Sida Liu

Large language models (LLMs) are trained by processing huge amounts of text. They “learn” by reading, much like people do through reading and listening.

When kids develop thinking skills, they first learn language, and only much later—around third grade or so—do they begin to understand math and logical reasoning. Thinking in precise, step-by-step ways is challenging for young kids and likely tied to stages of brain development.

In Daniel Kahneman’s terms, “System 1” (fast, intuitive thinking) develops earlier, while “System 2” (slow, analytical thinking) comes later and requires more effort—and, in the context of LLMs, greater computational power.

Similarly, LLMs struggle with math. While they can understand scientific facts to some degree, they aren’t yet able to apply this knowledge with the exact, logical thinking needed to solve problems accurately.

To illustrate this issue more clearly, consider how we solve algebra. We hold expressions in mind and mentally manipulate them step by step. In contrast, LLMs function more like someone who talks through each step out loud, relying on what they just said to reason through the next step. If we tried this approach ourselves, it would slow down and disrupt our ability to perform math effectively.

For now, it’s best not to rely on LLMs for math or tasks that require precise logical reasoning—they’re simply not there yet.