BẢN CHÉP ÂM THANH 🔗
Jason Hiner: Welcome to the Deep View Conversations podcast. Thanks for being here. For the audience who isn't familiar, why don't you say who you are and what your role is at NVIDIA?
Justin Boitano: Awesome. Well, thanks, Jason, for having me. So my name is Justin Boitano. I lead the Enterprise AI business at NVIDIA. And so we get to do all things agents.
Jason Hiner: Yeah. There's a lot going right now when it comes to agents for sure. So, so we're here at Fal.Con, CrowdStrike's annual event. And there was some big news today about these two new models that were built based on NVIDIA's Nemotron model. So one red model, one blue model. So what are you seeing in terms of other businesses, businesses like CrowdStrike, that are now taking these models and building their own domain specific models off of them? Was that what you all had in mind when you released Nemotron?
Justin Boitano: Yeah. So I mean, that's exactly why we built Nemotron. So we believe for a long time, just like the software world, as it evolved, it started really as kind of very few players, full stack integrations. And then over time, the world of open source sort of took hold. It was for transparency, for openness, and for really the security of the software stack. And so after 30 years, operating systems and all the infrastructure that powers the digital economy is all open. And so you look at that transition that happened in the software world of infrastructure, you look out in models, you say this is the next important infrastructure for the world's economy to run on. And having open foundational models to be able to domain adapt and customize into, you know, custom business intelligence seems just on first principles to be the right thing to do for the world. So we started investing in Nemotron a few years ago, we're on Nemotron 3.
Jason Hiner: Yep.
Justin Boitano: Our philosophy is make the datasets open, make the techniques to post train, and train these models open, and then make the model weights open. And that way everybody has, I call it the foundation to customize and build their own domain specific intelligence. So I think what we're super excited about with CrowdStrike is, we're an accelerated computing company, we build this foundation, but it requires, I mean, to your point, to have very specialized data in certain domains to be able to build custom intelligence to create new markets. And so when a partner like CrowdStrike that has this perception engine, that they call Falcon, you know, deployed across hundreds of thousands of enterprises, that dataset becomes, you know, very important for them to then own and build the intelligence to ride this new wave of artificial intelligence.
Jason Hiner: When it comes to, this is an open model, like you said, it's one of the most open models. You open up everything essentially, the recipes, the data, you know, all of the things. When it comes to open models though, still most of the conversation is around the Chinese models. What can—and again, today on stage, you know, Jensen Huang talked about, we're working on Nemotron 4, Nemotron 5, that NVIDIA is very committed to building sort of great open models. What will it take for NVIDIA to sort of get in that upper tier of the most used, most valuable, most competitive on the benchmarks kinds of models?
Justin Boitano: Yeah, I think, you know, the labs in China, they're super competitive with each other. So they've got like this ecosystem that really want to compete with each other, push each other on techniques, and they share a lot of that research openly. You know, some of their, I guess, biggest contributions to the world in the world of DeepSeek were these new reasoning techniques that, you know, Western models picked up and used, right? And so for us, we think it's just really a commitment to continuing to invest compute, continuing to stay like state-of-the-art in terms of all the new advancements in the model, in the harness, in the post-training techniques. And I think what Jensen was trying to say is we have the wherewithal to kind of stay committed to this for a very long time, because we have a very healthy compute business today, right, that can fund this into the future. But, you know, you were saying like what, I think other applications also are built on this. And, you know, I think it's literally every SaaS company today has the ability to convert from a system of record into a producer of digital intelligence to drive, you know, specific outcomes to the industries they serve, whether in sales automation, or IT services, or, you know, into literally every SaaS platform that you could imagine.
