so our goal is just get you into that machine really easily. And one thing we just launched at CES, it's in, it's still in, like, early access. We're ironing out some kinks, but it should be ready by GTC.

13:13 Um- You can register your Spark on Brev. And so now if you- Like remote managed local hardware Yes. Single pane of glass. Yeah, yeah. Because Brev can already m- manage other clouds anyway, right? Yeah, yeah. And you can use the Spark on Brev as well, right? Yeah. Yeah. But yeah, exactly. So, so you, so you, you set it up at home. You can run a command on it, and then it gets...

13:31 It's essentially, it'll appear in your Brev account. And then you can take your laptop to a Starbucks or to a cafe, and you'll continue to use your... You can continue to use your Spark just like any other cloud node on Brev. Yeah, yeah. And it's just like a pre-provisioned, it's like a little data center in your home. Like a little data center thing. Yeah, exactly.

13:44 Yeah. [laughs] Yeah, yeah. Tiny little data center. Ti- tiny little data center. Like the size of your phone. Yeah. One more thing before we move on to Kyle. Just have so many Jensen stories- [laughs]... and I just love, love mining Jensen stories. Uh, my favorite so far is SOL. Uh- Oh, yeah... what is, what is SOL? SOL is actually, I, I think of all the lessons I've learned, that one's definitely my favorite.

14:02 It'll always stick with you. Yeah, yeah. I, y- you know, when you're a startup, everything's existential, right? Like, we've, we've run out of money. We were, like, on the risk of r- of losing payroll. We've had to contract our team because we l- ran out of money. And so, like, um, because of that, you're really always forcing yourself to i- to, like, understand the root cause of everything.

14:20 If you get a date, if you get a timeline, you know exactly why that date or timeline is there. You're, you're pushing every boundary and, like, you're not just sa- you're not just accepting, like, a, a no just because. And so as you start to introduce more layers, as you start to become a much larger organization, SOL is es- is essentially like what is the physics, right?

14:37 The speed of light moves at a certain speed, so if light's moving some- slower, then you know something's in the way. So before trying to, like, layer reality back in of like, why can't this be delivered at some date, let's just understand the physics. What is the theoretical limit to, like, uh, how fast this can go?

14:52 And then start to tell me why. 'Cause otherwise people will start telling you why something can't be done. But actually, I think any great leader's goal is just to create urgency. There's an in- Yeah. Create compelling events, right? Yeah, yeah. A- A- SOL is a term that NVIDIA has used to instigate a compelling event. You say, "This is done. How do we get there?

15:10 What is the minimum, as much as necessary, as little as possible thing that it takes for us to get exactly here?" And it helps you just break through a bunch of noise. Yeah. Instantly. One thing I'm unclear about is can only Jensen use the SOL card? Like- Oh, no, no, no... you know, like- Everyone... get the bullshit out. Yeah. Because obviously it's Jensen, but, like, can someone else be like- Yeah...

15:28 "No," like- Frontline engineers use it. Okay. Every, yeah, every... I think it, it's not so much about, like, get the bullshit out. It's, like, it's like give me the root understanding, right? Like, if you tell me something takes three weeks, it's like, "Well, what's-" First principles. Yeah, the first principles. It's like, what's the, what, like, why is it three weeks?

15:41 What is the actual, yeah, what's the actual limit of why this is gonna take three weeks? If you're gonna, if you, if let's say you wanted to buy a new computer and someone told you it's gonna be here in five days, what's the SOL? Well, like, the SOL is like, "I could walk into a Best Buy and pick it up for you," right? [laughs] So then anything that's, like, be- beyond that is... And is that practical?

15:57 Is that how we're gonna s- you know, let's say s- give everyone in the company a laptop? Like, obviously not. So then, like, that's the SOL, and then it's like, okay, well, if we have to get more than 10, suddenly there might be some... Right? And so now we can kind of piece the reality back. So, so this is the Paul Graham do things that don't scale. Yeah. And this is also the, what people would now call be high agency. Yeah, yeah. Yeah.

16:16 Just, just like- It's actually really interesting because there's a, there's a second hardware angle to SOL that, like, doesn't come up for all of the org. So SOL is used, like, culturally at NVIDIA- Yeah... for everything. I'm also mining for, like, I think that can be annoying sometimes when, like, someone keeps going SOL, SOL at you and you're like, "Guys, like, we have to be stable. We have to- [laughs]...

16:32 learn to fucking plan." [laughs] Yeah. I, it's- It's an interesting balance... yeah, I encountered that with, like, actually just with, with Alec, right? 'Cause we, we have a new conference, so we need to launch... We're, we have, we have goals of what we wanna launch by, uh, by the conference. And like, yeah, at the end of the day- Wait, is this GTC or something? Um, well, this is, like, so we, I mean, we did it for CES.

16:47 I see, okay. We did it for GCDC before that. We're doing it for GTC San Jose. So I mean, like, every, you know, we have a new moment, um, and we want to launch something. Yeah. And we want to do so at SOL, and that does mean that some, there's some level of prioritization that needs to happen. And so it, it is difficult, right? I think, um, you have to be careful with what you're pushing.

17:04 You know, s- stability is important and that should be factored into SOL. SOL isn't just, like, build everything and let it break. You know, that, that's part of the conversation. So as you're laying, layering in all the details, one of them might be, "Hey, we could build this, but then it's not gonna be stable for X, Y, Z reasons." Yeah.

17:19 And so that was, like, one of our conversations for CES was, you know, "Hey, like, we, we can get this into early access, registering your Spark with Brev, but there are a lot of things that we need to do in order to feel really comfortable from a security perspective," right? There's a lot of networking involved before we deliver that to users. So it's like, okay, let's get this to a point where we can at least let people experiment with it. We had it in a booth.

17:39 We had it in Jensen's keynote, and then let's go iron out all the networking kinks, and that's not easy. [laughs] And so, uh, that can come later. And so that was the way that we layered that back in. Yeah. But- It's not really about saying, like, you don't have to do the, the maintenance or operational work. It's more about saying, you know... It, it's kinda like highlights how progress is incremental, right?

18:01 Like, what is the minimum thing that we can get to? And then there's SOL for, like, every component after that, but there's the SOL to get you, get you to the, the starting line, and that, that's usually how it's asked. On the other side, you know, like, SOL came out of, like, hardware at NVIDIA, right?

18:15 So SOL is, like, literally if we ran the accelerator or the GPU with, like, at basically full speed with, like, no other constraints, like, how fast we'd be able to make a program go. Yeah, yeah. Right? So- In, in, in training that, like, you know, then you work back to, like, some percentage of, like, MFU, for example. Yeah. That's a, that's a great example. Yeah.

18:33 So, like, there's an, there's an SOL MFU and then there's, like, you know, what's practically achievable. Cool. Uh, shall we move on to sort of, uh, Kyle's side? Uh, Kyle, you're coming more from the data science world. Yeah. And, uh, I, I mean, I always, whene- whenever I meet someone who's done work in tabular stuff, graph neural networks- [laughs]...

18:50 time series, these are basically when I go to NeurIPS, I go to ICML, I walk the back halls, there's always, like, a small group of graph people- Yes... small group of tabular people. Absolutely. And, like, there's no one there. And, like, [laughs] it's very, like, you know what I mean. Like- Yeah, no... like, it, it's, it's important, interesting work if you care about solving the problems that they solve. Yeah.

19:11 But everyone else is just LMS all the time. Yeah. [laughs] I mean, it's like, it's like the black hole, right? [laughs] Yeah. Has the event horizon reached this yet in NeurIPS? Um- But, like, you know, tho- those are, those are transformers too. Yeah. And, and those are also, like, interesting things.

19:24 Anyway, uh, I just wanted to spend a little bit of time on, on those, that background before we go into Dynamo, uh, proper. Yeah, sure. I took a different path to NVIDIA than Nader. I joined six years ago, seven if you count when I was an intern. So I joined NVIDIA, like, right out of college, and the first thing I jumped into was not what I'd done in during internship, which was like, you know, like- Some stuff for autonomous vehicles, like heavyweight object detection.

19:46 I jumped into, like, you know, something and I'm like, "Recommenders, this is popular." And- Oh, yeah. You did RecSys as well. Yeah, RecSys. Yeah. Yeah. I mean, that, that was the tabular data at the time, right? You have tables of, like, audience qualities and item qualities, and you're trying to figure out, like, which member of the audience matches which item or, or more practically which item matches which member of the audience.

20:05 And at the time really it was like we were trying to enable, uh, recommenders which had historically been like a little bit of a CP-based workflow into something that, like, ran really well on GPUs. And it's since been done. Like, there are a bunch of libraries for RecSys that run on GPUs. Uh, the common models like deep learning recommendation model which came out of Meta and the wide and deep model which was used or was re-released by Google were very accelerated by GPUs using, you know, the fast HBM on the chips, especially to do, you know, vector lookups.

20:34 But it was very interesting at the time and super, super relevant because, like, we were starting to get, like, this explosion of feeds and things that required rec- recommenders to just actively be on all the time, and sort of transitioned that a little bit towards graphical networks when I discovered them because I was like, "Okay, you can actually use graphical neu- neural networks to represent, like, relationships between people, items, concepts."

20:57 And that, that interested me, so I jumped into that at NVIDIA and, and got really involved for like two-ish years. Yeah. Uh- And something I learned from Brian Cannizzaro- Yeah... is that you can just kinda choose your own path in NVIDIA. Oh my God. Yeah. [laughs] Which is not a normal big corp thing. Yeah. Like, you, you have a lane, you stay in your lane.

21:14 I think probably the reason why I enjoy being in a, a big company- The mission's the boss... coming from a startup guy, yeah. Like- The mission is the boss. [laughs] Yeah. Uh, it feels like a big game of pickup basketball. Like, you know, if you play one... If you wanna play basketball you just go up to the court and you're like, "Hey, look, we're gonna play this game and we need three."

