Intelligence: Whence and Whither

By Akash Deshpande

Understanding and Changing the World: From Information to Knowledge and Intelligence. By Joseph Sifakis. Springer Nature, Singapore, May 2022. 171 pages, $49.99.

Both scientific and general interest in artificial intelligence (AI) have recently exploded around the world. While transformer neural networks and generative AI—exemplified by tools such as ChatGPT and Google Bard—are promising new technologies, it is important that we understand how they fit into the overall scheme of the physical and informational worlds. In Understanding and Changing the World: From Information to Knowledge and Intelligence, Joseph Sifakis (a 2007 Turing Award recipient) amasses a lifetime of knowledge as a deep thinker and effective practitioner of computer science—indeed, collecting a civilization’s wisdom—to provide a valuable framework that is rooted in the philosophy of science and society.

This densely rich yet concise book is hard to summarize. I recommend that interested persons read it thoroughly for its thematic development, insightful observations, and relevant sidelights on diverse topics. Here, I’ll attempt to describe a few key ideas to whet the reader’s appetite.

Sifakis follows the dualist tradition of separating physical reality from its models, the latter of which fall into the category of information. Information itself is purely logical in this context. It is a semantic relationship between a symbolic language and a set of concepts (not a statistical measure as in quantum mechanics or communication theory). Logical information requires a symbolic alphabet with rules of syntax that govern the combination of symbols. These combinations codify concepts and provide representations; rules of semantics then help us decode the representations and glean the referenced concepts. Together, these elements form a language. Physical reality exhibits a level of continuity and robustness that enables us to generalize observations into laws, thus making reality comprehensible. However, information and its transformation through computation are discrete, brittle, and therefore difficult to understand.

Knowledge is useful information that helps us understand a situation and act in order to achieve a goal. Understanding and Changing the World focuses on the how of knowledge and not on the why and what, which—as Sifakis explains—rest on rhetoric, mythology, and belief. He considers knowledge to be the how of things — e.g., how does a clam reproduce? How do you build a house? How can we avoid inflation? The answers to these questions are relationships between observations and concepts, and are the same irrespective of one’s belief system.

Sifakis’ text provides a hierarchy of knowledge: (i) Empirical (data-based) knowledge from observations of events and conditions in the real world, which may be processed by algorithmic or machine learning computation; (ii) scientific and technical (model-based) knowledge from reasoning and application; (iii) non-empirical knowledge from idealized concepts such as mathematics; and (iv) meta knowledge, or knowledge about knowledge that may be called wisdom. Empirical knowledge helps us predict, scientific knowledge helps us explain, and non-empirical knowledge helps us prove the validity of knowledge itself. Sifakis expands considerably on the development and application of knowledge at each level, including the roles of induction, generalization, modeling, abstraction, formalization, hierarchy, modularity, and emergence. He also discusses pitfalls like scientism, mystification of expertise, and research via vision (versus research via challenge), and examines the need for innovation ecosystems with research, industry, and startups.

Sifakis characterizes human intelligence as having two ways of thinking: slow and conscious or fast and automatic. Humans exhibit common sense and adaptability for assessing a situation and adjusting goals and plans, but we are constrained by cognitive complexity and can only conceive of relationships between a handful of parameters. On the other hand, weak AI (such as deep neural networks) has only fast and automatic “thinking” and limited (if any) common sense and adaptability, but it can access practically unconstrained parameter space to discover relationships between data.

To go beyond weak AI, we must pursue autonomous systems. These systems require perception (the ability to interpret sensory information), reflection (the ability to construct a model of the environment), and decision-making through strategic goal formulation and tactical planning — each of which is based on a knowledge repository of concepts, properties, and methods that is constantly updated via self-learning. Sifakis addresses the technical and social risks of strong AI systems that arise from issues of trust, explainability, and validity. He does not comment on whether strong AI can evolve from weak AI or if it must be designed from the outside, but his thoughts on this subject would no doubt be enlightening.

The second half of the book explores consciousness in great detail. Relevant to this section is the role of needs and feelings to help humans (and perhaps other living beings) formulate goals and prioritize actions. Needs—whether biological, psychological, or other—subconsciously translate to feelings of different types and intensities, which in turn actuate an individual towards consciously achieving an emotional equilibrium. Regarding consciousness, I’ll quote three rather beautiful passages from Understanding and Changing the World:

“How do we strike a balance and self-control between the conscious and the subconscious? For an athlete or artist, the subconscious is a key factor for their performance. All transcendental experiences are characterized by this temporary loss of the conscious, whether they are religious, erotic, artistic, or relating to creativity, where the subconscious play[s] an important role. An intervention by the conscious could lead to confusion and hesitation and damage the quality of performance … No emotion seems authentic when it is fully controlled by the conscious.”

“We must reassess the role of consciousness in our lives. Many of the wonderful things that people do go beyond conscious control. We can develop skills through intuition and practice in order to communicate with the world and access cosmic mystery.”

“Know thyself. It is a prerequisite for harmony with ourselves in order to achieve a happy life. It is the first and last major imperative for humans. All the others follow, the way theorems follow from axioms. In our personal lives, knowing is not enough. Unmanageable knowledge can become an unnecessary burden that not only does not help, but also sometimes clouds our judgment. Gaining knowledge through the game of freedom is the meaning of life. When the cycle of life closes, our conscience must have discovered and fully understood the rules.”

In the final sections of the text, Sifakis dwells on society and its gnoseological, technical, ethical, and economic values; on institutions, democracy, meritocracy, and bureaucracy; and on the decline of societies through the depreciation of values, decay of language, degradation of concepts, and tendencies like lying, self-convenience, and corruption — “leading to the triumph of idiots.”

Understanding and Changing the World is a highly accessible, must-read work for both the computer scientist and the general reader. The book provides context for recent advances in AI and offers a roadmap for its future direction, both technically (through autonomous systems) and socially (through regulatory considerations). Most of all, it presents the reader with a framework for understanding the individual and society within a complex technological world that continually evolves towards the future.

Akash Deshpande holds a Ph.D. from the University of California, Berkeley and currently works on internet infrastructure at Google. He first met Joseph Sifakis 30 years ago and the two have kept in touch sporadically ever since.