For the select/deselect training data, I assume this means choose which context to pay attention to?
Each ML model, LLM and otherwise, is a combination of matmul operations & nonlinear activation functions on static weights. My understanding of your "ignoring training data" is to change the vector values of the neural network, which is part of what happens during fine tuning.
Curious why telling an LLM to speak like a character, then using few shot examples to anchor the model in a certain personality/tone doesn't suffice? Is it really the training data (meaning the response strays to random nonsense) or is it that the instructions are not good enough?
LLMs won't destroy human thought since LLMs are an average approximation of human thought. Sure, this might elevate those who are fresh and are just looking for generic copy, though the best writers are secretly just the best thinkers, as writing is a medium to exercise thought.
I'm a bit biased, having built an AI writing tool myself (https://zenfetch.com), though it's for this very reason that we aren't interested in generating new content on your behalf. We simply want to make it easier for you to recall information to augment your work.
Yeah... at this point I just assume that anyone who can't see any negatives to AI has a financial incentive that depends on their not understanding it.
Very cool! Have you considered local LLMs for a free/lite tier?
I do agree that 3rd party LLM services (e.g. OpenAI) are much more powerful. For basic tasks (e.g. summarization, Q&A), local LLMs seem to perform decently.
We have considered it, though most of our users use Zenfetch for things like
1. Research across an industry
2. Generating new content
3. Connecting ideas / building networked thought
For that reason we were deterred from using local models. Hopefully the Mistral team and similar keep pushing out new developments to get comparable with GPT-4 levels
Zenfetch offers the ability to chat with a webpage, like Arc, while also pulling in references from your previously saved content. These references are used to augment the answer to the question.
The goal is to help reinforce material you've read in the past and form associations with current learnings.
Thanks for sharing this, markdownload looks super useful. I use obsidian myself and have tried a few bookmarks managers. They are all great, though I often find myself forgetting information from a few months back unless I actively build networked thought.
We built Zenfetch so that I can offload that work and have Zenfetch remind me of important information when I need it. Today that's done through emails and remind me prompts, and we have plans for helping recall more information in the near future :D
Each ML model, LLM and otherwise, is a combination of matmul operations & nonlinear activation functions on static weights. My understanding of your "ignoring training data" is to change the vector values of the neural network, which is part of what happens during fine tuning.
Curious why telling an LLM to speak like a character, then using few shot examples to anchor the model in a certain personality/tone doesn't suffice? Is it really the training data (meaning the response strays to random nonsense) or is it that the instructions are not good enough?