AI Models Still Far From AGI-Level Reasoning: Apple Researchers

Apple researchers are studying the reasoning abilities of advanced AI models, questioning their progress toward artificial general intelligence (AGI). They found that leading large language models (LLMs) struggle with complex reasoning, despite advancements. The researchers created puzzle games to test these models beyond standard benchmarks, focusing on their ability to think. Their findings revealed that current models can't effectively generalize reasoning and lose accuracy with increased complexity. These LLMs often exhibit inconsistent and shallow reasoning, sometimes generating correct answers initially, then failing later. The study suggests that these models may mimic reasoning patterns without truly understanding them, which is a significant obstacle for achieving AGI. The researchers' work challenges optimistic predictions about AGI's near-term availability. AGI, the pursuit of human-level intelligence in machines, is still a distant goal. OpenAI and Anthropic CEOs have expressed confidence in achieving AGI soon. The Apple researchers' research highlights the need for deeper understanding of LLM limitations. They found these models encounter fundamental barriers to generalizable reasoning.