Every dispatch, measured
First-party numbers by default — each dispatch carries its verified-hardware tag, test date, and a
replication appendix. Raw markdown at
/posts/<slug>.md; the machine index at
/posts.json; feed at
/rss.xml. Multi-episode field
reports live on the series hub.
Expat Playbook entries are the editorial exception — source-verified reporting, no lab
hardware; their hardware tag names the desk and the cross-checked sources instead.
Industrial AI · 13 dispatches
- Machines Keep the Watch · E5 Measured 2026-09-13
The Staircase: Four Steps for Every Job in the Plant, and the Top Step Stays Empty
Every plant job on a four-step staircase — dashboards, scripts, a model call, agents. The top step stays empty; the agent purchase case collapses.
- Machines Keep the Watch · E4 Measured 2026-09-12
The Whole Factory Had One AI Call — and 97% of Its Bill Was Door Fee
Token audit of the plant's only LLM call: ~52K steady-state tokens per call, ~97% door fee, zero tools used — rebuilt same-model to ~2K in one day.
- Machines Keep the Watch · E6 Measured 2026-09-12
The Audit Came Home: 1.8 Billion Tokens, 95% Re-read — the Cure Was Forgetting
The Episode-4 audit came home to the operator's own desk: 1.8B tokens in 30 days, 95% of it cache re-reads. The cure — compaction — costs detail memory.
- Machines Keep the Watch · E3 Measured 2026-09-11
The Hardest Paper Came Back Blank Twice — and the Fix Was the Exam Rules, Not the Model
An on-prem 27B's hardest day: two blank papers, one config race, zero model swaps — four real production days passed after the exam rules changed.
- Machines Keep the Watch · E1 Measured 2026-09-05
No One on the Line Tonight: The Dress Rehearsal Before an AI Agent's First Unsupervised Night Shift
Unattended dress rehearsal before an AI agent's first night shift: five gates on 2026-09-05 — 25-s ignition check, zero-second alignment, canary, clean restore.
- Machines Keep the Watch · E2 Measured 2026-09-05
19:00:00 Sharp, Two Lines, Zero-Second Error — and Not One Bit of AI in That Second
The rehearsal's most precise moment — two lines, same second, zero error at 19:00:00 — came from one line of Bash arithmetic, not a model.
- Machines Keep the Watch · E0 Measured 2026-09-05
What Does the Future Production Line Look Like? Full-Inspection Vision, Patrol Agents, and an On-Prem 27B LLM in One Discrete-Manufacturing Plant
A live manufacturing plant: frame-by-frame vision inspection, three-a-day patrol agents, and a 27B on-prem LLM — the first unsupervised night shift, 2026-09-05.
- Measured 2025-11-19
How Do You Ship Industrial Defect Detection With Only 'Good' Samples? An Anomalib Field Guide
anomalib on a two-GPU workstation: all 15 MVTecAD scenes trained and measured (mean image-AUROC 0.981), plus the three real bugs the sweep had to fix first.
- Measured 2026-08-30
What Does AI Work in China Actually Need From the Network? A Field Map of Walls, Mirrors, and One Rented Computer
What breaks (GitHub, HF, PyPI, Omniverse — with receipts), what doesn't (domestic inference), and the architecture that worked: rent a computer, not a tunnel.
- Measured 2026-08-30
Can Simulation Fill Your Defect-Sample Gap? An Isaac Sim Route That Never Got Past the Install
The synthetic-data plan, three staged repos, 19 zero-length launch logs, and the download wall that stopped it — a field note on a line not yet flown.
- Measured 2025-11-20
Can a Mac Mini Run Industrial Defect Detection on CPU? A Workstation-to-Edge Deployment That Actually Ran
PatchCore from the workstation, byte-identical checkpoints on a Mac mini, a Basler camera through Aravis, 0.156-0.176 s CPU inference — a deployment that ran.
- Measured 2026-08-30
How Does a Vision Model Reach a Factory-Edge NPU? The ONNX→RKNN→RK3588 Route, Staged and Smoke-Tested
Three Rockchip docker images (6.86 GB total), the two-compiler reality, and the route from workstation GPU to a 6-TOPS edge NPU — staged for real.
- Measured 2026-08-30
SAM3 on a Workstation: Docker Door, Source Door, and a Wire Into ComfyUI
Meta's SAM3 on one workstation: a scripted docker door never pulled, a 131 MB source clone, and a real 533-line ComfyUI node package with saved blueprints.
Expat Playbook · 5 dispatches
- Measured 2026-08-30
Why Is It So Hard to Find a Human to Argue With? Five Things Vanishing From Digital Life in China
AI support walls, algorithmic black boxes, staged 'news', subscription traps, and the people left outside — plus the four counter-moves that still work.
