Video: NVIDIA Engineers Nader Khalil and Carter Abdallah on AI Agents and Open Models

Indexé : 2026-05-21

[00:00:08][cheering and applause] >> Welcome to the Washington AI Network podcast from Pub Key. LET'S HEAR IT.

[00:00:17]THAT'S WHY AND THEY'RE NOT cheering for me, they're cheering for two fantastic guests that we're so honored to have courtesy of Nvidia. Thank you gentlemen so much. Let's hear it for Nader and Carter.

[00:00:32]>> [cheering] >> So that's Carter Abdulla and Nader Khalil. By the way, three Arabs on the stage, it's another headline.

[00:00:41]Come on, a cheer for that.

[00:00:44]Thank you very much and incredible technologists and engineers. We are so glad that you guys are here. We're glad that you're in town for the AI Expo and I saw you speak today, but let's start first with Brev. Let's tell everyone about Brev first.

[00:01:00]Yeah, so Brev is a developer tool that makes it really easy to use Nvidia's GPUs. It started with a simple idea that if people are turning their GPUs off at night and they're in such high demand, we could go to say a data center in Singapore and get supply and hook it up for Americans during the daytime when the supply crunch is happening. And so it focused on connecting a bunch of clouds all in one place and then we saw what do people want to do with that?

[00:01:24]Uh they want to run, let's say right now, Open Claw or it was maybe fine-tuning a model. And so we could actually set up the software for a developer as well. So we would get you a GPU, we'd deploy and it would the software would be set up. So when we give it to you it's just ready to go.

[00:01:37]And uh we were acquired by Nvidia in July 2024. So it's been >> Let's hear it for them.

[00:01:44]Those are words many in the room would like to hear.

[00:01:48]>> [laughter] >> Uh so it was great. So Carter and I were on the founding team. I was the co-founder and CEO. Um Carter and I go way back. Oh yeah, tell the story.

[00:01:56]Yeah, so uh a lot of people ask, you know, how do I know because I was the the first time in the founding engineer of a of a startup and that's a that's a big jump to make. Obviously, it's a bigger one to found one, but when you're joining that early, it's like, how do you know whether it's going to succeed or not succeed? And the answer is you don't. But Nader and I, like you said, go way back. I grew up in Albuquerque, New Mexico and when I was seven and Nader was 12, my dad recognized him in a in a Starbucks in Albuquerque, New Mexico from his how to learn Arabic videos on YouTube. It's a non-reproducible story. You can tell somebody that this is how you meet, you know, the person that you're going to, you know, build a company with.

[00:02:32]>> in summer camp and I learned about YouTube and I was also learning how to read and write Arabic. And so, put two and two together, I started making some videos. Not a lot of Lebanese people in Albuquerque, so that alone >> [laughter] >> That definitely helps.

[00:02:44]So, Brad, okay, so you explain what you're doing, but for those of us who don't code and don't know, it sounds to me like what you're saying is you're making an available AI available to anyone to use. Totally. Anyone who wants to use it and I think that audience itself has been growing, right? As the tooling gets better, it becomes more approachable.

[00:03:09]And so, it's a constant game of how do you build a good user experience, right?

[00:03:12]A good user experience means understanding your user and what's great is the user has been changing. And so, it used to be a tool that was very technical and it's progressively getting less technical because everyone can use agents and agents are a very technical audience. Carter, let's talk about agents because even in the last week, the ability for agents to talk to each other and dare I say Bitmoji and all the rest, um, and to to make not just make a lot of decisions, but I don't know I don't know the technical word to say, um, so many different steps all together. What do you call that?

[00:03:49]Yeah, I mean, I would say agents is is the right term, right? Agents comes from the word that we use with humans, which is agency, the ability to make and act on decisions. And to me, it's it's actually well understood. Like we we know how these things work. I think that uh if you look at just the model itself, that's not dissimilar to, you know, like the brain of a human, but you're so much more than a brain. If you were to do great work, you would need to use tools like a like a microphone and be able to talk to other people to be able to take on more complex problems. And so when you give an LLM, the the brain essentially, tools, and maybe you give it memory, and then you give it access to the browser so it can go look up real-time information, that ultimately is what an agent is. And of course, that is more powerful than just the model, but we're just giving it more tools, and we're doing so in a in a safe way, hopefully. And of course, that increases the capabilities of the system that you're building. And as the model gets better, so too will the agent. And as the the harness, which is the term that's used as the system around the model to give it access to again, the tools and the >> new term? Cuz when you talked about it today, I hadn't heard of it. And everyone else, did you guys all hear about harness?

