
The overthinking of LLM
I asked Deepseek to draw the Schlegel diagram of a hexahedron — should be a very simple task right? Just 2 squares and link up their vertices. And yet I noticed something interesting: it tends to overthink and end up outputting nonsense answers when DeepThink is on (which is supposed to be more powerful and “clever”), but getting the task done correctly when DeepThink is off. Huh. So this is the prompt I gave Deepseek:...