Biology, Buddhism, and AI: Care as the Driver of Intelligence

Diverse intelligences may work together, and be organized in layers. In natural systems, one finds numerous levels of organization, from fundamental physical particles at the bottom, up to the biosphere, or beyond the universe. This layered structure has been identified as a layered cake of “reductive levels” [132], “levels of mechanisms” [133], or the “multilevel selection theory” of cooperation in biology [134,135].

The evolution of life on Earth has seen major evolutionary transitions, characterized by individuals which could previously replicate independently, cooperating to form a new, more complex life form [136,137]. For example, archaea and eubacteria formed eukaryotic cells, and cells formed multicellular organisms. A major evolutionary transition involves two steps: first, the formation of a cooperative group; second, transition to a new level of organism, with division of labor, interdependence, and coordination of the parts. This new level constitutes a biological layer [90,138]. This may also translate into a debated framework for studying downward causation [29,139], where higher-level entities or properties may exert causal influence on [140] lower-level ones [54,141].

One system or level of organization may signal its stress to another one, which may, as a consequence, drive it in such a way that this second system may express a behavior that ends up reducing the stress level in the original system. One may view this phenomenon as a communication system being established between two systems, which allows them to drive each other, sometimes bidirectionally, thus effectively exchanging signals to mutually offload their own frustration. Eventually, for this method to work, a protocol needs to be established between the systems, so that the conditions lead to the right course of actions for the second system to reduce stress in the first system. If successfully established—i.e., if the signal manages to drive system B so as to reduce the stress in system A—a system may be seen as offloading its stress onto another system, as this stress in the first system converts into some structure in the signal sent to the second system, which the second system may then interpret to generate behavior helpful to the first system to achieve certain given objectives. Once established, such protocols may be understood and treated as foundational factors for the creation and evolving of artificial “Bodhisattva agents”. Effectively, the stress acts as a real-time indicator of care—the progress with respect to an objective function for a given system, since minimizing a given parameter may be generalized to optimizing for any well-defined goal.

In keeping with the Bodhisattva model, let us notice that this phenomenon may occur between levels of organization as well, with one layer signaling its stress to the next level of organization. Although single bacterial cells may be driven by a level of stress connected to local sugar levels, the tissues collectively formed by those cells may be concerned with evolving into the right shape and size instead.

In AI, goals are what allow one to explicitly drive a system’s learning. Every AI technology has a mathematically well-defined goal, often a labeled dataset, but sometimes a different paradigm. In supervised learning—which are the most commonly used machine learning techniques—a goal is defined by choosing a training set with determined labels. For example, if one wanted to identify whether a picture contains a cat, the training set may be a set of pictures each labeled as containing a cat or not. The learning system will then be trained to output whether an image contains a cat or not, by being rewarded when giving the correct response. Most of deep learning, neural networks, decision trees, random forests, and logistic regression are all training based on labeled datasets to define their goal objective. In another family of machine learning techniques known as unsupervised learning, a goal is defined too, although in a slightly different or perhaps more subtle manner. In unsupervised techniques such as clustering (k-means, etc.), the goal is very well-defined, as some mathematical cost function to minimize. An example of cost function may be the Euclidean distance between points within the same cluster, and the opposite of the distance between points not within the same cluster. In reinforcement learning, one would define a function to reward or penalize actions, which also determines a goal for the learning algorithm. Such technology has been used recently with a lot of success to play games such as chess and Go, where a sequence of actions is required to reach a particular goal such as a checkmate. The reward function continuously gives hints along the way to determine the correct learning path.

