https://andrewroseit.com/https://andrewroseit.com/posts/https://andrewroseit.com/findings/https://andrewroseit.com/tests/https://andrewroseit.com/hypotheses/https://andrewroseit.com/basics/https://andrewroseit.com/posts/e21n-paired-question-first-multikey-4b-512k/https://andrewroseit.com/posts/e21m-paired-kv-cache-precision-4b-256k/https://andrewroseit.com/posts/e21l-kv-cache-precision-multikey-4b-256k/https://andrewroseit.com/posts/e21k-question-first-multikey-needle-4b/https://andrewroseit.com/posts/e21j-multikey-needle-4b-context-curve/https://andrewroseit.com/posts/e21i-multikey-needle-9b-448k-4b-704k/https://andrewroseit.com/posts/e21h-strict-needle-9b-480k-4b-704k/https://andrewroseit.com/posts/e21g-9b-448k-one-rtx-3080/https://andrewroseit.com/posts/e21f-4b-704k-fully-resident-9b-262k/https://andrewroseit.com/posts/e21e-needle-depths-1m-2b-512k-4b/https://andrewroseit.com/posts/e21d-needle-512k-on-3080/https://andrewroseit.com/posts/e21b-upstream-fa-convert-buffer-cap/https://andrewroseit.com/posts/e22-real-query-attention-concentration/https://andrewroseit.com/posts/e21-long-context-gpu-fit-3080/https://andrewroseit.com/posts/e20-self-training-8b-math500/https://andrewroseit.com/posts/e19-self-consistency-math500/https://andrewroseit.com/posts/e17-failure-taxonomy/https://andrewroseit.com/posts/e18-coordinate-preconditioning/https://andrewroseit.com/posts/e17-math500-baseline/https://andrewroseit.com/posts/e16-gsm8k-truncation/https://andrewroseit.com/posts/idea-c-draft-difficulty/https://andrewroseit.com/posts/idea-a-contrastive-vote/https://andrewroseit.com/posts/e15-reasoning-baselines/https://andrewroseit.com/posts/e14-humaneval-quants/https://andrewroseit.com/posts/two-reasoning-ideas/https://andrewroseit.com/posts/op-offload/https://andrewroseit.com/posts/spec-decoding/https://andrewroseit.com/posts/32-user-dip/https://andrewroseit.com/posts/q6k-baseline/https://andrewroseit.com/posts/ssx-dropped/https://andrewroseit.com/posts/dropped-ideas/