Demis Hassabis, CEO of Google DeepMind, highlighted significant inconsistencies in AI’s problem-solving abilities during a Google for Developers podcast. He noted that while Google’s Gemini models, powered by DeepThink, excel in complex mathematics—such as winning gold medals at the International Mathematical Olympiad—they also struggle with basic high school math problems. This phenomenon reflects what Hassabis describes as “jagged intelligence,” a term resonant with Google CEO Sundar Pichai’s notion of “AJI” (artificial jagged intelligence), indicating highly specialized strengths but notable weaknesses. Hassabis emphasized that improvements require more than simply increasing data and computing power; essential reasoning and planning capabilities need to be developed. He also acknowledged ongoing challenges in achieving true artificial general intelligence (AGI) within the next decade, aligning his views with other AI leaders like OpenAI’s Sam Altman, who cited limitations like lack of continuous learning. Addressing issues like hallucinations and misinformation remains critical for AI’s progression to human-level reasoning.
Source link
Share
Read more