This week, a handwritten solution to LeetCode #73 "Set Matrix Zeroes" took about 20 minutes — in 2026, with AI coding tools now mainstream, that timing itself tells a story: the muscle memory for classic algorithm problems has not yet been replaced.
What This Is
"Set Matrix Zeroes" is LeetCode #73: given an m×n matrix, if any element is 0, set its entire row and column to 0 — in place, without allocating a new array. The trap is that zeroing during the first pass introduces new 0s, which then expand the zero range on every subsequent pass.
The standard approach is "mark first, modify later": use two arrays, or reuse the first row and first column, to record which rows and columns originally contained 0, then zero them in a second pass. The notebook we reviewed shows the O(1)-space optimized version — using the first row and column as markers, then handling those two edges last. The solution fits in under 20 lines. The core trap: distinguishing "original 0s" from "marker 0s" — the order cannot be reversed.
Industry View
The training value of hand-coding classic algorithm problems is being reassessed as AI coding assistants go mainstream.
Those who argue the practice must continue: AI excels at pattern-matching templates, but on in-place problems that demand precise state management — like using the first row and column as markers here — it still slips up. Understanding the logic of "treating edges differently from the interior" is the prerequisite for judging whether an AI's output is correct.
The opposing view is equally blunt: a growing consensus holds that in 2026 engineering practice, problems like this barely appear in production code. Forcing engineers to hand-write a classic problem in 20 minutes is, in essence, hiring 2026 engineers by 2018 standards. If AI can deliver a more reliable solution in seconds, the marginal returns of "hand-coding ability" are shrinking.
What deserves attention: the disconnect between algorithm interviews and real engineering capability is not new — but AI tools sharpen the contradiction. Do companies need people who can tell whether an AI's answer is right, or people who can simply produce the AI's answer?
Impact on Regular People
For enterprise IT: code review increasingly looks like "finding bugs in AI output" — solid grounding in data structures and edge handling now matters more than writing code itself. The thought process behind this problem (mark first, then zero, then handle edges) is a microcosm of that review capability.
For individual careers: for engineers currently job-hunting or switching roles, the training value of algorithm problems is shifting from "can you write it" to "can you judge its correctness within 30 seconds after AI produces an answer." The training focus is moving toward transfer.
For the consumer market: for non-technical managers and business owners, the specific problem is irrelevant — but the trend it surfaces is. More and more technical decisions once made by humans are becoming "humans reviewing AI output." The relevant job requirement is shifting from "knows how to write" to "knows how to review."