Induction, not tool use, as the claimed divide between reasoning models and base LLMs

A post argues that modern reasoning models predict instructions or a reasoning chain that produces an answer, and that this shift—not symbolic tool use—is what distinguishes them from base LLMs. It claims the shift gives reasoning models substantial fluid intelligence while base LLMs have almost none. The post says similarly sized or smaller reasoning models saturated the ARC 1 benchmark in 2025, while LLM performance remained around 10–15% as of October 2026.