<- all tokdocs
2026.10.12@alberta.nyc

@alberta.nyc's engineer holds a four-month estimate through every AI pitch; the four months is unsourced, but the only randomized trial on this question found AI made experienced developers 19 percent slower, not faster

AI
Watch on TikTok

View on TikTok ->

The artifact is a 23.87 s vertical sketch, 1080x1920 at 29.97 fps (r_frame_rate 2997/100), delivered as format id bytevc1_1080p_1573096-1 in a QuickTime/MOV container that ffprobe reports as mov,mp4,m4a,3gp,3g2,mj2 with two streams: HEVC Main profile, yuv420p, SDR, 715 frames at 1,502,408 bps, and HE-AACv2 audio at 44,100 Hz stereo, 518 frames, 64,492 bps. Overall bitrate is 1,573,163 bps and the file is 4,694,318 bytes, or 4.48 MiB. The video stream runs 23.857191 s and the container and audio stream both run 23.871995 s. The post timestamp is epoch 1791813750, which is 2026-10-12T14:02:30Z. The uploader is @alberta.nyc, uploader id 6757499075077555206, channel nickname "Alberta Tech". Audio is credited to track "original sound" by artist "Alberta Tech", so there is no licensed music bed. The description is 63 characters and duplicates the title exactly, including the trailing space: "And if you disagree with 4 months, you don't work in corporate ". At capture time, the mtime of metadata.json, 2026-10-19 00:30:25 local, the counts stood at 500,200 views, 57,600 likes, 786 comments, 2,647 reposts and 2,708 saves. I read all twelve extracted frames, frame_001 through frame_012, sampled at 0.5 fps and therefore covering roughly 0 s to 22 s, and the transcript is 119 words across 19 Whisper segments in English. Frames 001 and 002 show the author in a light wash denim jacket over a cream ribbed mock-neck, clear acrylic glasses, seated in a grey fabric cubicle with a wall clock, a manila folder wall rack and a blank paper calendar grid pinned behind her; a black sticker caption across the upper third reads "If engineers gave accurate eng estimate" and the burned-in subtitle at 0 s reads "it would take to build". By frame_003 at roughly 4 s the character has changed to a mauve linen button-down with pink-streaked hair worn down, and the subtitle reads "Still 4 months." with a numeral, which is where the on-screen caption and the Whisper transcript diverge, since Whisper wrote "Still four months." The sticker caption is present in frames 001, 002 and 003 and gone from frame_004 onward. The two characters are separated by a horizontal flip rather than a cut to a second person: the wall clock sits in the upper left in the manager frames 001, 002, 006, 008 and 011 and in the upper right in the engineer frames 003, 004, 005, 007, 009, 010 and 012, and the clock numerals read mirror-reversed in every single frame. The remaining subtitles I read verbatim are "Oh, that would be cool." (004), "No, it'll still" (005), "build the whole thing" (006), "Then it will take" (007), "it's an experienced" (008), "Then it will be an" (009), "to clean up." (010), "without looking at it?" (011) and "will be broken." (012). No terminal, editor, dashboard, chart, benchmark table, citation card or working app appears in any of the twelve frames, and no source is shown or named on screen for the four-month number at any point. The spoken argument is a two-hander: a product manager asks how long an Android version of the app would take, the engineer answers four months, and the engineer then holds four months against "what if you used AI", raises it to six months for "let the AI build the whole thing and then just kind of clean up any issues", raises it to a year for "tell the AI it's an experienced software engineer", and closes by saying the app will be broken if you ship unreviewed output. On the transcript-end check, the final segment is " Hmm." timestamped 23.240 s to 24.080 s, which runs 0.152 s past the 23.928163 s duration of audio.mp3 and 0.208 s past the 23.871995 s in-container audio stream; that is a boundary overrun of one sixth of a second on the segment end marker rather than invented sentences, and no segment text after the audible dialogue was fabricated.

