CORRECTIONS

Corrections log

Errors we found and fixed in published articles, newest first (7). Each fix was applied to both the Japanese and English versions. See our editorial policy.

  1. · The Linear Algebra Under LoRA and RAG — Eigenvalues, Low Rank and Vector Search, Hands On

    Was: Gave vector_cosine_ops as the way to normalize vectors in pgvector.

    Now: Now says normalization is done with l2_normalize; vector_cosine_ops is the operator class for cosine-distance indexes.

    Source: pgvector: Open-source vector similarity search for Postgres (README) github.com

  2. · Choosing a Data Structure — Arrays, Hashes, Trees and Heaps

    Was: Grouped SentencePiece with WordPiece as greedy longest-match tokenizers.

    Now: Now names only WordPiece as longest-match; SentencePiece's unigram model is a different problem.

    Source: Fast WordPiece Tokenization (Song et al., 2020) arXiv:2012.15524

  3. · Choosing a Data Structure — Arrays, Hashes, Trees and Heaps

    Was: Stated the heap property as "each parent is smaller than its children".

    Now: Now says each parent is less than or equal to its children.

    Source: heapq — Heap queue algorithm (Python documentation) docs.python.org

  4. · Database Internals — What Happens Behind a Single Line of SQL

    Was: Described the default random_page_cost of 4.0 as assuming costly disk-head movement, and said lowering it on SSDs is standard practice.

    Now: Now follows the documentation: 4.0 assumes most random reads are cached and is deliberately lower than the raw difference; lower it when data fits in memory, raise it when caching is less effective or on magnetic disks (2 places).

    Source: PostgreSQL Documentation: Query Planning — random_page_cost postgresql.org

  5. · Attention from Scratch — The Heart of the Transformer, Explained Visually

    Was: Said the original paper showed that a language model can be built from this mechanism alone.

    Now: Now says the paper showed this mechanism alone is enough for a sequence-transduction model, with experiments on machine translation.

    Source: Attention Is All You Need (Vaswani et al., 2017) arXiv:1706.03762

  6. · The Cost of Matrix Multiplication — Where Almost All of AI's Compute Goes

    Was: Gave the condition for the 6N approximation too narrowly, as "sequence length much smaller than d_model".

    Now: Now gives the paper's condition: d_model > s/12, i.e. sequence length below roughly 12 times d_model (2 places).

    Source: Scaling Laws for Neural Language Models (Kaplan et al., 2020) arXiv:2001.08361

  7. · The KV Cache from Scratch — The Heart of Fast Inference

    Was: Stated without qualification that decoding is memory-bandwidth-bound rather than compute-bound.

    Now: Now says it is memory-bandwidth-bound at small batch sizes, and that compute time dominates as the batch grows (4 places, including the summary).

    Source: Efficiently Scaling Transformer Inference (Pope et al., 2022) arXiv:2211.05102