Beyond Narrow Benchmarks
The Central Thesis
AGI is a Human-Level Architectural Challenge.
The term "AGI" is often ambiguous. We define it rigorously, leveraging the established Cattell-Horn-Carroll (CHC) Theory of human intelligence. General intelligence is not monolithic, but a composition of distinct, interdependent abilities.
1. The Ten Pillars of Cognition
The AGI Scorecard (0-100%)
K - General Knowledge
Commonsense, Science, History
RW - Reading & Writing
Decoding, Comprehension
M - Mathematical
Arithmetic, Algebra, Calculus
R - Reasoning
Fluid Intelligence, Deduction
WM - Working Memory
Active Attention (Short-Term)
MS - L-T Memory Storage
Continual Learning
MR - L-T Memory Retrieval
Fluency & Precision
V - Visual Processing
Perception & Reasoning
A - Auditory Processing
Speech & Sound
S - Speed
Reaction Time & Fluency
2. The Critical Deficits
Why Current AI is Not AGI
Current LLMs exhibit a "jagged" cognitive profile, confirming they are still Narrow AI.
Critical Flaw 1: Amnesia in L-T Storage
Problem: Models cannot stably acquire or store new experiential information. They lack plasticity.
Critical Flaw 2: Capability Contortions
Costly "workarounds" creating an illusion of intelligence.
Massive Context Windows
vs. Lack of L-T Memory (MS)
RAG (Retrieval Augmented Gen)
vs. Lack of Retrieval Precision (MR)
3. The "Jagged" AI Profile
Uneven Progress & Critical Bottlenecks
Application of this framework reveals a highly uneven cognitive profile in contemporary models, showing rapid progress but critical bottlenecks.
| Model | K/RW/M | R/WM/MR | V/A/S | Total AGI Score |
|---|---|---|---|---|
| GPT-4 (2023) | 18% | 6% | 3% | 27% |
| GPT-5 (2025) | 29% | 15% | 13% | 57% |