The four-month figure is unsourced, and the closest real benchmark is an overrun rate rather than a duration

At 2.52 s the engineer says "Four months." No source is spoken, and none of the twelve frames shows a roadmap, a sizing document, a JIRA board or any artifact the figure could be read from. The video's own description leans on social proof instead of evidence: "And if you disagree with 4 months, you don't work in corporate." That is a comedy premise, not a claim with a provenance, and I am labelling the four months Unverified.

There is no published industry statistic for how long an Android port of an unspecified app takes, because the quantity depends entirely on the app. What the literature does measure is how far estimates miss. Magne Jørgensen and Kjetil Moløkken-Østvold reviewed that question in "How Large Are Software Cost Overruns? A Review of the 1994 CHAOS Report", published in Information and Software Technology volume 48 issue 4 in April 2006. Their paper takes apart the Standish Group's widely quoted 189 percent average cost overrun and writes: "All surveys of cost estimation accuracy in the relevant period and countries that we were able to find suggest average cost overrun in the range of about 30%." Their Table 1 lists the three surveys that support it: Jenkins 1984, 23 software organizations in the USA, 34 percent; Phan 1988, 191 software projects in the USA, 33 percent; Bergeron 1992, 89 software projects in Canada, 33 percent.

The same paper explains why the 189 percent figure should not be used as the benchmark it is usually treated as. The authors asked the Standish Group how projects were sampled and record the answer: "The response to our research method question was that providing this type of information would be like giving away their business for free." They then quote version 3.0 of the CHAOS report on its own selection process: "We then called and mailed a number of confidential surveys to a random sample of top IT executives, asking them to share failure stories." A sample built by asking for failure stories is not a sample of projects.

One benchmark was suggested to me as a fallback for this video and it does not fit. The Bureau of Labor Statistics JOLTS program measures job openings on the last business day of the month, hires for the entire month, and separations split into quits, layoffs and discharges, and other separations, across a sample of roughly 21,000 establishments. The BLS JOLTS FAQ contains no time-to-hire or time-to-fill measure at all. JOLTS is the wrong instrument here in two ways: it measures hiring flows rather than engineering durations, and this video is about a build estimate, not a recruiting timeline.

"Still four months" is the one line in the sketch that matches a randomized controlled trial

At 3.36 s the manager asks "What if you used AI?" and at 4.52 s the engineer answers "Still four months." This is the strongest claim in the video and the only one with a direct experimental analogue.

Joel Becker, Nate Rush, Elizabeth Barnes and David Rein published "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" to arXiv on 12 July 2025, revised 25 July 2025, as arXiv:2507.09089. The abstract states the design and the result: "16 developers with moderate AI experience complete 246 tasks in mature projects on which they have an average of 5 years of prior experience. Each task is randomly assigned to allow or disallow usage of early 2025 AI tools. When AI tools are allowed, developers primarily use Cursor Pro, a popular code editor, and Claude 3.5/3.7 Sonnet." The outcome: "Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%. Surprisingly, we find that allowing AI actually increases completion time by 19%--AI tooling slowed developers down."

The sketch's engineer is conservative by comparison. She says the estimate does not move. The trial found the estimate should move in the wrong direction for exactly the population she represents: an experienced engineer working in a codebase she already knows. Two limits on the transfer are worth stating. The trial covers 16 developers on open-source issues in repositories averaging five years of their own history, not a greenfield Android port, and it measures tools at the February to June 2025 frontier. Neither limit rescues the manager's position, because the direction of the only randomized evidence runs against a speedup rather than toward one.

The six-month number for "let the AI build it and clean up" is unsourced, but the cleanup tax it describes is measured

At 10.74 s the manager proposes "What if you let the AI build the whole thing and then just kind of clean up any issues?" and at 12.72 s the engineer answers "Then it will take six months." The 1.5x multiplier has no source in the video and I am labelling it Unverified. The existence of a cleanup tax is a different matter and is measured in two places.

