It’s intelligent! (Who’s afraid of AGI?)
Artificial General Intelligence (AGI) raises hope and concern in equal measure. Capable of competing with humans across all areas of cognition, it promises to transform the world—but at what price? We dive into an uncertain future, which holds both promises and threats.
Artificial General Intelligence, by Julian Togelius, MIT Press, 2024.
AGI: What are we talking about?
We have five years left… According to Ray Kurzweil, the big shot of Google-made futurism, Artificial General Intelligence (AGI) will come to be by 2029. While these predictions date back to 2005, Kurzweil was recently interviewed and stands by his prophecy, even suggesting he wouldn’t be surprised if the great leap happened sooner.
What great leap are we talking about? What exactly do we mean when we talk about AGI? For Julian Togelius, AGI is the digital Holy Grail—a system capable of doing everything a human can, but without the headaches or sudden cravings for coffee. It’s AI with unparalleled flexibility, capable of solving diverse problems across varied contexts. This stands in stark contrast to today’s machines, the “narrow AIs,” which excel only within their limited domains: playing chess, recognizing cats in photos, and that’s about it.
Not all experts agree on the definition : Ben Goertzel, one of the pioneers of the AGI concept, describes it as a universal and conscious brain, while researcher Murray Shanahan envisions it as a machine with ultra-adaptable skills but lacking deep understanding.
However, they all agree on one point: AGI will not be a mere enhancement of generative AI, which has revolutionized how we create new content—texts, images, music, or videos. As powerful as it is, generative AI remains a specialized form of intelligence: it can produce original outputs, but it doesn’t understand the concepts it handles.
Ask ChatGPT
And it will candidly admit: “My expertise is limited to what I was trained for: providing relevant and useful answers based on a vast corpus of text. However, I cannot autonomously adapt to completely new or unfamiliar tasks. My decisions are not guided by internal goals but by probabilistic models.”
If you ask whether it could evolve into AGI, it responds pessimistically: “Becoming AGI would require major technological advances beyond what my current architecture allows.”
Togelius, for his part, also takes a cautious stance, pointing out that significant technical hurdles remain: insufficient computing resources, inefficient learning methods, and data too limited to train a machine capable of excelling in the unexpected. For now, AGI remains a distant dream.
© Copyright Business Digest - All rights reserved
