Okay, I’ll admit the title is a bit of an overstatement. But once you filter out developers who aren’t using AI at all, there’s a real possibility that people with lower token usage are actually more productive.

The LOC Analogy

In the 1960s–80s, a metric called LOC (Lines of Code per Man-Month) was widely used to measure programmer productivity. It led to absurd behavior — developers avoiding libraries just to keep their line count high. Bill Gates famously said: “Measuring software progress by lines of code is like measuring aircraft construction progress by weight.”

Now replace LOC with TOC — Tokens of Code — and you have the modern equivalent.

The Context Trap

LLMs have no inherent memory. Chat apps simulate memory by stitching conversation history back into every prompt. As a session grows, the context expands — until something has to give.

The solution is context compression: summarizing accumulated history into a few sentences. For casual chat, this works fine. For coding, it’s a real problem — critical instructions you gave earlier (“don’t do it this way”) get thrown away in the compression, and the model starts making the same mistakes again.

Worse, longer contexts cause attention dilution. The model has to attend to more, spreads thin, and starts focusing on irrelevant details while losing track of what matters. A massively long context often produces worse code, not better.

The Leaderboard Problem

Climbing the Tokenmaxxing leaderboard is easy: keep loading large documents or asking sweeping, open-ended questions. But as context grows longer, the model thinks slower, attention scatter degrades code quality, and you end up in a feedback loop — worse code → more bugs → more tokens spent on fixes.

Meanwhile, developers who work in small, deliberate steps — breaking tasks down, reading each generated file, continuously refining the model’s understanding — produce cleaner code with fewer bugs, and consume far fewer tokens doing it.

Token consumption is not a measure of productivity. It may actively harm a company’s engineering culture.

Read the full article on HackerNoon: https://hackernoon.com/why-the-tokenmaxxing-leaderboard-might-be-backwards