[00:05:02]See?

[00:05:03]Yeah, I think there's a, you know, the way that I used to use ChatGPT, for example, is I would go and paste code that I wanted it to edit. And then when it modified my code, I would copy it, and I'd go back into my code editor and paste it. So I was essentially moving the file manually back and forth. Okay.

[00:05:21]>> What's different about an agent is rather than I me having to put my code into ChatGPT every time, I can actually put an agent into my code base.

[00:05:30]I can essentially take ChatGPT and put it into my code base. So rather than me have to move the files, it can see the files. It can see how the files work with each other, and it can think through it. And so it's just a simple idea that if you put the agent in different places, it has access to the environment. And that makes it more useful cuz you don't have to be the person essentially moving things back and forth. And so in you know, harness everything could be a buzzword.

[00:05:56]So chat GPT is a harness. It's just essentially a harness is just you know, the thing that's the interface that you're using the model through. And so your code editor if it has AI inside of it is a harness, right? In 2025 people were building their own harnesses.

[00:06:10]I think the thing that got really exciting this year and why it feels like we hit you know, an unlock is that the harnesses got so good that you can actually just use them rather than have to build your own agent from scratch.

[00:06:21]And that's why I think we're hearing the word agent and the word harness so much.

[00:06:24]Well, also the way you're describing it that any any agent can work with any other agent. How does that why? Yeah, I mean think about how humans work with each other, right? Like we have a team of engineers and we all work on the same code base. We all build Brev.

[00:06:41]And so there's one code base that's called Brev. And what happens is I'll pull some code, I'll modify it and I'll put it back for everyone.

[00:06:48]And in the same way you can plug agents into the existing workflows that we already have. All of the work we do is human to human collaboration. So plugging in agents in order to do some of our more mundane tasks just makes a lot of sense.

[00:07:01]What's the craziest thing anyone ever tried to offer you to get a GPU?

[00:07:08]What's the craziest thing?

[00:07:10]Um Like tickets to a football game or something. That's what I'm like Super Bowl tickets or something.

[00:07:16]I would offer that. Yeah, they're definitely in high demand. I can't think of the craziest thing someone's offered me, but we did offer you know, in the early days of Brev we were trying to make it really easy to get GPUs. And so I remember at one point literally in the user interface you would say that you want a GPU and purchase it. And people are used to this taking a lot of time for you to get. And so behind the scenes I have this fun video of me and the CTO just literally phone calling to try to secure the GPU while there's a loading wheel for some user.

[00:07:45]And then my favorite was of course the the progression of generations has happened incredibly fast and so we don't have to go too far back when there was the the the Ampere series of Nvidia GPUs, the A100s and the H100s, the the Hopper series was just coming out and so we would have a tile for people who wanted to request an H100 on demand on Rev, but we did not have capacity. And so if you clicked that you wanted to deploy an H100, you got Rick Rolled.

[00:08:12]>> [laughter] >> Oh man, I remember this.

[00:08:16]Okay, I don't know what you mean. What happened?

[00:08:18]>> [laughter] >> EXPLAIN EXPLAIN EXPLAIN RICK ROLLING.

[00:08:24]They didn't know. They didn't know. Who here knows what Rick Rolling is? Okay, I can Google it.

[00:08:30]>> [laughter] [clears throat] >> You would have a YouTube video of Rick Astley Never Gonna Give You Up that would play when Okay, I got it now. somebody was clicking a link that they thought was going to do something and so they thought they were going to get an H100 but of course that they were listening to Rick Astley.

[00:08:47]>> It became an internet meme. I don't know why. And so it was that you know, if you got to that it's like you you should have known better or something. And so at the time the H100 was so in demand that like no one had them. And so for our little startup to have had one would have was crazy. So we Rick Rolled. It was actually a problem cuz I remember we did eventually get H100 capacity, but it wasn't flying off the shelves and something was wrong and then we did some usability testing and we saw that we were still Rick Rolling people.