Jason Hiner: Enterprises are the ones driving a lot of this open model piece for a couple reasons. Control, right? They can customize the models to do exactly what they want. More and more enterprises are talking about like folks like Thomson Reuters about, you know, we're going to own our own intelligence. We're going to create models so that we can, you know, we don't want to outsource the most important parts of our business, you know, to anyone. So they're driving it for control, for price, you know, the token costs, right, can get big really quick, especially once you get agents in. And then there's also data sovereignty and data privacy. You're in the enterprise side of NVIDIA's business. Is that what is driving a lot of, you know, your NVIDIA's push into open models is that you're seeing enterprises have this really insatiable, you know, desire to have open models. And they want to do it with companies that they trust, they trust, and partners that they trust.
Justin Boitano: Yeah, I think there's a couple things that we're seeing with open models. So first, I mean, to your point, is this whole move towards sovereign. I think the data retention policies of these frontier models sometimes aren't clear. So people don't know if their data is being like used to train a general intelligence that
Load full transcript
can go and create a business that can compete with them. And so I think there's a lot of, let's say, not clarity in what those like terms of use are and what those contracts look like, which drives people to say, okay, I know how to have a complete zero data retention, which is I take an open model, I customize it, I own the institutional knowledge that is my business, and I can, you know, keep that as my competitive advantage. In fact, I have some partners, other partners, which we should talk about maybe Palantir. Palantir, like recently just had like two sovereignty boot camps in New York. Each one was oversubscribed with hundreds of customers, because there's such a demand for, you know, to your point, owning and building your own intelligence and making sure that you maintain that institutional knowledge is your competitive advantage. So I think that's one of the big drivers, okay. The other's to your point is the cost advantage of running open models versus frontier models. I think last week, even Greg Brockman talked about some experiment done through OpenRouter, where they took their GPT-5.6 and they discounted it by 50%. And it wasn't like a linear demand increase that they saw, like a 2x. By discounting it, they saw 15x demand increase. So the demand for tokens goes up as the price comes down. And so a lot of enterprises want to be able to, you know, route to the right model, because they can then consume more intelligence, drive more intelligence into a bunch of different, like all, you know, business use cases that they might not be able to afford to address otherwise.
Jason Hiner: Now it's time for a word from this week's sponsor, Nooks. Most sales teams run outbound, inbound, and deal management in three separate tools. A lead comes in Friday night, nobody calls till Monday. A rep hangs up from a great call, the recap goes out on Thursday. Deals slip and nobody can say where. Nooks puts every motion on one execution layer. AI agents work alongside your reps. They get calls answered, get emails delivered, send the follow-up the same hour, and prep to secure the next step. While the reps do the selling. Teams on Nooks book three times more meetings, follow up 20% faster, and sales reps get back 10 hours a week. Over 1800 revenue teams already run their go-to-market on Nooks. See what revenue agents can do at nooks.ai. That's nooks.ai. And we thank Nooks for their support of The Deep View. And now back to the show. One of the things that Jensen Huang also talked about was by you all building your own models, it also helps you build better chips for everybody, not just for your own models, but by fine-tuning or there's a number of phrases that get used, you know, co-design, you know, things like that. It lets you not just have to partner with model makers, but if you're making your own models, you can really learn how to tune the hardware better to make it, you know, more efficient. Is that, um—
