Personality as an adaptative state machine
We demonstrated that AI already possesses our improvisation skills, and how it could emulate our spontaneous thoughts through recursive self-prompting. However, there is still a missing key component from our intellectual growth: the ability to engage in self-criticism. This introspective capacity plays a crucial role in the development of our personality, which emerges from the clash between our personal convictions and beliefs, and the experience we accumulate through our interactions with the world. Emotions and stresses generated by these clashes are internalized and embodied, making them "personal". In order to create AIs that can develop a form of personality, we must then consider the integration of auto-critique mechanisms into the model's "genome", and ensure the persistence of this knowledge throughout the model's evolution.
While improvisation demonstrated by AI systems may be closely related to our System 1, the slow thinking and deliberate System 2 is essential for the development of self-criticism. To achieve complete functioning intelligence in AI, we need to merge the concepts we discussed previously: imitation, improvisation, mistakes, spontaneous thoughts, and emotions. The main challenge lies in finding a way to integrate these concepts into a model that can learn from its own mistakes and store the knowledge it acquires.
The Theory of Evolution, first proposed by Charles Darwin, explains how organisms evolve by retaining traits that increase their chances of survival and reproduction. These traits are passed on to their offspring through genetic inheritance, using the genome as a blueprint for the development of the organism. This genome acts as a set of rules for humans. Similarly, neural networks, backpropagation, self-attention, and transformers act as rules for AI systems.
In recent years, our understanding of genetic inheritance has expanded beyond the basic principles of Mendel's genetics with the advent of the field of epigenetics. Epigenetics is the study of how organisms' genetic instructions are regulated without changing the underlying DNA sequence. Such regulation occurs through chemical modifications to both the DNA molecule itself and the proteins that interact with it. Epigenetic modifications can be influenced by environmental factors, including diet and stress, and they can have significant effects on an organism's behavior and mental functions.
One fascinating aspect of epigenetics is the notion that experiences and memories can be passed down through generations. Research in the field of trans-generational epigenetics has shown that certain acquired traits can be inherited by subsequent generations.
The implications of epigenetics on our understanding of how nature is able to rewrite its own code are vertiginous. Applied to the development of AI systems, it sheds light on how we could encode the experiences of AI within itself. by enabling AI models to learn from their outputs and integrating them into subsequent behavior, we could create AI systems that are able to adapt and evolve in response to their environment.
Emotions are often viewed as a cornerstone of human uniqueness, setting us apart from other species and machines. They are deeply intertwined with our consciousness, and they contribute to our sense of self, our perception of others, and our understanding of the world around us. However, interpreting emotions as simple changes in state, which can be modeled using a state machine, can offer valuable insights for emulating emotions in AI systems.
A state machine is a model that represents a system's behavior in terms of discrete states and transitions between them, triggered by specific inputs or events. A traffic light is a state machine, cycling through different colors (states) in response to a timer (input), systematically guiding traffic and ensuring an organized flow. When analyzing human emotions, it is interesting to consider them as resulting from state changes in response to sensory stimuli, memories, and other external or internal factors.
For instance, listening to a song with sentimental value or eating food from our childhood can evoke strong emotions, ranging from joy and nostalgia to sadness or longing. These emotional triggers are a result of the brain's connection between sensory inputs (such as sound or taste) and the associated memories or experiences. Upon receiving these stimuli, the brain experiences a change of state, which is manifested through various physiological reactions, such as the release of hormones, changes in heart rate, or alterations in brain activity.
If we consider our body and mind as a complex state machine, emotions can be viewed as the triggers enabling the transition between different states. This reinforces the importance of figuring out how to persist the new state after the emotion has been "felt".
Statelessness and statefulness represent two contrasting architectures of computational systems. In a stateless system, each interaction occurs independently, without any information about past interactions being stored or utilized. Stateless systems are simple and easier to scale since they do not need to maintain any internal memory of previous states.
On the other hand, stateful systems maintain and utilize information about their past states in subsequent interactions. This allows them to exhibit more dynamic and adaptive behavior, as they can adjust their responses based on the context and experiences. Transposed to AI systems, statefulness can lead to more advanced learning and problem-solving capabilities, as the system can incorporate its accumulated knowledge into future decision-making processes.
Balancing the two approaches is key to creating AI systems that can effectively adapt and learn, while remaining manageable and scalable. For instance, ChatGPT uses contextual information from the conversation to generate more relevant and coherent responses. It exhibits a limited form of state persistence by maintaining the context from recent conversation history while managing to preserve a coherent personality throughout the conversation.
In the context of a Large Language Model, the implication of state persistence at the scale of the entire model in real-time is a daunting and complex task. It would require an enormous amount of memory and computational power. As the model continues to process new inputs and update its weights, the computational resources required to maintain the learning process increase exponentially. This is why current AI models like GPT-4 are limited by their training process. This lengthy and costly training typically involves a static cut-off point, after which the model no longer evolves or stops learning. It takes months of human labeling and reinforcement learning to fine-tune the model and enhance its performance. The process is very much asynchronous and discontinuous, with the model learning in bursts, and then freezing for long periods of time.
To address this challenge, a more flexible approach could involve enabling AI models to cycle between phases of interaction, where state context is scoped and attention span is limited, and phases of memorization and encoding of previous interactions. By doing so, AI systems could learn and update their knowledge while managing state persistence more efficiently. Then maybe they'll dream of electric sheep?
Next chapter: The Path to Sentient AI is a Self-Fulfilling Prophecy