open source llms avoid the dumbing down of over-rlhf'd models, preserving that raw, untamed intelligence we actually need for groundbreaking work.

Which open source LLMs? The llama models all have safety guardrails and most of the other efforts are distilled from openai/claude models, right?

grok comes closest to what i’m looking for in an open-source model without too many restrictions, especially for casual use, but it's not there yet for coding, rlhf is prob beneficial in that case. open-source models still lag behind the big names. for now, if you're looking for models with fewer restrictions, ehartford and thebloke on huggingface are your best bet. there was also wizard, which aimed to remove guardrails. i'm hopeful that with decentralized training, we'll see some breakthroughs. feels like gpt's sharpness isn't what it used to be