ls articles --sort created

Articles

63 entries indexed

  1. Displaying Spotify Artwork on a Pixoo 64Displaying the currently playing Spotify track's artwork and title on a Pixoo 64 using a Raspberry Pi and a small pixel font.
  2. Releasing Bekko Embedding, an Ultra-Compact Multilingual Text Retrieval ModelBekko Embedding provides two ultra-compact multilingual text retrieval models with strong retrieval quality and practical CPU inference.
  3. cdd for herdrA cdd integration for jumping to working directories from herdr tabs, panes, workspaces, and agents.
  4. herdr-tiny-fingers: tmux-fingers-style Copying for herdrherdr-tiny-fingers is a small plugin for copying URLs, SHAs, UUIDs, and similar text from a herdr terminal screen with short keys.
  5. trn: A Fast Local Translation Command for macOS Tahoetrn is a macOS CLI for Apple's on-device Translation framework, with a low-latency mode about 12x faster than high-fidelity mode in a local README test.
  6. OpenAI API-Compatible Access Without Additional API Billing via CodexA Codex-authenticated OpenAI API-compatible server for Responses, Chat Completions, and image generation. Within the Codex subscription scope, it can be used without additional API usage billing.
  7. Japanese Full-Text Search in SQLite and DuckDB with VaporettoHow to add Japanese full-text search to SQLite and DuckDB with Vaporetto, including extensions, a browser demo, and BM25 search examples.
  8. Building a Machine Learning PC with Two RTX 5090 GPUsNotes from building a two RTX 5090 machine learning PC in Japan, including power supply constraints, cooling, parts, and multi-GPU training tradeoffs.
  9. Looking Back on 2025A personal 2025 retrospective covering a new child, life in the countryside, a new car, and work on AI and information retrieval.
  10. OpenProvence: A Model for Removing Irrelevant Sentences Before Passing Text to an LLMOpenProvence prunes irrelevant sentences from retrieved text before passing the remaining context to an LLM.
  11. Evaluating the Japanese Performance of Embedding Gemma 300M with JMTEBA google/embeddinggemma-300m benchmark on JMTEB v1, comparing its Japanese embedding performance with multilingual and Japanese models.
  12. JFWIR: A Large Japanese Information Retrieval Dataset Built from Japanese FineWebJFWIR is a 64-million-pair Japanese information retrieval dataset built from FineWeb2 Edu Japanese with varied query types and hard negatives.
  13. Evaluating the Japanese Performance of Qwen3 Embedding with JMTEBA Qwen3-Embedding-0.6B benchmark on Japanese JMTEB tasks, compared with Japanese embedding models and OpenAI embeddings.
  14. Releasing Small, Fast, and Practical Japanese Rerankers: tiny, xsmall, small, and base v2Tiny through base v2 Japanese rerankers designed for competitive quality with practical latency on CPUs and Apple silicon.
  15. query-crafter-japanese: A Model for Generating Queries for Information Retrievalquery-crafter-japanese provides small Apache-2.0 models that generate retrieval queries from documents for synthetic information retrieval datasets.
  16. FineWeb2 Edu Japanese: A High-Quality Educational Japanese DatasetFineWeb2 Edu Japanese is a filtered Japanese educational web dataset with 120M records and about 89.3B tokens, built from FineWeb2.
  17. Releasing a Japanese StaticEmbedding Model for Practical 100x Faster Text Embeddingsstatic-embedding-japanese is a fast non-Transformer embedding model for Japanese and English text, evaluated on JMTEB.
  18. Looking Back on 2024A personal look back at 2024, covering life, building a house, work on AI and information retrieval, and technical projects.
  19. Releasing Japanese SPLADE v2, a Strong Retrieval Model for Texts Under 512 Tokensjapanese-splade-v2 is an improved Japanese sparse retrieval model with strong JMTEB scores for documents up to 512 tokens.
  20. Releasing Japanese BERT RetroMAE Models and Evaluating Them on Downstream Retrieval TasksJapanese BERT models pretrained with RetroMAE, with released weights and an evaluation of the pretraining effect on JMTEB retrieval tasks.
  21. How to Build a SPLADE Model: Japanese SPLADE Technical ReportBuilding a Japanese SPLADE sparse retrieval model, including tokenizer issues, training implementation, evaluation, and the YAST trainer.
  22. Releasing a High-Performance Japanese SPLADE Sparse Retrieval ModelA Japanese SPLADE sparse retrieval model, with retrieval and reranking evaluations against dense embedding models.
