I don't want to be expressing too much cynicism, but just a heads up: there is nothing special about MCP and Skills; those just work on top of whatever fine-tuning that a lot of off-the-shelf models are capable of.
🇳🇱
@manlycoffee Well yea there is something special about it, it gives the LLM's a standard way to interact with the world. Thats proven to be pretty powerful
I mean, calling it "special" gives the illusion to people that MCP is something only those seasoned in fine-tuning LLMs can deploy.
But other than that, sure, it is powerful, with lots of emergent properties.
hmm interesting take. Maybe im disconnected with the common perspective, but i wont why people associated MCP with fine-tuning, they arent even related.
They could be.
LLMs are used for so many NLP purposes, but sometimes, we really need to sharpen an LLM's saw to focus solely on "tool calling", and a lot of tool calling involves NER. Without the fine tuning, you get lots of "hallucinations" (bad results).
That said, a lot of off-the-shelf LLMs have already been trained to handle NER just fine, and fine-tuning is not really that necessary.
Im confused what you mean here. Ive been workin gin the AI 20 years, not saying your wrong, just trying to understand what your saying...
MCP/skills are just any mechanism the LLM has to interact with the world. NER is the ability to classify one "thing" as another. "Jeff Freeman is a scientist" for example.
While NER is very important in NLP for sure,and yes we can fine-tune encoder-only llms to be better at NER I still am not grasping why your lumping fine-tunning + NER in with MCP, MCP is largely unrelated to NER and fine tuning, am I missing some context or something?
@manlycoffee Ahh ok that makes more sense, yes EAE can be helpful in figuring out what agent should do what, among other things. Fine-tuning can improve EAE in specialized cases (not strictly needed in general) and that can lead to situations where MCP/skills will be better utilized (by ensuring the proper agent is using the proper skill).
Though to be clear, fine-tuning is neither needed nor routinely used for this purpose. Not to say you cant, or that it wont help, just, usually a powerful generalized model is good enough for most things in that regard.
Personally I use MCP quite extensively in anything I do with LLMs and cant think of a case of an LLM that used skills that I had to fine-tune to improve its skill use.
@freemo
I think I'm using the wrong term.
I'm referring to "argument extraction".