https://hoijunkim.github.io/https://hoijunkim.github.io/blog/https://hoijunkim.github.io/blog/attention-weights-drawn/https://hoijunkim.github.io/blog/backprop-is-a-ledger/https://hoijunkim.github.io/blog/blocks-saved-memory-not-time/https://hoijunkim.github.io/blog/changing-the-task-inverted-it/https://hoijunkim.github.io/blog/counting-the-prefill/https://hoijunkim.github.io/blog/cut-the-cache-keep-the-beginning/https://hoijunkim.github.io/blog/en/attention-weights-drawn/https://hoijunkim.github.io/blog/en/backprop-is-a-ledger/https://hoijunkim.github.io/blog/en/blocks-saved-memory-not-time/https://hoijunkim.github.io/blog/en/changing-the-task-inverted-it/https://hoijunkim.github.io/blog/en/counting-the-prefill/https://hoijunkim.github.io/blog/en/cut-the-cache-keep-the-beginning/https://hoijunkim.github.io/blog/en/filling-the-seat-cost-cache/https://hoijunkim.github.io/blog/en/further-gradient-worse-loss/https://hoijunkim.github.io/blog/en/gate-removed-won/https://hoijunkim.github.io/blog/en/init-decides-whether-training-starts/https://hoijunkim.github.io/blog/en/kernels-learned-line-endings/https://hoijunkim.github.io/blog/en/korean-typography/https://hoijunkim.github.io/blog/en/learning-rate-has-a-ceiling/https://hoijunkim.github.io/blog/en/measuring-the-tail/https://hoijunkim.github.io/blog/en/model-does-not-pick/https://hoijunkim.github.io/blog/en/multi-head-what-splitting-buys/https://hoijunkim.github.io/blog/en/ninety-six-percent-was-waste/https://hoijunkim.github.io/blog/en/no-amount-of-zeros/https://hoijunkim.github.io/blog/en/normalisation-erases-the-init/https://hoijunkim.github.io/blog/en/one-recurrent-step/https://hoijunkim.github.io/blog/en/pooling-throws-away-position/https://hoijunkim.github.io/blog/en/positional-encoding-puts-order-in/https://hoijunkim.github.io/blog/en/predicting-characters-end-to-end/https://hoijunkim.github.io/blog/en/pulling-out-the-gates/https://hoijunkim.github.io/blog/en/quartering-the-weights/https://hoijunkim.github.io/blog/en/read-and-write-cost-differently/https://hoijunkim.github.io/blog/en/same-budget-gru-wins/https://hoijunkim.github.io/blog/en/shallow-stacks-use-it-all/https://hoijunkim.github.io/blog/en/stacking-layers-added-nothing/https://hoijunkim.github.io/blog/en/stacking-ninety-six-blocks/https://hoijunkim.github.io/blog/en/steps-favoured-the-transformer/https://hoijunkim.github.io/blog/en/tensor-is-one-dimensional/https://hoijunkim.github.io/blog/en/the-better-draft-was-slower/https://hoijunkim.github.io/blog/en/the-same-computation-n-times/https://hoijunkim.github.io/blog/en/training-loss-is-not-performance/https://hoijunkim.github.io/blog/en/two-thirds-of-the-parameters/https://hoijunkim.github.io/blog/en/waste-cut-time-not/https://hoijunkim.github.io/blog/en/where-it-starts-failing/https://hoijunkim.github.io/blog/en/why-minibatch-wins/https://hoijunkim.github.io/blog/en/wider-window-worse/https://hoijunkim.github.io/blog/filling-the-seat-cost-cache/https://hoijunkim.github.io/blog/further-gradient-worse-loss/https://hoijunkim.github.io/blog/gate-removed-won/https://hoijunkim.github.io/blog/init-decides-whether-training-starts/https://hoijunkim.github.io/blog/kernels-learned-line-endings/https://hoijunkim.github.io/blog/korean-typography/https://hoijunkim.github.io/blog/learning-rate-has-a-ceiling/https://hoijunkim.github.io/blog/measuring-the-tail/https://hoijunkim.github.io/blog/model-does-not-pick/https://hoijunkim.github.io/blog/multi-head-what-splitting-buys/https://hoijunkim.github.io/blog/ninety-six-percent-was-waste/https://hoijunkim.github.io/blog/no-amount-of-zeros/https://hoijunkim.github.io/blog/normalisation-erases-the-init/https://hoijunkim.github.io/blog/one-dataset-two-views/https://hoijunkim.github.io/blog/one-recurrent-step/https://hoijunkim.github.io/blog/pooling-throws-away-position/https://hoijunkim.github.io/blog/positional-encoding-puts-order-in/https://hoijunkim.github.io/blog/predicting-characters-end-to-end/https://hoijunkim.github.io/blog/pulling-out-the-gates/https://hoijunkim.github.io/blog/quartering-the-weights/https://hoijunkim.github.io/blog/read-and-write-cost-differently/https://hoijunkim.github.io/blog/same-budget-gru-wins/https://hoijunkim.github.io/blog/series/after-training/https://hoijunkim.github.io/blog/series/not-attention/https://hoijunkim.github.io/blog/series/seeing-deep-learning/https://hoijunkim.github.io/blog/shallow-stacks-use-it-all/https://hoijunkim.github.io/blog/stacking-layers-added-nothing/https://hoijunkim.github.io/blog/stacking-ninety-six-blocks/https://hoijunkim.github.io/blog/steps-favoured-the-transformer/https://hoijunkim.github.io/blog/tensor-is-one-dimensional/https://hoijunkim.github.io/blog/the-better-draft-was-slower/https://hoijunkim.github.io/blog/the-same-computation-n-times/https://hoijunkim.github.io/blog/training-loss-is-not-performance/https://hoijunkim.github.io/blog/two-thirds-of-the-parameters/https://hoijunkim.github.io/blog/waste-cut-time-not/https://hoijunkim.github.io/blog/where-it-starts-failing/https://hoijunkim.github.io/blog/why-minibatch-wins/https://hoijunkim.github.io/blog/wider-window-worse/