All these machine learning techniques, among many more, include a well-defined goal. In other classes of algorithms, the goal might be implicitly defined as part of the environment or the simulation in use. In evolutionary computation or genetic algorithms, for example, a population of mathematical functions is constantly evolved and selected so as to perform best on a given problem. The choice of the problem will determine the goal, most often in a “fitness function” that defines the reward, ultimately used to choose the path to learning for the algorithm. At first, these examples may feel somewhat distant from the way we view biological entities, including humans, having goals. However, they have in common that they drive the entity toward their future states. The notion of stress mentioned above should be similar in that respect, since it drives the system in a certain trajectory in time and space. Next, we explain a way to characterize the constraints over the states of agents and their goals, which correspond to two types of light cones, one over physical states, and the second over goals. We argue that the latter is equivalent to care.

To further clarify our description of care light cones, let us consider a simple example using the game of chess, an abstract strategy game where players take turns moving their pieces to achieve the goal of checkmating the opponent’s king. In the Figure 2 diagram, if we were to consider an agent only playing chess, each point of the space would correspond to a certain board position. This board position would be possible or not to reach (in the future) or having been reached (in the past) from the current position according to the rules of chess, based on whether it is located inside or outside the physical light cone. The points in the blue care light cone may for example be a board position which contains a checkmate. Points too distant in space or time to be affected by the agent might be for example a chess player aiming at performing a ladder checkmate, while the pieces necessary for it, rooks and queen, are all missing from the board. Some states in the physical light cone of possible moves in a game might be beyond the light cone of care because the player is incapable of reading far enough in advance or into the past, but they still remain achievable through a certain sequence of moves respecting the rules of chess. Our chess player will not calculate or consider moves that are out of their care volume. A hypothetic perfect player in a solved game, who would see all possibilities in the game, would have their care light cone extend infinitely over the whole space of possible boards, while an empty care light cone might mean the agency of the entity ceases to exist.

Earlier, we also mentioned that artificial intelligences may emerge or evolve in various substrates, adapting and constrained by them, which in turn creates different types and degrees of intelligences. Diversity in types of “intelligences” in our example of chess would lie in the fact that chess players display various styles in playing the game, in terms of style or paths to victory. The substrate is typically a human brain, but instead of playing the game themselves, one may write an algorithm to play the game in their stead, which would typically run on a machine and possess various types of strategies and various degrees of success against certain opponents, defined both by its code and the machinery running it. Diversity in terms of chess skill, if not evident, is exemplified in the light of the game not always needing to be transitive [142], i.e., for any triplet of players A, B, and C, A being consistently stronger than B and B being stronger than C does not necessarily mean that A will be consistently stronger than C. An anecdotal example is a series of matches of Tal vs. Spassky. The years around Tal’s winning the World Championship, 1954–1965, he scored 2-9 in wins against Spassky. One of the wins came from a lost position, which could have made it a 1-10. Spassky’s skill level peaked after that period, but Tal still went 5-0 against him after 1965. The same way chess is non-transitive for human players striving for victory determined by the rules of the game, so are other environments where agents must strive to achieve specific goals under different rules. This goes to illustrate how agents may display a high diversity of paths to achieve their goals and reduce their levels of stress.

Chess is a typical domain where hybrid play (two or more players playing as one) is commonly found, either as a combination of several human players, several AI engines, or a combination of both. This may yield a simple case study of a system where agents with mixed stresses may stretch their respective care in such a way that the system as a whole performs better at the game from their interaction, and relates to the impending enrichment of our world with a broad diversity of evolved, designed, and hybrid agents [4].

Novel technologies are exploiting the plasticity and interoperability of life to create novel living beings, such as hybrids, chimeras, cyborgs, brain-computer interfaces, etc., resulting from living tissue tightly integrated with designed inorganic interfaces [39,40], biological robots [42,43,44], neuroprosthetics [143,144,145,146,147,148], and hybrots consisting of living brain tissue instrumentized to control artificial new bodies [1,43,45,46,47,48,49]. Our future will involve a highly diverse space of novel beings in every possible combination of evolved cellular material, designed engineered components, and software. How do we know what we should expect from intelligences in unconventional embodiments? How do we relate to them, and what do we owe them, in a moral sense? Current distinctions that rely on the origin (evolved or designed) or composition (biological vs. technological) of agents will not survive the next couple of decades. These novel, unconventional beings will not resemble any other familiar touchstone in the evolutionary phylogenetic stream. In contrast to Western philosophies rooted in an essentialism of the Garden of Eden, where Adam named a standard set of animals, Buddhist concern for “all sentient beings” is suited to the astronomically large option space of possible beings.

