This post is also available in:
עברית (Hebrew)
A recent study by Apple researchers has raised serious doubts about the capabilities of today’s most advanced AI systems, particularly those designed for complex reasoning. The findings suggest that current-generation large reasoning models (LRMs) are still far from achieving the kind of general intelligence often associated with artificial general intelligence (AGI).
In their technical paper, titled “The Illusion of Thinking: Understanding the Strength and Limitations of Reasoning Models via the Lens of Problem Complexity,” Apple tested leading LRMs such as OpenAI’s O1/o3, Claude 3.7 Sonnet Thinking, DeepSeek-R1, and Google’s Gemini Thinking. Instead of traditional AI benchmarks, the researchers subjected these models to a progression of increasingly complex and non-standard problems designed to test their reasoning ability in unfamiliar contexts.
The results were stark: although LRMs perform reasonably well on simpler tasks, their accuracy collapsed when faced with problems of growing complexity. Apple’s team concluded that these systems—despite being branded as “thinking” models—fail to develop true general problem-solving skills. Instead, they tend to reproduce learned patterns without generalizing their reasoning strategies across new scenarios.
“Accuracy ultimately collapses to zero beyond certain complexities,” the researchers noted, arguing that these models merely mimic reasoning rather than engaging in it. This finding runs counter to recent claims from some AI industry leaders, including OpenAI and Anthropic, who have publicly stated that AGI could be achieved within the next few years.
From a technological perspective, the implications are significant. The study points to a structural limitation in how LRMs are built and trained, raising concerns about over-reliance on large language models for tasks requiring robust, adaptive reasoning.
Critics within the AI research community have echoed Apple’s conclusions, warning that LRMs may not be the pathway to AGI that some have envisioned. For now, it appears that the leap from impressive pattern recognition to true intelligence remains a work in progress—reminding technologists and policymakers alike that claims of near-term AGI may be premature.