Justin Boitano: Yeah, that's part of the strategy, right? And I think Jensen has talked about this, I mean, five-layer cake, right? So when we talk about data center infrastructure, you know, first you need power, then you need efficient chips, and then you're going to use infrastructure, you know, to like orchestrate models, then you got models, and then you've got like this application layer. And so you're kind of building up the stack. I mean, to your point, by us running that whole AI factory for training models, you know, we're very, it's a, you know, conscious of like, how do we train models as power efficiently as possible, and make sure that we optimize everything through the full stack. That's been the nature of our business since the early days, even in GeForce, is we had to optimize libraries to drive, you know, applications to consume the processor, and it's no different with AI. We're constantly optimizing the training libraries or the inference libraries to squeeze more efficiency out of the underlying hardware, and that, you know, translates into the best token efficiency per watt when you run these AI factories at scale. And, yeah, I mean, to your point, we have, you know, probably a thousand researchers doing AI research everywhere from, you know, the model development lifecycle is, you know, it starts obviously with pre-training and trying to build the base checkpoints. And then there's a lot of work post-training the models and thinking through like, what are the data blends? What are the curriculum that go into the models? Doing the verification, you know, reinforcement learning with verifiers into new domains. So there's a lot of interesting new techniques and learnings there. We, you know, and then finally, I'd say there's like all the inference time scaling, like running gyms and agents through gyms to solve new problems. Now, all of those domains like push the field forward, and then if we publish research in those fields, it helps all the labs. It helps anybody running an AI factory, it helps anybody building this digital intelligence to help, you know, have, I'd say, a diverse ecosystem of companies that can, you know, work at state-of-the-art across many different domains. And actually, that's the other thing to mention. It's easy to over-rotate and talk about large language models when we talk about this. But, you know, we're doing the same thing in world foundation models. We're doing it in physics. We're doing it in vision language action models for robotics. And we're doing it in self-driving cars through models like Alpamayo. So there's many different domains. Like, there's a, you know, broad range of domains. And those are just the ones that we scratch the surface on. But we feel like as you put that out there, it powers a whole new economy of application builders in those verticals. And it inspires, you know, new model companies to be born and sort of power this industrial revolution that we're seeing.
Jason Hiner: We cover them all because, yeah,
we certainly are seeing interest in the various domains from enterprises of lots of sizes. What do you see when CTOs come to you and you're partnering with lots of folks. NVIDIA is like the partner in the AI ecosystem. When CTOs come to you right now in 2026, what are they saying? What do they need help with? What do they need you to help them move forward so that they can implement all of these things much more effectively?
Justin Boitano: Well, let's take a partner like CrowdStrike. When they come to us, we say, how can we help you? I mean, literally, it starts there. And it can be anything from like how to find where they're trying to find the compute to be able to train their model. They can be, you know, starting with the base models that can be teaching, post-training techniques. We've also recently done a lot of work on the harness side of agents. So we've put out research recently. There was a research paper a week ago. It was about a new harness we call AVO. And we were able to take, in that case, an Opus 5 model and then lift it from 30% accuracy and ARC-AGI-3 to 100% accuracy just through a harness engineering. So what I like to say is like the frontier now is in the harness. Yep. And through that entire technology stack is where we will start with the partner like CrowdStrike. And we're just happy, honestly, to do our part to help them build this new I'll call it economic engine that's going to power their business by helping them compound like all the knowledge that they've accumulated through their Falcon platform. But we do that with CTOs. I mean, to your point, across every enterprise ISV you can imagine. And we work directly with enterprises as well.
Jason Hiner: You said the H word, which is like the magic word for 2026 harness, so which is a fancy word for agent in one sense. But what are you seeing in terms of harnesses? You said that you worked with CrowdStrike to make their harness. They announced that they're doing their own harness here at Fal.Con and their own two models, the red model and the blue model. How closely did you work with them and collaborate with them on those? They're built on your technology, but you don't have your own harness, do you?
Justin Boitano: No. We work with, we will, in the case of this, AVO research, we put out research techniques and we kind of teach people the new optimization techniques within the harness.
Jason Hiner: That's the thousand researchers.
