Don't be fooled–LLMs don't reason

Commenters debated whether large language models genuinely reason or merely simulate reasoning. Several participants (cmiles8, zer00eyz, reliablereason) argued LLMs are fundamentally prediction engines that assemble plausible token sequences rather than hold internal beliefs or engage in principled deliberation. Others (dahart, f6v) noted humans also confabulate and produce post-hoc rationalizations, questioning whether that difference is decisive. A different set of voices (meindnoch, nodja, dataviz1000) suggested LLMs approximate reasoning by narrowing sampling spaces, using chains of thought to explore options, or internalizing reasoning during reinforcement learning; Verdex and ph4rsikal framed LLMs as capturing a kind of social or cultural intelligence or implementing global-workspace-like behavior.

Opinion split around definitions and remedies. EarthBlues drew a line between instrumental reason (LLMs may achieve) and objective reason (LLMs lack), while coreyh14444 and Kim_Bruning advocated training for more rigorous, inspectable reasoning or building smaller convergent models. rkagerer proposed prompting and subagent architectures to expose assumptions and confidence, whereas pu_pe suspected commercial motives behind prescriptive approaches. Others urged treating these systems as fast brute-force prediction tools and combining sampling, agents, or hybrid methods to get practical results. The debate centers on whether current architectures can be made to reason transparently or are intrinsically limited to useful imitation.