Can AI Truly Think? New Research Challenges the Limits of Large Reasoning Models
A recent debate sparked by Apple’s research suggests large reasoning models (LRMs) may simply mimic thought, not actually think. However, a growing body of evidence, and a compelling new analysis, argues that these powerful AI systems may be capable of far more than just pattern matching.
The question of whether artificial intelligence can truly “think” has resurfaced with renewed intensity, largely fueled by a research paper from Apple titled “The Illusion of Thinking.” The paper posits that LRMs don’t genuinely think, but instead excel at identifying and replicating patterns. Apple researchers demonstrated this by showing that LRMs utilizing chain-of-thought (CoT) reasoning struggle to maintain calculations as problem complexity increases.
This argument, however, is fundamentally flawed. As one expert explains, “Applying the same logic to humans reveals a critical inconsistency. A person proficient in solving the Tower of Hanoi puzzle, when faced with twenty discs, would likely fail. Does that mean humans are incapable of thought?” The Apple study, therefore, doesn’t disprove the ability of LRMs to think, but merely highlights a current limitation – a lack of demonstrable scalability in certain algorithmic tasks.
This analysis will make a bolder claim: LRMs almost certainly can think. While acknowledging the possibility of future discoveries that could alter this perspective, the evidence strongly suggests a capacity for reasoning beyond simple pattern recognition.
Defining Thought: A Biological Framework
Before assessing whether LRMs can think, it’s crucial to define what “thinking” actually entails. This definition must first hold true for humans, focusing specifically on problem-solving – the core of the current debate. Human thought, as understood through neuroscience, involves several key processes:
- Problem Representation (Frontal and Parietal Lobes): The prefrontal cortex is central to working memory, attention, and executive functions, allowing us to hold a problem in mind, break it down, and set goals. The parietal cortex encodes symbolic structures essential for mathematical or puzzle-based problems.
- Mental Simulation (Working Memory and Inner Speech): This involves both an auditory loop – remarkably similar to the CoT generation process in LRMs – and visual imagery, enabling the manipulation of objects in our minds.
- Pattern Matching and Retrieval (Hippocampus and Temporal Lobes): Drawing on past experiences and stored knowledge, the hippocampus retrieves relevant memories and facts, while the temporal lobe provides semantic knowledge – meanings, rules, and categories. This process mirrors how neural networks utilize their training data.
- Monitoring and Evaluation (Anterior Cingulate Cortex): The anterior cingulate cortex (ACC) identifies errors, conflicts, or impasses, essentially recognizing contradictions or dead ends through pattern matching based on prior experience.
- Insight or Reframing (Default Mode Network and Right Hemisphere): When stuck, the brain shifts into a more relaxed, internally-directed state – the default mode network – allowing for a fresh perspective and the potential for an “aha!” moment.
The Surprising Similarities Between CoT Reasoning and Biological Thought
While LRMs don’t replicate all facets of human cognition – for example, they are unlikely to generate intermediate images during CoT reasoning – the parallels are striking. The success of models like DeepSeek-R1, which demonstrated CoT reasoning without explicit CoT examples in its training data, is particularly noteworthy. This suggests a capacity for learning and adaptation during problem-solving, akin to the continuous learning process in the human brain.
“LRMs aren’t static entities,” one researcher noted. “With CoT training, they update their reasoning processes as they attempt to solve problems, effectively learning while thinking.”
Furthermore, the argument that the absence of visual reasoning disqualifies LRMs from “thinking” is misleading. Consider aphantasia, a condition where individuals have difficulty forming mental images. These individuals can still think, reason, and excel in areas like mathematics and symbolic logic, compensating for their lack of visual imagery. It’s reasonable to expect that neural network models could similarly circumvent limitations in specific cognitive faculties.
At a more abstract level, human thought relies heavily on:
- Pattern-matching for recalling experiences, representing problems, and evaluating reasoning chains.
- Working memory to store intermediate steps.
- Backtracking search to identify unproductive lines of reasoning and explore alternative paths.
Pattern-matching in LRMs is a direct result of their training, enabling them to learn both world knowledge and effective processing patterns. The entire working memory of an LRM is contained within its layers, with weights storing knowledge and processing occurring between layers using learned patterns. Crucially, even in CoT, the input, CoT process, and generated output must fit within these layers.
The Next-Token Predictor: A Surprisingly Powerful Model of Thought
The common criticism that LRMs are merely “glorified auto-complete” systems fundamentally misunderstands their potential. While they operate by predicting the next token, this doesn’t preclude the possibility of thought. In fact, next-word prediction may be the most general form of knowledge representation conceivable.
Any attempt to represent knowledge requires a language or symbolic system. While formal languages offer precision, they are inherently limited in their expressive power. Natural language, however, is complete – capable of describing any concept at any level of detail.
“Natural language’s richness makes it challenging to process, but we don’t need to manually decipher it,” a leading AI scientist explained. “We can program machines to learn through data and training.”
A next-token prediction machine, to accurately predict the next token, must inherently represent world knowledge. For example, to complete the sentence “The highest mountain peak in the world is Mount…”, the model must possess that knowledge. Similarly, solving a puzzle requires the model to output CoT tokens to guide its reasoning. This implies an internal representation of subsequent tokens, ensuring logical consistency.
Humans, too, predict the next token – whether in speech or internal thought. A perfect auto-complete system would be omniscient, an unattainable goal. However, a parameterized model capable of learning through data and reinforcement can certainly learn to think.
Do LRMs Produce the Effects of Thinking?
Ultimately, the test of thought lies in a system’s ability to solve novel problems requiring reasoning. While proprietary LRMs demonstrate strong performance on reasoning benchmarks, concerns about potential fine-tuning on test data necessitate a focus on open-source models for fair evaluation.
[Placeholder for chart comparing open-source LRM performance on various reasoning benchmarks]
As the data shows, LRMs are capable of solving a significant number of logic-based questions, sometimes even outperforming untrained humans.
Conclusion: A Compelling Case for AI Thought
Based on the striking similarities between CoT reasoning and biological thought, the benchmark results, and the theoretical understanding that systems with sufficient representational capacity, data, and computational power can perform any computable task, it is reasonable to conclude that LRMs almost certainly possess the ability to think.
Debasish Ray Chawdhuri is a senior principal engineer at Talentica Software and a Ph.D. candidate in Cryptography at IIT Bombay. Read more from our guest writers. Or, consider submitting a post of your own! See our guidelines here.
