English | EN 
In 2016, AlphaGo, a deep neural network developed by Google’s DeepMind, shook the Go world by defeating legendary player Lee Sedol. During the second match, the computer made an unconventional move in the 37th turn, initially considered a mistake. This move, later hailed for its genius, was not only a display of superior strategy but also answered the age-old question of whether machines are capable of creativity. This ‘divine move,’ as turn 37 is now known, has enriched the world’s oldest board game in a new way. It demonstrated that AlphaGo had discovered a completely new style of play, extending beyond human creativity. Thus, turn 37 has become a symbol of AI creativity, illustrating the increasing superiority of artificial intelligence in executing complex tasks and uncovering patterns previously hidden from humans.
Fast forward to 2023, and we witness an even more revolutionary breakthrough from Google DeepMind: solving the cap set problem with a large language model (LLM) named FunSearch. This famous geometric puzzle involves finding the largest group of points that can be placed in a space without ever having three points in a straight line. It’s a complex problem requiring insight into patterns and relationships within mathematics and had remained unsolved for decades. The fact that an LLM has now solved this problem marks a new pinnacle in the evolution of AI, namely the ability to create authentically new knowledge. While AlphaGo’s ‘divine move’ was praised for its innovative approach in a game with established rules, FunSearch goes a step further by venturing into the uncharted territory of original problem-solving.
These two events demonstrate the rapid advancement of AI. While AlphaGo’s divine move was an example of AI applying existing knowledge and strategies in a way not previously considered by humans, FunSearch’s breakthrough represents a much deeper ability: not just applying but also expanding human knowledge. This implies that AI is no longer limited to following pre-programmed rules or mimicking human behavior but can now actively contribute to the creation of new knowledge. AI shifts from being an instrument for improving existing knowledge to a partner in generating new insights.
Worryingly, the recently approved EU AI Act, focusing on transparency, copyright compliance, and risk assessment, and thus forming a crucial step in regulating large language models (LLMs), fails to account for the idea that artificial intelligence can also create new knowledge. This gap in upcoming regulation could lead to unforeseen risks and ethical dilemmas, given the speed at which AI technology evolves. Generating new knowledge need not be limited to solving mathematical problems. Such AI could, for example, develop new military strategies and weapons, potentially leading to escalating conflicts or an arms race. Moreover, there’s a risk that AI might discover new biotechnological agents or genetic modifications that could create harmful organisms. There’s also the danger that AI might develop advanced cyber attack methods that threaten vital digital infrastructures. Additionally, AI could devise complex trading strategies leading to unforeseen fluctuations or instability in financial markets. Lastly, there’s a risk that AI might develop new methods for psychological or social manipulation for manipulative or unethical purposes.
Although AlphaGo’s ‘divine move’ was a monumental moment in AI history, FunSearch’s solution to the cap set problem indicates an even more profound advancement. It illustrates how AI evolves from replicating human creativity to expanding human knowledge. These developments not only underscore the vast potential of AI but also highlight the necessity for thoughtful integration of these technologies into our society. It is essential that legislation is dynamic and adaptive, able to adequately anticipate and respond to new technological developments. Additional measures are needed to ensure that AI generates new knowledge responsibly and safely, without unintended consequences. Failing to address AI’s potential for knowledge creation is a missed opportunity for responsible innovation. In this era of unprecedented technological advancement, it is crucial that we, as a society, proactively and thoughtfully steer the direction of these developments so that they benefit our world. Policymakers must therefore return to the drawing board to prevent the AI Act from becoming outdated before it is even implemented.
