Culture Is All You Need

The pursuit of Artificial General Intelligence (AGI) aims to engineer systems capable of defining and achieving goals across diverse contexts and scales. This paper introduces a novel engineering framework inspired by the Flynn Effect (1), interpreted as systemic phenomena (2), and by emergent intelligent properties of LLMs at scale (3, 4). We propose that the cultural accumulation of knowledge and functionality, compressed in language, code, art, and other cultural artifacts, is tappable by humans, contributing to growth in measured human intelligence (HGI) across generations. Recent LLMs owe much of their success to attention architectures, which optimize their ability to identify useful contexts and abstract them using statistical weights learned from language and code (5). In a vast system such as the universe, finite General Intelligence (GI) must use various strategies to conserve compute, in order to succeed. Humans and LLMs alike leverage knowledge amassed in language through heuristics, experience, empirical testing, logical extrapolation, mathematical modeling, and other strategies to parsimoniously model their environment and act within it. We posit that an engineerable, evolving cultural framework can be instrumental for developing AGI. We formalize culture from first principles, using the ‘Engineering Stance’(6) to outline a formal architecture grounded in concepts such as context, agent, and strategy. Defining the distribution of contexts as the combinatorial set of all possible system subsets, highlights the role and usefulness of their partial descriptions (fuzzy predicates). Contexts are key for GI to navigate a system’s inherent complexity and uncertainty. Our framework proposes methods to tokenize, compress, and abstract contextual knowledge and functionality for reuse. Morality, doubt, diversity, conservatism, exploitation, exploration, breaking down tasks and chunking them, are first class citizen foundations of our architecture, modulating all, learning, plans and actions. This is a paradigm shift from current alignment efforts that rely on data curation (12), static core models, post training deletions (13) and fine tuning (14) to achieve alignment. Instead, we argue for embedding these principles directly into the architecture to achieve antifragility and scalability. Compassion and truth should inform morality; one cannot hope to mitigate or defend against malicious intent without first understanding it. As Santayana noted, “Those who cannot remember the past are condemned to repeat it” (15). We propose that a diverse culture, with mechanisms for both conservation and innovation, can give rise to moral, responsible, antifragile AGI.