Research papers are written for the people who already read the last twenty. These are the same papers with that missing context put back — every assumed term, every unexplained design choice, and every quiet trade-off written out beside the original text.
each reading keeps the original text on the left · underlined phrases open explanations on the right · explanations link onward, so you can go as deep as you need
A 2.8T-parameter Mixture-of-Experts model with 104B activated parameters and a one-million-token context window. The annotations cover the architecture (delta attention, attention residuals, latent MoE), the training recipe, and the infrastructure that makes a model this sparse trainable at all.
A framework for proving protocols secure in the Lean theorem prover: a general-purpose symbolic (Dolev–Yao) semantics with user-defined equational theories and threat models, plus a trace-invariant proof methodology in the style of DY*. The annotations unpack the symbolic model, the labeling discipline, and what a machine-checked protocol proof actually buys you.
An algebraic attack that spends input degrees of freedom to hold Poseidon's partial-round S-boxes at fixed values, collapsing 22 of 23 partial rounds to degree 1 — and solving CICO-1 over 28 of 31 rounds of the KoalaBear instance Ethereum is evaluating. The annotations cover the HADES strategy, the CICO problem, degree accounting, and why the fix cannot be a better MDS matrix.