  23. Running Japanese Tokenizer Models with text-embeddings-inferenceHow to run Japanese embedding models that lack tokenizer.json on Hugging Face text-embeddings-inference by adding a dummy fast tokenizer.
  24. Releasing High-Performance Japanese Rerankers, and What Rerankers AreJapanese search reranker models, their evaluation results, and how reranking improves retrieval quality after initial vector or keyword search.
  25. Technical Report on Building Japanese RerankersA technical report on training Japanese CrossEncoder rerankers, including data construction, hard negatives, model variants, and evaluation results.
  26. After reading Human-in-the-Loop Machine Learning: a data-centric and suggestive bookA review of the Japanese translation of Human-in-the-Loop Machine Learning, focusing on why its data-centric view of active learning, annotation, and collaboration with annotators is valuable in the current AI era.
  27. ColBERT reaches e5-large-level performance on a Japanese RAG taskAn evaluation of the Japanese pretrained JaColBERT model on the AI-Ou Q&A retrieval benchmark, with performance slightly below multilingual-e5-large.
  28. Evaluating OpenAI's new text-embedding-3-small on a RAG taskAn evaluation of OpenAI's text-embedding-3-small on Japanese Wikipedia Q&A retrieval, comparing 512-dimensional embeddings with multilingual-e5 models.
  29. Summarizing same-date diary entries from secon.dev and sending them to DiscordA small OpenAI and LCEL system that summarizes years of same-date diary entries from secon.dev and posts them to a private family Discord channel.
  30. Understanding LangChain Expression Language (LCEL)LCEL is LangChain's recommended way to build chains. This article explains the basic behavior of Runnable, RunnableSequence, RunnableParallel, dict syntax, invoke, and RunnablePassthrough step by step.
  31. Looking back on 2023A personal 2023 retrospective covering work, travel, daily life, technology, social networks, and plans for 2024.
  32. Solving the first AI-Ou quiz competition with vector search onlyUsing vector search over Japanese Wikipedia passages for the first AI-Ou quiz competition, with a comparison of Japanese embedding models on Q&A retrieval.
  33. Training a Q&A + RAG-focused LLM with SFT, making 4-bit quantized models, and exceeding GPT-3.5 with a 7B modelFine-tuning and 4-bit quantization of rinna's youri-7b-instruction for Japanese Q&A over RAG context, with quality, speed, and memory comparisons against GPT-3.5 and GPT-4.
  34. Building Japanese Wikipedia embeddings and a FAISS index for RAGEmbeddings and FAISS indexes for about 5.5 million Japanese Wikipedia passages, prepared for RAG retrieval and question-answering experiments.
  35. Measuring speed, data size, and accuracy for vector search algorithms and quantization parametersA benchmark of FAISS vector search settings, including IVF, HNSW, and product quantization, with a focus on recall@1, @3, and @5 for RAG systems where top-N retrieval quality matters.
  36. Making Transformers inference 1.6 to almost 2 times faster with CTranslate2CTranslate2 embedding inference through hf_hub_ctranslate2, delivering about 1.6x faster GPU and 1.9x faster CPU performance with almost no accuracy change.
  37. Embedding conversion performance on Apple Silicon GPU (MPS)A multilingual-e5-small embedding benchmark comparing Apple Silicon MPS on a MacBook Air M2 with RTX 4090, Colab T4, and CPU execution.
  38. Starting Weekly AI News: automated summaries with clustering and GPTHow a weekly AI newsletter clusters articles with multilingual-e5-small and automatically generates topic titles and summaries with GPT.
  39. Implementing and trying gzip + kNN text classification from the paper that beats BERTAn implementation and evaluation of the gzip-based NCD + kNN text classifier on Japanese and English datasets including livedoor news, MARC-ja, and AGNews.
  40. Generating answers from images with ChatGPT 3.5 and extracting information through BLIP-2 promptsA note on using BLIP-2 with ChatGPT 3.5 for image-based answer generation when the task fits, and more importantly, on extracting image information through prompts to BLIP-2.
  41. Enjoying Stable Diffusion again from a technical perspectiveNotes on Stable Diffusion techniques available through stable-diffusion-webui, including ControlNet, LoRA, textual inversion embeddings, and checkpoint merging.
  42. Another major benefit of LoRA: switching task models instantly while sharing GPU memoryLoRA is usually discussed as a cheaper training method, but it also lets multiple task-specific adapters share one LLM base model in memory. This article shows how to switch adapters with Hugging Face PEFT.
