COOPER E.
Independent AI Researcher and Neural Systems Scientist
San Diego, California
jcooperkai@gmail.com
https://jcooperkai.com/

PROFILE

Cooper E., known online as jcooperkai, trains neural networks and investigates
what is hidden inside them. The work runs from operating systems and perception
stacks written from first principles to controlled measurements of what a
trained network actually uses. Results are published whether or not they are
flattering, including the ones that came out negative.

RESEARCH FOCUS

Mechanistic interpretability - causal states, interventions, probing critique
Computability - Busy Beaver, Turing machine enumeration and deciders
Learning systems built without frameworks - NumPy, from-scratch training
Autonomous perception - segmentation, tracking, path prediction
Honest evaluation - controls, negative results, calibration

RESEARCH NOTES

These are independent drafts in progress. None is peer-reviewed, and each is
labelled as unfinished on the site itself.

1. Complete Decision of Small Turing Machine Classes by Tree-Normal-Form
   Enumeration with Translated-Cycle and Backward Deciders, 2026
   Four classical classes decided with no residue: every machine in BB(2,2),
   BB(3,2), BB(2,3) and BB(4,2) is proven either to halt or never to halt.
   Maxima recovered are 6, 21, 38 and 107 steps, matching the published values,
   with no parameter tuned against those answers. Includes a soundness defect
   found in my own backward decider, described in full because the failure mode
   was silent.

2. Short-Horizon Price Direction Is Not Predictable: A Negative Result on
   258,737 Real Minute Bars, 2026
   Walk-forward AUC 0.5105 against a 0.5 baseline across 11,497 out-of-sample
   windows, with purged expanding-window validation and metrics implemented
   from scratch. Reports two defects in my own harness that manufactured large
   fictitious profits, and the zero-edge control that caught both.

3. Decodability Is Not Mechanism: Separating What a Network Encodes from What
   It Uses, 2026
   A channel decodable at 100 percent accuracy contributes 0.9998 bits to a
   correlational description of a layer and 0.0006 bits to an interventional
   one, a separation of roughly 1600x that is stable across five seeds.
   Includes a negative result: selecting the number of states blindly failed
   under two independent principled criteria.

SELECTED PROJECTS

1. HisokaOS
   A 32-bit x86 operating system written from scratch in C and assembly, with
   no Linux or BSD underneath. Boots on its own kernel to an interactive shell:
   paging, GDT/IDT, remapped PIC, PIT, PS/2 input, VGA console, ATA disk,
   RTL8139 networking, a RAM filesystem, preemptive task switching, and 27
   bundled applications. Roughly 10,000 lines across 115 source files.

2. The Lazy Driver
   An autonomous-driving perception and path-prediction stack over ordinary
   dashcam footage. Drivable-surface and lane segmentation, tracked vehicles
   with metric distance estimates, sign reading, and a predicted driving line
   clamped to the segmented road boundary.

3. Bill and Oneiro
   Two local-first systems. Bill wraps a local model in durable memory,
   retrieval, planning, grounding and verification, alongside bill2, a
   from-scratch predictive concept graph written without ML frameworks and
   evaluated on held-out data across ten measured iterations. Oneiro teaches a
   NumPy pixel-art generator continuously from an independent teacher model, so
   the generator never trains on its own output. Both require Ollama installed
   locally; neither has a hosted fallback, by design.

PROFILES

X: https://x.com/jcooperkai
GitHub: https://github.com/jcooperkai-sys
arXiv: https://arxiv.org/search/?query=jcooperkai&searchtype=author

Last updated: July 29, 2026