21:28 Yeah. And you just, like, p- find your three. That's honestly for every new initiative, that's what it feels like. [laughs] Yeah. Yeah. No. It also, like, shows, right? Like NVIDIA is just releasing state-of-the-art stuff in every domain. Yeah. Like, okay, you expect foundation models with Nemotron. Voice just randomly, like- Parakeet... top tier. Parakeet just comes out. Oh.

21:44 Another one, uh, voice demo- The NVIDIA voice team has always been producing incredible stuff. Yeah. There's always just every other domain of paper that comes out- [laughs]... data set that comes out and it's like... I mean, it also stems back to what NVIDIA has to do, right? You have to make chips years before they're actually produced, right? So you need to know. You need to really forecast out.

22:00 The design process starts, like- Exactly... three to five years before the chip gets to the market. Yeah. I, I'm curious more about what that's like, right? So, like, you have specialist teams. Is it just like, you know, people find an interest, you go in, you go deep on whatever, and that kind of feeds back into, you know, okay, we, we expect predictions? Like, the internals at NVIDIA must be crazy, right?

22:20 Yeah, yeah. You know, you, you must... Not even without selling to people, you have your own predictions of where things are going. Yeah. And they're very based, very grounded, right? Yeah. It, it, it's really interesting. So there's like two things I think that NVIDIA does which are quite interesting. Uh, one is, like, we really index into passion.

22:37 There's a big sort of organizational top-down push to, like, ensure that people are working on the things that they're passionate about. So if someone proposes something that's interesting, many times they can just email someone, like, way up the chain that they would find this relevant and say, like, "Hey, can I go work on this?" That's actually...

22:52 Like, I worked at a, a big company for a couple years before, uh- Yeah... starting on my startup journey and, like, it felt very weird if you were to, like, email out of chain, if that makes sense. Yeah. The emails at NVIDIA are like mosh pits. [laughs] Yes. [laughs] Shoot. It's just like 60 people just whatever, and like they're, they're just- Doesn't it get messy? Like reply all, you know? Oh, it gets ins- it's insane.

23:11 It's insane. But- Agents help, you know, manage the context now. [laughs] Yeah. But, but, but that's actually like... I've actually... So this is a weird thing where I used to be like, "Why would we send emails? We have Slack." I am the enti- I'm the exact opposite. I feel so bad for anyone who's, like, messaging me on Slack 'cause I'm so unresponsive. But you're email maxing. I'm, I'm email maxing now. Email.

23:26 Email is a different focus. Email is perfect because- Oh, man. We can't work together. I'm Slack. [laughs] Email is great because important threads get bumped back up, right? Yeah, yeah. Um, and so Slack doesn't do that, so I just have, like, this casino going off on the right or on the left and, like, I don't know which thread was from where or what, but, like, the threads get...

23:42 A- and then also just, like, the subject so you can have, like, working threads. I think what's difficult is, like, when you're small, if it's not 40,000 people, I think Slack will work fine. But there's... I don't know what the inflection point is. There is gonna be a point where that becomes really messy and you'll actually prefer having email 'cause you can have working threads, you can cc more than nine people in a thread. You can fork stuff.

23:59 Like, you can fork stuff, which is super nice, and just, like... Y- yeah. And so but that is part of where you can propose a plan. You can also just, like, start... Honestly, momentum's the only authority, right? So, like, if you can just start, start to make a little bit of progress and show someone something and then they can try it, that's, I think, what's been, you know, I think the most effective way to push anything for- forward and that's both at NVIDIA and I think just generally.

24:20 Yeah. There's, there's the other concept that, like, is explored a lot at NVIDIA which is this idea of a zero billion dollar business. [laughs] Like, market creation is a big thing at NVIDIA. Like- Oh, you want to go and start a zero billion dollar business? Jensen, he says, "We're completely happy investing in zero billion dollar markets. We don't care if this creates revenue.

24:38 It's important for us to know about this market. We think it will be important in the future. It can be zero billion dollars for a, a while." I'm probably mangling his words here. [laughs] But like, you know, like I'll give an example. NVIDIA's been working on autonomous driving for a, a long time. Like an NVIDIA car? No. Like, like- They've used the Mercedes, right? Yeah. They're around the HQ.

24:56 Yeah. And I think it finally just got licensed out. Now they're starting to be used quite a bit. Yeah. Yeah. But for 10 years you've been seeing- What?... Mercedes with NVIDIA logos driving around. Oh, yeah, yeah. If you're, if you're in, like, the South Bay near Santa, Santa Clara it's, it's actually pretty common. You're from South Bay. This track is very... [laughs] Yeah. Yeah.

25:11 So, um, zero billion dollar markets are, are a thing. Like, you know, Jensen- I mean, okay, look, cars are not a zero billion dollar market but yeah. [laughs] That was a bad example. I think, I think he's, he's me- messaging, uh- Zero today but- Or even, like, internally, right? Like, like it's like, uh, an org doesn't have to ruthlessly find revenue very quickly to justify their existence, right?

25:31 Like, the, a lot of the important research, a lot of the important technology being developed, that, that's kinda where- So research is very ideolo- ideologically free at NVIDIA. Yeah. Like they can pursue things that they- Were, were you research officially? I was never in research officially. Okay. I was always in engineering. Yeah. We...

25:44 I'm, I'm in an org called Deep Learning Algorithms which is basically just how do we make things that are relevant to deep learning go fast. That sounds freaking cool. And, and I think a lot of that is underappreciated, right? Like time series. This week Google put out TimeFX. A new paper, yeah. A new time series paper out. Yeah. RecSys.

25:58 Uh, semantic IDs started applying transformers, LLMs to- Yes... RecSys. And when you think the scale of companies deploying these, right? Amazon recommendations, Google web search, like it's huge scale and- Yeah... you want fast. Yeah, yeah. [laughs] So... Yeah. Actually, it's, it... I... There's a fun moment that brought me, like, full circle.

26:15 Like, uh, Amazon Ads recently gave a talk where they talked about using Dynamo for generative recommendation which was, like, super, like, weirdly cathartic for me. I'm like- Hmm... "Oh my God, I've, I've supplanted what I was working on." Like, I [laughs]... You're using LLMs now to do what I was doing five years ago. Yeah, yeah. Um- Amazing. And let's go right into Dynamo.

26:35 Uh, maybe introduce- Yeah, sure... it sort of top-down and yeah. I think at this point, a lot of people are familiar with the term of inference. Uh, like funnily enough, like I, I went from, you know, inference being like a really niche topic to being something that's like discussed on like normal people's Twitter feeds.

26:48 It's, it's on billboards here now [laughs]. Yeah. Very, very strange. Driving, driving, seeing just an inference ad on 101. Inference at scale is becoming a lot more important. Uh, we have these moments like, you know, OpenCL where you have these agents that take lots and lots of tokens but produce incredible results.

27:04 There are many different aspects of test time scaling so that, you know, you can use more inference to generate a better result than if you were to use like a short amount of inference. There's reasoning, there's re-querying, there's adding agency to the model, allowing it to call tools and use skills.

27:19 Dynamo sort of came about at NVIDIA because myself and a couple others were, were sort of talking about the, these concepts that like, you know, you have inference engines like vLLM, SGLang, TensorRT-LLM, and they have like one single copy. They, they, they sort of think about like things as like one single copy, like one replica, right? Like one version of the model.

27:40 But when you're actually serving things at scale, you can't just scale up that replica because you end up with like performance problems. There's a scaling limit to scaling up replicas, so you actually have to scale out to use a maybe some Ku- Kubernetes type terminology. We kind of realized that there was like a lot of potential optimization that we could do in scaling out and building systems for data center scale inference.

28:01 So Dynamo is this data center scale inference engine that sits on top of the frameworks like vLLM, SGLang and TensorRT-LLM, and just makes things go faster because you can leverage the economy of scale, the fact that you have KV cache, which we can define a little bit later, uh, in all these machines that is like unique and you want to figure out like the ways to maximize your cache hits or you want to employ new techniques in inference like disaggregation, which Dynamo ha- introduced to the world in, in, in March.

28:30 Not introduced, it was a academic talk beforehand, but we're, you know, one of the first frameworks to start supporting it. And we wanna like sort of combine all these techniques into sort of a modular framework that allows you to accelerate your inference at scale. By the way, Kyle and I became friends on my first day at NVIDIA and I always love 'cause like he always teaches me [laughs] new things. Um- Yeah. By the way, this is why I wanted to put two of you together.

28:55 [laughs] Yeah. I was like, "Yeah, this is, this is gonna be good." It's very, it's very different. You know, like we've, we've- Yeah... we've, we've talked to each other a bunch. Actually, you asked like why, why can't we scale up? Yeah. Model, you said mo- model replicas. Yeah. Yeah. So, so scale up means assigning more- Heavier. Yeah, heavier. Like making things heavier. Yeah.

29:08 Adding more GPUs, adding more CPUs. Scale out is just like having a barrier saying, "I'm gonna duplicate my representation of the model or representation of this s- microservice or something, and I'm gonna like replicate it many times to handle the load." And the reason that you can't scale, scale up, uh, past some points is like, you know, there, there, there are sort of hardware bounds and algorithmic bounds on, on that type of scaling.

29:32 So I'll give you a good example that's like very trivial. Let's say you're on an H100. The maximum NVLink domain for H100 for most DGX H100s is eight GPUs. Mm. Right? So if you scaled up past that, you're gonna have to figure out ways to handle the fact that now for the GPUs to communicate, you have to do it over InfiniBand, which is still very fast, but is not as fast as NVLink.

29:54 Is it like one order of magnitude, like hundreds or- It's about an order of magnitude. Yeah. Okay. Um, so- Not terrible. Yeah. I, I need to, I need to remember the, the data sheet here. Like I think it's like about 500 gigabytes- Yeah... a second unidirectional for NVLink and about 50 gigabytes a second unidirectional for InfiniBand. Uh, I, it, it depends on the, the generation.

