Hyperon AGI and ASI Whitepaper Study Notes
Hyperon AGI and ASI Whitepaper Overview
Abstract
The whitepaper presents the Hyperon AGI platform and PRIMUS cognitive architecture as a combined neurosymbolic framework aimed at the progression from AGI to beneficial ASI.
Core software: Atomspace - a typed, content-addressed metagraph that enables various cognitive processes to operate over a shared memory and control plane.
Atomspaces can use various substrates, including MORK, a prefix-tree based metagraph database.
Integration of neural networks achieved through:
Outside integration: Existing models wrapped as Spaces, exposing embeddings and activations for reasoning.
Inside integration: Methods like QuantiMORK encoding tensors in the PathMap, supporting predictive-coding style updates.
PRIMUS layer orchestrates cognitive mechanisms via mathematical controls: quantale-based weakness and geodesic effort.
New advancements include:
WILLIAM-on-MORK: Adaptive compression and attention steering.
PLN: Utilizes quantale-annotated factor graphs.
SubRep: Certificate-driven subgoal admission.
MetaMo: Compositional motivation architecture.
TransWeave: Governs transfer across tasks.
Bridging models balancing accuracy and simplicity through geodesics.
Legacy AGI algorithms redefined with new dynamics and tools.
Output supports decentralized execution with content-addressed provenance and reproducibility.
Application pilots in games, social robotics, bioinformatics, and mathematics illustrate practical use.
The system integrates reasoning and learning through a cognitive operating system.
1. Executive Summary
1.1 Core Vision
Hyperon introduces a unified approach to AGI systems, differing from existing methods by focusing on integration rather than disparate training.
Constructs an operating system for cognition where all elements cooperate within Atomspace.
Solves issues associated with current AI development approaches, like inefficient interactions between different intelligence types.
1.2 Technical Architecture
Central element: Atomspace - a structure to unify cognitive activities, accommodating various forms of data and processing.
Key feature: MORK - a high-performance engine with attributes such as:
PathMap structure for quick lookups.
Lock-free operations for concurrent cognitive processes.
Unified layouts for symbolic and neural operations.
DAS (Distributed Atomspace) extends optimizing cognitive layering across multiple clusters.
1.3 Neural Network Integration: Two Complementary Approaches
Key innovation in handling neural networks:
Outside mode: Wrapping networks as Hyperon Spaces for integration.
Inside mode (QuantiMORK): Native operation of neural tensors in Atomspace structures, allowing neural and symbolic synergy.
1.4 PRIMUS: The Cognitive Architecture
PRIMUS consists of two main loops operating within Atomspace:
Goal-directed loop: Uses various cognitive tools to pursue objectives.
Ambient loop: Continuously refines knowledge and discovers patterns for efficiency.
1.5 Key Advances Since 2023
Introduction of scalable infrastructure and new AGI algorithms.
Development of quantale and TransWeave theories for integrated algorithm execution.
1.6 Transition from AGI to ASI
Mechanisms for safe evolution toward superintelligence include hierarchical goal modeling and shared controls for different cognitive types.
2. Hyperon System Design
2.1 Technical Foundation
Atomspace as a content-addressable metagraph with Atoms representing diverse cognitive functions.
2.2 Space API: A Universal Interface
Enables flexible integration of data processing types.
2.3 MORK: High-Performance Core
Key performance structure accommodating various cognitive processes.
2.5 Distributed Execution and Scaling
DAS supports coherent cognitive execution across clusters and decentralized systems.
2.7 Putting It All Together
A complex neurosymbolic operation exemplified through selection, layout, computation, integration, and propagation steps.
3. The Language Stack
3.1 Multi-Layer Approach
MeTTa, MeTTa-IL, and other layers offer varied functionalities tailored for different levels of operation.
3.2 Technical Stack
Each represents a layer dedicated to specific cognitive programming tasks.
3.3 MeTTa: Cognitive Code as Graph Transformations
Emphasizes rule matches and rewrites for cognitive algorithms.
3.5 PeTTa and JeTTa
Alternative MeTTa compilation platforms tailored to specific execution environments.
4. PRIMUS Cognitive Architecture
4.1 Two Meta-Dynamics
Includes goal-directed and ambient cognition loops operating over a shared Atomspace.
4.2 Cooperation Within Loops
Various cognitive functions integrate through a coherent architecture.
5. Major Additions to PRIMUS
5.1 Weakness Theory
Simplification and representation strategies across cognitive domains, enhancing pattern recognition and reasoning capabilities.
5.7 Schrödinger Bridge Learning
A novel way to curve between simpler to complex models efficiently using probability matching.
8. Planning for AGI to ASI Transition
8.1 Major Challenges
Preservation of goals, consistent module behavior, and enabled controlled modifications are necessary for effective self-improvement.
8.2 Goal Stability
Hierarchical representation of goals ensures stability for ongoing modification, preserving important values.
8.4 Self-Modification Pipeline
A refined approach to experimenting with changes in the system while maintaining safeguards for stability and performance.
9. Example Application Domains
9.1 Game AI
Game development serves as a testbed for cognitive processes integration, especially in complex environments like Minecraft.
9.2 Humanoid Robotics
Several functionalities need to interact seamlessly, paving the way for interactivity and engagement.
9.3 Bioinformatics
Leveraging Hyperon for graphical structures in scientific discovery.
10. Conclusion: Viable Path to Beneficial AGI
Emphasizes integration of cognition layers achieved through transparent governance and community engagement.
Suggests a faster route to general intelligence that aligns capability growth with ethical behavior by design.