If consciousness is an evolved function, the immediate question is what exactly is the currency of evolutionary selection in terms of traits and functions? In almost all of these kinds of arguments there is an explicit requirement that there is on average (or slightly greater than on average) value to survival that results in the maintenance and promotion of the relevant functions. In my Berggruen essay, I argued for a primarily social role to consciousness. Consciousness is a central monitoring framework for the complex web of social interactions that creates a reflective model of a person (and partially and uniquely in certain other species) that can be used to evaluate and plan for sexual pairing and other life choices related to status within social hierarchies. There are other hypotheses, as well, like the idea that predator-prey planning and avoidance is enhanced by a central consciousness experience, including some intriguing work on dreaming (when there is no active consciousness) that shows enhanced dreams for game players who are in the role of being prey within the game context as well as other forms of cognitive activation.
Abstractly, evolution is a distributed adaptation and learning algorithm that is the only robust solution to the complexity of natural environments. Wasteful though it may be, it is the invisible hand that drives forward enhanced prediction and survival using the knobs of genetics and the social relationships that are an extended phenotype in social species. There are a few theories of abstract learning that can be brought to bear on this topic, with the obvious candidate being inductive optimality via Kolmogorov complexity: minimize the model parameters to bottleneck against overtraining and avoid overfitting. This is central to all distributed learning but has layered complexity when considering the how to predict larger, more distal patterns, both temporal and spatial in activation and extent. Large language models as predictive frameworks are a quite literal example of how to blend attention (and sometimes using recurrence mechanisms) to data cooccurrences in order to improve the likelihood of a viable longer-range word prediction.
In my Berggruen essay, I suggested that future social interactions might be of such low scrutability that we get a parallel to the inherent problem of ethical consequentialism that makes the outcomes of almost any complex response incalculable. The alternative is to use more granular rules that bucket together trends and signals combined with working responses that are tentative. This is our patchwork bridge for managing uncertainty and it also raises the broader question of whether reductionism has intrinsic limits for unpacking external phenomenal experiences that render concerns about determinism and free will largely moot.
So we have, as evolutionary currency, simple survival once again as the source for consciousness, but that survival is contingent on tuning not just for a given niche but a highly variable set of potential adaptive surfaces, from predation and prey, to sociality and reproductive success, and also potentially against parasitism in co-evolutionary races. A contemporary philosopher can easily argue against the necessity of a qualitative experience of this consciousness but there are some intriguing sub-theories that suggest proximate drivers for a kind of “theater of mind.” For one, the need for bucketing or schematization of environmental cues in managing attention and responses might be the only adaptive path to high behavioral plasticity given limited supercomputer real estate. This idea arises in the theory of unlimited associative learning, valence, and agency as ultimate drivers for conscious evolution.
An intriguing implication of all of this theory and science is that there is nothing blocking artificial general intelligence (AGI) or even machine consciousness except for the contingencies of wetware, evolutionary motives, and perhaps some architectural considerations about large-scale connectivity. Most contemporary AI research around LLMs is not focused on those topics, however, because incremental improvements on test sets is the currency of the industrial interests in AI. But it is inevitable that the greater scale of simulation fabrics will lead, in turn, to more alternative experimentation and theories in pursuit of this curious problem. “I think, therefore I am” is evolving through a transitional species of the purely computational, “It computes, therefore it thinks,” before moving through a punctuated equilibrium into “I evolved, therefore I am, and I occasionally think.”