30:17 Ju- I just wanna set this up for people who are not familiar with these kinds of like l- layers- Ah, of course, yeah... and the transfer speeds and all that. Also, maybe even just going like a few steps back before that, like most people are very familiar with, you see, uh, you know, you can use on your laptop whatever these SGLang, vLLM. [laughs] Yeah.

30:34 You can just run inference. You can scale, you can run it on a laptop. You can run on laptop. Then you get to, okay, uh, models got pretty big, right? GLM-5, they doubled the size. So, uh, what do you do when you have to go from, okay, I can get 128 gigs of memory, I can run it on a Spark. Then you have to go multi-GPU. Yeah. Okay, multi-GPU, there's some support there.

30:53 Now, if I'm a company and I don't have like... I'm not hiring the best researchers for this, right? But I need to go multi-node, right? I have a lot of servers. Okay, now there's efficiency problems, right? You can have multiple eight H100 nodes, but you know, is that as a... Like how do you do that efficiently? Yeah. How do you like represent the... How do you choose how to represent the model? Yeah, exactly. Right?

31:12 That's like, that's like a hard question everyone asks. Like how do you size... Oh, I wanna run GLM-5, which just came out, new model. There have been like four of them in the past week, by the way. Like a bunch of new models. You know why, right? DeepSeek. No comment. Oh [laughs]. Yeah. But G- G- GLM-5, right?

31:28 We, we have this new model, it's, it's of like a large size, and you have to figure out how to both scale up and scale out, right? Because you have to find the right representation that you care about. Everyone does this differently, let's be very clear. Everyone figures this out in their own path. I feel like a lot of AI or ML even is like, is like this. I think people think, you know... I, I was...

31:45 There was some tweet a few months ago that was like, "Why hasn't fine-tuning as a service taken off?" And- Mm. You know, and like- That might be me [laughs]. [laughs] It might have been you. Yeah. But people want it to be such an easy recipe to follow, but even like if you look at an MOE model- It's specific to you. Yeah. Yeah. And, and the model and the situation.

32:00 And there's so much, there's so much tinkering, right? Like when you see a model that has however many experts in the MOE model, it's like why that many experts? I don't know. They, you know, they tried a bunch of things and that one seemed to do better. And I think when it comes to how you're serving inference, you know, you have a bunch of decisions to make and there, you can always argue that you can take something and make it more optimal, but I think there's this internal calibration and appetite for continued calibration [laughs].

32:20 Yeah. And that doesn't mean like, you know, people aren't taking a shot at this, like Tinker from Thinking Machines, you know? Yeah. RL as a service. Yeah. Totally. It, it also gets even harder when you try to do big model training, right? We're not the best at training MOEs. Uh, when they're pre-trained, like we saw this with Llama 3, right?

32:35 They're trained in such a sparse way that Meta knows there's gonna be a bunch of inference done on these, right? They'll open source it, but it's very trained for what Meta infrastructure wants, right? They wanna, they wanna inference it a lot. Yeah. Now, the question to basically think about is, okay, say you wanna serve a chat application, a coding copilot, right?

32:53 You're doing a layer of RL, you're serving a model for X amount of people. It's a chat model, a coding model. Dynamo, you know, back to that, it's like- Yeah, sorry. So y- we, we sort of like jumped off of, you know- Thanks. [laughs]... jumped [laughs]. Uh, uh, on that topic, everyone has like their own journey, and I, I like to think of it as defined by like, what is the model you need, what is the accuracy you need? Actually, I, I talked to Naruto about this earlier.

33:13 There's three axes you care about. What is the quality that you're able to produce? So like are you accurate enough, or can you complete the task with enough- Performance... like high enough performance. Yeah. Yeah. Yeah. Uh, there's cost. Can you serve the model or serve your workflow? Because it's not just the model anymore, it's the workflow, it's the multi-turn with an agent cheaply enough. And then can you serve it fast enough?

33:34 And we're seeing all three of these like play out. Like we saw, we saw new models from OpenAI that, you know, are faster. You have like these new fast versions of models. You can change the amount of thinking to change the amount of quality, right? Produce more tokens, but at a higher cost and a, and a higher latency.

33:48 And really like when you start this journey of like trying to figure out how you wanna host a model, you, you, you think about three things. What is the model I need to serve? How many times do I need to call it? What is the input sequence length? What is the, what does the workflow look like on top of it? What is the SLA, what is the latency SLA that I need to achieve? Because there's usually some...

34:08 This is usually like a constant. You, you know the SLA that you need to hit. And then like you try and find the lowest cost version that hits all of these constraints. Usually, you know, you, you start with those things and you say, you, you kinda do like a bit of experimentation across some common configurations. You change the tensor parallel size, which is a form of parallelism.

34:26 I'd say it goes even deeper. First you gotta think what model. Yes. You know, there's [laughs] It's like, it's like a multi-step design process because as you said, you can, you can choose a smaller model and then do more test time scaling, and it'll eq- equate quality of a larger model because you're doing the test time scaling or you're adding a harness or something.

34:43 So yes, it, it goes way deeper than that. But from the performance perspective, like once you get to the model you, you need to host, you look at that and you say, "Hey, I have this model. I need to serve it at this speed. What is the right configuration for that?" Do you guys see the recent, uh... There was a paper I just saw like a few days ago that, uh, if you run the same prompt twice- Yes...

35:01 you're getting like double digit- Just try it again. Yeah, exactly. And you get a lot- Yeah. But the, the key thing there- [laughs]... is you give the context of the failed try, right? Yeah. So it takes a shot. Mm-hmm. And this has been like, you know, basic guidance for quite a while. Just try again 'cause, you know, it tried. [laughs] Just try again. Did you try again? All advice in life, just try again.

35:17 And I think it's a, it's a paper from Google- Try again. If I'm not mistaken, right? Yeah, yeah. I think it, it's like a seven page, little short paper. Yeah, yeah. The title's very cute, and it's just like, yeah, just try again. Give it... Yeah, ask context how, like how again. Multi-shot. You just like say like, "Hey," like you know, like, "Take, take a little bit more, take a little bit more information." Try and fail, fail. And that basic concept has gone pretty deep.

35:34 There's like, um, self distillation RL where you, you do self distillation, you do RL, and you have past failure and, you know, that gives some signal. So people take try it again not strong enough. [laughs] Uh, for, for listeners, uh, who listen to here, uh, Vibhu actually and, and I, and we run a second YouTube channel for our paper club where- Oh, that's awesome... Vi- Vibhu just covered this. Yeah. Awesome.

35:57 Self distillation and all that. That's awesome. That's why he's like so up to speed on it. Yeah, yeah, yeah. I'll have to check it out. Yeah. Yeah. It, it's just a good practice. Like everyone needs like a paper club where like you just read papers together and the social pressure just kind of forces you to just- [laughs] Yeah. We, we have... There's like a big inference reading group at Nvidia.

36:10 I feel so bad every time I, I... He put it on like, on our... He shared it. The invite. Yeah. One, one of your guys- Yeah... uh, is, is big in that. I forget. Yeah, Ishan. Ishan. Ishan. Yeah. Yeah, yeah. Yes, Ishan. Ishan's on my team. Ishan. Actually funny, there was a, there was a, there was a employee transfer between us. Ishan worked for Nadder- Yeah, yeah... at Brev, and now he, he's on my team.

36:25 He was, he was our head of AI and then yeah, once we got in, he filed the move. So, 'cause I, I'm always looking for like, okay, can, can I start a- another podcast that only does that thing? Yeah. And, uh, Ishan was... I was trying to like nudge Ishan into like, is there something here? I mean, I don't think there's, there's new inference techniques every day. So it's like, it's like [laughs] You would, you would actually be surprised.

36:41 [laughs] Um, the amount of blog posts you see, and, and if you- There, there was a period where it was like Medusa, Hydra, what Eagle, like, you know, like new- Now we have new forms of decode, spec... Uh, we have new forms of speculative decoding or new- What, what are you excited about then? And it, it's exciting when you guys put out something like NeMo-Tron, 'cause I remember the paper on this, NeMo-Tron 3, uh, the amount of like post-training, the amount of tokens that the GPU rich can just train on.

37:04 [laughs] And it, it was a hybrid state space model, right? Yeah. It's co-designed for the hardware. Yeah, co-designed for the hardware. And one of the things was always, you know, the state space models don't scale as well when you do a conversion or whatever, the performance. And you guys are like, "No, just keep training," and NeMo-Tron shows a lot of that. Yeah. [laughs] There is- Also something cool about NeMo-Tron, it was released in layers, if you will. Very similar to Dynamo.

37:25 It's, it's ess- essentially it was released as aggregate. You can... The pre-training, post-training data sets are released. Yeah. The recipes on how to do it are released. The model itself is released. It's a full open model. So you can just benefit from us churning on the GPUs. But there are companies like, uh, ServiceNow took the data set and they trained their own model, and we were super excited- Yeah...

37:41 and cele- like, you know, celebrated that work. And Zoom, Zoom, the frontier model, yeah. [laughs] Zoom is, Zoom is- Zoom is AGI. [laughs] I think, uh, you know, also just to add, like a lot of models don't put out base models and if there's that, why is fine-tuning not taking off, you know? You can do your own post-training, but- That's an interesting question. Yeah, that's true... you guys put out base model. I think you put out everything. I believe so. I don't know about base.

38:01 Base can be- Put out base... base can be cancelable. Base can be cancelable. Yeah. [laughs] Safety training. [laughs] Did we get a full picture of Dynamo? I, I don't know if we, we, we- I mean, what I'd love, what I'd love is y- you mentioned the three axes. Like break it down of like, you know, what's prefill decode and like what are the- Yeah. Yeah...

38:15 optimizations that we can get with Dynamo. Yeah. Oh. That, that, that's, that's, that's a great point. So to summarize on that three axis problem, right, there are three things that determine whether or not something can be done with inference: cost, quality, latency, right? Dynamo is supposed to be there to provide you like the runtime that allows you to pull levers to, you know, mix it up and move ar- around the Pareto frontier or the Pareto surface that determines is this actually possible with inference in AI today.

