Recommendation Reasoning and Its Path Toward Future AI

The author begins by questioning why recommendation systems require reasoning and how this reasoning relates to the development of strong, general AI, emphasizing that future AI should possess inference, analogy, abstraction, imagination, and consciousness.

To build confidence toward strong AI, the article suggests studying recent AI conference papers and the work of leading research labs, urging practitioners to look beyond immediate tasks and seek deeper insights that may lead to the "path" of ultimate AI.

Recommendation systems are described as personalized shelves that rank items for users, but the author argues that merely presenting ranked items is insufficient; a true recommendation system should also provide explanations, akin to a persuasive salesperson.

Explanation is defined as a two‑sided process involving objective facts and subjective interpretation, requiring both deterministic (discrete, structured) and probabilistic (uncertain, ambiguous) elements. The article contrasts symbolic logic and probabilistic graphical models as two AI paradigms for building explanatory languages.

Graph networks are highlighted as a bridge between deep learning and knowledge reasoning, offering relational inductive biases that enable interpretable, compositional reasoning over knowledge graphs.

The piece illustrates explanation generation with a concrete example: linking a user’s interest in "Iron Man" to a recommendation of "Ant‑Man" via a knowledge‑graph path through concepts like "superhero" and "Marvel".

It further discusses the limitations of treating explanation as a simple shortest‑path problem, emphasizing that the underlying force field (causal dynamics) must be inferred from data, not just the static graph structure.

The relationship between causality, reasoning, and explanation is examined, noting that causality captures objective world dynamics while reasoning provides the subjective tools to model and explain those dynamics.

Intervention and counterfactual reasoning are introduced as essential components of causal learning, illustrated with examples such as A/B testing and thought experiments about roosters and sunrise.

Finally, the author connects these ideas to human cognition, proposing a pyramid of subconscious, conscious, and attention layers, and relates them to System 1 (deep learning) and System 2 (symbolic reasoning), concluding with recommendations for further reading on explainable AI and consciousness.