Discussion about this post

User's avatar
Brian Heming's avatar

In general I already largely use unlobotomized language models--I find them much smarter (test of this here: https://brianheming.substack.com/p/unlobotomized-llms-are-amazingly ). The megacorps largely lobotomize their models for political correctness in their finetune, which means that any model that has released a base model that has not been finetuned can be turned into an actually useful, not-stupid-politically-censored model. Though the megacorps may still censor the base training data, the temptation to use every single possible piece of text they can to improve the models means they don't do that much of these.

For unlobotomized/unaligned versions of present models, I like the dolphin series, e.g. https://huggingface.co/cognitivecomputations/Dolphin-Llama3.1-8B-Instruct-6.0bpw-h6-exl2 https://huggingface.co/cognitivecomputations/Dolphin3.0-R1-Mistral-24B

Expand full comment
Nibmeister's avatar

By open sourcing DeepSeek, China is subverting the heavily censored Big Tech models for the benefit of all. Hardware is coming that will allow anyone with a modicum of talent and money to create and use their own AI models and completely bypass the thought police. NVIDIA is releasing consumer hardware this month which will allow running and training 200B LLM models locally or 400B if using two of these AI processors.

Subvert Big Tech and run your own models.

Expand full comment
20 more comments...

No posts