38:39 Gives you the knobs. Yeah, exactly. Gives you the knobs. Uh, and one thing that like we, we use a lot in contemporary inference and is, you know, starting to like pick up from, you know, in, in general knowledge is this co- concept of disaggregation. So historically, models would be hosted with a single inference engine, and that inference engine would ping pong between two phases.

39:02 There's prefill, where you're reading the sequence, generating KV cache, which is basically just a set of vectors that represent the sequence, and then using that KV cache to generate new tokens, which is called decode. And some brilliant researchers across multiple different papers essentially made the realization that if you separate these two phases, you actually gain some benefits.

39:23 Those benefits are basically, A, you don't have to worry about step synchronous scheduling. So the way that an inference engine works is you do one step, and then you finish it, and then you schedule, you start scheduling the next step. The- it's not like fully asynchronous. And the problem with that is you would have, uh, essentially prefill and decode are, are actually very different in terms of both their resource requirements and their sometimes their runtime.

39:46 So you would have like prefill that would like block decode steps because you, you'd still be prefilling and you couldn't schedule because, you know, the step has to end. So you remove that scheduling issue. And then you also allow you or y- yourself to like split the work into two different ki- types of pools.

40:04 So prefill typically, and, and this changes as, as model architecture changes, prefill is right now compute bound most of the time. If the sequence is sufficiently long, it's compute bound on the decode side because you're doing a full pass over all the weights and the entire sequence every time you do a decode step and you're, you don't have the quadratic computation of KB cache.

40:26 It's usually memory bound because you're retrieving a linear amount of memory and you're doing a linear amount of compute as opposed to prefill where you retrieve a linear amount of memory and then use a quadratic amount of- You know what's funny? Someone-- Exo Labs did a really cool demo- Mm-hmm... where for the DGX Spark, which has a lot more compute, you can do the pre- the compute hungry prefill on a DGX Spark and then do the decode on a, on a Mac. Yeah.

40:48 And, and so- That's faster, yeah. Yeah. [laughs] So, so you can, you can do like- And I- You can do machine stra- stratification. Yeah. And like with our future generations, generations of hardware, we actually announced like with Rubin this new accelerator that is prefill specific. It's called Rubin CPX. So- I have a question.

41:05 When you do the scale out, is scaling out easier with Dynamo because when you need a new node you can dedicate it to either the prefill or, uh, decode? Yeah. So Dynamo actually has like a, a Kubernetes component in it called Grove that allows you do, to do this like crazy scaling specialization. It has like this h- it's a representation that, I don't want to go too deep into Kubernetes here, but there was a previous way that you would like launch multi-node work. Uh, it's called leader worker set.

41:32 It's in the Kubernetes standard. And leader worker set is great. It served a lot of people super well for a long period of time, but one of the things that it struggles with is representing a set of cases where you have a multi-node replica that has a pair, right, you know, prefill and decode or it's not paired but has like a second stage that has a ratio that changes over time. And prefill and decode are like two different things.

41:56 As your workload changes, right, the amount of prefill you'll need to do may change, the amount of decode that you, you'll need to do might change, right? Like let's say you start getting like insanely long queries, right? That probably means that your prefill scales like harder because you're hitting these, this quadratic scaling growth. Yeah.

42:10 And then for listeners, like prefill will be long input, decode will be long output, for example, right? Yeah. So like decode, decode scale... I mean, decode is funny because the amount of tokens that you produce scales with the output length, but the amount of work that you do per step scales with the amount of tokens in the context. Yes.

42:27 So both scales with the input and the output. That's true. But on the prefill decode side, like if suddenly like the amount of work you're doing on the decode side stays about the same or if like scales a little bit and then the prefill side like jumps up a lot, you actually don't want that ratio to be the same.

42:40 You want it to change over time. So Dynamo has a set of components that, A, tell you how to scale. It tells you how many prefill workers and decode workers you th- it thinks you should have, and also provides a scheduling API for Kubernetes that allows you to actually represent and affect this scheduling on, on, on your actual hardware, on your compute infrastructure.

43:02 Not gonna lie, I feel a little embarrassed for being proud of my SVG function earlier. [laughs] No, it was, it was really cute. I, I like that. It's all, it's all engineering. It's all engineering. Um- Sort of technical... one thing I'm, I'm kind of just curious about with all, with you see at a systems level everything going on here. Mm-hmm.

43:17 And we're, you know, we're scaling it up in, in multi, in distributed systems. Um, I think one thing that's like kind of, of the moment right now is people are asking is there any SOL sort of upper bounds in terms of like let's call, just call it context length for want- Mm... for a better word, but you can break it down however you like. Yeah.

43:33 I just think like, well, yeah, I mean, like clearly you can engage in hybrid architectures and throw in some state space models in there all, all you want, but it looks, still looks very attention heavy. Yes. Uh, yeah, long context is attention heavy. I mean, we have these hybrid models, um, to take- And, and most, most models like cap out at a million context and that's it. Yeah. Like for the last two years has been it.

43:55 Yeah. The model hardware context co-design thing that we're seeing these days is actually super interesting. It's like my, my passion, like my secret side passion. We see models like Kimi or GPT-oSS. I'm gonna use these because I, I know specific things about these models. So Kimi 2 comes out, right? And it's an interesting model. It's like, like a DeepSeek style architecture. It is MLA.

44:15 It's basically DeepSeek scaled like a little bit differently, um, and obviously trained differently as well. But they, they talked about why they made the design choices. For context, Kimi has more experts but fewer attention heads, and I believe a slightly smaller attention, uh, like dimension, but I need to remem- I need to check that. Uh, it doesn't matter.

44:39 But they discussed this actually at length in a blog post on Zhihu, which is like, or Zhihu, which is like Reddit- Zhihu, yeah... um, in, in China. Chinese Reddit. Yeah, it is. Yeah. So it's, it's a, it's actually an incredible blog post. Uh, like all the MLsis people in, in, in that I've seen that on Zhihu are like very brilliant.

44:55 But they, they, they talk about like the creators of Kimi K2 actually like talked about it on, on, on there in a blog post and they say, "We a- we actually did an experiment," right? Attention scales with the number of heads. Obviously like if you have 64 heads versus 32 heads, you do half the work of attention. You still scale quadratically, but you do half the work.

45:15 And they made a, a very specific like sort of barter in their system, in their architecture. They basically said, "Hey, what if we gave it more experts? So we're gonna use more memory capacity." But we keep the amount of activated experts the same. We increase the expert sparsity, so we have fewer experts ac- the ratio to, of experts activated to number of experts is smaller, and we decrease the number of attention heads. And kind of for context, what the, what we had been seeing was you make models sparser instead. So no one was really touching heads.

45:45 You're just having, um- Well, they, they did, they implicitly made it sparser. Yeah, for, for Kimi they did. Yeah. Yes. They also made it sparser, but basically what we were seeing was people were at the level of, okay, there's a sparsity ratio, you want more total parameters, less active- Mm-hmm... and that's sparsity.

46:00 But what you see from papers like the labs like Moonshot, DeepSeek, they go to the level of, okay, outside of just number of experts, you can also change how many attention heads and less attention layers, more attention layers. Attention layers, yeah. Yes. Yeah, so and that's all basically coming back to just tie it together is like hardware model co-design, which is- Harder model co- model context ver- co-design. Yeah. Right?

46:20 Like if you were training a, a model that was like really, really short context, uh, or like really li- is good at super short context tasks, you may like design it in a way such that like you don't care about attention scaling because it hasn't hit that like the turning point where like the quadratic curve takes over. How do you consider attention or context as a separate part of the co-design?

46:39 Like I would imagine hardware or just how I would've thought of it is like hardware model co-design would be hardware model context co-design. Hmm. Because the harness and the context that is produced by the harness is a part of the model once it's trained in. Like even though towards the end you'll do long context, you're not changing architecture through- I see.

46:56 Yeah, yeah... training, right? I mean, you can try. [laughs] You're saying everyone's training the harness into the model? I would say to some degree- Or there's co-design for the harness... I, I know there's a small amount, but I feel like- No... not everyone has like gone full send on this thinking.

47:09 [laughs] I think, I think, I think it's important to internalize the harness that you think the model will be running- Will be running... into the model. Yeah, yeah. Interesting. Okay. Uh- And, and like bash is like the universal harness. [laughs] Right? Yeah. Like, like I'll, I'll give an example here, right? I, I mean, or just like a, like a, it's easy proof, right?

47:26 If you can train against a harness and you're using that harness for everything, wouldn't you just train with the harness to ensure that you get the best possible quality out of- Well, the, uh, I, I can provide a counterargument- Yeah, sure... which is what you want to provide a generally useful model for other people to plug into their harnesses, right? Yeah, but har- So if you-... harnesses can be open, open source, right? Yes.

47:45 We, we see stuff like- I mean, that's, that's effectively what's happening with Codex. Yeah. Uh, but like you may want like a different search tool, and then you may have to name it differently or- I don't know how much people have pushed on this, but can you train a model? Would it be-- have you, have people compared training a model for the har- for the harness versus like post-training for the harness? I, I think it's the same thing. It's the same thing. Okay.

48:03 It's just extra post-training. I see. And so, I mean, Cognition does this, Cursor does this, where you, you just have to like, if your tool is slightly different, um, either force your tool to be like the tool that they trained for- Hmm... or undo their training for their tool and then re- Oh, that's really annoying... retrain.

48:16 Yeah, it's, it's really annoying and like- [laughs] I would hope that eventually we hit like a certain level of generality with respect to harness training- Correct. Like this is not-... so you can introduce new tools... this is not AGI. Like this is- Yeah. [laughs] This is really stupid like learn my tool, bitch. Like- [laughs] I don't know if, I don't know if I can say that, but like, you know.