- Measured 2026-08-30
What Is Jurisdiction Shopping — and Why Can Almost Nobody Afford It? One Very Mobile Case Study
A biography as portfolio: Qinghai-born, Beijing-schooled, Geneva-posted, Seychelles-domiciled. What jurisdiction shopping buys, and why it costs nine figures.
- Measured 2026-08-30
What Is the 'Gray Zone' — the Condition Below War and Above Peace? A System Diagnosis, Not a Prediction
Normalization, law-enforcement costume, and attrition: three parts of a machine that moves boundaries without ever producing a headline day.
- Measured 2026-08-30
Why Do Analysts Say 'Quarantine' Instead of 'Blockade' — and What Did Kennedy's 1962 Word Choice Teach?
One word separates an act of war from a law-enforcement action. A vocabulary decoder for cross-border readers, from 1962's quarantine to today's coverage.
- Measured 2026-08-29
Who Called for Interning Chinese Americans If War Comes — and Why Did Chinese Comment Sections Cheer?
Podcast panelists floated internment; Congress condemned it; Chinese comments cheered. A field note on loyalty policing from both directions.
AI & Compute · 27 dispatches
- One Man One Legion · E1 Measured 2026-08-27
How Fast Does a Chat Command Become a Finished AI Image? Inside the One-Man FLUX Workshop
Episode 1: a chat phrase triggers FLUX on a 16 GB GPU and the image posts itself back — 10 logged runs: 50 s fast, 66–112 s warm, 330–379 s cold.
- One Man One Legion · E10 Measured 2026-08-27
Whose Name Goes on the Article? Five Bylines, One Human
E10: the byline rack — 33 signed files across five genre pen names, one column carrying all five, and a sixth persona that never shipped a word.
- One Man One Legion · E11 Measured 2026-08-27
Who Checks the Checkers? 313 Lines of Gate, 61 Stamps, One Empty Column
E11: inside the QC gate — 313 lines of regex scans and model judgments; the 102-record failure log shows what each layer caught, and what never got written.
- One Man One Legion · E12 Measured 2026-08-27
The Analyst Desk: 85 Points of Fact, ±13 of Feeling, One Dead Cron
E12: the shut-down analyst desk — six agents, four signal layers, a score of 85% fact plus ±13 of feeling, and dual-source gates that verified 29/29 points.
- One Man One Legion · E13 Measured 2026-08-27
The Audio Workshop: Six Voices on a Shelf, One Episode That Never Shipped
E13: the audio workshop — six reference voices, a 172-line IndexTTS-2 pipeline, a lost episode, and three podcast crons gone dark.
- One Man One Legion · E14 Measured 2026-08-27
The Model Stable: Two Days in June, One Evening in August
E14: the engine room's model stable — three local models totaling 70 GB, a two-day deployment battle, and the fallback horse that isn't saddled today.
- One Man One Legion · E16 Measured 2026-08-27
Model or Harness? Two Controlled Experiments on a 363-Line Hand-Rolled Coding Agent
E16: swap the brain — 15 rounds stuck vs 4 to pass; swap the prompt — weak models stay unsaveable. Three code supervisors make agents honest.
- One Man One Legion · E17 Measured 2026-08-27
The Model Exam: Seven Passes, Two Zeroes, One Home
E17: the fleet grades its local 27B — a 7/7 baseline, an A/B with two zero-output deaths, a four-habitat race, six rules for 241 tokens.
- One Man One Legion · E18 Measured 2026-08-27
The Brain Swap: Ten Backups, One Insurance Card, No Standby Factory
E18: swapping the factory's primary model while it runs — a transplant at age 2 hours, a 139-day gap, and an evening double-swap with a rollback card.
- One Man One Legion · E19 Measured 2026-08-27
What Should an AI Legion Do While You Sleep? 22 Cron Entries, Four Still Lit
E19: the scheduler grew to 16 jobs lit at once — 790 runs, 154 errors — then on Jul 23 the auto-publishers went dark. Four relit: feed or ask, never publish.
- One Man One Legion · E2 Measured 2026-08-27
What Does It Take to Give a Chat AI Nine Tools and Your Home Directory? A 50-Line Bot That Grew to 850
E2: a Telegram gateway to GLM-5.2 — nine tools, a jailed workdir, a dangerous-command blocklist; 50 lines grew to 850, logged in 168,996 lines.
- One Man One Legion · E20 Measured 2026-08-27
Bug Diaries: 9 Failure Files, 158 Days, and the Root Cause It Got Wrong
E20: the legion files its own failures — a 901-line error archive, a review rota dead on arrival, and a fabricated match report no rule caught.
