Editorial: Narrow and general intelligence: embodied, self-referential social cognition and novelty production in humans, AI and robots

Le résumé fourni par la source

It is useful to start with the review paper by Wu et al (2025) where the scene is set for AI generations that have unfolded over the last 7 decades from AI 1.0 to AI 4.0 . These developments have been driven by a triad of factors relating to algorithms and software; chip technology and computing power; access and storage of voluminous data in static and real time mode. AI 1.0 is ground zero, with algorithms aimed to fully direct outcomes mostly based on logic and rules based inference. This phase was accompanied by internet technologies such as search engines, digital automation and data processing. The authors categorize AI 2.0 to encompass Agentic AI, real time online bots and the advent of highperformance GPUs and vast labelled datasets. This gave rise to deep learning and reinforcement learning, with convolutional and recurrent neural networks achieving breakthroughs in vision, language, and control. The third AI 3.0 generation marks the embedding of digital intelligence in an embodied physical agency of robots, where AI operates in material spaces as in autonomous vehicles and other utilitarian cobots.The fourth generation AI 4.0 is set to coincide with aspirations of AGI with controversial notions of machine sentience, with self-capable, adaptive selection of goals and the wherewithal to evolve programs to achieve goals autonomously. Markose (2025) points out that the lack of AI alignment (Bostrom, 2014;Russell, 2019 ) with human values and goals encountered in agents capable of setting their own goals is not unique to AI, but is the problem that lies at the foundations of civil society. Conflicting or adversarial goals of agents and their accompanying actions that are inimical to life must be kept in check for the survival of the human condition. On the other hand, providing AIs that are value-aligned with some awareness, and capacity for moral reasoning, could make them safer and better able to recognise and mitigate risks (Wallach, 2008).The Markose (2025) perspective on genomic intelligence -underpinned by the algorithmic takeover of biology within a uniquely encoded system -is that there are lessons to be learnt on the alignment problem from the evolution of general intelligence in complex life. The view here is that alignment to life and the design of selfhood has been solved in formal ways that can be explicated using Gödel logic, with recent developments in cryptography with the blockchain. The principles involved here can be conjectured to maintain the immutability of original protein coding blocks against internal and external bio-digital adversaries within an evolvable and unbroken chain of life.It has become popular to refer to self-improving code-based systems as Gödel machines in AGI frameworks, which are necessarily end-to-end self-assembly programs as in life (Schmidhuber, 2006, Zhang et. al., 2025). Markose (2025) suggests that this misses the point of Gödel logic that is embedded in complex life first found in the adaptive immune system, AIS, of jawed fish 500 mya and latterly in the mirror neuron systems of primates. About 85% of expressed genes that can be identified as online self-assembly machines that create the morphology and phenotype of a multicellular organism can be viewed as its theorems. These are mapped offline in AIS 'Thymic Self' à la Self-Representation (Self-Rep) structures from Recursive Function Theory of Gödel-Turing-Post. The purpose of this is to recursively identify non-self codes, especially of digital adversaries wielding the negation operator, which are potentially uncountable infinity. A corresponding open-ended capacity to detect changes to self-codes -known to be found only in the AIS and the human brain in a process of prolific predictive coding -is empowered with the Recombination Activation Gene operators. In a bold hypothesis, Markose (2022Markose ( , 2023) ) states the Gödel Sentence to date known to have little relevance to the real world, is ubiquitous in complex life as a hashing algorithm (to adopt the language of blockchains) that enables embodied self-referential intelligence to detect any misalignment or negation of life's self-codes. This is accompanied by an arms race in novelty or surprises in a game with the viral/digital adversary, first identified by the game theorist Binmore (1987) in the archetype of Godel's Liar, to maintain the primacy of life codes. This self-regulation is achieved internally or by human external phenotypical interventions with human artifacts often in a structure of a perpetual Gödelian arms race.This nicely takes us to other papers that investigate self-regulation and embodied intelligence within humans and AI systems. The research paper of Verchure et. al (2022) investigates the self-regulatory processes not through code-based smart controls, which can suffer misalignment by attacks by internal or external bio-malware as per Markose (2025), but via the notion of allostasis, modelled by dynamical equations, whereby multiple physiological parameters are monitored and controlled "to maintain the stability of the integrated self rather than its parts". In particular, they consider how the mammalian brain conducts allostatic regulation of action, as an extension of the principle of homeostasis, using a predictive and adaptive multi-layered control architecture (see also Prescott and Jimenez-Rodriguez, 2025). They deploy an allocentric synthetic agent in a virtual environment and test the dynamical properties of the neural mass allostatic model with internal needs such as heat and hydration to be fulfilled in 3 scenarios. These relate to (1) open field rodent behaviour, (2) where adaptation in navigation is needed, and (3) when criticality reset optimizes the interoceptivedriven decision making process. They find, that though environmental stressors challenge the capacity to fulfil the agent's internal needs, the neural mass model with its self-regulatory dynamics achieves a robust balance in this regard.The perspective paper by Caucke et al ( 2022) explores how our understanding of the prolific capacity of social cognition in humans can help build the same capacities in robots. They review well known theories on embodied self -with self-knowledge -both from the interoceptive internal environment and the external environment, via the sensory motor cortex that undergrids physical situatedness. The use of self-knowledge as the basis of social cognition, empathy and action prediction of other similarly wired-up conspecifics and the strategic necessity of the Sally-Ann problem of false beliefs relating to perception of negation -are discussed. The authors are keen to emphasize that as human social cognition depends on some degree of individual autonomy, remote or externally controlled robots do not engage in social cognition. Likewise, they state swarm robots that can self-organize along a welldefined and externally limited action set do not have autonomy in the choice of goals or actions. They touch on the fundamental problem of coordination and cooperation when robot behaviours are mutually predictable by robots themselves via good internal models of the other. This requires that the robots do not engage in unpredictable actions that are adversarial or disruptive of what is mutually predictable. While specific robots can have their autonomy limited in order to be cooperative, as indicated in the seminal work of Binmore (1987), digital adversaries cannot be eliminated in general and robots like humans must be capable of detecting Liars/adversaries and enter into arms races with them to preserve autonomy of self. In Ryan's (2025) perspective paper, the embodied and ecological approach to intelligence, with the former based on the framework of the Learning Intelligent Decision Agent (LIDA), is used to understand novelty and improvisation in music. For this Ryan draws on the Jeff Pressing model which entails the knowledge base comprised of cognitive units of objects

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.