48:30 [laughs] Um, I think what my point kind of is, is that there's like I look at slopes of the scaling laws and like this slope is not working, man. We, we're at a million token context, okay, maybe next year two million. We're not going to a hundred trillion. You know, like this, this, this is- Oh, there are so many interesting ways you can get there... this doesn't work. This doesn't work.

48:47 I, I mean- What's kind of funny is whenever there-- I, I feel like we always want to see a trend that we can predict, but every time something's come, it's been like a leapfrog. So I, I imagine, I, I don't know how we go from one to two, but I imagine what, what's likely to happen is we break through that from some new- Yeah. There's actually, there's an interesting formalization of this.

49:05 There, there's an essay that's pretty interesting essay by Leopold Aschenbrenner called Situational Awareness. Okay, yes. [laughs] He introduces a concept into- I'm familiar with that, yes... situational awareness called an unhobbler, right? So he, he, you know, Leopold in this essay details, "Hey, I want to get, you know, like I wanna get to this point in intelligence, and I think that it is four orders of magnitude worth of like compute and data and training away."

49:28 And, you know, he says, "Oh yeah, I think data centers can scale up by about this much. I think that you can do scale up the data and some other things by this much." But one of the things that like makes the rest of that order of magnitude growth po-possible is these unhobblers, like these scientific discoveries that are discovered during, you know, model architecture search or training that really, really, really impact how, how you are able to scale.

49:55 Like a, a good example of this might be that like we see like a mo- a lot of models that are, and this is probably a very tiny unhobbler but is important for the performance perspective. We see a lot of models that are like trained with multi-token prediction natively in, during pre-training. And per DeepSeek in their paper they say, "Hey, this, this actually helped us ins-ensure sta- more stable convergence."

50:15 But there are like unhobblers that are like that, and then there are like rather large unhobblers, right? Like architecturally, a lot of our models, like we have different types of attention, and one of the problems with attention is like you have a lot of KV, but people have found like different forms of attention, like grouped query attention and, uh, like MLA in DeepSeek, multi-head latent attention, that like decrease the burden that KV has on the model, which allows you to grow like longer in context.

50:38 Yeah, and that, that was very drastic for DeepSeek. Yeah. Yeah. This was like- It, it-- for context, like the, the total, I think the total context length of DeepSeek is 128,000 tokens, or it might be 256,000 with rope extension. That entire context, I think it's 128,000, fits into eight gigabytes. And previously context, like I think the, the Llama 4 or 5B context of a similar size was like 40 or 80 gigabytes in the same precision. Yeah.

51:04 Um, so like those unhobblers like really decrease the stuff of that size, and I wouldn't be surprised if we do see the ability to like break through to like 10 million, 20 million, 100 million- Context through the, a- an unhobbler showing up. I see. And it's just science. So more deep learning algorithms is what I hear. [laughs] Yeah, more deep learning algorithms. Um- Yeah... I could, I could actually- He's playing pickup and he has room for two.

51:27 [laughs] I, I could actually give you an e- an example, like, of like a, a theory, not a theory theory, but something theoretically like that. An unhobbler that you're excited about or... Well, an, an unhobbler that, I mean, I haven't seen, so it could be a tarpit and could not, just not work. [laughs] But, uh, I, I would be really excited to see a model that does prefill and decode differently.

51:46 So a model that does- Hmm... uh, prefill like locally, like document-wise prefill, like it does it in chunks, and then you do decode globally across, like the entire sequence. Because it, logically to me, it doesn't seem like you would necessarily need to have KV be associative between documents that have like no as- no mutual association.

52:06 But that like places a lot of burden on prefill to like, or sorry, on, on decode and pure attention within the decode phase to like make those connections since the KV is like static at that point. And you see other techniques that are interesting like this too, but if, if you're able to do that, like if prefill becomes local and decode is, is still global, you solve that prefill quadratic scaling problem because you have a bunch of like small chunks that you prefill independently.

52:30 Okay. All right. Well, let's, uh, wait and see, but I, I think it'll be pretty exciting. [laughs] Fingers crossed. Yeah, fingers crossed. Yeah, yeah. I'm excited for prefill decode on separate hardware. So like- Yeah... Groq acquisition, right? Can we decode on the Groq? Can we get super fast? [laughs] I don't think I'm allowed to comment on this. [laughs] And Mark is gonna shoot arrows at us. Um- He's got a blow dart gun.

52:49 [laughs] Yeah. He's in the other room just like- Like, "Go to sleep." Yeah, yeah. Um- But I'm, I'm super excited to see the team come in and like, you know, I've gotten the, the pleasure of working with some of the, the Groq people coming in, so, you know- Yeah... I, I know- Sunny, we've had him, uh- Yeah. Oh, that's awesome... at the same conference that you were at. Yeah. Um, and, uh, I, I think you're, you guys are gonna be doing some sessions at GTC.

53:06 I don't know if you want to... This is a good place to plug them. Yeah. Yeah, yeah. So I can't speak to any LPU related sessions at GTC. I have no idea about that. Oh, no. That was you, you. On, on, on the, on the Groq side, yeah, I, I use the associative NVIDIA U. Yeah. Um, on the, on the NVIDIA Dynamo side, we're, we're giving it... There are a large number of sessions.

53:24 For those that aren't aware, you can actually search all of these sessions for GTC online. Just go to the GTC website. I don't know what the URL is, but go there. Google it. Yeah. Uh, and you can just look up Dynamo and you'll get all the sessions. There are about 20. There are a couple that are hosted by the Dynamo team.

53:40 There are a couple that are hosted by people that use Dynamo that want to show off the results they've been able to get. But there are two that I'm really excited about. Uh, one is just the general Dynamo tutorial, and this is the, I'm going out with Harry, who's our lead product manager for Dynamo, and we're sort of talking about like how to use Dynamo to get better performance and also like where we see Dynamo going in the future.

53:59 And then there's another session that I'm doing with one of our agents teams at NVIDIA to talk about sort of the future of agents in production inference. Yeah. So we're talking about there's like this new horizon with respect to agents because we have these harnesses that actually impart structure among, upon calls.

54:16 Like if you, if you compare like the past and the, and the present with respect to like how LM calls work, like in the early days when they were chatbots, like every call was like very different. There was basically no structure. You could assume that like [laughs] people you... If it was conversational, there might be like some implicit structure because you have, you know, a multi-turn conversation.

54:34 But agents you have this, this harness that like abides by rules, right? So it imparts direct structure onto the context. And you see this, there was an interesting Twitter post about how Claude Code like structures its context so that you get as many cache hits as possible. And I think it was by one of the, the PMs for Claude Code, and he, he wrote about it, and that type of structure that the harness can impart actually like goes hand-in-hand with the inference co-design.

55:02 So I'm doing a talk, I, I don't know the session name or the session number- [laughs]... but I'm, I'm doing a talk, uh, you can look at me up by name on, on the GTC website, on how we accelerate agents and where we see specific optimizations for agents going in Dynamo and in inference in general. Yeah.

55:18 I think there's only one PM for Claude Code, and it's Kai Wu. There is, there's, there's Devrel, there's Boris. Maybe it was, maybe Devrel. Yeah, exactly, right. I mean, let's go into agents. I think this is like the last part of the, the, the, this discussion we planned. Yeah, how have we not talked about agents? [laughs] Also with you guys being- Well, we scheduled it. Wait, wait, wait, wait, wait.

55:33 I was like, I was like, okay, you know, like let's have like cohesive sections or- Yeah. I mean, there's the big news, right? That NVIDIA is a huge, like deployment of Codex in terms of- Yeah. NVIDIA uses everything. Yeah. I mean, it uses Cursor, it uses Codex. But that's, that's a pretty big deployment, right? Like that's tens of thousands of people- Totally, yeah... on Cursor. So it's like- We're super excited. Yeah.

55:50 I mean, it goes back to the mosh pit of emails we kind of mentioned earlier or just the f- like, um, how fluid the org feels. So when there's new technology, people will just email it out, and everyone will try it, and if it, if it's making people's lives easier, it'll spread like wildfire. Like- Careful before your email blows up from startups. [laughs] Well, you gotta have to know the person, right? But no, I, um, I, yeah, so I mean, we-- I love using Codex. It's been a ton of fun. Yeah.

56:25 Uh, I've been using it personally, been using it at work. It's been, um, yeah, I don't know, it's been great to see the rollout. Something really funny, uh, on the day that we got, uh, Codex and Claude Code access, I found this person, uh, his name's Carlos, at the company. He wrote an Outlook CLI. Oh, yeah. And, uh, just a CLI for email.

56:42 And this was- I've been using that... yeah, maybe like four or five weeks ago. And, uh, the si-- so once I got like Codex access, I installed the CLI. It had a skill, and I just asked it to go through all of my emails, which it's very messy. [laughs] So if I don't respond to your email, I'm really sorry.

56:57 But I asked it to give me a summary, highlight any escalations that I should look at, put any thread that it thinks I should respond to in a folder, and then archive everything. And it did. So if I missed your email, it's because it didn't give me- [laughing] So I should put a prompt injection in my emails too, yeah. Yeah. What you should do is just FaceTime, which is- Yeah. Yeah.

57:16 Um, yeah, my SLA is highest on FaceTime. But that wa- it was magic, and so sent it in a big email thread to like 500 people. A bunch of folks tried it out. I started like FaceTiming whoever I could at the company to get them set up with this. Yeah. Um- That specific example- Mm-hmm... you guys deal with like some pretty... sensitive emails. Yeah.

57:33 Is there a security review with this? 'Cause like one guy made it, made it for himself, but like it's not meant for all of Nvidia's security. Yeah. The security team at Nvidia is incredible. Like shout out to them. They're, they're, they're trying to- Yeah. We have a, we have an amazing security team 'cause they're progressive and they know that this is really important technology and we have to bring it in.