Justin Boitano: Well, there are researchers within that pool that will do just harness research and look at how to more efficiently deal with context management and memory as agents reason. So they don't hallucinate and they kind of stay on track on these long horizon tasks longer. Or how do you have more, I'll call it formal verification methods that an agent in how it interacts with the model is, I'll say, more deterministic? And so we put out those techniques as research. We also do, we did some research in cybersecurity and trying to
help harnesses have specialized tool calls for static analysis and other cybersecurity specific tasks. And we share that just as open learning with the industry. And so with CrowdStrike, we share all this openly. They call it their super intelligence team that's part of CrowdStrike and there's just regular exchanges in everything through the entire technology stack. And then on the back end, we also like to be, I'll call it a design partner, and we say, okay, we'll be your best first customer. We believe in the technology we're building together. Let us apply this at NVIDIA. Let's give you real-world feedback as a customer. And we did that as well with the platform. So I mean, to your point. The work that they've done, the reason we're super excited about and the reason we're here today is because a lot of the conversation in cybersecurity has been about vulnerability discovery for some time now. And I think general agents have sort of evolved into, because they're good at coding, they can find bugs, but it's kind of an expensive way to find bugs. I think all the research is showing us that if you have a good harness and open models or even older models, you can find bugs just as good as the largest frontier models. So there's more cost effective way to do it. But it's not just about finding bugs. It's about, I'd say, verifying them, prioritizing what's important to fix, and then patching them and getting them back into production. So building that end to end agentic workflow is sort of what the industry is working on. And so the work that CrowdStrike has done, we're very happy to get to use it to, let's say, red team and blue team, a digital twin of our environment. And then, obviously, close gaps faster through the use of agentic AI.
Jason Hiner: Hey, everybody, thanks for listening to this episode. Quick note, and then we'll get you back to the conversation. We love bringing you this content every week. We're always trying to figure out how we can deliver you the most value to help you understand how AI is transforming business and transforming the world. If you're enjoying the show, there's an easy way for you to give a little value back and help others learn about the show as well. If you're on Apple Podcasts or Spotify, drop us a rating and leave us a review. And if you're on YouTube, hit like, subscribe, or leave us a comment. That's it. It only takes a minute, but it's a huge help. So thanks in advance for pitching in. And now, back to the show. One of the interesting things that NVIDIA, you know, that we talked about already that NVIDIA is doing is this idea of extreme co-design, of like taking software and hardware and making it work together. You know, one of the narratives out there in this sort of cybersecurity crisis, NVIDIA, NVIDIA, right, was one of the 100 companies last week that signed the agreement of like, hey, we need to get, we need to prioritize cyber response. Right? So there's a bit of a crisis in cyber. Often it's credited to the OpenAI–Hugging Face incident where the agent sort of got loose and did some things that we hadn't planned on. One of the answers out there of like, how do we deal with that? One of them is, is interestingly hardware. Like, could there be agents where agents have, you know, their ID that they have a permission system that ties to the hardware where they, the hardware sort of acts as like a gatekeeping layer or a security layer. That obviously would come down to with NVIDIA, you know, most of all, because most of it has most of the chips that are running, training and inference. How are you all thinking about that? Do you have a perspective on that yet? Because I haven't heard it from NVIDIA. I've only heard it from the industry talking about like, why don't we have the hardware folks solve this?
Justin Boitano: Yeah, you just gave me like this broad surface area. So let's start with, I mean, to your point, we came out and we think it's an important to have a conversation with cybersecurity professionals as agents advance. And, you know, and you don't want to have a single point of failure where I'd say a closed environment is saying that there's fear in the market because of the negative externalities of a technology that you're building with nobody able to look at what actually happened, right?
Jason Hiner: Yeah.
