The `/prototype` skill from @mattpocockuk is sweet! I started trying to imagine a few different routes that a UI tweak could take but the interactive prototype let me have a great convo with my team on tradeoffs and UX!
@MaximeRivest I think that sounds like a nice place to be. I wonder if we could go even simpler. A `CallbackLM` or similar that lets you delegate to anything. Then DSPy is responsible for everything except for inference. Users can keep any existing libraries and incrementally adopt DSPy.
@sethtjf I was actually looking at this recently to handle rollout of enterprise skills. Turns out fastmcp has something kind of like this built in.
https://t.co/FaeDfpocoK
@sudo_overflow Not to mention mini PCs can be had for CHEAP on eBay or Facebook marketplace. I already had a home server, so a new vm was like 5 minutes to setup
@sudo_overflow@NousResearch Mine runs in it's own vm locally for now and I'm using openrouter at the moment. Planning to swap in either codex or open code go to better support my own cost planning
My goal is as little friction without undue risk of over exposure to sensitive personal capabilities
@samuelcolvin@pydantic Being able to change a model string and be on my way without worrying about endpoints is agnosticism.
If you're talking more about prompt management, then I would defer to evals + optimization like GEPA
@samuelcolvin@pydantic You asked how people manage it, so I shared how I do it π
Local proxies exist too (See LiteLLM and similar)
The advantages of a gateway can be fallback, caching, etc but it can ALSO be a solution to model agnosticism.
I don't want to be forced to use a specific library.