The Illusion of Speed

S omewhere , right now , in a Slack channel you'll never see, a Sr. IC is shipping a PR she'll regret by Monday.

The code looks clean. The tests pass . The model wrote most of it while she sipped her coffee. She clicks merge. She feels fast.

She spends the next four days fixing it. Subtle bugs at the boundaries. Inconsistent naming. A pattern that doesn't match the rest of the codebase. A null check that's almost right , but not quite. By Saturday morning she's on a video call walking through the diff line by line with a teammate who's trying not to look annoyed. By Saturday afternoon she's debugging a race condition in production instead of making pancakes for her kids. The pancakes happen, eventually. The bug gets fixed , eventually. By the time she closes the followup PR on Monday, she has spent four full days fixing what she shipped in twenty minutes.

Twenty minutes to ship. Four days to finish.

This is the most expensive lie AI is currently telling us, and we love the lie because it's shaped like exactly what we wanted to hear. Faster .

Here's the thing the keynote doesn't mention. Generation is fast. Completion isn't . The two were never the same , but typing speed used to be a meaningful constraint, which made them feel roughly proportional. They aren't anymore. Generation outpaces completion by orders of magnitude, and the brain registers the burst of output as progress before any of the work that matters has actually happened.

***

The 90/10 inversion.

Software has always had a 90/10 rule. Ninety percent of effort for the first ninety percent of work . The last ten percent, the edge cases and integrations and parts that don't quite match , took the remaining ten. Predictable. Plannable. Fine.

AI inverts the curve. The first ninety percent now takes ten percent of the time. Sometimes less. The dopamine is front-loaded. The last ten percent has not sped up, because taste, context, and judgment still run on human time. The cleanup just got promoted from footnote to job description.

This sounds like good news . The hard part is shorter . It isn't. The hard part is now most of the part. Most of your day is correcting, not producing, and correcting doesn't pay the same dopamine . You spend Saturday on a race condition instead of pancakes, and at the end of it you haven't built anything. You've just kept something from breaking. The work is real. The applause is missing.

***

The streaming spectacle .

It's 7 PM. You close the laptop . You try to list what you got done today. The list is shorter than you expected. You were at the desk for ten hours. You watched the token count climb. Where did the day go , you think , and then , with a small dread , I just watched .

Watching the token count climb feels like working . It isn't. You're spectating.

The model writes , you watch , the screen fills , your brain registers momentum. By the time the response finishes, you feel like you have been working for the last forty seconds. You haven't. You've been receiving output. Across a day of dozens of prompts , the cumulative effect is hours of spectating that registered as hours of working. By 7 PM you can't tell the difference.

***

The " should be done by now" delusion. A manager pings a lead engineer at 4:47 PM on a Thursday. Hey, quick one . The Transactions API fix from this morning , can you do something similar for the Accounts API ? Shouldn't take long, right?

The api fix took the original engineer twenty minutes to generate and three days to actually finish. The manager saw the twenty minutes . He didn't see the three days . To him, the work took twenty minutes, because that's when the visible part stopped. There is no graceful way to explain this in Slack without sounding like you're making excuses .

This conversation, in some form , is happening in every engineering organization right now. Managers ask. Clients ask. You ask it of yourself , the next morning, looking at a half-finished cleanup , wondering why the thing that took thirty seconds to generate is still not shipped. The real work happens in your head, while you decide whether the model's output is correct , sufficient, integrated, shippable. Invisible work doesn't get budgeted, so the speed of the visible work resets

everyone's expectations of how long the invisible work should take.

The same trick hides in latency. The thirty-second wait for a model response feels long . It is also laughably short compared to how long real thinking takes. So you don't think during the wait. You stare. You take what came out. The thinking that used to happen while you were typing now has nowhere to live.

***

The compounding traps.

You feel fast , so you commit to more . Your perceived capacity grew faster than your real capacity, and the gap shows up later, usually at quarter-end, when several things you said yes to need to ship at once and several of them aren't actually done .

PR rates go up. Review capacity does not. Reviews go shallow. Bugs slip through that wouldn't have. The team is generating more and shipping less correctly. Faster generation produces slower correct shipping .

That sentence should be on a poster in every engineering manager's office. It is the thing teams feel daily and can't quite name .

Estimates rot. Stakeholders, calibrating to the AI speed they've heard about at conferences, quote shorter timelines than the work supports. Your timeline was never typing-bound. Most knowledge work isn't. So the estimates come in short , you miss them , you look slow , the tool gets the credit , and you get the blame. Cool tool.

***

The deeper move .

Speed in AI is throughput of artifacts , not throughput of value . AI Generation outpaces judgment by orders of magnitude , and judgment is where the value sits.

The feeling of going fast is real. The velocity itself is an illusion.

The productivity story everyone is telling has confused velocity with throughput. Velocity is how fast something is moving. Throughput is how much actually gets done, shipped, alive in production at 3 AM on Sunday without a pager. AI dramatically inflates velocity . It barely changes throughput for most knowledge work, because the bottleneck was never typing speed. The bottleneck was judgment, and judgment doesn't get faster because the thing under judgment got generated more quickly.

The honest version of the AI productivity claim is narrower than the keynote. AI compresses the parts of work where typing was the bottleneck . It does not compress the parts where thinking is. For most jobs, most of the time , thinking is the bottleneck. So AI compresses a smaller fraction of your work than you were promised, and the parts it doesn't compress are now a larger share of your day. You haven't gotten slower. The work just got heavier in the part that always took the longest.

***

Twenty minutes to ship . Four days to finish.

The Sr. IC from the opening is a composite. Her Saturday is real . It's happening to someone , somewhere , right now. Could be your team. Could be the team across the hall . The pager will go off . The followup PR will get written. The pancakes will eventually get made. And the next time the AI generates a clean-looking diff in twenty minutes , she'll probably click merge again , because the speed feels good and the four days feel theoretical.

This is the texture of work now. Fast, and behind. Producing , and not finishing . Generating , and not shipping . The dopamine of writing a feature in twenty minutes , then the slow grinding work of making it actually work . Work nobody sees , because nobody is watching the cleanup.

That's the illusion of speed .

***