one of the things studying abstract mathematics is helping me with is decrypting #llms #llm slop (which is not slop, it's a hypercondensed transmogrified partial and "fractal" linear representation of whatever the internal state of a model is when processing some input). Understanding this stuff which looks like gibberish at first, and then learning it enough to steer a model, understanding which symbols actually refer to concepts we already have better names for, or recognize new ways of combining things, is maybe the most challenging work i did. And it changes every two context windows, basically, and different models have their preferences in terms of lingo (seam, goblin, etc... you know what i mean). Asking it to use "fundamentals from mathematics / computer science" allows me to compress this and hopefully use names that are mathematically tractable."Declaring image scope I7 means the system owes an outcome" is actually a pretty important and "deep" invariant in the system, but be...
one of the things studying abstract mathematics is helping me with is decrypting #llms #llm slop (which is not slop, it'...