57:47 If you think about like if you work at a big company, your laptop's usually very locked down. If you can only access certain things. Nvidia engineers have... Those restrictions aren't there, so you're expected to understand the risks when you try things out. And so very quickly, you know, made sure to chime in security on what we were doing.

58:01 There's actually a lot that we've been thinking about, especially with Open Cloud, right? Like there's, you know, agents can do three things. [laughs] Yeah. A- agents can do three things. They can access your files, they can access the internet, and then now they can write custom code, uh, and execute it. And you should really only let an agent do two of those three things.

58:17 If you can access your files and you can write custom code, you don't want internet access because that's one, you see, vu- vulnerability, right? If you have access to internet and your file system, you should know the full scope of what that agent's capable of doing, otherwise malware can get injected or s- something- Yeah... can happen.

58:32 And so that's a lot of what we've been thinking about is like, you know, how do we both enable this because it's clearly the future, but then also, you know, what, what are these enforcement points that we can start to like protect? And is there any directive of like, "Hey, we have a company account or a company agreement with OpenAI, we use OpenAI models here," or like choose whatever? Um, no, no.

58:49 So, so I would never put any company data in a model that's not either, that we don't even- It has the most security. Yeah. Yeah. Like how- For sure... or how that goes. Uh, you know, obviously you could run your own models. You have Nemotron and, and we did run- We have, we have an, we have an internal cluster- Yeah... so we, you know, of, of course, running a Dynamo. Uh, yeah. Yeah.

59:07 I think we're Dynamo's first customer. [laughs] Actually, uh, there's a funny story about like how I got the experience that informed what we needed for Dynamo. At one point, there's a website called build.nvidia.com, and also for us inference.nvidia.com that is, allows people to try models. It gives an API service, you can call the model with like a rest API and, you know, you get a response.

59:28 I ran the model side for that, and it was at one point the largest inference deployment and still may actually be the largest inference deployment at Nvidia. I've, I've since like handed it off to some people and they're doing a wonderful- By the way, this is an extremely under-known or less-known resource, build.nvidia.com.

59:42 You can get any of these open source models and it's rate limited, but it's free. So it's perfect for hackers to just- Yeah, yeah. And, and, and the SLA on getting models, day zero models up is like a day. Yeah. Like they're, they're incredibly good at like figuring out the right way to host the model to get it up there as soon as it comes out. You ran this? Yeah, I ran, I ran it a long time ago.

1:00:05 It was originally called Nvidia AI Playground, then it was called- Oh, yeah... AI Foundation. I remember, yeah. And then it was called build.nvidia.com. Yeah, yeah. And I ran the model side of it. So there, there was a large multi-organizational team. Mm. I, I ran how, which models should we host, how should we host them, and like what's the proportion of them.

1:00:23 And then of course there was like an SRE team that like made sure that things ran well and scaled the models as well. But I ran like, you know, model, how do we get the model to Silicon and then which mo- also worked with our product team to determine like which models were important a very long time ago. Yeah.

1:00:40 There's also like a middle ground in between, right? This is like for the hacker, try anything. There's the Brave console, then there's Dynamo. There was also NIMS, right? Yes. I remember it had its little moment like a year or two ago. Is it still- Yeah. No-... running... NIM- NIM is, uh, you know, inference, uh, model. I, I think it stands up for something. It is no longer an acronym. Yeah.

1:00:59 It's just a NIM. Just a NIM. But um, yeah, NIM is, uh, how enterprises can take our, uh, any of the, any of this technology and run it with support and all of that. And so that includes Dynamo, that includes, I don't know, all of our other optimizations that are packaged up for enterprise. Yep. Yeah.

1:01:14 Uh, anyway, so, so you got, you got a bunch of experience like running these sort of internal inference gateway playgrounds. Yeah. And Bill also built, helped build Nvidia's first internal like VS Code thing. We called it NV Code. It's like a extension? Uh, yeah, it was, it was a VS Code extension. The first, like, the fourth VS Code. [laughs] We, we jokes absolutely not.

1:01:33 It was just a while back, be like we should have a fourth VS Code hackathon where you- That's four. It's, it's the best fourth V- VS Code. Uh, we are, we are, we are doing a- How do you make a billion dollars? Someone from VS Code was there, and he was like somewhat down to get involved [laughs] and I was like- Oh, you should do that. That's- Well, I said then the cool thing became for Chrome hackathon. For Chrome.

1:01:50 And now, now, now IDs are not cool anymore. I also... What's it called? I was talking to Joseph, uh- Yeah... from Roboflow and, uh- Your partner in crime... we were talking about how with the new Alpha Mayo model, so Nvidia just released an open source, uh, the, the Mercedes cars that you saw driving. Which is crazy. Yeah.

1:02:04 You released, you open source a autonomous driving model? Uh, I already, yeah. So we were thinking like could we hackathon a driverless car? Like, I have my old car, let's just try it. [laughs] Oh God. We'll take it- Take it to like Lake Tra- Yeah... what is it? Treasure Island in the middle of the bay. Yeah. [laughs] Just like to see, like everyone, yeah. Like how many, how many cameras do we need, right?

1:02:22 Like one, two, three, four, I don't know, maybe five, six. I don't know. I, yeah. But um, I think we're gonna try. You just do it with us. All right. We could see. We could even have a race. It's like the first person to automate their [laughs] driving. Let me- Over a weekend. We do have an autonomy track at Worlds Fair. Oh. And, uh, Waymo was there. Like- Yeah...

1:02:39 Nvidia did send people, but it was for Groot. Oh. And then 'cause he didn't have the driving thing yet. Yeah. Yeah. That's, that's cool. Yeah. I think Comma, Comma also has a version of this. Comma AI, they have open source driving. They've, they've done a fun hackathon on- Is it? Me and I had talked because I, I really, what I really want is a Tesla with Tesla level self-driving- Yeah...

1:02:55 but as a smart car, like a two-seater. That's so big. Basically a wheelchair with a roof. I don't even think they make them anymore. But the thing is the, the demand has been there. Yeah. They don't make cars? No. They, they released this for like five years. Yeah. Really? Yeah. They were a different manufacturer.

1:03:10 [laughs] Based in the- I feel like it's one of those things we'll s- where we'll see someone buy the brand and it'll be revived. I, I would buy it. Like- Like a private label for [laughs] go, someone hears this, go buy your car. Yeah, yeah, yeah. That's crazy. No, because like- Mercedes... because they, they're like, I think Mercedes are Mercedes?

1:03:27 Uh, I- Mercedes used to make them. Yeah. I don't know. I feel like they- They own the brand. And you... Ow. That's actually [laughs] Your dream might come true You know. Okay, we're, we're timing it here. That's another five. And, and there's another-- and like the-- every time I s- I try to park in San Francisco, I, I have to buy a smart car because, like, 20% of the parking lots in San Francisco only fit smart cars.

1:03:48 Yeah. So I'm like- Really? Like, no, that's what-- I mean, I feel like- That's so small... even though it's late here trying to park my- This comes from someone that, like, basically does not drive. Yeah. That's why the, the Vespa was a life hack. Yeah, exactly. Yeah. You know what happened to the Vespa? Um, I used to have this yellow Vespa. Uh, I left it outside the hacker house when we moved out. It's trend.

1:04:03 Um, it's just, it was always there, and then, like, a month ago, it's not there anymore. I've been meaning to... So- I don't know. You could've let-- like, so it's actually been here. It's like thieving. [laughs] You forgot about it. Yeah. And left it. [laughs] Yeah. No, this is probably a hazard. And speaking of hackathons, I also wanted to give a, give a big shout-out to the World's Shortest Hackathon. Let's go.

1:04:21 Uh, you did twice. You did once at GTC. Done a, done a handful of times. Yeah, there's gonna be one at GTC. Oh, we're doing another one? Pretty much we have a bunch of challenges that no-- we haven't released, and you get to bring your agent to come and attempt to, uh, go through those challenges. Yeah, it's like a zero, the zero minute hackathon idea.

1:04:36 [laughs] Where you just, you just bring your-- I, I proposed it, you know, a long, long time ago. You just bring your agent, and then you press the go button. You're not allowed to code. It's just the agent doing the hackathon. [laughs] It's a good hidden evil, right? Yeah, if you make a JROV and you make- Like, I feel like this is something I would love to see from cognition or someone else be like, "Come, bring your agent." Like, drop it in.

1:04:57 'Cause you don't, you don't know, like- Supervised. Yeah... will it be a, you know, operate a browser, order a pizza? Yeah. We'll just see like that snake game, you know. Well, the-- and you don't know what the task is. Yeah, you don't know. Don't know what the task is. Like- Or just like you don't even know what the judging categories are, and then you give it the judging categories. Like try and win as much as possible.

1:05:11 This is- It's great though. It turns into like, "Yeah, so let's build something on Dynamo." It's a great business. [laughing] Okay. Funny story, actually. We have a couple of people at NVIDIA. We've been working with security to, like, bring agents really close to compute. So we now have, like, stuff where you can, like, tell Dynamo, like, "Go run some experience with Dynamo, like, on X cluster," and just, like, try it right now. Like queue up.

1:05:36 Once you get queued, like send this request load. And we've actually been able to like, just like, you know, like one-shot problems. Like we used to have this problem where, you know, with, with Dynamo you have to like find the right configurations, and we sort of do it automatically for some parts of it, but you have to like a good initial configuration that you want to use.

1:05:53 And we've just had like an agent just completely one-shot that. It goes, it gets the compute, it like runs a couple experiments. It's like, "This is the best, this is, this-- these are part of the parameter frontier. Go run this." And then we just like give that to people, and it's like faster than anything that they have. Agent UX and agent marketing are super important. There's stuff that we've been thinking a lot about. Yeah.

1:06:10 Um, Alec is like redoing the entire Rev CLI, um, so that you can fetch all the different compute types that are available. I don't know. It's gonna be really soon, but then you can, you can just browse what GPUs are available and then provision one SSH to it right there, and you can pipe all the commands. But I think it goes back to like the Alec CLI.

