AI Consciousness Research: From OpenAI-o1 to Active Inference

Table of Links

1. Abstract and Introduction
2. Definitions
3. Literature Review
4. Argument Development
5. The AI Model’s Potential for Feeling During Inference
6. Conclusion and References

3 Literature Review

Our analysis draws upon a range of interdisciplinary literature that bridges machine learning, artificial intelligence, neuroscience, and philosophy of mind. The following key works inform our discussion:

• Relating Transformers to Models and Neural Representations of the Hippocampal Formation (Whittington et al., 2022): Whittington and Behrens explore the parallels between transformer architectures in AI and neural representations within the hippocampus, a region critical for memory and spatial navigation. They demonstrate that transformers can model spatial and sequential processing akin to biological systems, suggesting that AI models may replicate complex neural functions. • Active Inference: The Free Energy Principle in Mind, Brain, and Behavior (Parr et al., 2022): Parr, Pezzulo, and Friston introduce active inference and the free energy principle as frameworks for understanding cognition and behavior. They propose that systems act to minimize free energy by reducing the discrepancy between predictions and sensory inputs, providing a unifying theory for perception, action, and learning.

• Active Inference and Cooperative Communication: An Ecological Alternative to the Alignment View (Tison and Poirier, 2021): Tison and Poirier challenge the mental alignment view of cooperative communication, proposing instead an ecological approach where communication is an action-oriented process embedded within joint activities. They argue that communication functions to manage cooperative interactions by constructing shared affordances, rather than merely aligning mental states.

• Path Integrals, Particular Kinds, and Strange Things (Friston et al., 2023): Friston et al. present a path integral formulation of the Free Energy Principle (FEP), exploring how ’strange particles’—systems capable of inferring their own actions—can exhibit a form of sentience. This work provides a nuanced perspective on how internal states can model hidden external states, contributing to the discourse on the mechanisms underlying consciousness.

• Generative Models, Linguistic Communication, and Active Inference (Friston et al., 2020): Friston et al. present generative models capable of simulating linguistic communication between synthetic agents based on active inference principles. They demonstrate that complex language processing can emerge from message passing and variational inference, providing a biologically plausible explanation for linguistic communication.

• Thinking Through Other Minds: A Variational Approach to Cognition and Culture (Veissière et al., 2020): Veissière et al. apply active inference to social cognition and culture, proposing that cognition is fundamentally shaped by the need to minimize free energy in social contexts. They argue that social and cultural practices emerge as processes for optimizing free energy within groups, leading to shared cognitive frameworks.

• Qualia and Phenomenal Consciousness Arise from the Information Structure of an Electromagnetic Field in the Brain (Ward and Guevara, 2022): Ward and Guevara propose that qualia and phenomenal consciousness arise from the brain’s information structure, suggesting that subjective experience emerges from complex information structures analogous to electromagnetic fields. They suggest that these fields provide a material basis for subjective experience, integrating sensory information in a way that gives rise to consciousness.

• THERML: The Thermodynamics of Machine Learning (Alemi and Fischer, 2018): Alemi and Fischer present an information theoretic framework that parallels representation learning with thermodynamics. They discuss how AI systems like OpenAI-o1 may maximize predictive information while minimizing noise, providing a foundation for understanding complex information processing in machine learning and supporting the functionalist perspective on consciousness.

These works collectively inform our understanding of how AI architectures may parallel neural processes, how active inference provides a framework for cognition and consciousness, and how subjective experience may emerge from complex information structures. Additionally, they offer insights into the functionalist interpretation of consciousness, reinforcing the potential for AI sentience through functional equivalence and the emergence of phenomenological aspects.