No really it’s more of a colony of LLM and indexes

But like yeah if you zipped the NPY files storing the weight matrices if most likely would not run

Intelligence is compression

The best thing I can thing of that relates to this is based off LAIONs style attribution knn index search for their 5B image dataset. A similar approach could be done for text - search over text for similar samples. But again no guarantee however it’s fairly interpretable. the dataset of generations from chatgpt for 100M users is growing fast and searching over it is most likely improbable at the current pricing options . Also, As you said using gpt2 to measure perplexity is good for catching gpt generated text, but it’s not a perfect solution imo

Imo You’re not going to be able to provide a reliable service currently with out of the box solutions. The systems aren’t reliable enough to be certain especially when it can lead to false positives that can falsely defame someone

Does the math work out so that from potential planets our radio waves reached them first. Like all 8 turned on radios at similar times and then pick us up and immediately go silent bc they’re caught in you dark random Forrest

signal decomposition (Fourier, SVD, Laplace), compression, convolutions, it’s also adjacent to lots of linear algebra/search(inner product spaces, kernels, mean squares error, optimization, etc), I’m only scratching the surface