<- all tokdocs
2026.09.09@stay_human25

The 180% in This Video's Title Counts Commits, and the Paper Now Reports 240%

AI
Watch on TikTok

View on TikTok ->

The headline number in this video measures commits pushed to GitHub, not software that anyone uses, and the paper it comes from revised that number upward to 240% in the same month the video was posted. Helen Edwards, co-founder of the Artificiality Institute, posted this 101-second clip on 9 September 2026. She talks straight to camera against a wall of white bookshelves, four shelves deep and packed with nonfiction. The camera is mirrored, so the spines read backwards and most are unreadable, though a hot-pink copy of Who Are We Now? is legible over her left shoulder. There are no charts, no slides, and no screen recordings. A burned-in title card reads "Bottleneck moving to the edge - as we predicted" over the opening seconds, auto-captions run along the bottom for the rest, and a card reading "artificialityinstitute.org" replaces the title over the closing seconds. Everything substantive is spoken.

Where the 180% comes from

The paper is real and the video names it correctly. It 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, so "Wharton and MIT" checks out.

The numbers Helen reads out (100,000 GitHub developers, 40% for autocomplete, 140% for sync agents, 180% for async agents, 30% for releases) match the May draft as it was covered by Knowledge at Wharton on 8 September 2026 and by MIT Sloan on 2 September 2026. Murat Demirbaş's June 2026 review of the draft quotes the same figures and adds the per-generation release effects: 10.2% for autocomplete, 20.3% for sync agents, 30% for async agents. Given she posted a day after the Wharton piece, she was almost certainly working from it.

The current NBER abstract reads differently. It now says "more than 500,000 GitHub developers" and reports cumulative commit effects of 30%, 180%, and 240% for autocomplete, interactive agents, and autonomous agents. Projects come in at 80% and releases at 30%. So the sample grew fivefold, autocomplete's effect fell from 40% to 30%, the agent effects moved up, and the release figure of 30% held steady across both versions. None of that undermines the video's argument. It does mean the specific percentages she recites are already superseded, which is a normal hazard of covering a working paper in a week when it is being revised.

Commits are not output, and the video knows it

Helen's own framing is the correct one. She says "finished software, this is what matters" and puts the 30% release figure against the 180% commit figure. That is exactly the paper's structure. My objection is to the video's title and caption, which say "AI coding tools made developers up to 180% more productive." A commit is a unit of activity. An agent that writes more, smaller commits raises that number without producing more of anything. The paper's whole point is that the commit number is the least informative rung on the ladder, and the title leads with it anyway.

The spoken word "tripling" is also generous. A 180% increase is a 2.8x multiplier. Under the revised 240% figure it becomes 3.4x, so "tripling" lands closer to the newer number than the one she actually cites.

The App Store numbers are a different measurement than the 30%

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

Sequencing those two facts invites an arithmetic mistake. Going from 30,000 to 100,000 is a 233% increase. The paper's 30% release figure is a per-developer effect from a matched event study on a tracked cohort. The marketplace figure is a raw count of all new listings from everyone, including developers who were never in the sample and apps that are not tracked to any AI tool. They are not the same measurement and cannot be reconciled by subtraction. Helen does not claim they can, but she also does not separate them, and a listener will hear one continuous story.

The stronger version of the claim is the one she is actually making: more listings, no more users. That holds regardless of which denominator you pick.

The contradicting evidence the video does not mention

This paper is observational. Developers chose when to adopt each tool generation, and the authors match on observables to build a comparison group. That design cannot rule out that developers who adopt agents early were already ramping up activity for other reasons.

The randomized evidence points the other way on speed. METR's July 2025 trial randomized 246 real tasks across 16 experienced open-source maintainers working on repositories they averaged five years on. Allowing early-2025 AI tools made them 19% slower. The same developers estimated afterward that AI had sped them up by 20%. METR is explicit about the limits: 16 people, one setting, mature codebases with high review standards, and a snapshot of early-2025 capability.

METR's February 2026 follow-up re-ran the design with late-2025 tools and got point estimates that were still negative with confidence intervals crossing zero, roughly -18% for returning developers and -4% for new ones. METR itself says those numbers are unreliable because selection effects broke the design. Developers refused to work without AI, pay dropped from $150 to $50 an hour, and participants submitted mostly tasks where AI would not help. METR's own read is that developers are more sped up now than the early-2025 estimate suggested, and they are rebuilding the methodology.

The two studies are not in direct conflict. One measures wall-clock time on individual tasks for expert maintainers, and the other measures commit and release counts across a large observational panel. Both land in the same place on the question that matters here: the speed of writing is not the constraint on shipping.

What the video leaves out

The paper's central mechanism never makes it into the 101 seconds. The abstract reports an estimated elasticity of substitution of 0.23 between AI and human effort, which the authors read as strong complementarity. That is the quantitative form of Helen's argument about bottlenecks, and it is the most defensible thing in the paper because it is a structural estimate rather than a headline percentage.

The data source also goes unmentioned. The AI usage telemetry is Microsoft's, which means the tools being measured are largely one vendor's stack rather than the full market.

One claim I want to flag as looser than she states it. She says "the researchers tried AI reviewing AI code, and one of the authors said the review just isn't the same quality when the same system that wrote it checks it." The quote 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. I found no evidence in the abstract or in any of the coverage that the paper ran an experiment testing AI review of AI code, and I could not read the full PDF to check. "The researchers tried" is doing more work than the record supports.

Key Takeaways

  • The paper is real, correctly named, and correctly attributed to Wharton and MIT. The video's numbers match the May 2026 draft as covered by Knowledge at Wharton one day before posting.
  • The current NBER version reports a larger sample (500,000+ developers) and different commit effects (30%, 180%, 240%). The 30% release figure is unchanged across both versions.
  • The title's "180% more productive" measures commits. The paper's own finding is that commits are the least meaningful stage, which makes the title a strange choice for a video whose argument is exactly that.
  • The App Store figure (30,000 to 100,000 monthly releases) and the 30% release figure are different measurements on different populations. Presented back to back, they read as one number, and they are not.
  • METR's randomized trial found experienced developers were 19% slower with early-2025 AI tools while believing they were 20% faster. The video does not mention it. METR's own 2026 follow-up walks back the generalization.
  • Transcript quality: Whisper tracks the audio well on all the numbers, which is what matters most here, and it does not truncate. It garbles two lines. "None of that really worries" is missing a word, and "Now, you might, I don't find this surprising" is a fragment of whatever she actually said. The burned-in caption at 0:40 reads "The machine writes code" while Whisper produced "The machine writes this code faster than anything, or anyone can care about it," so the two disagree. The transcript also misses all on-screen text, including the opening title card and the closing URL.
  • Unverified: I could not access the full paper PDF, so every figure here comes from the NBER and RePEc abstracts plus institutional press summaries and one secondary review. I could not confirm that the paper ran any experiment on AI reviewing AI code. I did not check Helen's reference to "our own research" at Artificiality. Liyuan Yang's current affiliation appears in secondary sources as Boston University after MIT Sloan, and I did not confirm that against a primary page.

Resources

Published September 9, 2026. Writeup generated from a favorited TikTok.