AI more than tripled the code. Releases rose 30%. Usage didn't move. Where's your bottleneck now?
A new study puts numbers on something a lot of us feel: writing code got cheap, and almost nothing after it got any faster. I'd like to hear where the slowdown sits for your team.
I've been reading "Writing Code vs. Shipping Code," an NBER working paper by Mert Demirer (MIT Sloan), Leon Musolff (University of Pennsylvania) and Liyuan Yang. It follows GitHub developers before and after they adopted three generations of AI coding tools: autocomplete, interactive agents that edit alongside you, and autonomous agents that work on their own.
The headline from the revised version on NBER's site (September 2026): cumulative increases in commits of about 30%, 180% and 240% across those three generations. Then the gains shrink at every step down the line. The 240% in commits becomes about 80% more projects and about 30% more actual releases. An earlier May draft had different commit numbers, but the same shape and the same roughly 30% on releases.
The part that stuck with me is the app stores. The authors looked at four marketplaces (Apple's App Store, Google Play, the Chrome Web Store and SourceForge) and found a sharp rise in new apps, but no increase in total usage. More things shipped. People didn't use more things.
Their explanation is what they call a weak-link problem: AI sped up the parts it can do, and the human-led stages, such as reviewing changes, deciding what to release and getting it out the door, became the constraint. They estimate AI and human effort are strong complements, not substitutes.
One caveat worth saying plainly: it's a working paper, not peer-reviewed yet, and two of the authors disclose they consult for Microsoft, which owns GitHub, where the data comes from. I don't think that undermines a finding that is fairly unflattering to AI coding hype, but you should know it.
Here's why it lands for me. When AI tools made building faster at Murror, the honest result wasn't "we ship three times as much." It was that the slow parts moved. Deciding what not to build, reviewing a change that touches how someone feels about their day, writing release notes a person can trust, reading what users actually say afterwards. None of that got faster, and I'm not sure I want all of it to.
And the app store finding is the uncomfortable one for makers. If everyone can ship more, shipping more stops being an advantage. The bottleneck moves from "can we build it" to "does anyone need it, and will they find it."
So I'd love to hear from other makers, and please answer each other, not just me:
Where does work pile up for you now that the code comes fast? Review, QA, design decisions, launch, distribution, something else?
Has AI changed how much you ship, or only how much you start?
And if you've found something that unblocked the human part, a habit, a rule, a tool, what was it?


Replies
for us its review for sure, agents open way more prs than anyone can properly read. what helped a bit was making each pr do one thing and saying no to stuff before it gets built, not after