<- all tokdocs

Writing Code vs Shipping Code: What the 180% AI Productivity Figure Measures

By , founder of Mavric and Apso

View on TikTok ->

The "180% more productive" figure for AI coding tools comes from an NBER working paper by researchers at Wharton and MIT, and it counts commits pushed to GitHub. Software releases rose only 30%, and new apps did not attract more total usage. A TikTok from Helen Edwards (@stay_human25), co-founder of the Artificiality Institute, makes that point well, though the paper has since been revised and its numbers have moved.

What does the "Writing Code vs Shipping Code" paper find?

The paper is "Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools" by Mert Demirer, Leon Musolff and Liyuan Yang, NBER Working Paper 35275, issued May 2026 and revised September 2026. Demirer is at MIT Sloan and Musolff is at Wharton.

The numbers in the video match the May draft as covered by Knowledge at Wharton on 8 September 2026 and MIT Sloan on 2 September 2026. That draft studied 100,000 GitHub developers and found commit increases of 40% for autocomplete, 140% for sync agents and 180% for async agents, with releases up 30%. Murat Demirbaş's June 2026 review gives the release effects by tool generation: 10.2% for autocomplete, 20.3% for sync agents and 30% for async agents.

Did the numbers change in the revised paper?

Yes. The current abstract covers "more than 500,000 GitHub developers" and reports commit effects of 30%, 180% and 240% for autocomplete, interactive agents and autonomous agents. Projects rose 80%, and releases still rose 30%.

So the sample grew fivefold, the autocomplete effect fell from 40% to 30%, the top agent effect rose to 240%, and the release figure held steady.

Are AI coding tools making developers 180% more productive?

Not in the sense most people mean. A commit is a unit of activity, and an agent that writes more, smaller commits raises the count without producing more software. The paper's own structure treats commits as the least informative measure, with releases and usage further along.

Edwards makes that argument in the video, calling "finished software" what matters. The video's title still leads with the 180%.

Her word "tripling" is also generous. A 180% increase is a 2.8x multiplier. The revised 240% is 3.4x.

Did more apps mean more users?

No. Knowledge at Wharton reports that monthly new releases on Apple's App Store rose from roughly 30,000 before early 2025 to roughly 100,000 by April 2026, while total usage stayed flat or declined across all four marketplaces studied: the Apple App Store, Google Play, the Chrome Web Store and SourceForge. The abstract calls it "a sharp increase in the number of new apps but no increase in total usage."

The marketplace figure and the 30% release figure are different measurements. The 30% is a per-developer effect from a tracked group. The 30,000 to 100,000 jump is a raw count of all new listings from everyone. The conclusion holds either way: more listings, no more users.

Why doesn't more code mean more shipped software?

The paper estimates an elasticity of substitution of 0.23 between AI and human effort, which the authors read as strong complementarity. In plain terms, AI output still depends on human work like review and release decisions, so those become the bottleneck.

The video says "the researchers tried AI reviewing AI code." The quote behind it is real. Musolff told Knowledge at Wharton: "If the same AI that wrote the code also reviews it, that doesn't really solve the problem. The review just isn't of the same quality." That is an author's remark in an interview. The abstract and coverage do not describe an experiment testing AI review of AI code.

The AI usage data in the study comes from Microsoft telemetry, so it mostly reflects one vendor's tools.

What does randomized research say about AI and developer speed?

The NBER paper is observational, since developers chose when to adopt each tool. Randomized evidence points a different way on speed. METR's July 2025 trial assigned 246 tasks to 16 experienced open-source maintainers. With early-2025 AI tools they were 19% slower, yet estimated afterward that AI had sped them up by 20%.

METR's February 2026 follow-up found point estimates of roughly -18% for returning developers and -4% for new ones, with intervals crossing zero. METR calls those unreliable because developers refused to work without AI and submitted mostly tasks where AI would not help. Its own reading is that developers are more sped up now than the early-2025 result suggested.

The studies measure different things, task time for experts versus commit and release counts across a large panel. Both suggest that writing speed is not the constraint on shipping.

Bottom line

The 180% figure is real and counts commits. Releases rose 30% in both versions of the paper, and the extra apps did not bring more users. Use the release and usage numbers, not the commit number, when judging what AI coding tools deliver.

Sources

Published September 9, 2026. Based on a TikTok by @stay_human25. Full video notes (markdown)