so, i loaded the LLM called Athene on my pc and asked it a question, and it's writing the solution slower than a beginner two finger typist.

the blurb about it said that it beats GPT-4o at coding problems but right now it's burning a good 150W of power for about one word in three seconds, my CPU is flat out at 80% utilization.

however, if it manages to pull off what i asked it to, i will have a variable integer encoding that also sorts lexicographically, that is, the varint equivalent of big endian encoding. normally varints calculate forwards and the LSB is on the left but for this task, it has to be on the right. i don't really understand how it can actually work because a smaller number is shorter, so how can a longer value not disrupt the sort order of a collection of keys.

anyhoo. lol. i can see why they need so much power and you need to depend on giant data centers full of GPUs to decode the matrixes to get a usable experience. i can only assume that the GPU is maybe 2-4x as efficient as the CPU for power per computation, but has more memory bandwidth.