[00:09:12]>> [laughter] >> That's great. So Nvidia is making a huge push.

[00:09:18]Huge push into I feel like I lost control of this whole thing.

[00:09:21]>> [laughter] >> Huge push into open source.

[00:09:26]When you say open source, what do you mean?

[00:09:30]Yeah, so open source dates back I mean many many decades at this this And when we talk about open source when it comes to software, really what that means is doing things out in the open to where anybody can go view that code, anybody can contribute to that code, and that creates an ecosystem to where a bunch of, you know, developers across the world can contribute to this one project to make it better, to make it more secure, to do it in the sunlight.

[00:10:01]And so, as we talk about open source in AI, that comes in a number of different cases. It could be for the models themselves, and so Nvidia has the NeMo-Trons family of models, and we release the the weights, the the sort of math that encodes the model, how it thinks, why it thinks. We release the weights, but we also release the data that went into the model, we release the architecture, and what that allows it other companies to do, or even sovereign nations, you know, around the world, is to take an essentially an American open source AI model, AI stack, and customize it again, for security. I think that's incredibly important because everybody can can see what it is. If there is a vulnerability, it can be patched by people on on day one rather than lay dormant in code that that people do not understand or can't audit.

[00:10:51]And so, this has allowed essentially the the modern version of open source in artificial intelligence, which again is just the the same pattern that's been around since Linux, which was an operating system that was open sourced by a developer named Linus 30-plus years ago, and now powers over a billion devices worldwide, everything from things that are of course in the cloud to people's Android phones to the software that most government computers run on. Red Hat is actually from, you know, a Linux distribution made by a company that's in the open source.

[00:11:25]>> You know, I have a lot of empathy for, you know, policy makers and people in DC having to learn about open source and software, and honestly all of software engineering so rapidly. Like we had the the luxury of getting a decade plus to learn about all this stuff, but I think it's important to separate open source from AI. Open source is honestly it was like hippies using computers in California in like the 50s and 60s and it was this idea that if that software was your idea, if you were trying to you if we were all trying to figure out how to make sense of computers. They were new, we didn't know how to use them and so every little detail that you think about on your computer was actually really obsessed over by someone. If you think about like your cursor, the cursor on your computer is a slanted arrow. The reason why it's a slanted arrow is because some person in a research lab in like Stanford or Berkeley decided that or noticed that if you were building a cursor and it was moving and you wanted to click on something, if it was upright and we had really pixelated screens back then, then you would have a blunt that you were trying to point on. But if you have a if you have a very fat pixel, the corner is very pointy. So we angled the cursors because of old screens and that was someone's idea and because it was >> Or like maybe you can tell us whose idea that was and missed.

[00:12:35]>> [laughter] >> But all all these little details were just thought through and so open source was just the way that people in the early computing industry were able to share ideas aggressively and that's the same thing that happens now. But when you policy makers here, I'll speak for everyone, they when they hear open source, they think less controlled.

[00:12:53]>> Mhm.

[00:12:55]So I mean that couldn't be farther from the truth. The when when you see if you if there's a closed source project and you don't like the direction it's going, you don't have a choice.

[00:13:05]If there's an open source project and you think they're making a bad decision, one, you can propose the change and if that change gets rejected, you can clone it. There's a concept called forking in open source. So you fork the project and you make a variant that has your idea.

[00:13:20]And this is the benefit I mean open source is just about community, it's about getting ecosystem adoption and so the open source ecosystem is the best because the best idea wins. And so, we have really strong examples of this, like the way the internet stack is made.

[00:13:33]The internet stack actually encodes security as a as like a primitive. When you're when you're requesting anything from the internet on your computer, all of what you're requesting in the application are hidden. And these are open-source standards that it's leveraging. They're open-source American standards that it's leveraging because we were able to essentially evolve the standard out in the open. I was going to say, is the US the leader in open-source? I mean, absolutely. If you look at where all the standards that have been set, it's without a doubt America's the leader, and that's why I think it's important that we remain the leader.

[00:14:06]Continuing to contribute with Memotron models, like Carter was saying, the way that we release them all the way down to the data sets. Like, find me a company that's going to pay for data sets and then also release those.

