mbuchel-hn

attempted this approach before, fails in low level programming situations.

the hardest part of programming is the logic errors, as a result i tried a few times to use a formal verifier to write the pseudo code and have the llm translate it into the language i wanted. the problem with this however is that llms do not necessarily know the best techniques for speed and like to overcomplicate the problem/solution.

how does this language help with prevention of overcomplication?

does this work similar to airllm? i am wondering how it would handle something like quantizing kimi k3 on a budget of 8 gbs, or is that something you are not attempting to solve yet?

this reminds me of the whole situation with how llms cannot really solve logic problems, they only really memorized the answer, for example the riddle:

a mother and her son got in a car accident, the mother died, and the son was driven to the hospital. the doctor (who is the boy's father) says he cannot opperated because he is related to the boy. what is the relation between the boy and the doctor.

llms like to assume the doctor is the mother because of the fact that is the version that shows up in their training data.

for all 22 micro archs it takes about 9.4 hours for all opcode-dumps this includes hidden instructions that take a PC but do not produce text in the nvdisasm output, about an extra hour for a modifiers scan where we get all the possible modifiers that a single instruction can have, and than an extra half hour or so for the operands