The arrival of true artificial general intelligence (AGI) is anticipated, though significant hurdles remain, according to Google DeepMind CEO Demis Hassabis.
AGI’s Current Limitations
Speaking at an AI summit in New Delhi, Hassabis addressed whether current AGI systems can match human intelligence, stating, “I don’t think we are there yet.” He identified three key areas where existing AGI systems fall short.
Continual Learning
Hassabis explained that current systems lack “continual learning,” functioning as if “frozen” based on their initial training. He suggested that ideal systems would “continually learn online from experience, to learn from the context they’re in.”
Long-Term Planning
Current AGI systems also struggle with long-term planning. While capable of short-term planning, they lack the ability to plan “over the longer term, the way that we can plan over years,” Hassabis noted.
Consistency of Reasoning
A final limitation is a lack of consistency. Current systems can achieve high levels of performance in specific areas—even earning “gold medals in the international Math Olympiad”—but can simultaneously make errors in basic mathematics depending on how a question is posed. Hassabis argued that a “true general intelligence system shouldn’t have that kind of jaggedness.”
Hassabis previously indicated in a “60 Minutes” interview that true AGI could arrive within five to ten years. The AI Summit in New Delhi also included participation from OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Google CEO Sundar Pichai, and Meta’s chief AI officer, Alexandr Wang.
Frequently Asked Questions
What is AGI?
AGI is a hypothetical form of machine intelligence that can reason like people and solve problems using methods it was not trained in.
Is there disagreement about whether AGI has been achieved?
Yes, AGI is a disputed topic. The Databricks CEO stated that current AI chatbots already meet the definition of AGI, while others believe the goalposts are continually moved.
What are some of the challenges in developing AGI?
Current systems struggle with continual learning, long-term planning, and consistency of reasoning, according to Demis Hassabis.
As AI technology continues to evolve, what level of intelligence will be considered “general,” and how will we measure its arrival?