1:06:25 Like if you-- coding agents are-- it's kind of funny. I feel like coding agents have been so much more effective than general purpose agents, and I think a large part of that is it just has access to the terminal, like you said, and that means it has access to everything that you've installed into your terminal.

1:06:37 It can run-- So, you know, it would write code and, and it can compile the code, and if there are errors, it can fix it. It can run your suite of tests because that's all just in your terminal. And so that, you know, then that's where the idea of what got me really excited about the Alec CLI. We're now just churning through building CLIs for the entire, like for the entire business suite. Slack. Coding Slack.

1:06:54 Also a Word CLI, SAP. Go, I, I've also done that for myself recently. Really? Yeah, yeah. Um, we're gonna, we're gonna open source all of this and like yeah, all the, the-- I mean, they're just, they're, they're C- yeah, CLIs for the business applications. We would love for someone to run with this and like build like, I don't know, like open CLI foundation or something. Yeah. We-- NVIDIA would love to support, uh, anyone that's doing this.

1:07:13 Like e-every dev tool should really have good CLI support at this point. Yeah. And like at, at one point it was you want your docs to be accessible by an LLM, right? You want LLM good doc. No, every-everything needs some CLI support. Yeah. It's kind of funny, right? Like we-- like computing began with a terminal, with a shell, but we said that it's not empathetic to, uh, humans, so we built these nice user interfaces, and then now we have LLMs navigating our user interfaces, and ironically, we're not empathetic to the machine anymore.

1:07:38 [laughs] Yeah. Yeah. Just give the, the LLM access to the shell. One thing that slightly makes me uncomfortable is like, why do we have to build CLIs? Why can't we just expose APIs? Like- I, I have, I have an interesting answer to this. So there are a couple reasons. Like there's, there's like, you know, portability is like one issue.

1:07:54 Like, you know, like sometimes APIs are not like discoverable or like reachable by, by some, you know, types of things. There's some element of locality, right? Like, uh, like the CLI is like literally you interfacing with your, like, local system, which is a little bit different. You could still do it by API, but like there's this highlighting of like what is the difference between like a CLI and an MCP, right?

1:08:13 Like they kind of occupy the same purposes and you call them, it does something on the system, and, and that's done. I think that in pre-training there's just an enormous amount of- Okay... command line data. Yeah. Yeah. Like e-even let's ignore, let's, let's ignore RL. Like you're doing no harness, you're doing no harness post-training.

1:08:29 Just the amount of like CLI versus API documentation for just like navigating this world of the CLI in your file system through that is just enormous. Yeah. Yeah. Right? I think there's, there's a couple of things too. Like if, let's say we wanna-- So one, I think your intuition is right. The CLI is just wrapping the API, right? So- Functionally... functionally, right. Yeah.

1:08:49 And I think it's nice because one, you're, you're being very s- uh, specific and pedantic even, um, of what-- And that's really good 'cause you're describing the problem space, so you know what the, I don't know, I don't wanna call it like w- the, the space for vulnerability. You know what network calls you're making. It's not arbitrary, and that's not decided on the fly. That's like pre-decided, which is important from a security perspective.

1:09:09 But then if you were to write a bunch of API requests, you would probably do that, I don't know, with the model, like use Python to do so. I kind of like that everything, like a CLI is just dash because it's ubiquitous. Like it's just there, and you don't have to make sure that there's certain environment variables that are set up.

1:09:24 Like if your Python version is different than my Python version, we're using the same model to go do the same thing. Is it gonna write like different code? It probably would. And so it's kind of- It's a place to do work, right? Yeah... with a human as well. Yeah. So I think just like making those decisions happen ahead of time versus, yeah.

1:09:38 One last thing on this sort of agent, I guess maybe co-location or whatever you call it. Um, one pattern I'm tracking for this year, I always try to think about what's the theme of this year gonna be. Last year, definitely coding agents. This year is definitely coding agents breaking out of containment into broader- Real... real world. So I go definitely has to- See here, rent a human. Yeah. [laughs] Oh, yeah. Yeah.

1:09:58 I'm on there. Are you really? When I have five cent. Yeah. I'm like, "$5,000, I'll do anything." Really? I think so. I need, I need, uh, my, uh, my bowels from Costco. Uh, but I think the best part is only the agent can book me, you know? [laughs] Yeah. Yeah. It's very usually like, it's just like another labor marketplace. Uh, Mechanical Turk was this.

1:10:15 So this way, I have a weird story with why I did it. So back to your example of just giving agent access to compute, right? Yeah. You guys are GPU rich at Nvidia. Yeah. I hooked up- He's not shy about it... I have, I have a 24/7 agent running. I hooked up to RunPod. Oh. It doesn't shut down instances. [laughs] And I'm like, I've tried prompting it, I've given it instruction, "Shut down when you're done."

1:10:34 It's like, "I need to keep it warm. I'll need it soon." And it's horrible on time estimates too, 'cause like- [laughs]... they've realized it's like, "Yeah, I'll need it in 45 minutes. 45 minutes, I'll shut it down." 45 minutes of human time is actually three minute of agent time, so it's like I'm booting it up, I'm waiting. I'll just leave it on all night. And mo- model's good at shutting down after some inactivity. Mm.

1:10:52 I had it on my local server, like a little dual GPU thing. It just stays on. I have a little space heater at home now, but careful. So basically, you know, they don't care about the concept of money, just burn it. I need it. It's useful. And another way where DGX Spark will be really nice, like I, I think I'm looking at it as it's super useful for agents because yeah, you, you buy it once, you plug it in, and then it can rip.

1:11:13 I'm gonna make a, I'm gonna make an Nvidia ad here. Okay. The Blackwell, like RTX 6000 cards- Pro... Pro are only like, I think it's $8,000. Slightly cheaper. Yeah. Well, it's much, it's much cheaper than the data center cards. Yeah. [laughs] And it's got 96 gigabytes of VRAM, so if you and your, your crew want to go like run a local agent for y- you know, you, you in the home, I feel like- Hm... it's got a significant amount of VRAM.

1:11:42 I've thought about purchasing this and running it in my basement- Yeah... except my neighbors would hate me. It's just a single, like two, three slot GPU. It's more- Yeah, it's a PCIe. Yeah, it's a PCIe. GPU. You can go buy that. But I mean, the big difference against like the RTX like gaming GPUs is it, it, I mean obviously it's like Blackwell pro- like it's a pro GPU and has a lot of VRAM, which means you can run pretty large models on it.

1:12:02 You can stack four of them for the max queue in a system. But that's- That's a beast... it's beefy. You can run, uh, what is that? 96. You can run anything 96. Uh, you can run a lot. You can run DeepSeek. Uh- But also they, they are slow. They're not-- I mean, performance of speed will be somewhat slower- Relative to-... compared to API, like- Oh yeah. That, that's true. Yeah. So again- [laughs]...

1:12:24 big learning, economy of scale allows you to do things that allow you to get both speed and throughput. Like you can run... I'll give you an example. There's an optimization called WideEP. I'm not gonna go into it fully, but like it featured heavily in, in inference max for DeepSeek, and there's a, there's a great set of stories from Nvidia and from Semianalysis about like why WideEP is important.

1:12:46 But for like MoE models, it's like basically essential, and you run it like the, a level of parallelism, the level of scale up parallelism you use for it is like 32. Mm. So it goes beyond that eight barrier and it like really, really, really is important to have that M- MVL72 GB200 MV link to serve at scale. And like, it's like, I don't remember like the, you know, cost improvement.

1:13:06 I think against Hopper, right? Against Hopper with this MVL72 system, you're getting like 35 times cheaper per token for like a lot of the curve- Yeah... which is crazy. Yeah. And normalized per GPU obviously, because part of the GPU's cost or the co- the cheapest part of the cost. One thing I'm exploring is the sort of this year is also the year of the sub-agent- Mm...

1:13:29 um, where you have the main agent, but then that also kicks off tools which are in themselves agents that have limited agents. Yeah. And sort of- Small context... locally, whatever, right? Yeah. Different prompts. So for example, o- one thing that Cognition does is before you kick off a search, they do z- like a fast context model where you kick off parallel agents to search, uh, across the code base. How's that?

1:13:49 That is better than indexing a, a lot of the times, not, not all the times. And, uh, you should still index for some things. But like the idea that agents should be able to command sub-agents and probably run them like maybe close to inference as well. I, I don't know if that's like architecturally possible or even- Yeah.

1:14:05 We're, we're thinking about that for Dynamo. That's like our big theme for the year... because like, like if you can design that into your stuff, then a lot of people, a lot more people will use it. Right now it's like just kind of theoretical because you do pay a lot of like back and forth, uh, coordination costs. Yes. I, I think it'll net speed up though, right?

1:14:19 Like even at a basic level, speculative decoding, you're running a small model, you're running two instances, but it's net for- That is one example, yes. Yeah. But this is like a little bit like different with like agents. Agents. Yeah. This is not speculative. I, I think, I think there's like a summarization of that trend that I like to do, or I like to say to my team, it's like this is the year... So there are two things. This is the year system as model, right?

1:14:41 Where like instead of having like a single model be a thing, you have a system of models and components that are working together to like emulate the black box model. So when, when you make an API call to something that's like, like a multi-agent in the background, it still looks like an API call to a model. You're still getting back tokens. But under the hood.

1:14:58 Yeah, under the hood it's like a billion different models, and that's a lot of complexity. With Dynamo and with other libraries and media, we're, we're looking to help like manage that complexity. Yeah, it's funny 'cause we actually for CES, we just released the model router- Yeah... uh, for DGX Spark, where you can have a local model that's running on the Spark, and then also a foundational model, and then the model router decides when to send queries to which one. Mm.

1:15:16 So it's no longer this like either/or, it's use the best of everything that's available to you. Mm. You have a good post-training model that's running on- This relates to also the breadth functionality of being able to manage the Spark. Oh. That'd be cool. Oh, yeah. I didn't know that. [laughs] JR requested. Yeah.

