I place a lot of importance in my own projects that every agent session captures non-obvious learnings as rules and that skills learn from their executions. Interesting approach to bring this fully local! A bit unfortunate that the RAM requirements will lock some Macs out, but I guess that's not easily changeable right now.
So How do you prevent the continuously fine-tunes itself from drifting away from its original capabilities? The idea of training LoRA adapters from user corrections is interesting, and how do you balance personalization vs. catastrophic forgetting?
Well, when fine tuning it doesn't do it as one big adapter. It works like skills it learns a new skill and that becomes an adapter, that is then tested against a golden file that tests its core skills as well as the skills the ai wants to know and if it fails to complete the essential skills than the adapter gets dropped and the new training data considers the gaps in its capabilities.
Well, yeah it is fully local. Unless you want to bring in an MCP server from a frontier AI model to explain and give the solution (but that is in a heavy work in process). Thanks for the feedback and yeah the ram do lock some macs out!
I place a lot of importance in my own projects that every agent session captures non-obvious learnings as rules and that skills learn from their executions. Interesting approach to bring this fully local! A bit unfortunate that the RAM requirements will lock some Macs out, but I guess that's not easily changeable right now.
So How do you prevent the continuously fine-tunes itself from drifting away from its original capabilities? The idea of training LoRA adapters from user corrections is interesting, and how do you balance personalization vs. catastrophic forgetting?
Well, when fine tuning it doesn't do it as one big adapter. It works like skills it learns a new skill and that becomes an adapter, that is then tested against a golden file that tests its core skills as well as the skills the ai wants to know and if it fails to complete the essential skills than the adapter gets dropped and the new training data considers the gaps in its capabilities.
Yep, Thanks for your reply and it's helpful!
Well, yeah it is fully local. Unless you want to bring in an MCP server from a frontier AI model to explain and give the solution (but that is in a heavy work in process). Thanks for the feedback and yeah the ram do lock some macs out!