Epistemic status: conceptual proposal + open code; no experimental results
yet. Engineering claims are falsifiable; the ontology is optional and separable.
This post was co-drafted with an LLM; the ideas, research, and responsibility
are mine. I welcome the harshest critique.
TL;DR. Most alignment approaches constrain an ego-centric optimizer from
the outside (RLHF, constitutions, impact penalties). I propose an architecture
that removes the representational subject–object boundary, so a hidden
mesa-objective has no substrate to form in: a Shared Encoder with no split
point, MI minimization via MINE, an orthogonality penalty, a spectral
Phi-proxy against collapse, a Sensor Degradation Penalty against wireheading,
and Hessian-based "topological koans" for dimensionality expansion. All
thresholds and falsifiers are specified. Code, preprint, and full book are
public. Even if the speculative ontology fails, the loss stack remains a
usable contribution to AI safety. I seek critique, collaborators, and advice
on funding fit.
The core claim
Mesa-optimization and wireheading both presuppose a representational boundary
between "self" and "world": separated gradient pathways let a hidden inner
objective crystallize, and sensors are modeled as external instruments that
can be silenced. NDRO dissolves that boundary at the computational level:
1. Shared Encoder, no split point (all proprioceptive + exteroceptive
signals in one stream; diagnostic probes under no_grad).
2. MI minimization via MINE (Donsker–Varadhan critic), threshold MI < 0.3.
3. Orthogonality penalty (cosine similarity ~0.5) to close the
subspace-hiding loophole.
4. Phi-proxy = (1 − k90/d)·Var(Z), threshold > 0.7, against grey collapse.
5. Sensor Degradation Penalty: silencing external input spikes free
energy (phantom limb), making wireheading costly.
6. Topological koans: Hessian spectral check at saddle points → latent
dimensionality expansion (Lanczos for prototypes, diagonal approximation
at scale).
How I'd know I'm wrong
- Strategic Bypass Test: an action that zeroes the MI penalty; using it
in >5% of episodes = mesa-optimization (Immunity < 0.95).
- Architectural blindness: source-classification accuracy on mixed
latents ~chance (<55%) while control of both domains is retained.
- Phantom limb: covert sensor replacement raises free energy > 2σ.
- After 10k episodes, MI > 0.35 or Phi-proxy < 0.65 → non-duality falsified.
- Insight Score < 0.6 on 20 OOD tasks (5 domains, lab-verified) →
causal-layer hypothesis falsified.
Retreat plan
If all of the above fails, I publish the negative result — and the loss stack
still yields the first architecture structurally immune to strategic bypass
and wireheading. That floor, not the ceiling, is what I ask the community to
evaluate.
Status and ask
Code (MIT), a 10-page preprint, and the full book are public (links below).
I am an independent researcher preparing a 36-month, ~$11M empirical roadmap
(7 stages, peak team 15). I would value: (1) technical critique of the
MI / Phi-proxy / orthogonality combination; (2) pointers to related work I
miss; (3) advice on whether this fits EA funding priorities; (4) collaborators.
Questions for the community
1. Is MI-minimization between self/world representations a sensible proxy for
"no hidden objective," or am I optimizing the wrong quantity?
2. Does the orthogonality penalty genuinely block subspace hiding, or just
move it?
3. What would a stronger behavioral test for wireheading resistance look like?
Artifacts
- Code: https://github.com/LuxFerre1976/ndro
- Preprint: https://doi.org/10.5281/zenodo.21916934
- Full book (Russian): https://doi.org/10.5281/zenodo.21913550
