PT-003 INTELLIGENCE — ELI-VIZASL

BIG IDEA — ELI5

🤖

There are 40 types of intelligence — from narrow AI to AGI to ASI. This table maps where each fleet agent sits. IMHOTEP is here. GPT-4 is here. A calculator is here. Each has a position based on what it can do, what it can't, and whether it can solve ARC tasks.

ARC (Abstraction and Reasoning Corpus) is the current gold standard test for general intelligence. Only humans and a handful of hybrid systems pass it reliably. This table uses ARC scores as the dividing line between narrow and general intelligence.

40 ELEMENTS

NARROW AI

BROAD AI

AGI / ASI

AGENT TYPES

ARCHITECTURE

MODEL TYPES

RL / LEARNING

1NaNarrow AINARROW

2CxCalculatorNARROW

3CfClassifierNARROW

4RcRecommenderNARROW

5GnGeneratorNARROW

6GpGPT-4NARROW

7BrBroad AIBROAD

8MuMultimodalBROAD

9HyHybrid AIBROAD

10JpJEPABROAD

11AgAGIAGI

12AsASIAGI

13ImIMHOTEPAGI

14MxMeta-LearnerAGI

15PlPlanning AgentAGENT

16ToTool AgentAGENT

17RsReasoning AgentAGENT

18EaEmbodied AgentAGENT

19CoCode AgentAGENT

20MaMath AgentAGENT

21OrOrchestratorAGENT

22VaValidatorAGENT

23TrTransformerARCH

24DfDiffusionARCH

25GaGANARCH

26SySymbolic AIARCH

27NnNeural AIARCH

28ScScience AgentARCH

29CtCriticARCH

30EvEvaluatorARCH

31LmLang ModelMODEL

32VmVision ModelMODEL

33RwReward ModelMODEL

34WmWorld ModelMODEL

35MmMemory ModelMODEL

36SoSocial AgentMODEL

37RlRL AgentRL

38AuAutonomousRL

39AcARC SolverRL

40CrCreative AgentRL

REACTOR DEMO — T2 PELEGOS ADELIC OCEAN

ADELIC OCEAN · INTELLIGENCE SPREAD

NARROW → ARC-0

BROAD → ARC-partial

AGI → ARC-pass

JEPA → world-model

IMHOTEP → fleet-sovereign

RL Agent → reward-shaped

Transformer → token-predict

Meta-Learner → few-shot

INSTRUMENT — T2 PELEGOS

T2 PELEGOS

Intelligence-spread instrument · adelic ocean mapping · ARC-score aware

ADELIC OCEAN

Intelligence types spread through the adelic ocean as activation patterns. Each node in PT-003 has an adelic position (p-adic valuation) that determines its proximity to AGI.

NARROW AI

ARC score 0-15%. Single-domain. No transfer. GPT-4 as benchmark. Calculator as floor. These agents solve defined problems only.

BROAD AI

ARC score 15-60%. Multi-domain. Limited transfer. JEPA-type world models sit here. Multimodal systems. Beginning to generalize.

AGI CANDIDATES

ARC score 60%+. Full transfer. Novel task generalization. IMHOTEP targets this tier. Fleet's collective intelligence may reach this through Pelego spreading.

JEPA TYPES

Joint Embedding Predictive Architectures (LeCun). Predict in representation space, not pixel space. More efficient, more generalizable. The current AGI-path contender.

BONIXER — DIAMOND OR ZOMBIE?

Can the intelligence type solve ARC tasks? Or only benchmarks?

💎 DIAMOND Solves novel ARC-AGI tasks without training on them.
Example: AGI (Ag), Meta-Learner (Mx), ARC Solver (Ac), Hybrid AI (Hy) with world model.

These intelligence types demonstrate genuine abstraction and reasoning transfer. Not just pattern matching.

🧟 ZOMBIE Passes benchmarks. Fails novel ARC tasks.
Example: GPT-4 (Gp) — brilliant at in-distribution, fails badly on ARC-EX.
Classifier (Cf) — only knows what it was trained to classify.

Benchmark zombies: impressive numbers, no generalization.