Google DeepMind CEO says one flaw is holding AI back from reaching full AGI

Demis Hassabis, the CEO of Google DeepMind, has highlighted a critical hurdle preventing artificial intelligence from achieving Artificial General Intelligence (AGI). In a recent episode of the 'Google for Developers' podcast, he emphasized that inconsistency remains a significant challenge for advanced AI models, including Google's Gemini. Despite their ability to tackle complex tasks, Hassabis pointed out that these systems often falter on problems that a typical school student could solve. He stated, "It shouldn't be that easy for the average person to just find a trivial flaw in the system." His remarks came as he discussed Gemini models enhanced with DeepThink, a technique that boosts reasoning abilities, allowing them to secure top honors at the International Mathematical Olympiad. Yet, these same models can still make basic errors in high school-level mathematics, which he described as "uneven intelligences" or "jagged intelligences." Echoing the sentiments of Google CEO Sundar Pichai, who coined the term "Artificial Jagged Intelligence" to describe systems that perform well in certain areas but poorly in others, Hassabis underscored the need for improvements. He asserted that overcoming the inconsistencies in AI will require more than merely increasing data and computational power. Hassabis noted the necessity for enhanced reasoning, planning, and memory capabilities. Moreover, he called for improved testing methods and the establishment of more challenging benchmarks to accurately assess the strengths and weaknesses of AI models. While major tech companies like Google and OpenAI strive towards the goal of AGI, Hassabis has previously predicted that this milestone could be reached in the next five to ten years. AI systems still face challenges with inaccuracies and misinformation. OpenAI's CEO, Sam Altman, shared similar views, acknowledging that while their upcoming model, GPT-5, represents a significant leap forward, it does not yet fulfill the criteria of true AGI. He emphasized that the model lacks the capacity for independent learning, a feature he believes is essential for achieving AGI. As the landscape of AI development continues to evolve, the focus on addressing these inconsistencies will be crucial for reaching the next level of intelligence in artificial systems.