Lets say you’ve got a task like write a blog post. Instead of issuing a single command, have a GPT model plan it out. Something akin to

system: You are a planning AI, you will come up with a plan that will assist the user in any task they need help with as best you can. You will layout a clear and well followed plan.   
User: Hello Planner AI, I need your help with coming up with a plan for the following   task : {user_prompt}   

So the now LLama2-70B generates a plan that has steps in it that are numbered. Next, you can regex on the numbers and than pass that along to the worker model that will execute the task. As LLMS write more than humans and add in additional details that LLMS can follow, the subsequent LLMs will do a better job in executing the task rather than if you asked a smaller model write me a blog post about 3D printing D&D minis. Now go replace the task of writing a blog post with whatever it is you’re doing and you’ll be getting results

Llama2-70B for generating the plan and than using CodeLlama-34B for coding, or LLama-13B for executing the instructions from LLama-2-70B

Currently in the process of exploring what other models to add once LLama2-70B generates the plan for what needs to get done

THIS is exhibit A of why Open Sourced local LLMS are the future

What model was used with that prompt for bootstrapping the data for a training set? Did you then take all that data and fine-tune it on the model used to bootstrap the initial dataset?

Its been the opposite and I’ve watched my 3090 age like fine wine

you gotta look at the internal memory clocks and data transfer rates on the GPUS and what you’re going to see is that only the XX80 and XX90 cards have enough memory bandwidth to transfer all that vRAM so the 4060 with all that vRAM can’t actually move that much memory around