The 2025 Stack Overflow Developer Survey asked developers about their frustrations with AI tooling. Across 31,476 respondents on that question, 66 percent selected "AI solutions that are almost right, but not quite" and 45.2 percent selected that debugging AI-generated code is more time-consuming. The same survey's accuracy-trust question, answered by 33,244 respondents, breaks down as 3.1 percent highly trust, 29.6 percent somewhat trust, 26.1 percent somewhat distrust and 19.6 percent highly distrust the accuracy of AI output.

GitClear's "The Maintainability Gap: 2026 AI Code Quality Research", published January 2026 over 623 million analyzed changes from 2023 to 2026, measures the shape of the mess rather than the hours it costs. Its headline movements are copy/paste up 41 percent, block duplication up 81 percent, from 40.3 to 73.0 duplicated blocks per million changed lines, error-masking constructs up 47 percent, two-week code churn up 15 percent, cross-file function calls down 35 percent, refactoring line moves down 70 percent, and refactoring down to 3.8 percent of changed lines year to date in 2026 from 21 percent in 2022. The report's own framing is "The throughput is real, but so is the debt it accrues."

None of that produces a six-month number from a four-month baseline. It establishes that generated code arriving faster and being refactored less is a documented pattern, and that the cleanup the manager waves off as "just kind of clean up any issues" is the part developers report as the expensive half.

Telling the AI it is an experienced engineer does not reliably produce an over-engineered mess; the research says it reliably produces nothing

At 13.72 s the manager asks "What if you tell the AI it's an experienced software engineer?" and at 16.76 s the engineer answers "Then it will be an over-engineered mess that will take a year to clean up." This is the one place where the sketch is wrong in a specific, checkable way, and the correction cuts in the joke's favour rather than against it.

Mingqian Zheng, Jiaxin Pei, Lajanugen Logeswaran, Moontae Lee and David Jurgens published "When 'A Helpful Assistant' Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models" to arXiv on 16 November 2023, as arXiv:2311.10054. Their design: "We curate a list of 162 roles covering 6 types of interpersonal relationships and 8 domains of expertise. Through extensive analysis of 4 popular families of LLMs and 2,410 factual questions, we demonstrate that adding personas in system prompts does not improve model performance across a range of questions compared to the control setting where no persona is added." Their closing judgement is blunt: "while adding a persona may lead to performance gains in certain settings, the effect of each persona can be largely random."

The video's premise is that the persona string is a lever that pushes output in a predictable direction, toward over-engineering. The measured finding is that the persona string is not a reliable lever in either direction. The joke lands either way, since a manager reaching for a magic prompt is the target of the bit, but the stated mechanism is not what the paper found. Two scope notes: the paper measures factual question accuracy across 2,410 questions, not code architecture quality, and it predates the agentic coding tools the video is implicitly about, so it does not rule out that persona text shifts code style. It does rule out treating "tell it it's a senior engineer" as a known, directional effect.

"Ship what the AI builds without looking at it" and the one finding that is now industry-wide

At 19.20 s the manager asks "What if you just ship what the AI builds without looking at it?" and at 21.82 s the engineer answers "Then the app will be broken." The 2025 DORA State of AI-assisted Software Development report, published 23 September 2025 across nearly 5,000 technology professionals and more than 100 hours of qualitative data, measures the industry version of that exchange.

Three of its findings bear directly on the line. First, adoption is near-universal: "90% of survey respondents report using AI at work. More than 80% believe it has increased their productivity." Second, trust is not: "30% report little or no trust in the code generated by AI." Third, and most relevant, the report separates speed from stability. On throughput it reports "Unlike last year, we observe a positive relationship between AI adoption on both software delivery throughput and product performance," and immediately after: "However, AI adoption does continue to have a negative relationship with software delivery stability."

