Skip to content

METHOD.md

Latest commit

 

History

History
53 lines (41 loc) · 2.36 KB

METHOD.md

File metadata and controls

53 lines (41 loc) · 2.36 KB
Raw

Method

Problem Target

This topic tests whether selected AI scaling and architecture-efficiency quantities can be represented in a UET-style information framework. The current hardening pass targets measurable scaling and sparsity diagnostics first.

Evidence Lanes

Lane Files Current role
Scaling-law benchmark scaling_laws.json, GPT3_Scaling_Laws.csv, Research_AI_Scaling_Audit.py primary verifier lane
Sparse architecture diagnostic deepseek_moe_data.json, UET_AI_Core.py, Research_AI_Scaling_Audit.py primary verifier lane
Entropy-learning engine UET_AI_Core.py implemented but not benchmark-validated
AI detective/cross-topic reasoning Research_AI_Detective_V2.py excluded from primary claim; depends on 0.1 galaxy data
Alignment/ethics simulation Research_Alignment_Equilibrium.py future exploratory lane
Consciousness/developmental AI Research_Consciousness.py, Code/05_Developmental_AI/ future exploratory lane

Variables

Symbol / field Meaning Unit
L test-loss proxy dimensionless
N parameter count count
D training tokens count
C compute PF-days or FLOP-derived proxy
alpha_N, alpha_D, alpha_C scaling exponents dimensionless
kappa_macro current UET macro proxy used for comparison dimensionless
active_fraction active parameters divided by total parameters dimensionless

Procedure

  1. Load topic-local scaling-law constants and model metadata.
  2. Fit a simple log-log exponent from GPT3_Scaling_Laws.csv.
  3. Compare the fitted exponent with stored alpha_N.
  4. Compare MoE active fractions with dense active fractions.
  5. Check whether the current kappa_macro=0.1 proxy is close enough to alpha_N to support a constant-identification claim.
  6. Write a machine-readable artifact with hashes, metrics, thresholds, blockers, and limitations.

Domain of Validity

The current method applies only to the topic-local scaling/sparsity benchmark package. It is not a proof of AI alignment, ethics, consciousness, or a general physical law of intelligence.

Dependency Policy

Research_AI_Detective_V2.py depends on topic 0.1 galaxy data and must not be used as the primary evidence for topic 0.24 until it is explicitly modeled as a cross-topic dependency artifact.