Notes
Field notes
Shorter writes and launch notes. For how-tos, prefer Guides.
-
Introducing AsyncQA: A Local Model That Files Bug Reports You Can Check
A long-running QA agent driven by models on a 12 GB RTX 5070. It explores a web app, verifies every suspected bug independently, and attaches the evidence and the arithmetic behind each confidence score. Tested against a shop with 11 planted bugs.
-
The Hardest Style: A Line-Art Short, and the Best Video the 5070 Can Make
A 12-second black-and-white line-art film about a father, a toddler, and a flower — the deliberately worst-case style for these models — used to find the real ceiling of local video on a 12 GB card.
-
Chasing Sharpness: Two-Pass Rendering, FLUX.2 klein, and a Judge That Needed Glasses
The first local generations were soft. Two fixes measured on the same 12 GB card — a second denoising pass with the model you already have, and a new-generation 4B model — plus a judging bug worth knowing about.
-
Gus Moves: Your Own Photos Are the Best Video Seeds
The highest motion score of the entire series came from animating a real photo of a real dog — locally, in eight minutes, without the picture ever leaving the machine.
-
How Long Can a Clip Get? Duration, Health, and Why Cartoons Don't Help
Pushing local video generation to 15 seconds on 12 GB: no memory wall in sight, cost scales linearly — and the limit that actually bites is temporal health, measured with contact sheets.
-
The Toolkit Learns to See: Local Image Generation on the Same 12 GB Card
Standing up ComfyUI on the RTX 5070 and benchmarking SDXL against Z-Image-Turbo — measured seconds per image, measured VRAM, and a fox-off the newer model wins.
-
The Card Learns to Move: Local Text-to-Video on 12 GB
LTXV 2B and Wan 2.2 5B measured on the RTX 5070 — seconds of compute per second of video, peak VRAM, and the actual clips, embedded and unretouched.
-
The Reference Shot: What 12 GB Video Actually Looks Like at Its Best
After a series of honest disappointments, the lever that finally produced a compelling clip: anchor the motion model with the best still the card can make. Image-to-video, measured, embedded.
-
Thirty Seconds, Two Ways: One Shot Collapses, Eight Shots Make a Film
Pushing a single generation to 30 seconds produces literal nothing — an orange shimmer. Composing eight 4-second shots produces a 31-second film with a plot, a consistent character, and measured motion in every cut. The film is embedded.
-
Your Upscaler Is Better Than Its Score: Metrics, Eyes, and a Local Judge
A ground-truth upscaling benchmark where the objective metric picks the wrong winner — and a local vision model, framed as a harsh critic, sides with human eyes.
-
Introducing Bakeoff: A Fair Fight Between Local Models
Run one written spec through several local models unattended, on identical baselines, and compare what they actually build — part of the local-llm toolkit.
-
I Gave Two Local Models the Same Take-Home. Neither Shipped.
A controlled experiment running qwen3-coder:30b and gpt-oss:20b unattended against an identical spec on a 12GB RTX 5070 — what they built, what broke, and why four of the five failures were mine.
-
Case study: business coaching crew in about 7 minutes
Wall-clock timings for a business-coaching crew in CrewDefine → Zero, plus why a specialized crew differs from a generic chat model for repeatable advisory work.
-
A map of the Lab Z stack
The technologies we actually touch across CrewDefine, Zero, DataMaker, and pgLens — from YAML crews and multi-agent runtimes to FastAPI, React, Postgres, and synthetic data.
-
Introducing DataScribe: Describe Data, Then See It as a Galaxy
An open-source tool that turns a plain-language conversation into a reproducible dataset—and renders the result as a living constellation so you notice things a table never shows you.
-
Introducing CrewDefine: Design Multi-Agent Crews Through Conversation
An interactive CLI that turns natural conversation into production-ready agent configurations for LabZ.
-
What's That Font? AI-Powered Typography Detection
An open-source tool that uses Claude Opus vision to identify typefaces in any image—with confidence scores, visual analysis, and instant access links.
-
Introducing pgLens: See Your Postgres Clearly
A lightweight, open-source tool for exploring PostgreSQL databases with intelligent, type-aware interfaces.
-
pgLens Now Supports SQLite
pgLens expands beyond PostgreSQL—explore SQLite databases with the same intelligent, type-aware interface you already know.
-
Biomimetic AI Architectures
What fungal networks teach us about distributed intelligence.
-
High-State, Low-Presence Systems
Designing AI that acts without demanding attention.
-
The End of Dashboards
Why static visualization is the bottleneck in modern decision-making.