[00:14:17]Okay. You today, your presentation over at the AI Expo was super cool when you told the story and showed us how you handled the parking tickets.

[00:14:28]Sorry to bust you out, but he has five parking tickets pending.

[00:14:32]But from me, I thought it was the best example of what Agentech is today.

[00:14:38]Can you walk these guys through it?

[00:14:39]Yeah. So, I have Open Closet up on my phone.

[00:14:43]Well, it's running on a DGX parked at home, but I can talk to it through Telegram. And I just asked it. I know I'm really bad about following up with things, so I know I have some parking tickets. And so, I asked it to find any parking tickets in San Francisco, and I gave it my license plate. And right away, it got stuck with the CAPTCHA, which is when, you know, the the internet's evolved to make sure that bots don't use the internet, which is actually really hostile to agents. And so, um I asked it to try to solve the CAPTCHA, figure out how. And so, it's tried a couple of methods. It used OCR, object character recognition, and it was failing. So, the you know, CAPTCHAs are made pretty well at this point. And so, then I just said, "You know what? Send me a screenshot." And >> And you're just saying this just like I would say it into Exactly.

[00:15:25]>> And you know, I I think a lot of the way that you should be using agents is trying to do things. The first time you do anything, just think about like an intern on your team or something. You're going to work a lot as you get them up to speed on something, but once they know it, it's it's done.

[00:15:39]And and conversely on that point, I think that interacting with these these large language models or these agents is is a different pattern I think for a lot of people to understand their true capabilities because we're so used to trying a product and if it doesn't work for us, we assume the product is broken or maybe we need to try a different product or a different vendor in the space. But in reality, you know, these are more of an iterative process as what matters describing to where if it doesn't work the first time, ask it why.

[00:16:06]If it says, "Oh, I need more access to this." or it says, "Well, I tried this."

[00:16:10]and you can say guide it. Okay, well, actually what you should be doing is this instead. And the instructions are are The amazing part about that is we all speak natural language. And so we've all done this to people and coached them.

[00:16:23]And so just reusing that skill that you've learned to to guide, you know, people, you can now use for these systems which were again trained on a bunch of human text. That's what they are. They're next token prediction for human text. Right. So you you try it 20 Was it 22 times it tried to break through CAPTCHA, right?

[00:16:41]>> Yeah. And then it tried other things.

[00:16:44]And then eventually something worked. So that's I mean, it struck me as That's like everything out there on the internet or available or agents. Like I guess we don't even say the internet anymore. You can't say internet. Um you know, came in and you just said, "Keep trying to do this." Yeah, absolutely. I mean, I think >> a different level of agency is my point.

[00:17:07]That's not agency. That's like the whole web. Well, you know what's funny? It it it feels like there is some crazy unlock with this agent, right? But it really wasn't. It When it's When I'm able to iterate with it, all that it knows is the context of what we chatted about. And so, that's literally just a text thread, right? If you wanted to show your friends some a situation at work, you want to you want them to You know, I I forward an email that to Carter, he reads it, suddenly we can talk about the context of that email.

[00:17:34]And it's no different with these with these LLMs. If you go to ChatGPT without without an A If you go to ChatGPT and you paste an email, it'll have all the context and it could it could suggest things to do. So, it's simply about giving access to tools and giving access to context that the that the LLM could have. And then suddenly that becomes an agent, but it's hard because the word agent implies that there's something crazy, but it's really just direct access to the things that you would like it to manipulate. So, for example, I gave my Open Claw a prepaid debit card.

[00:18:04]Uh that way it can pay my tickets. I know that at some point it's going to reach a credit card form and it should put some credit card details in. I don't want to give it my credit card.

[00:18:11]>> So, you said you said your credit card information and then it searched and found it. Well, no. No, no. So, I um I wanted to have a credit I wanted to have the ability to make transactions online, but I don't want to give it my credit card just cuz it's, you know, that's my credit card and for safety.

[00:18:27]And so, I um I made a I essentially it's like a kid. I gave it an allowance.

[00:18:32]I needed to give it an allowance of $500 and no nothing more where it would sting if it lost it, but it would be really useful if it used it. And so, that's why I thought of a prepaid debit card. I actually tried to I went to my bank and I tried to open an account for my Open Claw, but they wouldn't let me. They were just like, "Listen, we don't We don't know what this means."

