We built SOH because system one models need a different runtime than system two models.
After building and running SOH, I'm thinking that we might only need different types of system one models and System One Harness to process many types of work, with system two harnesses handling planning, in-the-loop key decisions, and final review. That could be more efficient and effective than training one model to do everything.
Think of it as a reflex-and-brain split: the system two harness and model work like the brain, while System One Harness (SOH) and system one models work like reflexes, handling the majority of jobs without deliberation, and the brain only jumps in when something is uncertain or fails.
What I'd like to learn from the community: what use cases would you like to see from System One Harness, and how would you integrate SOH with a system two harness? Where do you see the reflex-and-brain split breaking?
We built SOH because system one models need a different runtime than system two models.
After building and running SOH, I'm thinking that we might only need different types of system one models and System One Harness to process many types of work, with system two harnesses handling planning, in-the-loop key decisions, and final review. That could be more efficient and effective than training one model to do everything.
Think of it as a reflex-and-brain split: the system two harness and model work like the brain, while System One Harness (SOH) and system one models work like reflexes, handling the majority of jobs without deliberation, and the brain only jumps in when something is uncertain or fails.
What I'd like to learn from the community: what use cases would you like to see from System One Harness, and how would you integrate SOH with a system two harness? Where do you see the reflex-and-brain split breaking?