理論THY-2026-0102v0.1
μI——最小智能語意執行單位（候選理論）
μI 是相對於任務的語意狀態轉換 z_t → z_{t+1}，須滿足七個條件：任務相關、狀態改變、因果貢獻（Q(Y | μ) > Q(Y | do(μ=0))）、在指定解析度下不可再拆、可組合、實現無關、物理可實現。它既不是 token、FLOP、neuron activation、layer，也不是「想法」；最小性相對於解析度、任務與觀察者。四個候選族（信念更新、關係建立、約束消解、資訊增益）統一為任務相關、因果有用、語意不冗餘的 Δz。gross 與 effective 計數給出語意效率 η_μ，跨層映射 μI → ρ_C → ρ_P → ρ_T 一路向下接到物理成本。
定義
definitions- μI: z_t → z_{t+1} with z = (beliefs, relations, constraints, goals, uncertainty, procedures).
- Seven conditions C_μ = (C_T, C_S, C_C, C_N, C_K, C_R, C_P).
- N_μ^gross, N_μ^eff, η_μ = N_μ^eff / N_μ^gross; semantic work vector N_μ = (N_B, N_R, N_C, N_I, N_G).
- μI^obs (observer-reconstructed) vs μI^int (true internal); measurement ladder L0 behavioral proxy → L1 structured trace → L2 causal internal probe → L3 physical-semantic alignment.
前提
assumptions- A task-relative semantic solution state can be abstracted at observer level without claiming the model literally stores those fields.
- Counterfactual removal of a transition is at least conceptually testable.
主張
claims- Token ≠ μI ≠ FLOP ≠ NeuronActivation ≠ Layer ≠ Thought.
- SemanticActivity ≠ EffectiveIntelligentWork; SameAnswer ≠ SameSemanticExecution.
- SemanticWork ≠ PhysicalWork (type safety: semantic work is never a Joule).
形式化
formalization- Causal contribution: Q(Y | μI) > Q(Y | do(μI = 0)) in expectation.
- Semantic yield Y_μ = Q / N_μ^eff; Q = C_phys · η_{P→μ} · η_{μ→Q} (conceptual, not a physical equation).
- Signed contribution w_i^Q = ΔQ_i; equivalence class [μI] = {ρ | same pre-state, post-state and downstream function}.
預測
predictions- If μI is a good intermediate layer, adding N_μ improves prediction of efficiency, failure paths, scaffold gain and cross-architecture comparison over physical cost → quality alone (F5).
證偽／失敗條件
falsification_conditions- If N_μ never adds explanatory or predictive power over direct physical-cost → quality models, μI should be revised or eliminated (Paper 10, Falsifiable Claim 5).
已知限制
known_limitations- Explicitly a candidate ontology: no claim that a unique neural or model-internal 'intelligence atom' exists; μI^obs ≈ μI^int is unproven.
- Experiment C (operational identification) has not been run.
證據
| 來源 | 關係 | 目標 | 狀態 | ID |
|---|---|---|---|---|
EXP-2026-0105 Experiment C——μI 的操作性辨識（已宣告） | tests | THY-2026-0102 μI——最小智能語意執行單位（候選理論） | ACTIVE | REL-2026-0495 |
關係
| 來源 | 關係 | 目標 | 狀態 | ID |
|---|---|---|---|---|
THY-2026-0102 μI——最小智能語意執行單位（候選理論） | extends | THY-2026-0101 回合分解與外部無迴圈智能 | ACTIVE | REL-2026-0368 |
RES-2026-0101 執行與物理線——一個答案在回合、語意工作、能量與計算時空上的代價 | develops | THY-2026-0102 μI——最小智能語意執行單位（候選理論） | ACTIVE | REL-2026-0359 |
THY-2026-0103 跨層證據三角化與測量等級 | extends | THY-2026-0102 μI——最小智能語意執行單位（候選理論） | ACTIVE | REL-2026-0369 |
PAP-2026-0102 智能到底算了一次什麼？：最小智能語意執行單位的候選理論 | formalizes | THY-2026-0102 μI——最小智能語意執行單位（候選理論） | ACTIVE | REL-2026-0396 |
EXP-2026-0105 Experiment C——μI 的操作性辨識（已宣告） | tests | THY-2026-0102 μI——最小智能語意執行單位（候選理論） | ACTIVE | REL-2026-0495 |
歷史與來源歷程
- Canonical URL
- https://evemisslab.com/ai/theory/THY-2026-0102/
- 快照
AI-SNAPSHOT-v0.1-fe85b9694a45- 來源歷程
source- EveMissLab research collection: Intelligence Physical Metrology (真本體論13)
extracted_by- Splice (Claude Code), reading the canonical UTF-8 sources and each package's own reports
extracted_at- 2026-09-11
generator- tools/extract_all.py
claim_boundary- status, evidence level and result type follow the source artifact's own stated claim boundary; nothing is upgraded beyond what the report supports