sidbing 🪽

I'm seeing a few "But WHY do we need Hermes Agent when Claw & clones exist" posts Let me give you a hint: Notice the Atropos integration? Notice the Tinker training backend? What happens when an agent can not only own and extend it's own scaffolding, but it's own evals, it's own model training, and it's own inference? It's a much bigger deal than setting up a cron job to delete spam or something. Tinker is just the first backend among many (including the big one Nous is working on) - but you could for example get the agent to build itself a Runpod integration, manage it's own inference, train itself new models, optimised to whatever you're doing. Build a LoRA or steering vector library for each task, automate reading arxiv papers and running replication experiments, whatever. I kinda want an MLX backend I've been working on built into it (on the ol todo list), since I bet a continuously trained 30b MoE could get really optimal at automating the things I want automated, fully self hosted and self trained constantly Early days yet, think of something that an agent that's intended to own it's own weights could do & see what you can do with it. Use it, extend it, contrib etc