[00:18:51]Oh, yeah. They probably have no idea what you're talking about.

[00:18:54]>> is that it has its own bank account and when it's empty I can refill it cuz now when it uses the money I have to go get a new prepaid debit card. So, I don't know. That'll come soon. I see. What about you? Do you have a debit card through Open >> I don't I don't have a debit card, but I I also have an Open Claw that's running on my DGX Spark at home and I really use it for uh you know, as an executive assistant, right? I don't have one and I don't think anybody in this room would say oh yeah I've got too much time on my hands. There's too little to do for me in my day-to-day life. And so I use it to do things that candidly I'm not great at doing like managing my calendar. And so what I do is I did I gave it its own email and its own calendar and I say hey you know I'm going to have a podcast with Tammy at you know 6:00 p.m. tonight at PubKey and it'll send me a calendar invite. And so now it's like oh it moved you know the location moved the time moved and it'll move it. And then you know I can have it you know read the emails that come inbound and say oh you know I think these are the ones I think you should follow up with. And so it's so simple right? It's something that I think everyone would benefit from in their daily life and now you know it's allowing me to to do more be more present with the people that I'm physically supposed to be with and I'm able to get more work done. It's not you know replacing anything that I already did I'm able now to actually just do more. That sounds great I got to say I like the idea of doing more and less email right? I think everyone agrees with that.

[00:20:21]Tell me about the day you got the call from Nvidia saying we're going to acquire you.

[00:20:26]>> [laughter] >> Right? We want to know. Yeah it was really funny I mean you know after some period of negotiating I actually feel really bad. So my sister was doing a PhD at USC and it was her PhD defense. And so we were all my so I flew to San Diego which is where my family's from and we're all driving up to LA to go watch her and I'm just on back and forth with you know lawyers and communications and emails and whatever and then I get the phone call which is the final offer and I accept. And this is about 30 minutes before my sister goes and defends her PhD. And so my mom and you know we're an immigrant family so but you know I've always seen my parents work so hard and my My is waiting she's like so how did it go? And I didn't want to say anything because I didn't want to steal my sister's shine. And so, we're all sitting home like, "How did it go?" And I was just like, "Yeah, we accepted the offer." And my mom starts screaming.

[00:21:15]She's so happy. My sister's like, "Cool.

[00:21:17]Thanks, Jensen."

[00:21:18]>> [laughter] >> How about you?

[00:21:22]Uh it was it was fun cuz Nvidia brought the the the team inside. So, we all worked out of a a hacker house, like a Victorian standard, you know, house in San Francisco, but we're all, you know, in the the front, which was the living room, which we converted into an office.

[00:21:36]And um we convened the the team. And uh you know, certain people have heard murmuring. Certain people have had their theories. I, for example, speaking of public calendars, could see his calendar. And and there was a a code name that was not hard to decode.

[00:21:53]>> [laughter] >> A lot of meetings.

[00:21:55]>> Jensen, you were cheating.

[00:21:57]A lot of meetings with a lot of lawyers.

[00:21:59]And he's always gone, and that's really weird. He's normally very present right now. And uh he says, "So, guys, I like I want to, you know, tell you guys that uh you know, starting on this date, we're we're all going to be going to Nvidia." And uh as somebody who, you know, uh I have a photo of me not at a GTC, but at a at a gaming convention um when I was uh 15 years old, I built my first gaming PC with with an Nvidia graphics card. And I have a photo of me playing uh with uh the Nvidia logo above my head. Um and then to be able to to go into the office and actually be able to contribute now to this company that was a very big part of my childhood. Uh particularly culturally for for, you know, how I probably the reason I got into computer engineering um was truly incredible.

[00:22:43]Well, you guys are very lucky. Aren't they incredible?

[00:22:46]Thank you [applause] so much. Jensen and Carter, I'm so glad to meet you. Thank you for all you've taught us. Thanks for coming to Washington. The work you do, we look forward to seeing you many times more. Don't you want them to come back?

[00:22:59]>> [applause and cheering] >> Thanks so much.

[00:23:01]>> Thanks for having us. Thank you all for being here at Pub Theology, it's the Washington Hay Network podcast. See you soon.

[00:23:08]>> [applause and cheering] [music]