Artificial Intelligence (AI) is arguably the technology of this century. Machine Learning is used in countless applications, ranging from healthcare to engineering, strongly affecting our everyday life. AI powers the recommendation algorithms that shape our online experience, it is inside our phones and devices, and it is at the core of self-driving cars.
However, not all that glitters is gold. This technology is also characterized by a very peculiar trait: we understand it only partially. Most AI models work incredibly well, but we do not fully understand why—they are essentially black boxes. It’s like building and using a nuclear reactor without understanding the principles of nuclear physics; the outcomes can be unpredictable and dangerous. In this case, however, the risk is not atomic fallout, but rather phenomena such as opinion polarization or algorithmic discrimination. More recently, with the advent of Large Language Models (LLMs) showing human-like behaviors and cognitive capabilities, even Matrix-like scenarios are no longer unthinkable. Consequently, advancing our understanding of AI is a central issue that we must address as soon as possible.
In my research, I approach this challenge from two different perspectives. On one hand, I apply techniques of Statistical Mechanics to understand the functioning and properties of Artificial Neural Networks. On the other hand, I use Network Theory and concepts from Complexity Science to explore how these models, particularly LLMs, can interact and form artificial, human-like societies.
Large Language Models have profoundly impacted our daily lives serving, for instance, as writing tools and coding assistants. However, in addition to serve as assistance tools, LLMs have the potential to revolutionize entire research ares. One example is the potential integration of LLMs in social sciences, that could mark a transformative era in this discipline. Their application extends from answering surveys to serving as experiment participants, and notably, in agent-based simulations (ABMs), where their impact could be profoundly significant.
The study of the behavior of LLMs in isolation, while informative, offers a limited perspective on the potential dynamics and capabilities of these models in more complex, interactive settings. In human societies, many of the most significant processes and phenomena are not the result of individual actions but are collective outcomes. For instance, cultural trends, economic shifts, and social movements are all emergent properties of numerous individuals interacting with each other, often in complex and unforeseeable ways. In the very same way, we solve most challenges and problems working as teams, not as individuals. These ideas have already lead to the development of LLMs teams, where several LLMs interact and cooperate to perform complex tasks. When multiple LLMs interact, they may exhibit behaviors and generate outcomes that are not predictable or directly coded into any single agent or LLM. Instead, they arise from the complex interactions and interdependencies between multiple LLMs. In a way, each LLM contributes to a collective “intelligence” or pattern that is more sophisticated and unpredictable than its individual capabilities would suggest.