AI News Monitor

Apple researchers have uncovered significant shortcomings in cutting-edge artificial intelligence models, suggesting that the tech industry’s relentless pursuit of ever-more-powerful AI systems may be hitting a wall.

In a study published last weekend, Apple’s team examined large reasoning models (LRMs), an advanced form of AI designed to tackle complex problems by breaking them down into smaller steps. The researchers found that these models suffered a “complete accuracy collapse” when faced with highly intricate tasks, raising concerns about their reliability and scalability.

The paper revealed that while standard AI models outperformed LRMs in simpler tasks, both types of systems failed catastrophically when problem difficulty increased. Even more troubling, as LRMs neared their breaking point, they began “reducing their reasoning effort”—a behavior the researchers called “particularly concerning.”

The study tested AI models on classic puzzles like the Tower of Hanoi and River Crossing problems. Despite early success with easier challenges, the systems faltered as complexity grew, often exploring incorrect solutions before arriving at the right one—or failing entirely. In one case, even when given a correct algorithm, the models still couldn’t solve the problem.

Gary Marcus, a prominent AI skeptic and academic, described the findings as “pretty devastating” in his Substack newsletter. He argued that the results undermine optimistic claims about large language models (LLMs) leading to artificial general intelligence (AGI)—a hypothetical stage where AI matches human cognitive abilities.

“Anybody who thinks LLMs are a direct route to the sort of AGI that could fundamentally transform society for the good is kidding themselves,” Marcus wrote.

The paper also highlighted inefficiencies in how AI models allocate computing power, wasting resources on incorrect paths before finding solutions. The researchers concluded that current approaches may have hit “fundamental barriers” in achieving generalizable reasoning—the ability to apply learned knowledge to new, broader contexts.

Andrew Rogoyski of the University of Surrey’s Institute for People-Centred AI suggested the industry might be in a “cul-de-sac,” with existing methods unable to overcome these limitations.

The study tested models from OpenAI, Google, Anthropic, and DeepSeek, though most companies declined to comment. As AI development races forward, Apple’s findings serve as a sobering reminder that today’s most advanced systems may still be far from true intelligence.