The Cost of Calculators (and Other Helpful Crutches)

The Cost of Calculators

I grew up doing math in my head. Partly by preference, partly because in Kathmandu in the early 2000s, you didn't always have a calculator in front of you. You learned to estimate first. You felt the magnitude of a number before you computed it precisely. You developed a sixth sense for when an answer was wrong before you could articulate why. This wasn't discipline. It was the accidental formation of intuition through friction. People who rely heavily on calculators arrive at correct answers just as fast. Sometimes faster. But they struggle to explain why those answers make sense. The computation is outsourced, but so is the understanding. This isn't a moral failure. It's a tradeoff. Calculators were meant to remove friction after intuition had formed. The problem begins when the tool shows up before the intuition ever has a chance to develop. We don't use calculus in everyday life, just as football players don't bench press during a match. Those tools exist to build capacity. But if someone jumps straight to symbolic manipulation without ever developing a sense for quantities and rates of change, the math never becomes real. It stays ceremonial.

LLMs and the Outsourcing of Reasoning

Large language models introduce the same tradeoff at a much higher cognitive level. When an LLM reasons through a problem, it produces something that looks like understanding. Often, it's correct. But if the human never engages in the intermediate reasoning, never struggles with the framing, never makes the wrong abstraction first, something subtle is lost. Not intelligence. Intuition. Just as mental math trains you to develop a feel for numbers, struggling through ideas trains you to develop a feel for reasoning itself. Knowing when something makes sense. When it's incomplete. When it's elegant versus over-engineered. This isn't a skill you use daily. It's the meta-skill underneath all the other skills.

When we hand that process to a model too early, we risk becoming fluent operators of tools we don't actually understand.

Tools Should Compress Insight, Not Replace Its Formation

Calculus is powerful because it compresses deep intuition about change into a precise language. Strength training works because it builds capacity that shows up implicitly on the field. Calculators are useful because they accelerate computation once number sense already exists. LLMs should occupy the same role. They are extraordinary compression engines for insight. But compression only works if there is something meaningful to compress. If the human never develops intuition first, the output may be useful, but the understanding never compounds. What we lose is not daily competence. It's the ability to reason independently when the tool fails, misleads, or subtly hallucinates. And that cost is harder to measure than tokens or latency.

Why This Matters for Builders

When we design AI systems that infer everything instead of being told what matters, we repeat the same mistake at an architectural level. We substitute brute force context consumption for clarity. We confuse sophistication with depth. I watched this happen at EkLine. We built an AI documentation system that tried to infer everything from context. It was impressively complex. It was also outperformed by a simpler system that was told explicitly what to pay attention to. The inference engine was a calculator that had never learned arithmetic.

Intuition must be earned before abstraction is useful.

LLMs don't eliminate the need for thinking. They raise the bar for where thinking should happen. If we skip that step, we don't just build fragile systems. We slowly train ourselves out of the habit of understanding.