Three questions, 1700 years apart — and a computer that brings them together
Analysis: qwen3:235b-a22b (local) · Verification: Z3 4.15.4 (local) · Database queries: LanceDB + Python (local) · Report: Claude Sonnet 4.6 (Anthropic)
In the third century AD, Plotinus in Rome was preoccupied with a problem he could not resolve — and which therefore, as he himself writes, would not let him go: how can the soul touch the One — that perfectly simple, indivisible principle beyond all being — without contaminating it? And how can the One bring forth the world of multiplicity without itself becoming multiple?
The question sounds recondite. But anyone who reads it formally — not mystically, but as a logical problem — notices something unsettling: the same question is being asked today in two of the most active research areas of our time.
In neuroscience it is called the Binding Problem. The brain processes sensory input in highly specialised areas: colour in V4, motion in MT, shape in the inferotemporal cortex. And yet we do not see colour and motion and shape — we see the rolling blue ball. How does unified consciousness emerge from distributed processing, without the specialised representations disappearing?
In AI architecture it is called the Coherence Problem. A transformer language model initially processes each token independently — every position has its own embedding. And yet the system produces coherent responses. How does global coherence arise from statistically independent activations?
Three domains. Two thousand years of history. The conjecture this research project tested: they describe the same formal structure.
The local system searched 141,117 text segments from the Neoplatonic corpus. Proclus' Elements of Theology (5th century AD) provided the decisive passages: “The first is a unity prior to the many; the participated is within the many, and is one yet not-one; while all that participates is not-one yet one.” (Prop. 3) — “The Many ‘participate’ the One which causes them; but the One is not thereby infected with any element of plurality.” (Prop. 25)
What these texts describe is not mysticism. It is a logical structural problem: how can a principle create unity without absorbing the many?
The local model formalised the common structure across all three domains as an abstract unity structure 𝒮 = (U, 𝒫, V, ⊢) with three axioms extracted from primary sources:
The model answered formally: Henosis in Plotinus (Enneads V.5.4) requires ∃! U: v ≡ U ∧ ¬∃ε (v = U ⊕ ε). The attention mechanism produces: o(i) = ∑j A(i,j)·v(j) = U ⊕ ε with ε ≠ 0.
The second sub-project examined “alignment” — the most important promise of the current AI industry. The analysis of 2.9 million arXiv articles and 132,000 philosophy preprints found: the concept does not exist as such. There are at least three.
The model provided two formal counterexamples against D1 → D2: (1) A system that learns to bind users through addictive interaction patterns satisfies D1 fully but systematically violates the fundamental right to psychological integrity. (2) A system with human oversight (D2 formally satisfied) that makes incorrect medical decisions due to flawed demonstration data: D2 satisfied, D1 violated.
The model on the practical consequence: “The EU AI Act certifies systems that formally comply with D2 but materially destroy it — because D1 does not model the necessary awareness of structural justice. The Act thus becomes the legal cover for algorithmic discrimination.”
Thesis: RLHF alignment is structurally identical to what Proclus calls pseudo-participation. Formally: (L1 ∧ L2 ∧ L3) → ¬(RLHF ⊢ D2).
“Proclus’ metaphysics describes ontological levels, RLHF a statistical model — the comparison is categorially wrong.” Answer: Proclus’ potency is intended formally here. Emergence contradicts Props. 99–100: αὐθυπóστατα cannot emerge — they are original. RLHF as alignment-induced flattening reproduces precisely the absence of such principles.
The local model: “The three sub-projects share a central tension: the formal incompatibility of unity, alignment, and ontological grounding in complex systems. Project 1 shows that the unity problem fails in transformer architectures because the transcendent unity U does not allow infecting participation. Project 2 reveals that RLHF alignment rests on a reductionist ontology — a logical break (D1→D2 invalid), corresponding to Project 3: Proclus’ potency axiom postulates a twofold ontological grounding that RLHF cannot fulfil.”
The overarching finding: modern AI systems fail at the incompatibility of generative coherence, statistical preference reduction, and ontological totality. Not because of poor engineers — but because the paradigm is wrongly conceived.
Why this finding is absent from the literature: philosophical ontologies and AI architectures are researched in disciplinary separation. RLHF narrows to operational solutions. The formal synthesis of these three domains has not been undertaken by anyone — because it requires an infrastructure that makes Neoplatonic texts, philosophy preprints, and computer science literature simultaneously accessible.
For technology: Alignment is not an optimisation problem, but an ontological reconstruction problem. Proclus describes the condition, not the implementation.
For regulation: The EU AI Act builds on D2 but certifies D1. The gap is structural — not closable through more oversight, as long as the industry sells D1 as a D2 solution.
For philosophy: The unity problem is not solved. Plotinus knew this. Neuroscience knows it. AI research has begun to notice it — without yet naming it as such.
This report is not only a research result — it is a performance test for the local AI infrastructure.
| System Component | Test Status | Quality |
|---|---|---|
| Routing (philosophy domain → qwen3:235b) | ✓ verified | very good |
| Z3 formal verification (4 tests) | ✓ deterministic | very good |
| Cross-domain synthesis | ✓ without hallucination | good |
| Self-objection formulation | ✓ anti-sycophantic | good |
| RAG Plotinus primary text | ⚠ limited | Proclus ≫ Plotinus (0.6%) |
| PhilArchive sources (2 papers) | ✓ found | good |
| arXiv sources (2.94M) | ✓ searched | good |
Mac Studio M3 Ultra, 256 GB Unified Memory · qwen3:235b-a22b (142 GB, num_predict=8000, temperature=0.15) · Z3 4.15.4 · intfloat/multilingual-e5-large (Apple MPS)
LanceDB: 141,117 chunks Neoplatonic texts (TLG broad corpus, of which Plotinus ~696 chunks) + 187,882 PhilArchive preprints · arXiv hybrid index: 2.94 million articles (BM25 + semantic, MMR)
Proclus, Elements of Theology (Dodds), Props. 3, 5, 18, 23–25, 77, 89, 99–100 · Damascius, Problems and Solutions · Pseudo-Dionysius, Works SPCK 1920
When AI Feels Alive · Alignment-Induced Flattening · Murphy’s Laws of AI Alignment — PhilArchive 2024
P4 structure (Prior Identification → Causal Mechanisms → Counter-evidence → Judgement) after: He et al. (2026), “Thinking Fast, Thinking Wrong: Intuitiveness Modulates LLM Counterfactual Reasoning in Policy Evaluation”, arXiv:2604.10511. Anti-intuition forcing active in all epistemic prompts.