In a recent discussion on Twitter, renowned AI researcher Yann LeCun offered a thought-provoking perspective on the current state of artificial intelligence development and the often premature calls for stringent controls on AI systems. His analogy comparing the premature control of AI to the overly cautious approach that might have been proposed for aviation in 1925 highlights a critical issue in the technology discourse today: the importance of understanding and realism in AI development.
LeCun’s tweet suggests that before we delve into “urgently figuring out how to control AI systems much smarter than us,” we need to first develop a system that surpasses the intelligence of a simple house cat. He argues that the urgency some express about controlling superintelligent AI systems reflects a “distorted view of reality.” This sentiment resonates with the historical development of other technologies, such as aviation, where progress and safety were achieved not through fear-driven restrictions but through years of careful engineering and iterative refinements.
LeCun’s commentary brings to light an important reminder: the evolution of AI is gradual and experimental. Comparing the capabilities of current AI to those of a house cat, he points out that significant time and effort are required to reach and eventually surpass human intelligence levels. This path is characterized not by sudden leaps, but by gradual, sustained improvements and adaptations.
Critically, he touches on a notion that is sometimes overlooked in the rush to regulate and control AI: the difference between knowledge accumulation and retrieval, which current AI systems excel at, and genuine intelligence. This distinction is crucial because it underscores that current AI systems, while advanced in specific tasks, do not yet possess the broad, adaptable intelligence that would necessitate the kinds of controls that some advocates fear.
I completely agree with this point. We are still far from the stage where AI needs to be tightly controlled. AI has not yet fully taken shape, and attempting to control it now is like trying to put the most secure and controllable reins on a newborn car…
It’s laughable, a waste of time, self-indulgent, and futile.
Many current AI safety experts might be fans of Isaac Asimov, imagining themselves as the drafters of the core principles for future worlds. They wonder if being the originator of the Three Laws of Robotics would make them the greatest person in the world. However, even though Asimov is one of the greatest science fiction writers and his works repeatedly demonstrate that such logic-based principles cannot be simply applied to intelligent beings, which are complex forms of life, they still fantasize about becoming godlike. They fail to understand that without technological advancement, you truly don’t know how to control technology. Pre-emptively optimized paths are often mistaken because they inevitably miss some developmental pathways.