Tech
AI Made Us Busier, Not Faster
Picture a developer at 11 pm. There's a half-finished feature on the screen, but instead of pushing it to staging, she's refactoring the auth module — again. The old version worked fine, but the new AI-assisted one is 40% shorter and, she's convinced, more elegant. She's been at this for six hours. The customer who asked for the feature three weeks ago is still waiting. Token usage: very high. Things shipped to users: zero. This is not a story about laziness. It's a story about a trap that everyone in tech is currently inside, and almost nobody is naming correctly.
A piece from Substack's i0exception (published June 30) circulated heavily on Hacker News last week, and the comment thread that followed was the most honest 400-post thread about AI and engineering productivity I've read in 2026. The core observation is uncomfortable: across teams, token usage is up, AI spending is up, adoption is up — and the rate of shipping, the thing that actually reaches customers, hasn't moved. We got better at the easy part. The hard part didn't get better. It got more obvious.
You optimized the downhill. The uphill didn't change.
AI is brilliant at the part of engineering that used to be the annoying slog. That's the problem.
Writing boilerplate is now instant. Generating CRUD endpoints, migration files, unit test skeletons — gone in seconds. Engineers who used to groan at those tasks don't groan anymore. What do they do with the recovered time? This is where the Substack piece puts its finger on something real: they slide toward more of the same. More features, more refactors, more internal dashboards that three people will use. AI optimized the downhill slope of coding — the path of least resistance — without moving the actual constraint, which lives upstream. Review. Verification. Understanding what's even worth building. The decision about whether to ship at all.
Why it matters: when you speed up something that isn't the bottleneck, you don't go faster overall. You just produce more inventory in front of the jam. The queue doesn't clear; it grows.
Liu Weipeng called this out before AI tools existed, writing about how we optimize our studying: the effort you don't put into the right place doesn't count, no matter how hard you work. In Dark Time, he uses the phrase "optimizing the wrong thing" — grinding at a sub-problem while the real constraint goes untouched. A student who spends twelve hours color-coding notes but never once does a practice problem is in the same trap. The notes look great. The exam result doesn't.
Engineers follow the path of least resistance — always
When AI makes code generation easy, engineers do more code generation. Not because they're bad engineers. Because gravity.
A construction crew gets a power drill. It's much faster at driving screws. So they drive more screws — including a few hundred screws on the parts of the building that were fine already. The wall that's actually load-bearing, the one where the structural engineer needs to sign off, is still waiting for the structural engineer.
The Hacker News thread had a comment that's stayed with me: "AI isn't the problem. Bad decisions and poor architecture are the problem. AI just lets you feel them faster." This is exactly right and slightly incomplete. The decisions weren't necessarily bad before AI. What AI did was remove the friction that used to slow engineers down on the wrong work — friction that accidentally forced reflection. When generating a feature takes four hours of grinding, you naturally pause and ask whether the feature is worth it. When it takes forty minutes, you just do it. The friction was a disguised forcing function. We removed it, and nobody installed a replacement.
The Substack author frames this as the "easy gradient path": engineers drift toward work where AI helps most, not work that matters most. More feature additions. More refactors of things that already work. More tooling and internal dashboards. All low-ROI. All measurable in tokens. None of it is the hard conversation with a customer you've been avoiding.
The constraint just moved upstream. It didn't disappear.
Faster code generation moved the jam from writing to reviewing, from building to deciding — from easy to hard.
Liu Weipeng has a clean way to think about this in Dark Time: true investment = time × efficiency directed at the right place. Optimizing input speed on the wrong segment gives you the illusion of productivity. The engineers generating more code are working hard — genuinely hard — but they're adding inventory upstream of a constraint, not clearing it. This is what the Substack piece calls the productivity trap, and it's invisible from the inside because every metric except the one that matters (value in customers' hands) is pointing up.
What the engineers getting most from AI actually do
They don't use AI to go faster at code. They use it to spend less time on code — so they can focus on what code can't do.
The Hacker News thread has a quiet cluster of engineers who report genuinely shipping more. Their pattern is consistent. They set goals before they open an editor. They treat code as a signal, not an endpoint — something you write to run against reality and then, often, throw away. And they discard code ruthlessly, because they know a feature that doesn't reach a customer is just warm inventory. The Substack post's framing for this: the real leverage is shortening the cycle to customer reaction, not making the existing cycle faster.
AI is best used as a reason to skip code you don't need, not as a reason to write more of the code you were already going to write anyway.
This maps cleanly onto what Dark Time says about time use: the engineers who win aren't the ones with the most hours logged or the highest token counts. They're the ones who are honest about where the actual constraint lives. When Liu writes about students who spend entire nights studying and still fail, he's describing the same trap — an enormous amount of sincere effort aimed at something other than the blocking problem. You can be exhausted and wrong at the same time. AI doesn't fix that. If anything, it makes the exhaustion arrive faster.
The line worth keeping
The bottleneck was never writing the code. It was knowing what code to write, shipping it before it went stale, and listening to what happened. AI didn't fix that. It just made it cheaper to miss it.
Framework from Liu Weipeng, Dark Time (暗时间) — true investment / optimizing the wrong thing / effort × direction. Reporting on i0exception's "The AI Productivity Trap" (Substack, 2026-06-30) and the associated Hacker News discussion. Figures on token usage, AI spend, and adoption growth are industry-reported directional trends, not audited data. Commentary, not professional advice.