Blog posts
2026-11 2 posts
-
xsm: A Message Channel for Agent Collaboration
Preserving existing work environments and agent autonomy while organizing collaboration that people can understand and intervene in.
-
Weight-Space Programming
How an LLM edited weights directly, and where it leads
2026-10 10 posts
-
What Happens to HNSW Recall After Adding Vectors? (1M+1M)
HNSW recall in pgvector and Qdrant after adding one million vectors
-
What Happens to HNSW Recall After Adding Vectors? (10K+10K)
HNSW recall in pgvector and Qdrant after adding ten thousand vectors
-
Minority Responses in Persona-Prompted LLMs: Following Architect of Silence
Minority responses lost in persona-prompted LLM surveys
-
How Managed Agents Takes an Agent Apart: Brain, Hands, and Session
The session, harness, and sandbox split in Managed Agents, with a workshop run
-
Model Providers Are Taking Over the Agent Runtime: Managed Agents and the OpenAI Agents API
Runtime abstractions in Managed Agents and the OpenAI Agents API
-
Using Claude Subscription Credits for the API
Scope and use of the API credits in Max and Team plans
-
In the Agent Era, the Scope a Specialist Software Supplier Can Choose
What a specialist software supplier can sell in the agent era
-
Seeds and Germination: Software That Evolves in the Field
The structure of software that adapts and evolves at customer sites
-
Every Organization Needs an Agent Runtime
General-purpose agent runtimes in air-gapped organizations
-
Separating the Worker from the Tool
The separation of worker and tool in the age of general-purpose agents
2026-09 2 posts
-
Can tgrep, CodeGraph, Mantic, or ripwire Replace ripgrep?
Four code search tools against ripgrep on eight repositories
-
ego-browser and Aside Fail Differently
How ego-browser and Aside fail on the same browser tasks
2026-07 2 posts
-
Driving a Free Colab GPU from the CLI
A free Colab GPU driven from colab CLI, with a PaddleOCR-VL example
-
Wiring codex to a Local Qwen — ollama, vLLM, and the developer role
Connecting codex CLI to a local Qwen on ollama and vLLM
2026-06 1 post
-
Speeding up LLM inference with MTP and diffusion
Speed and accuracy of MTP and diffusion decoding on Gemma 4 and Qwen 3.6
2026-05 29 posts
-
Requirements as Latent State
Requirements as latent states and natural language as their observation
-
Skill and Harness
The relation between skills that compress behavior and harnesses that judge it
-
Harness as Environment
The harness as the condition for agent adaptation
-
Memory as Adaptation
Memory layers and policy adaptation in long-running agents
-
Weights, Prompts, Codes as Parameters
Weights, prompts, and code as three layers of parameters in one policy space
-
Comparing Four LightRAG Variants — Same Root, Different Production Strategies
A source-level comparison of three LightRAG derivatives
-
PostgreSQL Lakehouse (2/2) — Distributed Processing and Citus Integration
Distributed processing engines and Citus integration for a PostgreSQL lakehouse
-
PostgreSQL Lakehouse (1/2) — When DuckLake Hit a Wall, pg_lake Was There
DuckLake's libpq collision and a pg_lake-based lakehouse
-
PG-Strom SSB Benchmark — Arrow FDW Comes Before GPU
PG-Strom GPU acceleration and Arrow FDW on the SSB benchmark
-
GraphDB Benchmark (2/2) — Workload Matrix and Final Recommendations
Workload performance and selection criteria for eight graph engines
-
PDF to Markdown — Five Tools Compared
Five PDF-to-Markdown converters evaluated on an academic paper
-
Markdown to Slides and PDF — A zsh Pipeline
zsh functions for markdown-to-slides and PDF with pandoc and XeLaTeX
-
Sharing Claude Code Sessions via Symlinked .jsonl
Session sharing across Claude Code accounts through symlinks
-
Apple Silicon LLM Inference — Five Backends Compared
Single and concurrent request performance of five LLM backends on Apple Silicon
-
LightRAG Without Apache AGE — Graph Storage in Recursive CTE
LightRAG graph storage on plain PostgreSQL recursive CTEs
-
Codex App Server Python SDK — JSON-RPC v2 over stdio
Installation and call structure of a Python SDK for codex app-server
-
Running Multiple Claude Code Accounts on One Machine
Multiple Claude Code accounts with CLAUDE_CONFIG_DIR and a shell function
-
Long-Context Evaluation — NIAH and Lost in the Middle
The limits of NIAH and measurements of Lost in the Middle
-
Should Databases Handle AI Too?
Five in-database AI systems and the abstraction boundary
-
autocorrect.zsh — 167 lines of zsh that fix failed commands
A zsh script that fixes failed commands with Gemini Flash and shell hooks
-
Replacing Elasticsearch with PostgreSQL
Korean hybrid search on PostgreSQL alone, compared with Elasticsearch
-
NVIDIA NIM API — Free Inference for GLM, Kimi, Nemotron, and Gemma 4
Usage and limits of the free NVIDIA NIM inference API
-
GraphDB Benchmark, Eight Engines (Part 1) — Decomposing the RCTE 290x Gap
The causes of the throughput gap between PostgreSQL RCTE and Apache AGE
-
Disk Utilities Through Two Surfaces — zsh Aliases and a Claude Skill
Disk utilities shared by zsh aliases and a Claude Skill
-
Claude Code Settings Sync and Troubleshooting
Settings sync and troubleshooting for multiple Claude Code accounts
-
10 OpenAI Models Through Quick Benchmarks — The Model Isn't as Smart as You Pay
Accuracy of ten OpenAI models on three reasoning problems
-
GitHub Models Inference API — Free Model Access Tested
Usage and limits of the free GitHub Models inference API
-
Experimenting with the GPT-5 Responses API Web Search Tool
Model-by-model behavior of the GPT-5 Responses API web search tool
-
Life As It Could Be
A guide to Artificial Life through the ALife 2025 exhibition
No matching posts.