The Paradox of Why AI Code Is Failing Us - 3 Pillars

The video’s paradox is that AI now writes code that is correct line by line, and that this is the problem. It borrows the Jevons paradox (William Stanley Jevons, The Coal Question, 1865): when a feature gets cheaper to build, more of the backlog clears the bar, so companies build more software rather than spend less on it. That only follows if demand for features is price elastic, which is the standard condition for a Jevons effect. The video asserts that elasticity; it does not measure it.

More software means systems that outgrow what a model can hold at once. The video walks through three workarounds, each with its own blind spot. A bigger context window costs more, since causal attention compares roughly n²/2 pairs of tokens, and models use the middle of a long input worse than its ends. That second point is well supported: Liu et al. named it in “Lost in the Middle” (TACL, 2024), and Chroma’s 2025 “Context Rot” report found all 18 models it tested got worse as input grew. Sparse attention saves cost by skipping pairs, so it can miss a link that sits inside the window. Retrieval finds only what the code references, so a coupling with no reference (a nightly report reading a table directly, which an AI change to discount logic silently breaks) is invisible to it. The conclusion is a loop: the tool that builds the systems grows less able to maintain them.

The loop is argued, not measured: the video cites no data on AI-written codebases. The nearest measurement is GitClear’s 2025 study of 211 million changed lines, which found copy-pasted code rising and refactoring falling over 2020 to 2024, though GitClear stops short of showing AI caused it.