seems likely that mythos has a n^2 pretraining phase and then an nlogn inference phase, which is done with meta-attention, or with some type of iterative refinement (diffusion?) of the attention mask, which allows long context lengths needed to chain vulnerabilities