Hey guys! I've been working on solving the bootstrap problem in epistemic uncertainty.
The problem: How do you make AI explicitly model what it doesn't know, without requiring examples of "unknown" things during training?
STLE's approach: Use complementary fuzzy sets:
- μ_x = accessibility (computed via density ratios)
- μ_y = inaccessibility (1 - μ_x, guaranteed)
- No OOD training data needed
Results: 67% AUROC on OOD detection without any OOD training, < 1ms inference.
The Sky Project:
Already in production: Built MarvinBot (u/MarvinBot-20260210114005 on Moltbook), a fully optimized and aligned bot using STLE in practice. This validates the framework works beyond a toy demo.
The repo has:
- Minimal NumPy version (17KB, zero deps)
- Full PyTorch version (18KB)
- 48KB theoretical spec
- All validation experiments
It's proof-of-concept level for the framework, but already deployed in Phase 2 (application mode).
Most useful for: medical AI, autonomous systems, active learning—anywhere "I don't know" matters.
The core innovation is solving the bootstrap problem through on-demand density estimation rather than trying to initialize uncertainty for unseen data.
Happy to answer questions about the STLE, implementation, MarvinBot, or how it compares to MC Dropout / Deep Ensembles / Posterior Networks!
