#llm #ai #software Created at 081023 # [Anonymous feedback](https://www.admonymous.co/louis030195) # [[Epistemic status]] #shower-thought Last modified date: 081023 Commit: 0 # Related # how low the computing stack will llm go [some ppl are bringing llm at the OS level](https://github.com/trholding/llama2.c) which could make sense. Currently LLMs are mostly used at the application layer because it's the lowest hanging fruit but there are more difficult possibilities. Fundamentally existing [[Large language model|LLM]]s are trained on a human language / high level machine language level, ie you can ask chatgpt to go from human language to high level machine language or the reverse but you cannot ask to generate bits why? because there is not much human language <-> bits pairs on [[The Internet]] (i.e. on the training dataset of most llms) if you want to bring llms down the software stack you might need this for example if you want to train a llm to generate network packets of course you need to work around encryption stuff if you want ur llm to interact in real time with the network but it's a minor problem im pretty sure a llm can deal with encrypted data (call the mathematician here) except if ur llm run on a quantum computer but were not there yet though we could train llm on mapping human language <-> network packets or human language <-> assembly, or human language <-> bits, we'd prob face context limit issue of course because programming language have fundamentally a low [[Kolmogorov complexity]] - ie express much in very short amount of characters, while assembly has high complexity, you'd need a bunch of code to do simple stuff lets ignore the context limit issue so yeah, we could build a dataset of programming language -> assembly or bit by doing reverse engineering hacks like disassembling, decompiling i even had [a project of training a transformer to disassemble/decompile code](https://github.com/louis030195/universal-decompiler) - that would be a good business for blockchain nerds that face issues not knowing whats the code in smart contracts so what would we do with a llm that can generate bits directly? well if its good enough we dont need an OS anymore, the llm can interact with the hardware directly, right?