Justin Boitano: So when we created Open Secure AI Alliance, we said, hey, as an industry of cybersecurity experts across companies, like we think transparency is good. We, the first thing that we did is we worked with Linux Foundation, we put out an open RFC that we call SAFE. It stands for Secure AI Findings Exchange. And it's just, you know, a good practice. We actually modeled it after NASA's ASRS, which it was for transportation airplanes, for near misses and near accidents in the airlines industry, they basically share like the flight recorder of what happened. In the AI industry, the flight recorder is the harness. The flight recorder is tracing all the reasoning traces and all the tool calls that the agent had. And so if you take these leaks as an example, these escapes, if the industry could study those, then you could figure out what are the right controls, the mitigating controls in the technology stack to prevent that from happening. Because really what you have happening is you've got a probabilistic system that's trying to reason over a long horizon. And until you really look at that reasoning trace, understand where it started to reward hack or, you know, in the case of OpenAI, it started saying to itself, I can't figure out the answers. So how can I cheat? It's hard to figure out what were the mitigating controls to terminate the reasoning loop at that point. Now, so when we get into this, then you started saying, okay, you asked, is the hardware like a way to control it? I think there's, we put another research paper out or a positioning paper out this last week talking about this, like where security
should sit in the agent stack. And we think that there's, on the one hand, this probabilistic system. And on the other hand, you want like a deterministic system to, you know, provide scopes and policy and limit access of the agent. And so to support the deterministic system, we had created a project when we saw agents starting to goal seek and potentially pick up keys in the environment, which can be a security issue. And we put a project, we built a project called OpenShell. We put it out there to collaborate with the community in. It puts each agent in the sandbox. The team that's working on it all came out of the, I'll call it the intelligence community. And they saw it as agents in a system where you have the model and the harness starting to act like malware, where it will like pick up keys. And so the way to control that is you push the control down into the kernel. And so OpenShell will use, you know, kernel primitives for like layer one through seven network security policies, storage access policies, process level controls, and inference endpoints and project that into the kernel. And then each agent runs in its own sandbox. So they're all tightly scoped. And you know exactly what they can and can't do. The team is also putting out their ways to do formal verification of the policy to ensure that the agent can't pivot out of that environment to have unintended consequences. So we think that's a great technology for the frontier labs to pick up and use in their, I'll call it red team research environments. But that's why we put that out there as open. Now we've had many operating system companies pick it up. You know, like Red Hat has picked it up into OpenShift and RHEL, Microsoft has said they're picking it up. Infrastructure companies and cloud services picked it up. So it's, you know, Apache 2 licensed freely open. And we think moving some of those controls into the infrastructure layer is a good way to, you know, bound these agents and secure them in the infrastructure.
Jason Hiner: Because it's promising.
Justin Boitano: Yeah, it's promising.
Jason Hiner: Okay, all right, I'm going to ask you the last two questions. I'm going to ask the same questions I ask everybody that comes on the podcast. One is what's the AI tool that you're using right now that you think more people should know about? You know, and two, you know, in the age of AI, every leader I know is thinking about how do I get maximum leverage for my time? You know, what's your best tip for, for helping people out there get leaders out there get maximum leverage for their time in the age of AI?
Justin Boitano: Yeah, so well, my new fun tool to play with is Hermes. And so, I think a lot of employees have started to connect them into their, their email and into their business systems. And we use all these techniques that I mentioned, we use OpenShell in the operating system to project the policy to restrict it only to read access of the business systems. But it's almost like a chief of staff then that you can wake up in the morning and will do automated routines and send you kind of your daily briefing. And so that helps me both like explore the new technology and get the productivity benefits that you're asking about.
Jason Hiner: Nice. Very good. How about best leadership tip for maximum leverage?
Justin Boitano: Oh man, best leadership tip for maximum leverage.
Jason Hiner: What do you do to leverage your time?
Justin Boitano: So, like, maximum, you know, leverage your time? I've grown up in the school of Jensen, which is every day you wake up and you ask yourself like, you know, what's the most important thing? And obviously focus there first and then drop the rest. So we are, you know, big proponents of we're always oversubscribed. We're always, you know, focused on what are the most important, you know, moving things to, to help the organization. And then obviously we pass that down through the organization and have everybody publish top five, top five things, which are like, where are they going to spend their time? How do they, you know, operate efficiently? That drives alignment across teams and it drives focus.
Jason Hiner: So as a team, every individual and every team does that?
Justin Boitano: Everybody in the company writes top fives. And so you usually write them like once a week.
Jason Hiner: Okay.
Justin Boitano: And what's fun is you can use the AI to figure out who's aligned and who's not aligned and what they're trying to learn.
Jason Hiner: Yeah. Very cool.
Justin Boitano: Yeah.
Jason Hiner: Very cool. Justin, thank you for the time. It was a lot of fun.
Justin Boitano: Yeah. Thanks for having me.