Humans are able to learn from few examples without considering the probability distribution of training data. Although Judea Pearl (2011 Turing Award) said that “Machines’ lack of understanding of causal relations is perhaps the biggest roadblock to giving them human-level intelligence”, we know that LLM have no concept of truth and are therefore intrinsically not trustworthy. I present interaction spaces theory (IST), a unified mathematical theory of complex systems which is able to embed cellular automata, agent based models, master equation based models, continuous or discrete dynamical systems, networked dynamical models, artificial neural networks and genetic algorithms in a single notion. In IST, we can mathematically define complex adaptive system: adaptation means minimizing suitable costs and, at the same time, maximally diversifying information exchanged by the interactions occurring in the system. I present cause-effect preserving functors to represent the link between agent’s interactions and environment’s interactions. This leads us to consider reinforcement learning (RL) by exploration to learn how entities in the environment interact, and RL by simulation to learn how to change an entity’s state using causal composition of interactions. The approach can be described as: compositional causal world models for learning from a given core knowledge, and corresponds to the following definition of intelligence: intelligence is the process of discovering causal relations (RL by exploration) in a given environment and to use them to simulate (RL by simulation) solutions of problems. This immediately opens realizable applications to ARC-AGI benchmarks, labeling of pictures considering few examples, applications in software engineering, teaching strategies (e.g. driving) from few examples, etc.