Post
121
CPU-Oriented Selective Inference for Persistent AI
Independent researcher and developer based in Japan, exploring efficient AI systems for modest computing environments.
This project studies whether structured state and selective execution can reduce unnecessary computation in long-video and multimodal tasks while preserving reliable behavior.
Current prototype
CPU-only Synthetic World prototype:
* 64×64 tile-native video
* Five development seeds, 360 schedule-offset cases
* W1/W2 quality gate: 99.17%
* W3 safe non-answer: 100%
* Deterministic replay: 100%
* Cold end-to-end wall-time ratio: 0.40090× versus matched full-pass
These are limited engineering measurements—not proof of universal 10× acceleration, trained-model superiority, real-world high-resolution performance, or consciousness.
Goals
Process task-relevant evidence, maintain persistent experience, select small specialized components, separate perception/memory/reasoning/verification, abstain when evidence is insufficient, and record costs and failures transparently.
Long-term goal: practical AI that can observe, remember, reason and adapt on affordable CPU and edge hardware.
The internal method is intentionally not described publicly. Technical evaluation is welcome through black-box demos, benchmark traces, profiling data and preserved failure cases.
Next: longer/higher-resolution video, occlusion and camera-shift tests, trained models, fair baselines, independent implementations, and CPU/memory/energy measurements.
Experimental project. Reproducibility and honest accounting come first.
Independent researcher and developer based in Japan, exploring efficient AI systems for modest computing environments.
This project studies whether structured state and selective execution can reduce unnecessary computation in long-video and multimodal tasks while preserving reliable behavior.
Current prototype
CPU-only Synthetic World prototype:
* 64×64 tile-native video
* Five development seeds, 360 schedule-offset cases
* W1/W2 quality gate: 99.17%
* W3 safe non-answer: 100%
* Deterministic replay: 100%
* Cold end-to-end wall-time ratio: 0.40090× versus matched full-pass
These are limited engineering measurements—not proof of universal 10× acceleration, trained-model superiority, real-world high-resolution performance, or consciousness.
Goals
Process task-relevant evidence, maintain persistent experience, select small specialized components, separate perception/memory/reasoning/verification, abstain when evidence is insufficient, and record costs and failures transparently.
Long-term goal: practical AI that can observe, remember, reason and adapt on affordable CPU and edge hardware.
The internal method is intentionally not described publicly. Technical evaluation is welcome through black-box demos, benchmark traces, profiling data and preserved failure cases.
Next: longer/higher-resolution video, occlusion and camera-shift tests, trained models, fair baselines, independent implementations, and CPU/memory/energy measurements.
Experimental project. Reproducibility and honest accounting come first.