In the absence of common markers (such as brain size), we must establish a rubric within which to compare truly diverse intelligences and set rational policies for proportional moral relationships between very different beings. One such is the measurement of the area of Concern: we can gauge the degree of each being’s possible radius of compassion, create only beings with large, outward-facing compassion capacity, and at the same time enlarge our own agency and intelligence by acting on the Bodhisattva vow.

How do we pick our goals? Try to dissolve goals? That dissolves the Self. Or enlarge goals to improve lives of other beings? Bodhisattva enlargement.

How to live a good life when binary notions of real and false are increasingly called into question and we are called upon to skillfully occupy the space between them?

Stress drives agents towards homeostatic goals—a concept central to Buddhism with its teaching of existence as dissatisfaction, duḥkha. Expanding one’s space of possible goals to face outwards, exhibiting compassion toward other agents’ goals, potentiates the increase of intelligence and thus the potential to identify better, more global solutions. In this scheme, compassion (in the sense of practical concern) and intelligence are tightly linked because the driving definition of an active agent is the bundle of processes that expend energy toward system-level goals. The scale and content of what a given agent measures, prefers, and seeks to implement define its cognitive sophistication. All beings (including humans, synthetic organisms, and engineered AI) can expand their cognitive boundary, working on the meta-goal of enlarging and turning outward (toward the struggle of other sentient beings) their capacity to Care. The Bodhisattva vow is an example of how this can be done and can be viewed through the lens of biology, cognitive science, AI, and Buddhism, which are surprisingly coherent in their emphasis on Care as a central invariant across diverse embodiments.

Here, we have considered a framework for defining intelligence in terms of stress reduction or stress transfer. We have discussed how this cognitive light cone model can be used for comparing instances of evolving intelligence in terms of their ability to identify stress and their alleviation at increasingly complex scales, and we have paid special attention to the way a Buddhist account of increasing intelligence through increasing care, specifically as associated with the Bodhisattva vow, may play out along the lines of this model. If we extend this model of intelligence into AI contexts, we might suggest that a radical expansion of the definition of goals, structured around the principle of care for the alleviation of stress, would be required for growth processes that might ultimately lead to systems of artificial general intelligence. Above all, we have identified Care as a central invariant concept across biology, AI, and Buddhism, to capture the motivation, stress, and goal-directedness of agents. This concept may offer some directions of improvement for both natural and artificial intelligence, by committing to expanding the cognitive boundary or the light cone introduced above.

The concept of a Bodhisattva, given its infinite goals and infinite care, provides us with a roadmap towards hyperintelligence, where the scope of goals and their quality/impact are constantly improving. The Bodhisattva vow is a critical point in the evolutionary or personal continuum of intelligence of any agent because it initiates a positive feedback loop and triggers a “great evolutionary transition” [75,90,149,150] in individuality. Strategies that focus on implementing the Bodhisattva vow are a path for enabling a profound shift from the limited scope of current AIs and their many limitations. Consistent with a central concept of Buddhism—commitment to seemingly unachievable goals—the building of agents capable of undertaking the Bodhisattva vow is a profound challenge. However, progress along this path is as essential for our personal efforts toward personal growth as for the development of synthetic beings that will exert life-positive effects on society and the biosphere. Above all, the concept of Care provides a strong and fundamental link between practical strategies that will enhance engineering capacities and a way to develop a mature system of ethics that will be essential for a future in which highly diverse sentient beings must coexist and thrive together.