Despite being inherently social creatures, humans appear to impose natural constraints on the number of direct relationships we can maintain effectively. New research indicates that AI may also be capable of collaborating effectively within significantly larger teams.
In the 1990s, renowned British anthropologist Robin Dunbar proposed that most individuals are capable of maintaining a maximum of approximately 150 social connections. Despite ongoing controversy over the effectiveness of the methods employed by Dunbar, his findings have become a widely accepted benchmark for determining the optimal size of human groups in business management.
Researchers have been pondering whether today’s Large Language Models (LLMs) are similarly limited when it comes to their capacity to facilitate collaboration among multiple individuals. Researchers found that highly successful fashion trends often involve collaborations involving teams of over 1,000 individuals, a significant increase from human participation.
“I was once utterly astonished,” said Giordano De Marzo, a researcher at the University of Konstanz in Germany. “With our advanced computational capabilities and available resources, we successfully simulated thousands of broker scenarios, yielding no indication whatsoever of a breach in the ability to form a community.”
Researchers tested the social abilities of Large Language Models (LLMs) by creating numerous scenarios featuring an identical model, assigning each a unique and randomly generated opinion. The researchers subsequently verified each individual’s opinions, seeking confirmation from their peers before inquiring whether they were willing to relinquish their personal perspectives.
The workforce found a direct correlation between the likelihood of achieving consensus and the efficacy of the underlying model. While smaller models like Claude 3 Haiku and GPT-3.5 Turbo struggled to achieve settlement, the 70-billion-parameter Llama 3 model surprisingly reached a settlement in fewer than 50 scenarios.
While GPT-4 Turbo is likely the most effective model among those tested by researchers, it’s possible that teams of up to 1,000 identical models could achieve a consensus outcome. The study was limited by insufficient computational resources, which prevented them from investigating larger teams.
According to Dunbar, larger AI models may potentially collaborate on a scale beyond what humans are capable of. “Their ability to gather diverse perspectives and reach a consensus seems more likely to happen quickly, especially when considering a broader range of opinions,” he said.
Research on AI collaboration continues to yield promising findings, as numerous studies have demonstrated that groups of artificial intelligences can excel in a diverse array of mathematical and linguistic tasks. Despite their ability to thrive in massive teams, the sheer computational cost of processing such a large number of scenarios might render the concept impractical.
Additionally, simply agreeing on one thing does not necessarily imply that it is proper; as Philip Feldman from the University of Maryland has taught. Despite the apparent predictability of identical mannequins forming a consensus, there is a significant risk that their collective response may not necessarily yield the optimal outcome.
Despite this, it seems plausible that AI brokers are better suited for large-scale collaboration, unencumbered as they are from organic limitations on speed and data capacity. While the effectiveness of current trends in utilizing this technology is uncertain, it’s plausible that future advancements will ultimately enable optimal results.
