Apple’s paper "The Illusion of Thinking" sparked debate by arguing that LRMs, when pushed to solve larger instances of puzzles, fail because they cannot carry out algorithmic calculations. The author counters by pointing out that even humans would stumble on a Tower‑of‑Hanoi with twenty discs; thus the failure only shows a lack of evidence against thinking, not proof of its absence. By dissecting human cognition—prefrontal problem representation, inner‑speech simulation, hippocampal memory retrieval, ACC monitoring, and default‑mode insight—the article maps each component to analogous stages in LRM CoT reasoning. While LRMs lack visual imagery and real‑time feedback, their training allows internal backtracking and shortcut discovery, mirroring human back‑tracking when a line of reasoning stalls.
The piece then tackles the next‑token‑prediction myth. It explains that natural language, the medium of LRM training, is a complete expressive system capable of encoding any concept. Therefore, a model that can predict tokens with high probability implicitly stores world knowledge and can generate intermediate reasoning steps, just as a human’s inner voice does. The article argues that, given sufficient parameters, data, and computation, such a system can learn to think, not merely auto‑complete.
Finally, benchmark results for open‑source LRMs are reviewed. On logic‑based tasks, many models outperform average untrained humans and approach specialist human performance. Coupled with the structural parallels to human cognition and the theoretical universality of next‑token predictors, the author concludes that LRMs almost certainly possess the ability to think, though future research may nuance this claim. The takeaway is clear: the limitations highlighted by Apple are not evidence of non‑thinking, but of current model capacities.
Want the full story?
Read on VentureBeat →