AGI DNA: A Generative Theory of Open-Ended General Intelligence

This paper proposes a generative research program for artificial general intelligence. Instead of treating AGI primarily as a final architecture assembled from memory, planning, tools, world models, and self-reflection, it asks whether a smaller set of generative constraints can repeatedly produce increasingly capable cognitive systems. The proposal connects artificial life, open-ended evolution, major evolutionary transitions, memory consolidation, world models, empowerment, cultural evolution, and recent self-improving agents. Its central abstraction is a cross-scale generative-selective-compression loop: systems generate alternatives, expose them to environmental or internal tests, retain useful structure, compress recurring regularities, and use the retained structure to generate further variation. Death and forgetting are interpreted as forms of structural turnover; sleep-like states as possible solutions to competition between online activity and internal maintenance; stochastic imagination as cognitive variation; culture as a second inheritance channel; and science as an error-correcting process for world models. The paper introduces Future Adaptive Potential (FAP), a diagnostic vector covering viability, controllable futures, expansion of actions and representations, transfer across unfamiliar worlds, and evolvability. It further proposes Bidirectional Genesis: a bottom-up program that starts from minimal digital physics and tests for the emergence of replication, evolution, cognition, and culture, combined with a top-down program that removes implementation-specific features from known human intelligence to identify functional invariants. Temporally firewalled historical prediction and synthetic universes provide independent validation. The paper states falsifiable predictions and identifies hidden teleology, metric capture, and anthropomorphic interpretations of consciousness as key methodological risks. AGI DNA, in this framework, is not a list of modules but a candidate set of generative invariants that enable a system to become progressively better at producing capabilities it did not previously possess.This record contains the English canonical version and the official Chinese translation, provided in PDF and editable DOCX formats.