A recent study from Apple challenges the widespread notion of Artificial General Intelligence (AGI) by examining the capabilities of current AI models. This research highlights the significant challenges faced by Artificial Intelligence (AI) when tasked with high-complexity reasoning, undermining the belief that AI is nearing human-like logical thought processes.
According to Apple’s findings, many advanced AI models, such as the large reasoning models (LRMs) from companies like Anthropic and DeepSeek, rely heavily on pattern-matching. Essentially, they mimic the data they were trained on, rather than applying genuine logical reasoning. These models stumble significantly when confronted with complex problem-solving tasks that demand structured logic or multi-step processes.
Interestingly, Apple found that these AI systems perform well on medium-complexity tasks. However, they falter on simpler tasks compared to standard large language models (LLMs), and completely collapse on high-complexity scenarios. Such challenges reveal the intrinsic limitations of transformer-based AI, debunking some of the industry’s AGI assertions.
The study meticulously designed experiments with mathematical and puzzle challenges to thoroughly test these AI frameworks. Apple discovered that, as problems increase in complexity, the cognitive effort of these AI models paradoxically declines. This indicates limitations in the models’ reasoning effort and scaling mechanisms, unveiling constraints in their capability to handle nuanced logic.
Furthermore, the research exposes a tendency for AI to ‘overthink’ simple problems, expending unnecessary computational resources. However, on encountering complex challenges, the models fail almost entirely, hinting at fundamental barriers to the much-touted AGI.