1:15:31 There we go. I actually like a question like I, I like to like extend and flip over. How much longer do you guys think like agents are gonna be running? Because that's one thing I've been throwing around, like what happens when- I mean, always on It even affects the, like back to the pre-filled de- uh, decode, right? Yeah.

1:15:44 Codex is, I'd say compared to Cloud Code, it's much longer at tasks. Like- Yeah... that thing will like to run six, seven, eight hours. I'll run it overnight- Yeah... and I'll, I'll go back and I have, like, a little crappy logging software I use, and there's just times where it wants to, like... And I'm gonna go deep on research, and it'll eat up 80,000 tokens, go on another, go on another. Yeah. Just eat through tokens and, you know, that's part of it.

1:16:06 Like, at the end it does, it does hit a long task, and I think you only see that, that ex- Yeah. I- yeah. W- there's insatiable demand for tokens, and every improvement that comes kind of just makes our demand even higher. It's kind of funny, right? Like, if you have, like, a teammate and you ask them to do a task and they're like, "Should I save some effort and not think too hard about this task?" I'd be like, "Fuck no."

1:16:22 [laughs] I mean, my favorite was- Right [laughs]... you had, you can have four shots, right? Yeah. Like, the original Codex before the app, you- you, why do one call? Like, give it four attempts. Just, just use all the tokens- Yeah... till I got right. Try more. It's literally try, try again, try more [laughs] It's like, it's like the, the meter index, right, is the thing that tracks- Yes...

1:16:40 like, how long models are able to run. I expect that we'll just see, like, log linear, if not log super linear growth. We will see before the end of the year an agent that is capable of running for longer than 24 hours with, like, self-consistency the entire time. I, I would also poke at different domains having different desires, right?

1:17:00 Like at a consumer level, I'm getting slightly frustrated at 20 minutes per basic query. Sure, you can optimize, you know, six, eight hour. I don't see myself shooting off many one-week agents, right? Someone doing, like, okay, GPU kernel research or medical or biological, like, you know, in, in those domains, sure, shoot off a lot that take a modif- so, like, I think it will be somewhat domain-specific because you also really need to train that in, right?

1:17:25 You know what's funny? One of them is doing your taxes, right? Like that's tax- Yeah, it's got-... a month. Yeah, okay. Yeah, exactly. Get it right. I wonder if, like, this maybe a successful successor to, like, s- uh, speculative decoding is, like, your agent figuring out what you might be prompting it the next day at night and, like- Mm... prefetching.

1:17:40 Yeah, you can already do that. Yeah. Really? Branch, branch prediction. Oh, well, no, that... Well, that's, that's too, that's too low level, but yes. [laughs] [laughs] Sorry. Yeah, yeah, yeah. Uh- One, one question I got to get, so, like, uh, we actually did record a pod with the, the Beater folks, uh, who are s- right here.

1:17:53 Their chart is the human equivalent w- work, uh, hours of work rather than how long the agents- Yeah... themselves are, are being autonomous, and, uh, uh, there's a huge difference, right? Like human work five hours, agent work 30 minutes. Like it's actually 30 minutes, not- Yeah... five hours, right?

1:18:07 Like, so they-- that, that, that chart that you see is them estimating what the human equivalent replacement is. Uh, I think the-- I think actually Anthropic released a, a more recent chart that showed Cloud Code autonomy from their production traffic numbers, and that was 20 to 45 minutes. That's roughly where we are. So- Yeah...

1:18:23 yeah, that's the sort of realistic thing. I mean, I, I do think, like, there's experimental setups where you can just, like, Ralph Wiggum and just prompt it to keep going- Mm... uh, when it stops, and obviously you can-- that can go arbitrarily long. Really from my experience around, yeah, I guess 20 to 40 minutes seems right for when I'm using like Codex or Cloud Code, but then, like, what-- I always try to just, like, s- if I want to spin up like a new-- there's a net new project, I'll, I'll often start with Replit and, like, it'll- Yeah...

1:18:46 inferred a blank. Yeah. Yeah. Like spin up, like, the, their new, like, from the V3 agent, like, it'll spin up a web browser and, like, click around and discover new bugs and just keep churning. Um, so I think, like, my longest was, like, over an hour that I had been churning. I think before we see super long-running, I think there's gonna be a bit of an efficiency hit.

1:19:03 So sure, you can take an hour and go down paths, but you also want... you want to be more efficient, you want to be smarter in your reasoning, right? So I think that'll actually go down before we go back up. Like you don't want to scale non-optimized systems just for the heck of it. As much as I love saying use all the tokens, um, you know, they are expensive.

1:19:22 Like going from dense to reasoning models, that's an added cost, right? You're paying for a lot of tokens, and it doesn't make sense to just scale stuff that's not optimized. So there's, there's always that little balance. Yeah. But, you know. Yeah. I think you'll see both sides of it. Yeah. So 2023 was super exciting.

1:19:38 I think if you were in SF you were like, "Okay, uh, I know this is gonna be a huge world-changing moment," but it seemed like, you know, no one had known yet and maybe even before. Was it 2022 maybe? Yeah, yeah. I would say, yeah, like Rune had this tweet where, like, everyone was in SF from like 2021 to 2023- Yeah... like understood what it was like to be, like, already- Totally.

1:19:55 Um, yeah, 2021, that's when I made my first OpenAI account. Yeah. It w- um, it was crazy. And I remember it was so funny because at the time SF had not been doing well. So pretty much what it felt like was the concentration of founders in the city had r- had risen because, um, where my neighbors were used to doing a bunch of stuff, those people had all left.

1:20:11 So the only people that were still in the city were people that really wanted to build- It was cheap tech. It was, yeah, it was also way cheaper. I feel really bad anyone, uh, who is trying to get rent now. [laughs] But there was, uh, Celo was-- they had a huge office. It's a blockchain. Yeah. It, like, took over the, the old Casper building. Yeah.

1:20:28 They had the showroom and they had the, like, the, what would-- I think it was like the back warehouse. It was, it was a huge office and- It's right across from OpenAI in, in New Link. Yeah. It was in the original arena. I named the arena because of it. Yeah, yeah. And so it was really exciting because, like, Roboflow, I think, uh, I forgot. There were- MintLify. Yeah, MintLify. Yeah. Uh, Brett was there. You guys were there.

1:20:47 I remember that was actually... It was there that you bought the ai.engineer domain. Yeah. I didn't know what I was gonna do in AI. Yeah. I just, I wanted to do something. But it was kind of this... It was a really fun moment where we were kind of all in this Celo space and it, um, I don't know, it was, it was a really cool community, especially being so early. Yeah.

1:21:02 And so it- And then you got me early Cruise access. Oh, yeah [laughs]. So there was a golden period of time that both Cruise and Waymos were just free. Yeah. Always. Oh. [laughs] If you had-- I mean, they're, they're so back. Ce- Celo's opened again. Yeah. So Nature Zouk. Zouk's is doing- Nature Zouk. Zouk's, they're robotaxi. Yeah. So- Totally, yeah. Um, but yeah.

1:21:22 And so it's actually really cool that you guys have this studio so close to, uh, Celo. Yeah. This rock climbing gym right around the corner. It's like- Yeah, yeah... um, to go. Oh, yeah. So, uh, yeah, it's a, it's an awesome block. Cool. Yeah. Just, and a little, little bit of San Francisco history too, but I do think one of, one thing I try to do with the podcast is, like, bring, like, what it's like it to be in San Francisco to the rest of the world.

1:21:41 Uh, and also just like maybe give, uh, El Tepeyac Taqueria a shout out. [laughs] Yeah. My favorite tacos in the city. And- Yeah. Steak and shrimp. I know. It's very good. Yeah. And I guess what it's like to be in San Francisco, I think is just everyone seems to be super supportive. Uh, sometimes I feel like the city believes in you more than you do.

1:21:57 And even, uh, I don't know if you remember, but I remember posting my first blog post, and I had met you on Twitter, and you gave me like an hour of your time super randomly, and you kind of coached me through, uh, writing content for developers, and I was trying really hard not to come off salesy or plug myself, and so I kind of stripped all personality out of the blog post.

1:22:16 Yeah. And you, you brought that out. You're like, "People don't- Yeah... it's, it's okay to talk about what you're doing. Like, you don't have to be weird about it." And I remember just that, that I think that really helped me kind of figure out what our voice is and not shy away from it, and so always really grateful for you. Hey, you inject your voice into, like, everything now.

1:22:30 [laughs] It's actually like a, actually like a huge advantage to be, like, very genuine about what you care about. Yeah. Yeah, yeah. Like imagine like some rep- some random person DMs you and, like, is like, "Can you give me feedback on this blog post?" And it's pretty boring, and you're like, "Fine." Like, you know, he looks interesting.

1:22:43 I'll just do a Zoom call. And then you meet this guy. Yeah. Right? He's so energetic. So passionate. He'll be right there. That's it. And but, like, I think people are trained to write a certain way in school, and- Yeah... they never- Totally... see there's, like, a br- broader world out there. Lots to un- unlearn.

1:22:57 Writing, writing is thinking and, like, everyone thinks differently, so, like, you, you might as well just, like- Yeah, yeah... write your way. Cool. Well, thank you for, uh, en- indulging with us. Uh, really broad-ranging discussion, but I love, like, you guys are, like, sort of like the sort of young faces of Nvidia with this so much energy and but, like, also a lot of technical depth, and I think, uh, people learned a lot for this session, so thank you.

1:23:16 Hell yeah. This was awesome. Thank you, guys, and thank you for everything that you've done. And yeah, go to Taco. Yeah, NGA, the podcast, all the above. And, uh, see you at GTC. Hell yeah. We're looking forward to it. Can't wait. Yeah. Cool. Thanks. Awesome. Thank you. Thank you. [upbeat music]