Designing a Curious Machine Intelligence That Actually Thinks

Dr. Karl Friston, a distinguished psychiatrist, neuroscientist, and pioneer of modern neuroimaging, is a leading expert on intelligence, both natural and artificial. I've followed his work with growing interest as he and his team uncover the principles underlying mind, brain, and behavior through fundamental physics and Bayesian inference. The natural world's complexity emerges from a few deceptively simple principles.

One principle is that natural systems seek to reduce surprise through exploration, minimizing what is called free energy—the energy available to do work. Thought, brain, and mind operate in an evolutionary manner, constantly improving predictions and reducing errors through feedback loops of perception and action. Natural systems seek to minimize "free energy", reducing surprise and finding the path of least resistance— closely related to the physics principle of "least action".

Bayes' Theorem explains how the probability of making a correct guess improves as we receive additional information. It uses conditional probability to update prior beliefs (priors) based on new evidence, refining them into more accurate models (posteriors). The probability of an event A given B is determined by the probability of B given A, along with the independent probabilities of A and B. This process allows new information to be systematically integrated, continuously improving our understanding of the world. This equation is elegant and powerful.

In this interview, Dr. Friston and I discuss how he became an expert in artificial intelligence (AI), what the different forms of AI are, and what makes the model he is developing different, and uniquely powerful.

Grant H. Brenner: Can you share a bit about your background, and how you came to work on artificial intelligence, specifically?

Karl Friston: My background rests on careful career choices and a series of serendipitous opportunities. I studied natural sciences at Cambridge University, picking specialties that became the foundations of later work; namely, psychology and physics. I then qualified as a clinical doctor, specializing in psychiatry at the University of Oxford. I started my research at the inception of human brain mapping, developing models of functional brain architectures to make sense of brain imaging data. This endeavor turned to theoretical neurobiology, using the same mathematical principles to understand how we make sense of the world and how this sense-making underwrites sentient behavior. The ensuing work was formalized as the Free Energy Principle, and applications like Active Inference. Both the empirical (complex system modelling) and theoretical (free energy principle) programs are committed to understanding our brains in health and disease. I like to think of myself as an expert in natural intelligence, with a particular perspective on artificial intelligence.

GHB: What are the different approaches to machine learning and artificial intelligence out there?

KF: Approaches to artificial intelligence come in two flavors. The first rests upon our understanding of how intelligent artifacts—like ourselves—work and various attempts to formalize this understanding mathematically and in silico. The disciplines here range across the physical and life sciences; from psychology to pathology, from computer science to social science. These approaches could be cast as aspiring to artificial intelligence that is biomimetic and, perhaps, neuromorphic.¹ This approach seeks principles, mechanisms and explanations that can be instantiated in artifacts to emulate intelligent behavior. If the first (biomimetic) approach is read as the ‘neat’ and principled approach, then the second ‘scruffy’ approach (Poirier, 2024) is machine learning.

Machine learning is an engineering endeavor: basically, optimize something until it can be optimized no further. The results of this approach are things of great beauty; for example, the generative AI and large language models that fascinate our friends and financial markets. The notion of optimization is central here; in the sense this technology rests upon neural networks inspired by the early days of the biomimetic approach. It soon became apparent that (recurrent) neural networks are universal functions of approximators (Wray & Green, 1995), which means one can, in principle, optimize anything.² But what to optimize? This is a deep question, which is seldom addressed in machine learning.³ In short, the biomimetic approach tries to explain intelligent behavior in terms of principles, methods and mechanisms. In contrast, the machine learning approach optimizes artificial artifacts to emulate natural intelligence.

GHB: What kind of problems can Genius — developed by VERSES — address in ways that other AIs, including the much-touted LLMs, cannot?

KF: The problems that Genius is designed to solve revolve around the limitations with large language models and other forms of generative AI; namely, their efficiency, explainability and reliability. Genius resolves these limitations through a commitment to the first principles that underwrite the biomimetic approach. These can be summarized as the tenets of active inference: namely, gathering evidence for generative models of the (digital or physical) niche in which an agent is embedded. The evidence in question can be read as adaptive fitness.⁴ Mathematically, this kind of evidence can be scored with variational bounds such as variational free energy. In consequence, explainability is assured through the use of a generative model that can generate consequence (e.g., content) from cause. Efficiency is underwritten—both in terms of information theory and thermodynamics—via the principles of least action inherent in the free energy principle (Friston et al., 2023). This efficiency translates into intelligent problem-solving that is cheaper, faster and better, in a well-defined sense (Friston et al., 2024). Because active inference does what it says on the tin: namely, infer by acting to resolve uncertainty. The requisite uncertainty quantification means that artefacts like Genius know what they don’t know — and could tell you how confident they are in their answers or recommendations.

This direction of travel for AI distinguishes itself in several ways. First, the generative model in active inference necessarily includes the consequences of action. This future-pointing aspect endows AI with a minimal but authentic agency; in the sense agents can entertain counterfactual futures and select among them to act (or recommend) optimally. But what is optimal in this setting? It is simply to maximize model evidence or adaptive fitness. Interestingly, this entails information-seeking behavior that equips artificial intelligence with a natural characteristic; namely, curiosity (Friston et al., 2017; Sun, Gomez & Schmidhuber, 2011). By understanding the principles of intelligent behavior, one eludes many of the concerns voiced about AI and machine learning. This is because the only objective of active inference is to model and anticipate exchanges with other agents, natural or artificial.

GHB: How do you see the role of AI in humanity’s future? What do you think the hopes and dreads, realistically, are?

KF: This is too big a question to answer here.⁵ Artificial intelligence has had—and will continue to have—a profound effect on our lived world. Should an inclusive ecosystem of natural and artificial agents⁶ emerge over the forthcoming decades, one might hope that it is sustainable. However, one thing we believe is that it must pursue a path of least action.