Building Better Minds: Artificial General Intelligence via the Cog' Architecture
Ben Goertzel with Cassio Pennachin, Nil Geisweiller & the OpenCog Team. August 2012 (preliminary draft). 48 chapters across 2 parts + appendices, stored in 52 chunks.
This is the historical foundation for the Hyperon/PRIMUS architecture. It describes CogPrime (the predecessor to PRIMUS) and the OpenCog system that preceded Hyperon. Most wiki cards covering PLN, MOSES, ECAN, pattern mining, attention allocation, and the cognitive synergy principle trace their conceptual lineage to this book.
Reading strategy: This book is not a direct description of the current Hyperon system — it describes the CogPrime/OpenCog era. Use it for: (1) understanding why current components exist, (2) the philosophical framework (patternism) underpinning the project, (3) detailed technical specifications of PLN, MOSES, ECAN that remain largely valid, and (4) the ethics and self-modification framework that evolved into the 2025 whitepaper's §8. For the current system state, use the Whitepaper 2025 Map instead.
| Theme | Chapters | Chunks | Cards Grounded |
|---|---|---|---|
| Philosophy & Intelligence Theory Patternism, general intelligence definitions, mind-world correspondence, cognitive synergy principle | 1–3, 8–10 | 1, 3–5, 10–11 | PRIMUS Full (cognitive synergy lineage) |
| Cognitive Architectures Survey SOAR, ACT-R, NARS, Cyc, DeSTIN, LIDA, Psi, MicroPsi, Global Workspace | 4–5 | 6–8 | OpenCog Legacy Full |
| CogPrime Design Overview Architecture, memory types, goal-oriented dynamics, analysis/synthesis | 6–7 | 8–9 | OpenCog Legacy Full, PRIMUS Full |
| Cognitive Development & Ethics Piaget stages, Friendly AI, Coherent Extrapolated Volition, AGI societies, eight ways to bias toward friendliness | 11–12 | 12–15 | Self-Modification and Safety |
| Knowledge Representation Weighted labeled hypergraphs, glocal memory, AtomSpace architecture, atom types, procedural knowledge (Combo language) | 13–15, 19–21 | 16–17, 20–23 | AtomSpace Full, Knowledge Substrates |
| AGI Preschool & Roadmap Virtual preschool design, curriculum, assessment, self-modification via supercompilation and theorem-proving | 16–18 | 18–19 | Self-Modification and Safety Full (Ch 18 lineage), Game Worlds |
| Cognitive Cycle Emotion/motivation (Psi model), ECAN (STI/LTI, Hebbian links, information geometry), economic goal selection | 22–25 | 24–26 | ECAN Full, MetaMo Full (Psi→MetaMo lineage) |
| Perception & Action Perceptual-motor hierarchies, DeSTIN deep learning, symbolic/subsymbolic bridge | 26–29 | 27–29 | Robotics |
| Procedure Learning (MOSES) IRC learning, hillclimbing, MOSES/EDA, hierarchical program learning, fitness estimation | 30–33 | 30–32 | MOSES Full, Reasoning and Search |
| Declarative Learning (PLN) PLN truth values and rules, spatiotemporal inference, inference control, pattern mining (Fishgram), concept blending | 34–38 | 33–35 | PLN Full, Semantic Parsing Full |
| Integrative Learning Dimensional embedding, episodic memory, map formation, procedure encapsulation | 39–42 | 37, 44 | PRIMUS Full (integration model) |
| Natural Language Psynese (inter-mind communication), Link Grammar, RelEx, NL comprehension/generation, embodied language | 43–47 | 38–42 | Semantic Parsing Full, Knowledge Substrates (Link Grammar) |
| Summary & Synthesis Argument for Cog' approach, synergies between components | 48 | 2, 43 | PRIMUS Full |
| Appendices A: Glossary. B: Category theory (functors for memory conversion). C: Hyperset model of consciousness. D: GOLEM self-modification. F: Neural inheritance & term logic. H: Propositions about Cog'/PLN | App A–H | 45–51 | Self-Modification and Safety Full (App D: GOLEM) |
| References | — | 52 | — |
The most valuable use of BBM for wiki enrichment is tracing how CogPrime components evolved into current Hyperon/PRIMUS ones:
| BBM / CogPrime | BBM Chapters | Current Hyperon/PRIMUS |
|---|---|---|
| AtomSpace (weighted labeled hypergraph) | 13–14, 19–20 | AtomSpace (retained), MORK (new substrate) |
| PLN (truth values, inference rules) | 34–36 | PLN on quantale-annotated factor graphs |
| MOSES (EDA + evolutionary programming) | 30–33 | MOSES/GEO-EVO with geodesic search |
| ECAN (STI/LTI, Hebbian links) | 23 | ECAN with fluid-dynamic extension |
| Pattern mining (Fishgram) | 37 | Stream-based pattern mining with I-surprisingness |
| Concept blending | 38 | Concept blending (retained, now TransWeave-compatible) |
| Psi motivational model | 22 | MetaMo (pseudo-bimonad replacement) |
| Cognitive synergy | 8 | PRIMUS cooperation model / TransWeave formalization |
| Self-modification | 18, App D | Five-stage pipeline with supermartingale potentials |
| Link Grammar + RelEx | 44–45 | Semantic Parsing (retained + Symbolic Heads) |
| DeSTIN perception | 26–29 | QuantiMORK (inside integration) |
| OpenCog framework | 19 | Hyperon (MeTTa + Space API + DAS) |
BBM content that could deepen existing wiki cards in a future enrichment pass:
| BBM Theme | Target Card | Priority | What to Add |
|---|---|---|---|
| Ch 8: Cognitive Synergy | PRIMUS Full | Medium | The original cognitive synergy argument — specific synergy pairs (PLN↔MOSES, ECAN↔PLN) that motivated the architecture |
| Ch 22: Psi Model | MetaMo Full | Medium | Detailed OpenPsi→MetaMo transition story with specific limitations of the Psi model that MetaMo addresses |
| Ch 12: Ethics | Self-Modification and Safety Full | Low | Historical ethics framework (CEV/CAV, eight friendliness biases) — already cited but not detailed |
| Ch 38: Concept Blending | No dedicated card | Low | Blending theory criteria (topology, web, unpacking, metonymic tightening) — matches whitepaper §6.5 gap |
| App B: Category Theory | AtomSpace Full | Low | Functor-based model of memory type conversions — historical precursor to the Space API abstraction |