They develop a fully functional human mind that uses AI to respond to millions of questions just like a human would.

In the race to develop a  general artificial intelligence (IAG) , companies such as  OpenAI  and various tech giants are investing billions of dollars into systems capable of replicating human thought. This quest focuses on creating AI that can perform a variety of tasks similar to a person.

Despite recent advancements, there remains no universally accepted definition for  IAG . Although current AI systems demonstrate remarkable skills, such as defeating chess champions and accurately predicting protein structures, they remain far behind the human mind in terms of nuanced reasoning and decision-making.

Recently, a groundbreaking system called  Centaur  has emerged, developed by an international team of scientists and led by cognitive scientist  Marcel Binz  from the Helmholtz Munich Research Center. This project delves into whether a language model akin to ChatGPT can behave like a human in psychological experiments.

The name  Centaur  reflects its mythological essence, embodying characteristics of both humans and horses. This system was trained to mimic the decisions and behaviors of real participants in over  160 psychological studies .

Unlike traditional models, which focus on a single skill, Centaur is designed to  generalize human behaviors  across various contexts. Binz explained, “We wanted to see if it was possible to imitate different mental functions in a single model without needing to build a theory from scratch for each task.”

To accomplish this, the team utilized the open-source  FLAME model  created by  OpenAI , allowing them to customize its internal operations. They enriched the system with over  10 million human responses , rewarding it whenever its choices aligned with those of human participants.

The outcome was a model that not only mimicked response patterns but also mirrored human reasoning in new scenarios. One of the standout features of Centaur is its  knowledge transfer  ability. When presented with new variations of tasks, such as substituting a  spaceship  for a  flying carpet , Centaur maintained the original search strategy, much like a human would.

Moreover, when faced with logic problems outside its training, Centaur not only gave human-like correct answers but also reflected common human mistakes. This sparked significant interest among experts.

Experts have labeled Centaur as “impressive.”  Russ Poldrack , a cognitive scientist at Stanford University, asserted that it’s the first model to replicate multiple tasks with unprecedented fidelity to human behavior. Another specialist,  Ilia Sucholutsky  from New York University, emphasized that Centaur surpasses traditional cognitive models.

Nevertheless, criticism exists.  Olivia Guest  from Radboud University in the Netherlands questioned the theoretical foundation of Centaur, stating, “Predictions do not imply an understanding of the mind’s workings.” Likewise,  Gary Lupyan  from the University of Wisconsin-Madison expressed concern, asserting that cognitive science should focus on understanding the mechanisms behind behaviors, not merely replicating them.

Acknowledging these critiques, Binz clarifies that Centaur does not aim to be a definitive theory of human thought. Instead, it serves as a platform for generating new hypotheses regarding how specific cognitive patterns emerge. Currently, his team is focused on expanding the  experiment database  fivefold to enhance the model further.