Everyone says they're faster
A claim that costs nothing to make and can't be checked, plus the part of the cycle nobody counts.
Standup, a few weeks ago. Somebody said a task that would have taken two days took four hours. Somebody else said they’re “probably three times quicker on anything greenfield.” A third person nodded and said something about a tenfold difference on boilerplate.
Nobody had measured anything. Nobody was lying either. And I sat there deciding not to say the thing I actually think, which is that after 14 months of using these tools every working day I have no idea whether I’m faster, and I don’t believe anybody else does either.
Nobody counts the review
Here’s the mechanical problem with all the numbers in circulation.
The thing that got dramatically quicker is generation. Producing the first version of something went from an hour to 90 seconds, and that’s real and it’s felt immediately and it’s the part people report.
The cycle, though, is generate, read, evaluate, correct, integrate. And the reading and evaluating are not faster. If anything they’re slower than they used to be, because reviewing plausible code that somebody else produced requires you to build your own model of what it should do and then check theirs against it, whereas writing it yourself only requires the first half.
I timed myself properly on two tasks in March, because I was getting annoyed about this. On one, a form with some awkward validation, the whole cycle was maybe 40% quicker. On the other, a piece of pricing logic, it was slower, because I spent an hour on a subtle wrongness that I’d have spotted immediately if I’d typed it.
Two data points, which is nothing. It’s still two more than the standup had.
Generation got faster. Generation was never the bottleneck. The bottleneck was deciding what correct means and checking whether you've got it.
The claim is costless and can’t be checked
Now the social mechanism, which is the bit I actually want to write about.
Saying “this made me three times faster” costs nothing. It can’t be falsified, because nobody logs their own throughput, and it signals that you’re capable and adaptable and on top of things.
Saying “honestly, I’m not sure it’s helped much” costs something. It sounds like you’re either behind or resistant, and both of those are bad things to sound like in 2025.
So the distribution of statements that get made is skewed upward, hard. And then everybody calibrates against the statements, which are the only data available, and concludes that everyone else has cracked something they haven’t.
Which is exactly the mechanism I wrote about in 2018 with contribution graphs. A visible signal that looks like a measure of competence, is actually a measure of something else, and produces a quiet anxiety in everyone who compares themselves to it. The green squares measured what kind of employment you had. This measures your willingness to make an unfalsifiable claim.
And it’s already been spent
The part that worries me more than the standup.
Organizations have been told, by vendors and by consultancies and by each other, that there’s a large productivity gain available. Some of them have planned on that basis: headcount that isn’t being replaced, timelines that assume it, hiring that’s been paused because a smaller team is now supposedly sufficient.
Which means the gain has been budgeted before it was measured. And if it turns out to be smaller than the projection, the shortfall doesn’t get reported as a bad forecast. It gets absorbed by the people doing the work, in the form of a workload that assumes a speed they haven’t got.
I’ve seen that shape before. It’s the 2021 story from the other end: fewer people, same clients, same deadlines, absorption that’s invisible because everyone’s competent and doesn’t complain immediately.
What I actually notice
For balance, and because I use these things constantly and would be annoyed if somebody read this as a rejection.
Genuinely, substantially better: anything I know exactly how to do and don’t want to type. Explaining an unfamiliar error. Enumerating edge cases I’d have missed. First drafts of tedious configuration. Reading an unfamiliar codebase, which has turned out to be the biggest one for me and the one nobody talks about.
Genuinely worse: anything where the failure is subtle. The confident wrong answer costs more than no answer, because no answer sends me to the documentation and a wrong answer sends me down a road.
And a third category I don’t know how to score: I take on things I’d previously have declined or deferred. The print stylesheet, the accessibility pass, the small refactor. Not because they’re faster exactly, but because the activation energy is lower and I start. That might be the real change and it doesn’t show up as a multiple.
Anyway, what I’d like, and I’m aware this is a modest ambition, is for it to be socially acceptable in a standup to say “I don’t know.” Not as a criticism of the tools. Just as an accurate report of the state of the evidence, which is that a massive change to how a whole industry works is 14 months old and almost nobody has measured anything at all.
If you’re reading this and quietly wondering why you don’t feel ten times faster: you’re probably fine. I don’t either, and neither, I suspect, does the person who said it 🥲