Towards observable, reliable, scalable AIhttps://jaesolshin.com/enCopyright 2026 Jaesol ShinAstro + @astrojs/rsshttps://jaesolshin.com/posts/mtp-diffusion-gemma4-qwen36/https://jaesolshin.com/posts/mtp-diffusion-gemma4-qwen36/MTP and diffusion inference on Gemma 4 and Qwen 3.6, fp8 on one H100Thu, 25 Jun 2026 00:00:00 GMTAILLMvLLMMTPSpeculative DecodingDiffusionGemmaQwenBenchmarkhttps://jaesolshin.com/posts/requirements-as-latent-state/https://jaesolshin.com/posts/requirements-as-latent-state/Why spec-driven development is a requirements-inference architectureWed, 27 May 2026 15:00:00 GMTAIAgentRequirementsSpec-DrivenHarnessFeedbackControlhttps://jaesolshin.com/posts/skill-and-harness/https://jaesolshin.com/posts/skill-and-harness/Why skills and harnesses overlap in implementationWed, 27 May 2026 15:00:00 GMTAIAgentHarnessSkillsPolicyFeedbackhttps://jaesolshin.com/posts/harness-layer/https://jaesolshin.com/posts/harness-layer/How harness design determines whether agents actually adapt.Tue, 26 May 2026 15:00:00 GMTAIAgentHarnessMemoryAdaptationRAGSkillsLLMhttps://jaesolshin.com/posts/memory-as-adaptation/https://jaesolshin.com/posts/memory-as-adaptation/Why agent memory moved from RAG storage to the foundation of policy adaptationFri, 22 May 2026 00:00:00 GMTAIAgentMemoryRAGSelf-ImprovementPOMDPEpisodicProceduralhttps://jaesolshin.com/posts/weights-prompts-codes-as-parameters/https://jaesolshin.com/posts/weights-prompts-codes-as-parameters/Weights, prompts, and code as parameters at different layers of a learnable policy spaceFri, 22 May 2026 00:00:00 GMTAIAgentLLMPOMDPDSPyHarnessSelf-ImprovementRLhttps://jaesolshin.com/posts/graph-db-benchmark-8-engines/https://jaesolshin.com/posts/graph-db-benchmark-8-engines/Eight graph engines measured across OLTP, memory, analytics, and differentiation queriesThu, 21 May 2026 00:00:00 GMTAIGraphDBPostgreSQLRCTENeo4jMemGraphFalkorDBpgRoutingGraphRAGBenchmarkLightRAGhttps://jaesolshin.com/posts/lightrag-variants-comparison/https://jaesolshin.com/posts/lightrag-variants-comparison/Source-level comparison of RAG-Anything, ApeRAG, and EdgeQuake as LightRAG derivativesThu, 21 May 2026 00:00:00 GMTAILightRAGRAG-AnythingApeRAGEdgeQuakeGraphRAGRAGGraphhttps://jaesolshin.com/posts/pgstrom-ssb-benchmark/https://jaesolshin.com/posts/pgstrom-ssb-benchmark/Why Arrow+GPU achieves 10x at SF=100 and Heap+GPU loses to CPU on wide tablesThu, 21 May 2026 00:00:00 GMTDBPG-StromPostgreSQLGPUOLAPBenchmarkArrowSSBcuVSpgvectorhttps://jaesolshin.com/posts/postgres-lakehouse-arch/https://jaesolshin.com/posts/postgres-lakehouse-arch/Distributed processing (Ray/Daft/Smallpond) and Citus + pg_lake FDW integrationThu, 21 May 2026 00:00:00 GMTDBPostgreSQLLakehousepg_lakeIcebergCitusRayDuckDBDaftpostgres_fdwhttps://jaesolshin.com/posts/postgres-lakehouse-pglake/https://jaesolshin.com/posts/postgres-lakehouse-pglake/Eight phases building a PostgreSQL-centered Lakehouse: DuckLake's libpq collision and pg_lakeThu, 21 May 2026 00:00:00 GMTDBPostgreSQLLakehouseDuckLakeIcebergpg_lakeDuckDBParquetSnowflakehttps://jaesolshin.com/posts/pdf-to-markdown-comparison/https://jaesolshin.com/posts/pdf-to-markdown-comparison/Five PDF-to-Markdown converters (markitdown, pdftotext, pymupdf, mineru, opendataloader-pdf) scored