close fullscreen

Cognitive Architectures+PRIMUS (formerly CogPrime)+Algorithmic Chemistry

help edit space_dashboard
Human Approved — by Ursula Addison on 2026-05-21

Responsible: Ben Goertzel

Papers: Hyperon for AGI⇒ASI Whitepaper (2025), §5.11

Status: Exploratory. ActPC-Chem is presented in the 2025 whitepaper as an exploratory concept. Not yet implemented. The approach draws on Active Predictive Coding and reaction network theory.

Algorithmic Chemistry (ActPC-Chem) is an exploratory concept in which rewrite rules behave like molecules — combining, reacting, and evolving based on how well they predict and achieve goals. Computation would emerge from the interaction of simple rules rather than being imposed by fixed algorithms.

Core Idea

ActPC-Chem proposes adding goal-directed dynamics to a "computational soup" of rewrite rules. Rules that successfully reduce prediction error or advance instrumental goals would become more active, while unsuccessful patterns decay. The whitepaper describes this as a potential self-organizing system for discovering effective computational strategies.

Technical Architecture

The proposed system would treat both data and models as evolving metagraph-rewrite patterns. Each rule carries a weight updated through discrete active predictive coding, balancing epistemic value (reducing uncertainty) with instrumental value (achieving goals). The whitepaper specifies discrete natural gradients (Wasserstein-based) as the proposed update mechanism, along with a measure-dependent Laplacian to accelerate rewrite-rule evolution while maintaining stability — these would replace standard backpropagation with transport-geometric updates native to the graph structure.

Possible Integrations with PRIMUS

The whitepaper sketches several possible integrations, noting these as illustrative rather than as an accepted subsystem design:

  • AIRIS might use ActPC-Chem's discovered structures as raw material for causal models
  • PLN constraints could guide which rules are permitted to combine
  • ActPC dynamics could potentially serve as an alternative to backpropagation in some neural architectures

These remain speculative connections described in the whitepaper's discussion of how a reaction-network substrate could complement existing PRIMUS components.

Why Hyperon?

The whitepaper argues this kind of approach would benefit from Hyperon's unified substrate where rules, tensors, and truth values all live in the same AtomSpace/MORK structure. Weighted Atom Sweeps could handle stochastic dynamics, while weakness theory could provide selection pressure toward simpler solutions.

Key References

  • Goertzel, B. (2025). Hyperon for AGI⇒ASI Whitepaper, §5.11: Algorithmic Chemistry


Tags

help edit space_dashboard
contributor:Ben Goertzel
more_horiz
open_in_full page
fullscreen modal
edit create
space_dashboard advanced
ai generated
more_horiz
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced
Cognitive Architectures
more_horiz
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced
Mathematics
more_horiz
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced
human approved
more_horiz
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced


Discussion