The artificial intelligence community is buzzing with a seismic shift in perspective on the capabilities of large language models (LLMs), spurred by a groundbreaking research paper from Apple and a scathing critique from AI skeptic Gary Marcus.

Apple’s recent study, published on their machine learning research platform, casts serious doubt on the notion that LLMs and their successors, Large Reasoning Models (LRMs), possess genuine reasoning abilities. Instead, the researchers argue that these models rely heavily on sophisticated pattern matching, often mimicking the appearance of thought without truly understanding or reasoning through complex problems.

Gary Marcus, in his latest Substack post on Marcus on AI, amplifies this critique with a sharp takedown of the AI industry’s overblown claims. He points to Apple’s findings as a potential “knockout blow” for the hype surrounding LLMs, humorously noting that the models are so flawed in their reasoning that his name has been turned into a verb— “to Marcus” something, implying a critical dismantling of flawed AI logic. Marcus argues that the illusion of thinking peddled by these models is a dangerous misconception for industries banking on AI for decision-making.

Unpacking the Illusion of Thought

Apple’s research, detailed in their paper titled “The Illusion of Thinking,” conducted exhaustive tests on frontier language models to evaluate their supposed reasoning prowess. Their findings are stark: there is no evidence of formal reasoning in these systems. Instead, behaviors that appear thoughtful are better explained by advanced pattern recognition, honed through massive datasets rather than any inherent cognitive process.

The implications are profound for industries like finance, healthcare, and legal services, where AI is increasingly deployed for tasks requiring nuanced judgment. Apple’s researchers warn that relying on LLMs for such tasks risks catastrophic errors, as the models can produce convincingly wrong answers when faced with novel or complex scenarios outside their training data.

A Call for Reevaluation

Gary Marcus, long a thorn in the side of AI optimists, uses Apple’s findings to reiterate his call for a fundamental reevaluation of how we develop and deploy AI. Writing in Marcus on AI, he argues that the field’s obsession with scaling up models—throwing more data and compute power at the problem—ignores the deeper issue of building systems that can truly reason and adapt. He suggests that hybrid approaches, combining symbolic AI with neural networks, might be the path forward.

This isn’t merely an academic debate. Businesses investing billions in AI solutions need to heed these warnings. The illusion of reasoning could lead to flawed strategic decisions or misdiagnoses in critical applications. Marcus emphasizes that the AI community must prioritize transparency about these limitations rather than marketing LLMs as near-human intellects.

Looking Ahead with Caution

As Apple’s research ripples through the tech world, it serves as a sobering reminder of the gap between AI’s promise and reality. The industry must grapple with whether current models can evolve beyond pattern matching or if a paradigm shift is needed.

For now, the critiques from Marcus on AI and the hard data from Apple’s machine learning research platform are a clarion call for caution. Companies and developers must approach AI with eyes wide open, recognizing that the “thinking” they see may be little more than a mirage.