Skip to content
MoonWorkLab
  • Local LLM & GPU
  • AI-Assisted Dev
  • Hardware & Desktop
  • Self-Hosting
  • About
  • Contact
MoonWorkLab
  • Self-Hosting & Infrastructure

    One Requirement, Three Different Builds: Read It, Tap It, Or Give Up

    ByMoonWork September 24, 2026November 4, 2026

    Reading a notification, tapping an approval, and capturing a protected screen required different designs. The lock-screen limit applies to the device and path tested.

    Read More One Requirement, Three Different Builds: Read It, Tap It, Or Give UpContinue

  • Local LLM & GPU

    The server was fine. SSH was fine. The firewall rule was fine. The block was on a screen nobody was looking at.

    ByMoonWork September 15, 2026November 4, 2026

    Remote requests timed out while local requests and SSH still worked. Inspecting the Windows desktop revealed a pending prompt; remote requests recovered after Allow.

    Read More The server was fine. SSH was fine. The firewall rule was fine. The block was on a screen nobody was looking at.Continue

  • Local LLM & GPU

    One More Offloaded Layer Coincided With a 29% Throughput Drop.

    ByMoonWork September 10, 2026November 4, 2026

    Moving an idle occupier from one GPU layer to two coincided with a 29% generation-speed drop. The 70MB counter delta does not establish the cause.

    Read More One More Offloaded Layer Coincided With a 29% Throughput Drop.Continue

  • Local LLM & GPU

    My Compressed KV Cache Ran 43% Slower at 7k, Then Crashed at 32k.

    ByMoonWork September 2, 2026November 4, 2026

    At depth 7,168, compressed cache reached 5.27 tok/s versus 9.31 for fp16 in the same fork. A later run exited with 0xC0000409; the crash cause is unconfirmed.

    Read More My Compressed KV Cache Ran 43% Slower at 7k, Then Crashed at 32k.Continue

  • Three features — keyboard navigation, bookmark notes, and the settings page — each marked built and working, but each reachable only by a path the developer knew. Thirty days of daily use found zero; one tester in about an hour found three.
    AI-Assisted Development

    I Used My Own Extension for 30 Days. One Tester Found Three Bugs in an Hour.

    ByMoonWork August 25, 2026November 4, 2026

    One tester exposed keyboard-focus and feature-discovery problems I had missed. A later correction separates fetch errors from unverified root causes.

    Read More I Used My Own Extension for 30 Days. One Tester Found Three Bugs in an Hour.Continue

  • Register sweep of two monitors: Monitor A returns seven different plausible values including a fictional input code of 17, while Monitor B returns 110 for every register and a maximum brightness of minus 128. Every call exited 0.
    Hardware & Desktop

    I Read the Same Value Five Times and Called It Verified. It Was Backwards.

    ByMoonWork August 25, 2026November 4, 2026

    An unsupported monitor input query returned stable values. A visual switching check contradicted my mapping and exposed the limits of repeatability.

    Read More I Read the Same Value Five Times and Called It Verified. It Was Backwards.Continue

  • Korean sentence meaning it does not OOM, translated by three backends: Apple on-device in 0.44 seconds reversed it to OOM does not work, while ollama and claude preserved the meaning
    Local LLM & GPU

    On-Device Translation Turned “It Doesn’t OOM” Into “OOM Doesn’t Work”

    ByMoonWork August 25, 2026November 4, 2026

    Ten technical Korean-to-English cases completed without errors, yet several outputs changed the meaning. I retained review instead of automating unchecked posting.

    Read More On-Device Translation Turned “It Doesn’t OOM” Into “OOM Doesn’t Work”Continue

  • An Idle Image App Held 2.65GB of VRAM. Closing It Made My Local Model 46% Faster.
    Local LLM & GPU

    An Idle Image App Held 2.65GB of VRAM. Closing It Made My Local Model 46% Faster.

    ByMoonWork August 19, 2026November 4, 2026

    One GPU occupancy comparison showed 28.15 versus 41.00 tok/s. A server restart separates the conditions, and the memory-spill mechanism was not directly observed.

    Read More An Idle Image App Held 2.65GB of VRAM. Closing It Made My Local Model 46% Faster.Continue

  • Same Python, Same Command — One Launch Path Passed, the Other Hit WinError 4551
    Self-Hosting & Infrastructure

    Same Python, Same Command — One Launch Path Passed, the Other Hit WinError 4551

    ByMoonWork August 17, 2026November 4, 2026

    Manual SSH import passed while a WMI launch failed with WinError 4551. Disabling a policy restored that setup; the exact launch-path mechanism was not isolated.

    Read More Same Python, Same Command — One Launch Path Passed, the Other Hit WinError 4551Continue

  • My Local Model Wrote the Code. The Driver Reported $0.66 per Unit, Not a Bill.
    AI-Assisted Development

    My Local Model Wrote the Code. The Driver Reported $0.66 per Unit, Not a Bill.

    ByMoonWork August 17, 2026November 4, 2026

    Seven successful units reported an average $0.659 API-price estimate. No metered bill was charged. The logs show orchestration volume, not a controlled savings comparison.

    Read More My Local Model Wrote the Code. The Driver Reported $0.66 per Unit, Not a Bill.Continue

Page navigation

Previous PagePrevious 1 2 3 4 Next PageNext

© 2026 MoonWork Lab — every number measured on hardware I own.
About · Contact · Privacy · Affiliate Disclosure

  • Local LLM & GPU
  • AI-Assisted Dev
  • Hardware & Desktop
  • Self-Hosting
  • About
  • Contact