RESEARCH & ENGINEERING NOTES
Hi, this is Hongjun. I'm interested in computation and optimization.
From categorical probability and decision theory to Jev’s typed API, Kev’s pointer head, benchmarks with stated measurement conditions, and a reproducible local workflow.
Derive the directional block mask and conditional linear scaling, then explain attention supervision, direct ranking, and the limits of the reported speedup.
Understand Nested Learning through explicit state updates, associative memory, optimizer traces, and the distinction between a general mathematical model and the Hope architecture.
Turn performance advice into testable hypotheses using Amdahl’s law, memory traffic, allocation behavior, queueing, and reproducible experiments.
Derive inverse-probability estimates for joins, explain why DISTINCT changes the sample space, and separate asymptotic convergence from finite-budget accuracy.
Why temporal prediction can look solved under easy negatives, how EdgeBank exposes memorization, and how modern temporal benchmarks define a reproducible protocol.
Derive retrieval-aware supervised training, distinguish oracle documents from distractors, and design an evaluation that exposes missing evidence and data leakage.
A unified mathematical view of graph decisions, from message passing and constrained actions to policy gradients, knowledge-graph traversal, and reliable evaluation.