A 256GB Mac shouldn't run out of memory. Mine did, three times, during the safe part of an AI benchmark. The one-line bug, and how I proved it from the panic.
Thanks for the comment. I try to do some experiments every few weeks since I invested so much into a MacStudio. I like to share what I’ve learned in hopes it will help others on the same journey.
The distinction you landed on is the one most agent demos skip: a panic during the unsafe part of a bench is a bug report, a panic during the safe part is a hardware story. Parallel local MOA is quietly brutal on machines too. When multi-agent pipelines get built for production use (I make custom AI agents at neu Intelligence, https://neu.ie) it is worth keeping orchestration and the heavy model calls on separate hosts: one runaway worker should degrade a queue, not take down the box the other workers live on. Curious whether the panics lined up with a specific model loading, or whether any worker could trigger it.
That's an interesting experiment!
Thanks for the comment. I try to do some experiments every few weeks since I invested so much into a MacStudio. I like to share what I’ve learned in hopes it will help others on the same journey.
The distinction you landed on is the one most agent demos skip: a panic during the unsafe part of a bench is a bug report, a panic during the safe part is a hardware story. Parallel local MOA is quietly brutal on machines too. When multi-agent pipelines get built for production use (I make custom AI agents at neu Intelligence, https://neu.ie) it is worth keeping orchestration and the heavy model calls on separate hosts: one runaway worker should degrade a queue, not take down the box the other workers live on. Curious whether the panics lined up with a specific model loading, or whether any worker could trigger it.