site.btaAI Must Learn From Experience, Say Researchers at Heidelberg Laureate Forum

Artificial intelligence must learn to teach itself from experience, like humans do, prominent computer science researchers at the Heidelberg Laureate Forum said on Monday. 

The forum at Heidelberg University will take place from September 14 to September 20. This 12th edition of the forum will bring together 208 early-career mathematicians and computer science researchers from 54 countries to engage in discussions on artificial intelligence, research ethics, mathematical advancements, and career development with 28 recipients of the world’s most prestigious mathematics and computer science awards.

The second day of the meeting began with talks by Sanjeev Arora (ACM Prize in Computing 2011), David Silver (ACM Prize in Computing 2019), Manjul Bhargava (Fields Medal 2014), and Richard Sutton (Turing Award 2024), all dedicated to the development of artificial intelligence.

The session was moderated by Tom Crawford, a mathematician at the University of Oxford and University of Cambridge, a science communicator with over 30 million views on his YouTube channel.

Sanjeev Arora noted that until 2022, progress was mainly in large language models, but after 2023 the key question became how AI can learn independently from its own experience. Currently, researchers discuss AI capabilities to learn from AI, not from human-created datasets.

According to David Silver, current AI relies on learning from human-generated data. The future lies in an era of experience, where AI’s actions and observations are rooted in its environment and planning is based on its own experiences, unlike today’s large language models. He compared human knowledge to fossil fuels, but said experiential learning will be the renewable energy of AI. Silver urged young scientists at the forum to take part in addressing AI’s biggest challenge how to learn from experience.

Manjul Bhargava spoke about algorithms that will support AI’s future development and ensure quality control.

Richard Sutton, described as the man who taught AI to learn by itself, offered a philosophical and ethical perspective on the future of artificial intelligence. He predicted that within a few decades we would understand the principles of intelligence and emphasized that this is a marathon, not a sprint. Sutton urged both young and established scientists to remain fully focused on advancing AI, without being distracted by applications, domains, large language models, or questions of discreteness and safety. 

"I have thought about intelligence all my life, for about fifty years, and one thing I decided early on was that core to intelligence is to have a goal and a purpose. If you show me a system that has no goal and no purpose, and ignores all its input, does not use its input in its learning processes, then I have a hard time thinking that is intelligent,” he noted.

According to him, the idea of decentralized cooperation will become the natural guiding philosophy in the future. He added that cooperation requires institutions, which may help in the short term but tend to become dictatorial. 

Sutton concluded that AI is not some alien technology, but one of humanity’s oldest aspirations. He expects people to understand intelligence well enough to cultivate it, noting that AI is not a threat to human intelligence but the inevitable next step in its evolution.