2024 was the year GGUF took off

I wish everything wasn't such a "tower of babylon" with so many independent incompatible model file formats, and then within any given model file format there's a possible set of a few model encodings / quantizations which might be used. But almost every inference SW accepts only one or a few model file formats to the exclusion of all the others. And almost every inference SW accepts only one or a few options for encodings / quantizations it will work with in some use case to the exclusion of all others. So what file format someone publishes a model in almost 100% dictates available quantizations and the combination of those almost 100% dictate which of 2-3 inference SW tools you can use and also which GPUs / accelerator hardware and software frameworks you can use. And most of the model file formats are not capable of (or I suppose intended to) preserve the full fidelity of the originally published model. e.g. any available auxiliary data / metadata etc. And you cannot in practice easily convert from model file format X to model file format Y and even worse is trying to convert quantization X to quantization Y even if the target one is roughly equivalent as or more coarsely quantized than the origin. So, tower of babylon. It seems as an overall ecosystem that is cross OS platform, cross accelerator type, cross inference SW type we could've standardized some kind of "container format" which at least can preserve metadata and data with fidelity about the original model and then has the flexibility to contain data resources encoded with various possible schemes to represent the quantized model, or metadata to represent a suggested / desired quantization that could be made dynamically (install time, run time, ...) from the superior / upstream model weights to enable their use in some way. GGUF has its positive attributes of being minimalistic for what it is designed to do -- support basically one type of inference SW library, but it's a leaf node which is de facto likely going to stay incompatible with basically all other inference SW ecosystems -- onnx, HF / safetensors, et. al. Larger file sizes of original upstream models deter many from just downloading them, though that is the superior choice in that it's not totally uncommon for GGUF options / conversions to require or desire to be changed over time -- maybe you upgrade your GPU and you'd rather run a Q8 than your previous Q4 -- maybe there was an upstream model metadata / tokenizer fix and now you need a new GGUF that reflects that. So you download the new GGUF 1x, 2x, 3x whatever over the time you want to use it. But if it was easier (or at least more commonly done) to convert from the original model to a quantization locally or even at run time then one would likely be able to keep more choice and currency of proper conversions and options over time for the price of having to download the true full model once.