Artificial intelligence is moving beyond chatbots and into systems capable of taking action — and according to NVIDIA engineers Nader Khalil and Carter Abdallah, that shift is happening faster than most people realize.
During the 75th episode of the Washington AI Network Podcast hosted by Tammy Haddad, Khalil and Abdallah explored the rapid evolution of AI agents, the importance of open-source models, and how everyday users can better understand the technology shaping the future.
The conversation, recorded live at PubKeyDC, brought together two engineers who helped build Brev, the developer platform acquired by NVIDIA in July 2024. Khalil co-founded the company, while Abdallah joined as a founding engineer.

Originally designed to simplify access to NVIDIA GPUs, Brev allowed developers to launch fully configured systems across multiple cloud providers in minutes.
“It focused on connecting a bunch of clouds all in one place,” Khalil explained. “So when we give it to you, it’s just ready to go.”
Abdallah reflected on the uncertainty of joining an early-stage startup.
“How do you know whether it’s going to succeed or not succeed?” he said. “The answer is you don’t.”
A major focus of the discussion centered on the role of open-source AI models and America’s position in the global AI race.
“If there’s a closed source project and you don’t like the direction it’s going, you don’t have a choice,” Khalil said. “If there’s an open source project and you think they’re making a bad decision, you can propose the change — and if that change gets rejected, you can clone it.”
Abdallah highlighted NVIDIA’s Nemotron family of open models, which provide not only model weights, but also training data and architecture, allowing businesses and governments to customize systems for their own needs.
Khalil emphasized the geopolitical significance of maintaining U.S. leadership in AI development.
“If you look at where all the standards have been set, without a doubt, America’s the leader,” he said. “That’s why it’s important that we remain the leader.”
The pair also broke down one of the industry’s buzziest concepts: AI agents.
“When you give an LLM tools, memory, and access to the browser, that ultimately is what an agent is,” Abdallah explained, comparing large language models to the human brain — powerful on their own, but far more capable when connected to systems that allow them to act.
Khalil shared a live example of an AI agent running on a personal NVIDIA DGX Spark system that successfully located and paid his outstanding parking tickets in San Francisco. The agent reportedly attempted 22 different methods to solve an internet CAPTCHA before succeeding, highlighting both the persistence and current limitations of agentic AI systems.
The discussion also touched on “harnesses,” the surrounding infrastructure that enables AI systems to operate more independently.
“The harnesses got so good that you can actually just use them rather than have to build your own agent from scratch,” Khalil said. “That’s why we’re hearing the word agent and the word harness so much.”
Despite the complexity of the technology, both engineers stressed that AI interaction is becoming increasingly intuitive for everyday users.
“Interacting with these models is a different pattern for a lot of people to understand,” Abdallah said. “If it doesn’t work the first time, ask it why. These are more of an iterative process.”
He added that communicating effectively with AI systems relies on a skill most people already have.
“We all speak natural language,” Abdallah said. “You can now use that same skill for these systems.”
The full episode of the Washington AI Network Podcast is available on YouTube, Apple Podcasts, Spotify, and other podcast platforms.










































































You must be logged in to post a comment.