Michael Gold

I got @openclaw working on my DGX Spark. After a lot of experimentation here is my setup: - Nemotron-3-super:120B model running in ollama - Node 24 binary installed to ~/.local - Increased agents.timeoutSeconds to 6000 - Decreased context length in Ollama and Openclaw to 65536

ComfyUI NYC Meetup is BACK! March 25th @ZeroSpaceNYC Speaker lineup: ✨ One-shot character workflows with @michaelgold ✨ Real-time collective imagination with ETHEREA by James Barnes See you there! 🫡

Speed run the posters at #NVIDIAGTC with me

@jason_haugh @openclaw A difference when running local models is to anticipate that you will only be able to run one worker at a time and plan accordingly. For example, if it’s running a background coding task, it will respond to me when that job is complete.

@michaelgold @openclaw Nice setup. I'm running 8 agents on a Mac Mini M4 Pro with Claude as the backbone. Curious how Nemotron handles multi-agent coordination at 120B. The context length tradeoff is real, I've had to tune that a lot on my end too.

I got @openclaw working on my DGX Spark. After a lot of experimentation here is my setup: - Nemotron-3-super:120B model running in ollama - Node 24 binary installed to ~/.local - Increased agents.timeoutSeconds to 6000 - Decreased context length in Ollama and Openclaw to 65536

Enabling verbose mode in telegram helps to see if the agent is still working (without going into the dashboard). You'll get little messages like these

I also reduced the context length in openclaw.json

Testing out NVIDIA's NemoClaw on my DGX Spark

And bumped up the timeoutSeconds way up to 6000 (jobs like coding do take some time)

@Gavriel_Cohen Yes. But found this issue that affects the telegram bridge. The bridge expects the sandbox name to be "nemoclaw" not the default name of "my-assistant" github.com/WuKongAI-CMU/N…

@michaelgold Did you get it running in the end?

Make no mistakes

"You learned to walk in simulation, right, Olaf?" "Oh yeah!"