The issue isn't getting LLM weights. Llama taught us that it's basically impossible to prevent people from distributing them if they choose to. You don't have to worry about "future access changes" too much, OpenAI knows there is no "undo" button once they publish weights.
The real issue is having SOTA, collapse-proof hardware for inference. Apple and Nvidia hardware are both reliant on drivers that can brick your server with an over-the-air update. AMD hardware has generally more resilient Mesa drivers that can theoretically survive a hostile OEM, but with fewer options for finetuning and training. Intel GPUs are a high-VRAM option but it's unclear how long they'll be supported in-software for. Everything is a system of tradeoffs.
The issue isn't getting LLM weights. Llama taught us that it's basically impossible to prevent people from distributing them if they choose to. You don't have to worry about "future access changes" too much, OpenAI knows there is no "undo" button once they publish weights.
The real issue is having SOTA, collapse-proof hardware for inference. Apple and Nvidia hardware are both reliant on drivers that can brick your server with an over-the-air update. AMD hardware has generally more resilient Mesa drivers that can theoretically survive a hostile OEM, but with fewer options for finetuning and training. Intel GPUs are a high-VRAM option but it's unclear how long they'll be supported in-software for. Everything is a system of tradeoffs.