Apple’s recent paper, “The Illusion of Thinking,” argues that LRMs lack true thought, citing failures on larger Tower‑of‑Hanoi problems as evidence. The author notes the flaw in equating a human’s inability to solve a 20‑disk puzzle with a lack of thinking; instead, it shows that even experts can stumble when tasks exceed mental capacity. The paper reframes the debate by asking what “thinking” actually entails and argues that if humans can represent problems, simulate solutions, retrieve knowledge, monitor progress, and reframe insights, then the same computational structures might exist in LRMs.
Thinking, as defined here, mirrors several brain systems: prefrontal cortex for problem representation, parietal regions for symbolic encoding, the hippocampus and temporal lobes for memory retrieval, and the anterior cingulate for error monitoring. LRMs, though not literally brain‑like, emulate these stages via pattern‑matching weights, multi‑layer working memory, and back‑tracking behaviors in CoT generation. Even without visual imagery, LRMs can construct symbolic “mental models” through textual representations, much as people with aphantasia rely on symbolic reasoning.
Empirical tests reinforce the theoretical argument. Open‑source LRMs evaluated on logic‑heavy benchmarks outperform average untrained humans and, in some cases, rival trained human baselines. The models’ ability to generate coherent CoT, recognize when a direct solution path fails, and seek alternative shortcuts mirrors human problem‑solving. These results suggest that, with sufficient training data and computational power, LRMs possess the necessary representational capacity to perform any computable reasoning task—they almost certainly can think.
Want the full story?
Read on VentureBeat →