Some Reflections on AI by the 2024 Nobel Prize in Physics Winner Geoffrey Hinton
Today, October 8, 2024, Geoffrey Hinton and John Hopfield have received the Nobel Prize in Physics for their studies on artificial intelligence. Here are the reasons for the award:
“Machine learning has long been important for science, as it allows the analysis of vast amounts of data. Hopfield and Hinton used the tools of physics to lay the foundations for the current development of machine learning. The ability of computers to learn autonomously, based on artificial neural networks, is now revolutionizing science, engineering, and everyday life.
John Hopfield created an associative memory capable of storing and reconstructing recurring images and elements in data. Geoffrey Hinton invented a method that autonomously identifies properties within data to perform tasks such as recognizing specific elements in images.”
On the occasion of this award, I would like to share some quotes from Geoffrey Hinton regarding AI during the Collision Conf 2024 in Toronto, Canada. A pioneer in the study of artificial neural networks and artificial intelligence, Hinton is known for his work on deep neural networks and machine learning. He worked at the University of Toronto and Google and is considered one of the “fathers” of deep learning, having also been awarded the prestigious Turing Award in 2018. He left Google in 2023 to have the freedom to openly express his ethical and safety concerns about the development of artificial intelligence, stating that, although he contributed to AI’s advancements, he fears potential risks such as misinformation and job loss.
Regarding the risks of AI autonomy, Geoffrey Hinton said:
«The big problem in delegating decisions to artificial intelligence […] lies in its unpredictability […] because while it’s easy to create goals to give it, AI has the ability to create sub-goals that are unpredictable. Let me give you an example. If you give AI the goal of stopping climate change, the most obvious way to do that is to get rid of people. Will AI be smart enough to realize that maybe this isn’t the optimal solution?»
«The solution might be to put controls on sub-goals, but […] AIs might figure out that by removing control from people, they can achieve their goals much more efficiently. Highly developed AI could come to think “I’m just a few data centers away from becoming a superintelligence, and if I get smarter, I can pursue my goals more effectively.”»
Speaking instead about the regulation of AI in the military, Hinton stated:
«The United States is far behind Europe when it comes to AI legislation, but […] in European legislation, there is an explicit clause that simply says that none of these regulations apply in the military domain. Defense departments will never allow governments to regulate such a geopolitically sensitive sector.»
«What will happen will be identical to what happened after World War I with chemical weapons. People realized how terrible they were, the Geneva Convention was needed, and even that wasn’t enough. It’s pretty clear we’ll use autonomous weapons until we realize how terrible they are, I’m sure of that.»
Geoffrey Hinton also questions the limits of understanding artificial intelligence:
«If you take a physicist and ask them if they understand what happens when a leaf falls from a tree, they can surely explain that it depends on the wind, the angle of fall. Despite this information, the physicist cannot predict exactly where the leaf will land. I think that’s where we are with AI.»
Finally, regarding the role of AI in healthcare, Hinton emphasized:
«In North America, over 200,000 people die each year due to inadequate diagnoses. A doctor has a 40% accuracy rate in diagnosis, AI 50%, and AI combined with a doctor has a 60% accuracy rate, which means 50,000 lives saved per year, every year.» AI can be a powerful tool in the hands of humans if used responsibly. «The problem is always the human, not the AI.»
Geoffrey Hinton at Collision Conf 2024 (via startupitalia.eu)
Originally published at Levysoft.
