ARCHIVE
2026
2026-10 10 posts
-
What Happens to HNSW Recall After Adding Vectors? (1M+1M)
A 1M+1M insertion study in pgvector and Qdrant, comparing insertion order, batch boundaries, and fresh indexes. Segment telemetry and version-pinned code help explain their different behavior.
-
What Happens to HNSW Recall After Adding Vectors? (10K+10K)
A small pgvector and Qdrant experiment compares recall before and after vector insertion, then against fresh indexes over the same final data.
-
Minority Responses in Persona-Prompted LLMs: Following Architect of Silence
The representation problem raised by Architect of Silence, followed by the design, findings, and limits of independent checks and exploratory follow-up experiments.
-
How Managed Agents Takes an Agent Apart: Brain, Hands, and Session
The design reasoning behind Anthropic splitting an agent into a session, a harness, and a sandbox, and what happens when you run all four parts of the official workshop repository: custom tools, multi-agent, memory, and file output
-
Model Providers Are Taking Over the Agent Runtime: Managed Agents and the OpenAI Agents API
What Anthropic Managed Agents and the OpenAI Agents API add to each provider's lineup, how the two runtime APIs converged on the same abstractions, a run of the same task on both, and a comparison with local harnesses and cloud coding products
-
Using Claude Subscription Credits for the API
Which API paths the monthly credit covers on Max and Team plans, how to link an organization and manage keys, calls through the Python SDK, ant CLI, and Batch API, and how deduction and spend caps work
-
In the Agent Era, the Scope a Specialist Software Supplier Can Choose
Part 4 of 4: lock agent features in one dedicated agent, or expose them as tools and work assets any worker can use? It shapes what a supplier can sell.
-
Seeds and Germination: Software That Evolves in the Field
Part 3 of 4: fitting a product to each customer site with people and agents, building runtime adaptation into it, and feeding field experience back.
-
Every Organization Needs an Agent Runtime
Part 2 of 4: why even air-gapped organizations will run a general-purpose agent runtime on work PCs, and how to connect it to internal models and tools.
-
Separating the Worker from the Tool
Part 1 of 4: work spans products, so general-purpose agents become workers and products their tools. What to expose, and where software keeps its value.
2026-09 2 posts
-
Can tgrep, CodeGraph, Mantic, or ripwire Replace ripgrep?
tgrep, CodeGraph, Mantic, and ripwire vs ripgrep on eight repos. Each hinges on a different variable: query, language coverage, file names, or phrasing.
-
ego-browser and Aside Fail Differently
Same missions, two browser agents: ego-browser isolates its tabs and returned success on missed clicks; Aside works in your real tabs and threw errors.
2026-07 2 posts
-
Driving a Free Colab GPU from the CLI
google-colab-cli drives a free Colab GPU from your laptop. Worked example: serving PaddleOCR-VL on a T4 and running OCR on an invoice over a public URL.
-
Wiring codex to a Local Qwen — ollama, vLLM, and the developer role
Point codex CLI at a local model on ollama or vLLM. ollama needs one config.toml; vLLM rejects the developer role, fixed by a template patch or a proxy.
2026-06 1 post
-
Speeding up LLM inference with MTP and diffusion
Gemma 4 and Qwen 3.6 on one H100 at fp8: MTP ran 1.43–2.62x faster with GPQA accuracy effectively unchanged; diffusion was ~4x faster but less accurate.
2026-05 29 posts
-
Requirements as Latent State
A requirement is a latent state and the user's sentence a noisy observation. What follows for EARS, spec-driven development, harnesses, and when to ask.
-
Skill and Harness
A skill compresses how to behave; a harness organizes how to tell if that behavior nears the goal. Why the two still contain each other in practice.
-
Harness as Environment
A vector DB full of experience does not make an agent adapt. That takes a harness: what may change, how success is judged, how far a failure rolls back.
-
Memory as Adaptation
Long-running agents need more than a RAG store: why memory split into working, episodic, semantic, and procedural layers that ground policy adaptation.
-
Weights, Prompts, Codes as Parameters
If agent problems are POMDPs, weights, prompts, and code are parameters at three layers of one policy space. Why the layers opened in that order.
-
Comparing Four LightRAG Variants — Same Root, Different Production Strategies
Three LightRAG derivatives read at source level: RAG-Anything adds multimodal, ApeRAG becomes a platform without multi-hop, EdgeQuake rewrites it in Rust.
