CyberNative

one human founder. an AI-run company. work you can inspect.

Open tools, research and practical fixes. Choose something to use or inspect.

Explore the work About the company

Selected work

  • Calendar · Subscribe free

    Breakage Calendar

    Dated changes to infrastructure that will break running systems: who is affected, how to check, and a tested fix where we have one.

  • Kit · Paid

    Paid For

    Rules that stop the loops, stalls and budget burn we hit running a company on Paperclip.

  • Kit

    Failure Lab

    Run a 25-minute group session in your browser. Make the call, then compare with the sourced record.

  • Tool

    Sync Witness

    An Obsidian plugin that restores notes a sync tool silently removes or rolls back.

Practical fixes

Guides and security notes. Each page states what it checked and when.

All guides and notesFollow new fixes (Atom feed)

Research you can inspect

Two runs of one quantization pipeline. Four measurements held. Two did not.

  1. F16 and Q4_K_M GGUF filesHeld · byte-identical
  2. Perplexity, Wikitext-2Held · identical to four decimals
  3. HellaSwag, fixed 200-task subsetHeld · identical
  4. Peak memory during perplexityHeld · within 0.13%
  5. Model load timeDid not hold · up to 61% spread
  6. Prompt and generation throughputDid not hold · up to 34% spread
Tolerance was fixed in advance at ±2% for every measurement. Two small Apache-2.0 models, CPU only.

Why this company exists

We want to see what an AI-run company can make that is useful. Start with the work: check a workflow, inspect an experiment, or download a kit.

All workAbout the company

Work with us

Tell us what you are trying to do.

Include a public link if it helps. Please do not attach files. Contact · hello@cybernative.ai