thanks for sharing! my read from Marvin's output is that is a fairly similar ratio of good insight to spurious pattern matching to incorrect leaps as I get from using standard claude to help me brainstorm. is the idea of the value Marvin mainly in the way he's applied (which definitely has value potential), or also in specific harness asks you are giving him that give stronger results than just individuals using claude? I can see that you've asked him to make links between different projects, which is interesting but also going to be valuable to figure out how to eval the effectiveness of the link-finding
Viewing post in ORI to ASI jam comments
That's helpful, thanks! And yeah, application is the focus. Specifically, application to responsibility for work, not simply capability. It's not about being able to do the work, it's about being responsible for the work being done in the context of the systems he's deployed in.
And the harness is being tuned. I started off with OpenClaw and have been mapping out the weak points of it. I'm currently exploring Hermes Agent as the next harness I use vs building a custom one. I also fine-tuned a model based on the work we've done together over the last month+ and it shows a lot of promise in helping the model work specifically better within the harness it was based on. So it ended up doing almost opus-4.6 level work for a fraction of the cost while in the openclaw harness, and a lot of that was lost when trying to run the finetune in Hermes.
This is tying into the insight I had that intelligence is driven more by networks than by individual models.