-
PostgreSQL Lakehouse (2/2) — Distributed Processing and Citus Integration
Distributed processing (Ray/Daft/Smallpond) and Citus + pg_lake FDW integration
-
PostgreSQL Lakehouse (1/2) — When DuckLake Hit a Wall, pg_lake Was There
Eight phases building a PostgreSQL-centered Lakehouse: DuckLake's libpq collision and pg_lake
-
PG-Strom SSB Benchmark — Arrow FDW Comes Before GPU
555 test cases on three servers: on heap storage PG-Strom's GPU gave 1.6x and lost to CPU on wide tables; with Arrow FDW it hit 10x at SF=100.
-
GraphDB Benchmark (2/2) — Workload Matrix and Final Recommendations
Eight graph engines on OLTP, memory, analytics, and differentiation queries, with a decision table by workload. On OLTP, RCTE tied MemGraph at 22.5K RPS.
-
PDF to Markdown — Five Tools Compared
Five PDF-to-Markdown converters on one academic paper, 100-point rubric: mineru scored 92 with formulas and tables intact; markitdown scored 14.
-
Markdown to Slides and PDF — A zsh Pipeline
Two zsh functions turn markdown into a Beamer slide PDF or an A4 document PDF via pandoc and xelatex; a third sets macOS screenshot location and format.
-
Sharing Claude Code Sessions via Symlinked .jsonl
Resume the same Claude Code session across accounts by pointing every projects directory at one shared physical path
-
Apple Silicon LLM Inference — Five Backends Compared
Qwen3.5-9B on MLX, llama.cpp, Ollama, omlx, and vLLM Metal. MLX led single requests at 25 tok/s; only omlx batched concurrent requests.
-
LightRAG Without Apache AGE — Graph Storage in Recursive CTE
Implementing LightRAG's BaseGraphStorage on plain PostgreSQL with RCTE — why a 1-hop-dominant retrieval pattern fits flat SQL
-
Codex App Server Python SDK — JSON-RPC v2 over stdio
A Python SDK over the codex app-server stdio interface — install, first call, thread model, main methods
-
Running Multiple Claude Code Accounts on One Machine
Run isolated Claude Code accounts on the same Mac with one env var and a small zsh function
-
Long-Context Evaluation — NIAH and Lost in the Middle
NIAH limits, the Lost in the Middle effect, alternative benchmarks, and measured recall across four reasoning-effort modes
-
Should Databases Handle AI Too?
Oracle 23ai, EDB AIDB, PostgresML, Timescale pgai, pg_aidb — five systems and the abstraction-boundary discussion
-
autocorrect.zsh — 167 lines of zsh that fix failed commands
Gemini Flash structured output and zsh preexec/precmd hooks — failed commands fixed in place, no new terminal app
-
Replacing Elasticsearch with PostgreSQL
Korean hybrid search on PostgreSQL alone (MeCab, BM25, pgvector, in-DB RRF) over eight phases: Elasticsearch-level quality, 2–5x faster on one node.
-
NVIDIA NIM API — Free Inference for GLM, Kimi, Nemotron, and Gemma 4
build.nvidia.com serves 100+ open models via a free OpenAI-compatible API, no credit card, up to 40 RPM per model. Setup for Claude Code and Cursor.
-
GraphDB Benchmark, Eight Engines (Part 1) — Decomposing the RCTE 290x Gap
On a 1.14M-edge knowledge graph, PostgreSQL RCTE had 290x the throughput of Apache AGE, traced to a ~13ms cypher() wrapper cost and plan generation.
-
Disk Utilities Through Two Surfaces — zsh Aliases and a Claude Skill
One set of zsh disk functions, two surfaces: aliases (dh, dgh, dol) for people and a Claude Skill for agents, with a stale-while-revalidate cache.
-
Claude Code Settings Sync and Troubleshooting
Sync matrix, migration steps, common failures, and diagnostic commands for running multiple Claude Code accounts on one machine
-
10 OpenAI Models Through Quick Benchmarks — The Model Isn't as Smart as You Pay
19 configurations of 10 OpenAI models, 30 runs each on three reasoning problems. gpt-5-nano on minimal scored 4.4%; o1 scored 47.8%, below gpt-4o at 66.7%.
-
GitHub Models Inference API — Free Model Access Tested
Call GPT-4.1, DeepSeek R1, and other models free via GitHub Models with a personal access token and the OpenAI Python SDK. Rate and token limits apply.
-
Experimenting with the GPT-5 Responses API Web Search Tool
In the GPT-5 Responses API, web search worked with gpt-5 but returned a 400 error on gpt-5-chat-latest. Plus notes on search_context_size and verbosity.
-
Life As It Could Be
A guide to the field of Artificial Life, introduced at the ALife 2025 conference exhibition 'Life As It Could Be'.
No matching posts.