against a seven-criterion 100-point rubricWed, 20 May 2026 11:40:00 GMTAIPDFMarkdownRAGIngestionminerumarkitdownpymupdfpdftotextopendataloader-pdfBenchmarkhttps://jaesolshin.com/posts/markdown-slide-pdf-pipeline/https://jaesolshin.com/posts/markdown-slide-pdf-pipeline/Pandoc plus LaTeX Beamer and macOS screenshot settings wired as zsh aliases — one markdown file emits slide PDF, document PDF, and configurable screenshotsWed, 20 May 2026 11:39:11 GMTDevelopmentzshMarkdownPandocPDFSlidesLaTeXContent WorkflowDotfileshttps://jaesolshin.com/posts/claude-code-session-sharing/https://jaesolshin.com/posts/claude-code-session-sharing/Resume the same Claude Code session across accounts by pointing every projects directory at one shared physical pathWed, 20 May 2026 11:30:37 GMTAIClaude CodeMulti-accountSession SharingSymlink.jsonlConfigurationDeveloper Workflowhttps://jaesolshin.com/posts/apple-silicon-llm-backends/https://jaesolshin.com/posts/apple-silicon-llm-backends/Benchmarking Qwen3.5-9B on Apple Silicon across MLX, llama.cpp, Ollama, omlx, and vLLM Metal — single-request throughput, prefill scaling, decode vs input length, and concurrency responseWed, 20 May 2026 08:36:00 GMTAILLMApple SiliconMLXllama.cppOllamaomlxvLLMBenchmarkInferenceQwen3.5GatedDeltaNethttps://jaesolshin.com/posts/lightrag-pg-rcte/https://jaesolshin.com/posts/lightrag-pg-rcte/Implementing LightRAG's BaseGraphStorage on plain PostgreSQL with RCTE — why a 1-hop-dominant retrieval pattern fits flat SQLWed, 20 May 2026 08:35:02 GMTAILightRAGPostgreSQLRAGGraphRAGApache AGERecursive CTEGraph StorageBenchmarkhttps://jaesolshin.com/posts/codex-app-server-sdk/https://jaesolshin.com/posts/codex-app-server-sdk/A Python SDK over the codex app-server stdio interface — install, first call, thread model, main methodsWed, 20 May 2026 08:13:38 GMTAICodexOpenAIPython SDKJSON-RPCGPT-5Coding Agentstdiohttps://jaesolshin.com/posts/claude-code-multi-account/https://jaesolshin.com/posts/claude-code-multi-account/Run isolated Claude Code accounts on the same Mac with one env var and a small zsh functionWed, 20 May 2026 08:01:28 GMTAIClaude CodeMulti-accountzshDotfilesProductivityConfigurationDeveloper Workflowhttps://jaesolshin.com/posts/niah-lost-in-the-middle/https://jaesolshin.com/posts/niah-lost-in-the-middle/NIAH limits, the Lost in the Middle effect, alternative benchmarks, and measured recall across four reasoning-effort modesWed, 20 May 2026 07:56:10 GMTAILLMLong-contextNIAHBenchmarkLost-in-the-MiddleRULERLongBenchEvaluationReasoninghttps://jaesolshin.com/posts/in-db-ai-five-systems/https://jaesolshin.com/posts/in-db-ai-five-systems/Oracle 23ai, EDB AIDB, PostgresML, Timescale pgai, pg_aidb — five systems and the abstraction-boundary discussionWed, 20 May 2026 07:39:17 GMTDBPostgreSQLAIRAGin-db-aipgaiOracleEnterpriseDBTimescaleSupabaseMongoDBClickHousearchitecturehttps://jaesolshin.com/posts/autocorrect-zsh/https://jaesolshin.com/posts/autocorrect-zsh/Gemini Flash structured output and zsh preexec/precmd hooks — failed commands fixed in place, no new terminal appWed, 20 May 2026 07:29:17 GMTDevelopmentzshGeminiTerminalAIShellProductivityHooksstructured-outputhttps://jaesolshin.com/posts/postgresql-replaces-elasticsearch/https://jaesolshin.com/posts/postgresql-replaces-elasticsearch/textsearch_ko (MeCab) + pg_textsearch BM25 + pgvector HNSW + DB-side RRF matches Elasticsearch on Korean search quality and runs 2–5× faster on a single node. Eight phases of measurement, head-to-head against ES, Qdrant, and Vespa.Wed, 20 May 2026 07:10:58 GMTDBPostgreSQLElasticsearchKoreanBM25pgvectorHybrid SearchRRFMeCabQdrantVespaRAGhttps://jaesolshin.com/posts/nvidia-nim-api/https://jaesolshin.com/posts/nvidia-nim-api/NVIDIA's build.nvidia.com offers 100+ models on H100 infrastructure for free. Plug it directly into Claude Code, Cursor, or any OpenAI-compatible coding agent.Wed, 20 May 2026 02:34:05 GMTAINVIDIANIMAPIDeepSeekfreeinferenceLLMcoding agenthttps://jaesolshin.com/posts/claude-code-sync-settings/https://jaesolshin.com/posts/claude-code-sync-settings/Sync matrix, migration steps, common failures, and diagnostic commands for running multiple Claude Code accounts on one machineWed, 20 May 2026 00:00:00 GMTAIClaude CodeMulti-accountSettings SyncTroubleshootingConfigurationMCPDeveloper Workflowhttps://jaesolshin.com/posts/disk-utils-zsh-and-claude-skill/https://jaesolshin.com/posts/disk-utils-zsh-and-claude-skill/The same disk-management logic exposed as zsh aliases (dh/dl/dol) and as a Claude Skill, with stale-while-revalidate caching in the shellWed, 20 May 2026 00:00:00 GMTDevelopmentzshClaude CodeSkillDotfilesDisk Managementstale-while-revalidateDev Workflowhttps://jaesolshin.com/posts/graph-db-benchmark-rcte-vs-age/https://jaesolshin.com/posts/graph-db-benchmark-rcte-vs-age/On a 1.14M-edge knowledge-graph workload, PostgreSQL RCTE beats Apache AGE by 290x — tracing the cypher() wrapper's 13ms cost and PG plan generation accumulationWed, 20 May 2026 00:00:00 GMTAIGraphDBPostgreSQLRCTEApache AGENeo4jGraphRAGBenchmarkLightRAGRecursive CTEhttps://jaesolshin.com/posts/openai-models-comparison/https://jaesolshin.com/posts/openai-models-comparison/I ran 30 trials per configuration across GPT-4, GPT-5, and o-series models using three reasoning problems. gpt-5-nano on minimal scored 4.4%. o1 scored lower than gpt-4o.Fri, 15 May 2026 18:46:52 GMTAIOpenAIGPTmodel comparisonlanguage modelsreasoning modelsperformance analysishttps://jaesolshin.com/posts/github-inference-api/https://jaesolshin.com/posts/github-inference-api/How to easily call modern AI models such as GPT-4.1 and DeepSeek R1 through GitHub's free model inference APIFri, 15 May 2026 18:43:29 GMTAIGitHubInference APIAPImachine learningGPT-4DeepSeekOpenAIhttps://jaesolshin.com/posts/gpt5-web-search-api/https://jaesolshin.com/posts/gpt5-web-search-api/An experimental record of implementing web search with OpenAI's GPT-5 Responses API, focused on tool-support differences between models and how parameters shape responses. The analysis centers on web search tool compatibility between gpt-5 and gpt-5-chat-latest.Fri, 15 May 2026 18:43:29 GMTAIGPT-5APIweb searchOpenAItutorialhttps://jaesolshin.com/posts/alife_summary/https://jaesolshin.com/posts/alife_summary/A guide to the field of Artificial Life, introduced at the ALife 2025 conference exhibition 'Life As It Could Be'.Thu, 14 May 2026 00:00:00 GMTResearchartificial lifealifecomplexity scienceemergenceconference