The Next ChatGPT? Testing NVIDIA’s Bold New AI Stack (NemoClaw)

The video begins with Jensen Huang's bold declaration at the NVIDIA GTC conference that Nemo Claw is the next major evolution in AI. This new stack is an update to the OpenClaw platform, aiming to capitalize on the massive uptick in autonomous agent usage. NVIDIA positions this technology as a direct competitor to the ubiquity of ChatGPT but with a focus on specialized stacks. The speaker introduces the goal of the video, which is to verify if Nemo Claw is as revolutionary as advertised. This section sets the stage by highlighting the high expectations surrounding NVIDIA's latest software release.

Nemo Claw is described as a secure environment designed to bridge the gap between powerful automation and the lack of oversight in standard AI agents. It functions by wrapping the agent in a 'safety sandbox' called Nvidia OpenShell, which acts as a managed infrastructure layer. The core of this system is a 'blueprint,' essentially a master Python script that governs the entire lifecycle and security guardrails of the AI. Every action, including file access and network requests, is subject to a declarative policy that triggers manual approval if a violation is detected. This design ensures that sensitive workflows remain protected while the agent performs its tasks.

The speaker demonstrates the practical setup process, starting with NVIDIA's dedicated Nemo Claw page and the 'Try Now' button. This leads to Brev, NVIDIA's preferred cloud GPU platform, which offers environments pre-installed with CUDA, Docker, and necessary drivers. To incentivize testing, NVIDIA provides $2 in free credits for new accounts, which is sufficient for a basic initial deployment. The user connects to the remote instance using the 'brev shell' command to begin the actual software installation. This part of the video emphasizes the accessibility of the hardware side of the stack.

The installation phase immediately reveals technical friction, as the default script fails to install Open Shell, requiring a manual download from GitHub. The user must provide a name for the agent, an NVIDIA API key, and select an inference model, with Nemotron being the recommended choice. Additional steps include configuring a Telegram bot token to facilitate communication between the user and the agent. Once the 'Nemo Claw Start' command is executed, a series of green checkmarks indicates that the agent's core processes are running. This section illustrates that despite the 'enterprise' branding, the current setup remains quite technical and prone to errors.

The speaker explores the Open Shell Manager, a TUI-based blueprint interface that allows users to oversee the system in real time. Almost immediately, a pending network request appears that must be manually approved for the agent to function. The workflow moves to launching a gateway to connect the agent to Telegram, but this proves to be a buggy and finicky process. Often, the user must manually kill previous gateway processes to get a new one to successfully start. This highlights a major bottleneck where the security layer interferes with the user experience and system stability.

A critical drawback discovered during testing is the extremely slow inference speed, with responses taking up to two minutes on Telegram. The speaker also encounters a '255 error code,' which requires exiting the shell and killing outstanding processes within the deployment container. To fix persistent connection issues, the user had to manually enter the Telegram token into the Open Claw configuration wizard. These 'little annoyances' added up to a significant struggle just to reach a baseline of functional stability. The segment serves as a warning about the current unrefined state of the software's performance.

In a final test, the speaker attempts to have Nemo Claw create a cron job for Hacker News articles, which reveals the 'tedious' nature of the tool. The agent's autonomy is 'kneecapped' because every individual outbound network call requires the user to switch windows and manually click 'approve'. It took thirty minutes of constant monitoring and 'babysitting' to get a simple three-minute task working correctly. While NVIDIA provides security policy commands, they are currently too limited to create sophisticated, automated rules for production environments. The video concludes that while the concept has potential, Nemo Claw is currently too unstable and complex for practical use.