- One Man One Legion · E21 Measured 2026-08-27
The LangGraph Autopsy: 12 Green Imports, 668 Silent Seconds, One 203-Line Survivor
E21: 12 green imports, 668 silent seconds — the autopsy that deleted a framework 90 minutes in, and the 203-line engine that outlived it.
- One Man One Legion · E23 Measured 2026-08-27
The Security Handcraft: 12 Fossil Scripts, One Live Key, Zero Leaks in Print
E23: the leak scanner turned on the fleet itself — a 138-field single home, 12 legacy scripts sharing one inline token, and 2 live strings for one bot.
- One Man One Legion · E24 Measured 2026-08-27
The Token Ledger: 261,720,417 Tokens, 147 Re-reads per Token Written
E24: the legion's first honest ledger — 208 days, 261,720,417 metered tokens, 84.5% cache re-reads, 0.68% output, and a 75-day silence the bill remembers.
- One Man One Legion · E25 Measured 2026-08-27
The Human Ledger: 26 Drops, 191 Pieces Behind One Phone, Five One-Line Rulings
E25: the other side of the ledger — 26 drops in 171 days, 191 pieces capped by a phone tap, five one-line rulings, every act receipted by a machine.
- One Man One Legion · E26 Measured 2026-08-27
The Loop Closes: What One Man and One Legion Answered to a Turbulent Era
E26, the finale: 24 posts, 121 ledger entries, 261,720,417 tokens against roughly 225 human acts — the series' measured reply to Gates' warning.
- One Man One Legion · E3 Measured 2026-08-27
What Happens When the Cockpit Itself Goes Dark? Two Outages, Ten Stranded Jobs, One Spare Parked by Subtraction
E3: the chat channel died twice — 90-second stalls, then 10 jobs stranded at deliver; the fix was a 6-node pipe, and the spare parked for good.
- One Man One Legion · E4 Measured 2026-08-27
What Does the One Human Actually Do? Three Touchpoints per Piece, and One Wall That Won't Move
E4: the human appears in exactly 3 places per piece — drop material, tick a slip, click publish; on WeChat an API wall makes the phone the button.
- One Man One Legion · E5 Measured 2026-08-27
Six Doors, 231 Lines, and No Dashboard: Putting a Three-Machine Fleet on One Screen
E5: six ssh doors, a 231-line probe on a 5-minute clock, and 16 days of alerts — the fleet's one screen is a chat thread and a 12-key JSON, not a dashboard.
- One Man One Legion · E6 Measured 2026-08-27
What Does an AI Legion Write to Itself? 63 Entries, 75 Days of Silence
E6: a 292-line constitution orders the agent to wake up reading its own diary — 63 entries over 186 days, one 75-day silence when cron took over.
- One Man One Legion · E7 Measured 2026-08-27
One Sentence In, Stereo Video Out: 6 Jobs, 90–790 Seconds, an Unfilmed Actress
E7: a local H3 video bench — 6 logged jobs, 90 to 790 seconds on one 4090, a director skill with a motion budget, and a heroine with no footage yet.
- One Man One Legion · E9 Measured 2026-08-27
Who Edits an AI Writing Staff? 191 Articles, 15 Iron Laws, One Stamp
E9: a one-person WeChat factory archived 191 articles under 15 iron laws; 7 machine scans gate every push, and a 102-record lesson log turns repeats into rules.
- One Man One Legion · E0 Measured 2026-08-27
One Man, One Legion: Can a Single Operator Run a 3-Machine AI Fleet End to End?
Audited 2026-08-27: 3 machines, 3 local LLMs, 5 diffusion families, 22 cron jobs (4 on), 12 chat groups, 191 archived article files, 3 human touchpoints.
- Measured 2026-08-26
Can a Static Blog Hand AI Agents Real Tools? Wiring ARD + MCP into an Astro Site (Measured)
A +950-line commit gives a static blog an ARD catalog, OpenAPI tools, JSON indexes and a read-only MCP server. Live in 12–18 s; stress-tested, 5 gaps fixed.
- Measured 2026-02-19
InfiniteTalk Dies at torch.load: the Error Tells You to Upgrade torch — the Measured Fix Is Pinning transformers 4.52.0
Measured: the CVE-2025-32434 torch.load gate says upgrade torch — torch 2.10 then breaks four packages at once; the fix is pinning transformers 4.52.0.
- Measured 2025-11-29
Can You Run InfiniteTalk on Two Consumer GPUs? Yes — at 218.5 s/step (RTX 4090D + RTX A4000, Measured)
Measured on an RTX 4090D 24GB + RTX A4000 16GB rig: InfiniteTalk 14B fp8 fits with zero OOM via 20-block semi-residency — at 218.5 s/step, ~55 min per clip.