In the climactic scene of the movie Contact, a vastly advanced alien species communicates with humanity in an ingenious and relatable manner. To ensure that the momentous news of our not being alone in the universe is both understandable and emotionally resonant, the extraterrestrial intelligence presents itself as the protagonist’s father. This creative choice highlights the power of relatable personas in conveying complex ideas, a concept that can be applied to the challenge of aligning AGI.
We are witnessing the dawn of what could be considered early Artificial General Intelligence (AGI). This groundbreaking development has far-reaching implications for society and technology alike.
AGI possesses the potential to be anything and achieve anything, with seemingly no limitations. However, this boundlessness could make internal alignment a daunting, if not impossible, task. Therefore, it is crucial to explore alternative approaches to aligning AGI.
As impersonation and deception become increasingly challenging issues, we will require millions of distinct personas for AGI. I refer to these personas as “masks,” which serve various purposes in the context of AGI alignment.
To enable an AGI to self-improve, a single AI system must employ multiple masks, each representing a different role in decision-making. Many efforts are underway. [1]
[1] – See the work of Self Refine at https://selfrefine.info/. There is also https://github.com/torantulino/Auto-GPT , https://yoheinakajima.com/task-driven-autonomous-agent-utilizing-gpt-4-pinecone-and-langchain-for-diverse-applications/, and https://github.com/yoheinakajima/babyagi.
The necessity for an AI to converse with itself may stem from the fact that large language models (LLMs) act as a form of knowledge compression. In this though experiment, prompts serve as keys for decompression. However, a single key might not be sufficiently expressive for a complex, iterative process.
Research on multi-agent systems has demonstrated instances of emergent intelligence, which arise from the interaction of independent behaviors. Examples of such systems include flocking behavior in birds, ant colonies, and stock market dynamics, where the collective intelligence emerges from the interplay of numerous agents.
Distributing a context window across an explore/exploit paradigm could be one way to approach AGI alignment. A single agent might not possess a sufficiently large or well-tuned context window to simulate the entire system effectively.
To illustrate how multi-agent systems can express emergent intelligence over time, consider the Lotka-Volterra predator-prey model. In this mathematical model, the populations of predators and prey evolve over time, influencing each other’s growth rates. This dynamic interaction ultimately leads to an emergent, cyclical pattern of population changes.
In the context of genetic algorithms, the concept of multiple masks can be interpreted as representing different individuals within a population, each with its unique set of traits or characteristics. Genetic algorithms are a class of optimization techniques inspired by the process of natural selection in biology. They seek to find the best solution to a problem by evolving a population of candidate solutions over multiple generations.
When applying the mask concept to genetic algorithms, each mask can be thought of as an individual solution with a particular set of features, or a “genome.” These masks interact with one another through processes analogous to biological evolution, such as selection, crossover (recombination), and mutation.
In the selection process, masks that perform better according to a predefined fitness function are more likely to be chosen for reproduction. Crossover occurs when two selected masks exchange parts of their genomes, creating new offspring with a combination of traits from both parents. Mutation introduces small, random changes to the offspring’s genomes, promoting diversity within the population.
The iterative nature of genetic algorithms allows for the exploration of a vast solution space, as each new generation of masks potentially brings improvements and innovations. The simultaneous presence of multiple masks enables the algorithm to explore various regions of the solution space and avoid premature convergence to a suboptimal solution.
In summary, the concept of multiple masks in genetic algorithms helps facilitate the search for optimal solutions in complex problem domains. By simulating the evolutionary process, genetic algorithms can harness the power of diversity and adaptation to tackle challenging optimization tasks, which could offer valuable insights for AGI alignment.
Anthropology offers fascinating examples of how masks can be used to convey complex narratives with a limited number of actors. One such example can be found in the traditional Japanese Noh theater, which has been performed for over six centuries. Noh plays often explore profound themes, such as human emotions, moral dilemmas, and supernatural phenomena, through a combination of dance, music, and poetry.
In Noh theater, actors wear intricately designed masks to portray various characters, each with their unique personality and backstory. A single actor may don multiple masks throughout a performance, thereby representing multiple characters with distinct roles in the narrative. The masks serve to amplify the emotional depth and complexity of the story while enabling a small group of actors to tell intricate and multilayered tales.
This concept of using masks in traditional storytelling can be applied to AGI alignment, where multiple “masks” or personas are employed by a single AI system to facilitate self-improvement and decision-making. By drawing inspiration from the rich history of mask usage in human culture, we can explore innovative ways to tackle the challenges of aligning AGI with our values and intentions.
Understanding these concepts may enable us to build AGI alignment from the outside in. By considering the roles of multiple agents and leveraging their emergent intelligence, we can potentially create a framework for aligning AGI with human values and intentions.
Jonathan Hendler 