Top 3 Tomek Korbak’s papers on Artificial Life

Introduction

Since 2010, it was already estimated that the number of scientific articles published per day is abysmal, even in specialized topics. Despite the proliferation of LLMs, today we know that these tools are not entirely reliable in obtaining scientific references. A crude fact follows from the above: even with search engines such as Google Scholar, it is impossible to review all the literature devoted to a particular topic or problem. That is why every time I find a good reference, one that answers exactly what I was wondering about, and that curiously is not the most cited on the field, I take on the task of promoting and sharing such information.

Tomek Korbak is a young scientist who was borned in Poland, where he studied philosophy and cognitive science as a BSc and MSc, respectively. Later, he did his PhD at Sussex, under the supervision of Chris Buckley and Anil Seth. Although today Tomek is a Senior Research Scientist at the UK AI Safety Institute working on Language Models, during his short career he has written interesting pieces on Artificial Life, writings that are not so well known (neither well recognized) by the majority of researchers in such a field. Today I am going to briefly present three of his papers that deserve to be shared with the community.

Korbak, T. (2015). Scaffolded minds and the evolution of content in signaling pathways. Studies in Logic, Grammar and Rhetoric, 41(1), 89-103.

Since the last century Ludwig Wittgenstein pointed out the relationship between language and cognition. However, it was until 2013 that Hutto & Myin proposed what is nowadays known as the Hard Problem of Content. In layman’s terms, the authors argue that naturalistic theories can’t use information-as-content to explain content, because it would presuppose the existence of contentful properties. In order to circumvent this problem, Hutto & Myin allude to radical enactivism and anti-representationalism, which leads them to conclude that basic minds don’t involve content.

Inspired by some classic writings of Humberto Maturana and Howard Pattee, in his paper Tomek argues that since Hutto & Myin’s thesis is based on the existence of content being underpinned by social and linguistic practices, then there is no reason why simpler communication systems based on joint action orientation, such as cell signaling pathways, should not also give rise to structured content. Tomek’s argument is valid even if we assume that language is not mediated by symbols, which drives him to conclude that content evolves spontaneously in complex regulatory systems.

“Now whatever it is that makes human linguistic practices give rise to content should also give rise to content in waggle-dancing honeybees, quorum-sensing in bacteria, and hormones in one’s endocrine system as they manifest the same lightweight language-like properties.”

Korbak, T. (2021). Computational enactivism under the free energy principle. Synthese, 198(3), 2743-2763.

Computationalism and enactivism have long been seen as two distinct perspectives when describing cognition. On the one hand, computationalism asserts that cognitive systems can be described as symbolic information processing, such that cognition is a computational process. On the other hand, enactivism alludes to the premise that cognitive systems are autonomous entities; this results in cognition being corollary of the interaction between an organism and its environment. Between these two opposite perspectives, we find the Free Energy Principle (FEP), a narrative proposed by Karl Friston that suggests the minimization of free energy as a mechanism to reduce uncertainty and maintain an homeostatic state in biological architectures.

An enactivist will point out that FEP explains cognitive systems as constantly self-organizing to non-equilibrium steady-state. A computationalist, on the other hand, will argue that as FEP claims that Bayesian inference underpins both perception and action, it entails a concept of cognition as a computational process. In this magnificent article, Tomek shows that if we assume that at least a weak version of the free energy principle is true, then we can merge enactivism and computationalism into a single philosophical scheme, which he baptizes as “Computational Enactivism”, and it turns out to be stronger than either computationalism or enactivism on their own.

“The computational enactivism defended in this paper differs in acknowledging and emphasizing active inference being a computational process (as opposed to only being describable as computing), while at the same time entailing a fully-fledged enactive theory of mind.”

Korbak, T. (2023). Self-organisation, (M, R)–systems and enactive cognitive science. Adaptive Behavior, 31(1), 35-49.

In the last century, several mathematical models were independently developed to explain the self-organization observed in biological structures. Among some of them, we find Ashby’s homeostat, Prigogine’s dissipative structures, McCulloch-Pitts’ neural networks, Maturana-Varela’s autopoiesis, Rosen’s (M, R)-systems, Kauffman’s autocatalytic sets, and of course, Ganti’s Chemoton. In the first part of his paper, Korbak undertakes the task of reviewing all the formal models of self-organization mentioned above and relating them to different narratives in modern cognitive science.

According to Korbak himself, (M, R)-systems are “general enough to encompass all kinds of mathematical structures (and formal or computational models)”. Therefore, the remainder of the paper is devoted to studying Rosen’s argument for the computability of (M, R)-systems, the canonical model in relational biology. Tomek analyzes in detail the biological and epistemological assumptions made by Rosen, arriving at the conclusion that his argument about the non-computability of life is invalid. As if that were not enough, Korbak offers a temporal parameterization of (M, R)-systems, which makes it possible to connect such a class of models to an important family of systems that assume the free energy principle through active inference. Unnoticed by the author, this also solves what some call “The Starting Problem” for relational models.

“To summarise, Rosen claimed that (M, R)–systems are not computable because each component of the system depends on other components. I have argued that this argument is misguided because a simple reformulation of (M, R)–systems is possible that both (i) captures all the biological and philosophical intuitions regarding the notion of self-organisation and (ii) can implement computations under an iterative scheme.”

Conclusion

Science is not a linear process. Any form of acquisition and production of scientific knowledge has its difficulty, even more so today where the quantity of articles carries more weight than their quality. Using the understanding acquired by thinkers who have come before us is a revolutionary act, as this guarantees the dialectical process that leads us to intellectual progress. Sharing with you these three pieces by Tomek is a grain of sand on the beach, a drop of water in the ocean, which although small, gives rise to a change of paradigm, where excellence in each piece of writing is more important every day.