Teoharie Catalin

Predictive Processing theories and the Free Energy Principle (FEP) face a dilemma when addressing the problem raised by the Extended Mind Hypothesis (EMH): on the one hand, few can deny that some objects, such as phones, computers, or other people, have a disproportionate role in the reduction of free energy; hence, it is natural to accept that the mind is extended. On the other hand, this seems to go against the carefully built framework in which Markov blankets (MBs) create an evidentiary boundary between the brain and the external environment, thus rejecting the claim that the mind can be extended to external objects. One approach would be to allow for permeable MBs, where boundaries can move, hence allowing external objects to become part of the mind. I find these solutions to be unsatisfactory because they go against the FEP framework, and I propose a different solution instead: human agents have the possibility of using Direct Input Variables (DIVs), which involves directly using the results of other models without performing the statistical inference themselves. Reducing free energy can be achieved either by statistical inference performed by the brain or by using DIVs. The use of DIVs removes the need to accept the EMH, preserves the integrity of the FEP framework, and opens new questions regarding how DIVs can be integrated within the FEP framework.

At first glance, the treatment of LLMs from the perspective of the Free Energy Principle (FEP) seems straightforward: LLMs do not fully operate under the FEP due to their lack of direct connection with the external environment and the incompleteness of their active inference. LLMs do not fully capture the consequences of their actions; hence, they will never acquire a human-like world model. While I agree that LLMs do not operate under FEP and have major limitations resulting from their lack of connection to the external environment, I argue that they capture an important part of our world model by grounding their responses in it. Although this grounding is mediated—and therefore imperfect—I attribute to LLMs a deeper relation with the external world than many have thus far acknowledged. I explore two main directions when evaluating the possibility of LLMs encapsulating a human-like world model. First, I evaluate the similarity of their behavior, with the underlying hypothesis that, under the FEP, similar behaviors exhibited in similar environments imply similar internal models. Second, I use research on mechanistic interpretability to explore whether human and LLM neural networks are similar. The conclusion is that LLMs might acquire part of our world model, even though this was not intended during their training.