  43. Analyzing the Iris dataset with ChatGPT's Noteable pluginAn evaluation of Noteable on the Iris dataset, generating plots, model comparisons, clustering, and dimensionality reduction notebooks through ChatGPT.
  44. Using ChatGPT's Noteable and WebPilot plugins to build a notebook that predicts OpenCALM 14B performanceUsing Noteable with ChatGPT and WebPilot to collect OpenCALM model data, build a notebook, visualize parameters and perplexity, and estimate a hypothetical 14B model.
  45. Quantizing fastText to build a practical 1.7 MB text classifierA 1.7 MB fastText classifier for filtering AI-related English articles, using quantization while retaining practical accuracy and recall.
  46. After reading Kaggle ni Idomu Deep Learning Programming no GokuiA review of the Japanese book Kaggle ni Idomu Deep Learning Programming no Gokui, which works well as a compact index of practical machine learning ideas for both Kaggle beginners and people already working with machine learning.
  47. Similar embedding search with SVM: an alternative to kNNHow LangChain's SVM Retriever works, with a comparison of kNN, SVM, and hybrid embedding search on AI News data.
  48. Launching AI News and how I used OpenAI behind itHow AI News collects AI, data science, and machine learning topics and uses OpenAI APIs to classify and summarize them in three lines.
  49. RAPIDS SVR and SVC: fast training without fine-tuning, evaluated on MARC-jaAn introduction to RAPIDS SVR and SVC, using neural-network embeddings as features without fine-tuning and evaluating the approach on the Japanese MARC-ja classification dataset.
  50. How secon.dev was implemented, December 2022 editionA late-2022 snapshot of secon.dev, covering Dropbox-synced Markdown, Next.js static builds, related-article generation, GCS image processing, and planned improvements.
  51. Kaggle Feedback Prize - English Language Learning: team gold medal, 15th place, and Kaggle MasterA 15th-place gold-medal solution for Feedback Prize - English Language Learning and the final result needed to reach Kaggle Competitions Master.
  52. Solo silver medal, 43rd place, in Kaggle Feedback Prize - Predicting Effective ArgumentsA solo 43rd-place silver-medal solution for Feedback Prize - Predicting Effective Arguments, covering what worked, what failed, and the experience of competing alone.
  53. Was the end of Japan's 2022 rainy season unusually early?A statistical look at Japan Meteorological Agency data to measure how unusual the early 2022 Kanto-Koshin rainy-season ending date was.
  54. Finding optimal weighted-ensemble coefficients with constrained least squaresUsing non-negative least squares to compute Kaggle ensemble blending weights automatically instead of tuning them by hand.
  55. My first Kaggle competition ended with a team gold medal, 8th placeA first Kaggle competition experience that ended with a gold medal in U.S. Patent Phrase to Phrase Matching, including how the team organized its collaborative work.
  56. Reading Basic Statistics by Kimio Miyakawa: statistics before machine learningWhy statistical foundations matter for understanding data, experiments, estimation, hypothesis testing, and machine learning model evaluation.
  57. Adding type hints to a Python project and getting value from type checkingAdding Python type hints and pyright to a machine learning project, with a low setup cost and immediately useful editor support and static checks.
  58. A CLI for finding similar documents in static site generatorssimilar-documents-cli computes TF-IDF and cosine similarity over Markdown or HTML files to generate related-article data for static sites at build time.
  59. When sharp fails to install on WSLOn WSL2, sharp tried to fall back to a local build because APPDATA pointed at a Windows npm cache path. Clearing APPDATA let the Linux prebuilt binary install correctly.
  60. Fixing WSL clock drift after sleep from the Windows sideA fix for WSL2 clock drift after waking from sleep, using Windows Task Scheduler to run hwclock inside WSL.
  61. NumPy cast overflow behavior can vary by environment and array sizeA NumPy casting issue where overflowing float32 values converted to uint8 behave differently depending on the environment and array length.
  62. Building a simple fully connected neural network with TensorFlow 2 without KerasA hands-on note implementing a simple feed-forward neural network with only TensorFlow APIs, without Keras, to understand layers, activation functions, losses, automatic differentiation, and manual training.
  63. Inferring Hiragana in the Browser with TensorFlow.jsA browser-based TensorFlow.js demo for recognizing handwritten hiragana, covering model size, Keras conversion, and the limits of importing Python-trained models into JavaScript.