The report's mechanism matches the engineer's answer: "AI accelerates software development, but that acceleration can expose weaknesses downstream. Without robust control systems, like strong automated testing, mature version control practices, and fast feedback loops, an increase in change volume leads to instability." Its summary framing is "AI doesn't fix a team; it amplifies what's already there."

Note what this does to the sketch's internal logic. DORA finds a positive throughput relationship in 2025 where the prior year did not, which means the manager's hope for a faster delivery pipeline has some support at the organizational level even though METR's randomized evidence found no speedup at the individual-task level. The engineer's flat "still four months" is right about her own task time and incomplete about the organization's delivery rate. What survives both measurements is her last answer: the acceleration shows up as instability when nobody reviews the output.

Key Takeaways

  • Unverified: The four-month estimate at 2.52 s has no source in the audio, the description or any of the twelve frames. It is a comedy premise, and no published statistic measures how long an Android port of an unspecified app takes.
  • Clarification: The nearest sourced benchmark for estimation is an overrun rate, not a duration. Jørgensen and Moløkken-Østvold (2006) found average cost overrun "in the range of about 30%" across Jenkins 1984 at 34 percent, Phan 1988 at 33 percent and Bergeron 1992 at 33 percent, and showed the Standish 189 percent figure came from a survey that asked executives "to share failure stories."
  • Context: BLS JOLTS is the wrong yardstick for this video. The JOLTS FAQ lists job openings, hires, quits, layoffs and discharges, and other separations across roughly 21,000 establishments, and contains no time-to-hire or time-to-fill measure. It also measures recruiting, while the sketch is about a build estimate.
  • Verified: "Still four months" at 4.52 s is consistent with the only randomized trial on the question. METR's arXiv:2507.09089, submitted 12 July 2025, found 16 experienced developers across 246 tasks were 19 percent slower with early-2025 AI tools, against a 24 percent forecast speedup and a 20 percent self-reported speedup.
  • Partial correction: The six-month figure at 12.72 s for build-then-cleanup is unsourced. The cleanup cost is real and measured: 66 percent of 31,476 Stack Overflow 2025 respondents named "AI solutions that are almost right, but not quite" and 45.2 percent named debugging AI-generated code as more time-consuming, while GitClear's January 2026 report over 623 million changes recorded block duplication up 81 percent and refactoring down to 3.8 percent of changed lines from 21 percent in 2022. Neither source yields a 1.5x multiplier.
  • Correction: The video says at 16.76 s that telling the AI it is an experienced software engineer produces "an over-engineered mess." Zheng et al., arXiv:2311.10054, tested 162 personas across 4 LLM families and 2,410 questions and found adding a persona "does not improve model performance" and that "the effect of each persona can be largely random." The persona string is not a reliable lever in the direction the joke claims, and the paper measures factual accuracy rather than code architecture.
  • Verified: "Then the app will be broken" at 21.82 s matches the 2025 DORA report, published 23 September 2025 across nearly 5,000 respondents, which reports a positive relationship between AI adoption and throughput and that "AI adoption does continue to have a negative relationship with software delivery stability," with 30 percent reporting little or no trust in AI-generated code.
  • Clarification: The sketch's two characters are the same person separated by a horizontal flip, not a second performer. The wall clock sits upper left in the manager frames and upper right in the engineer frames, and its numerals read mirror-reversed in all twelve frames I examined.
  • Context: The on-screen subtitle at roughly 4 s reads "Still 4 months." with a numeral, while Whisper transcribed "Still four months." The sticker caption "If engineers gave accurate eng estimate" is present only in frames 001 through 003 and gone from frame_004 onward.
  • Context: The final Whisper segment " Hmm." ends at 24.080 s, which is 0.152 s past the 23.928163 s audio.mp3 duration and 0.208 s past the 23.871995 s in-container audio stream. That is a segment-boundary overrun, not fabricated text.

Resources

Published October 12, 2026. Writeup generated from a favorited TikTok.