A Solo Builder Warns That Running a Vibe-Coded App Costs More Than Building It
Watch on TikTok
An 82-second talking-head video, shot indoors against a blind-covered window, with one static caption across all 41 sampled frames: "Vibe coding is difficult." No app, dashboard, analytics screen, or user count appears at any point. The creator never names the product. What they deliver is a fatigue testimonial: a reported 20-hour workday, five hours of sleep, a growing user base, and a warning to anyone thinking about releasing software they built with an AI assistant. The specific load the creator describes is the cost of operating software after launch, and that cost is documented in engineering practice that predates AI coding tools by many years.
What the creator actually said
The claims in this video are first-person and unverifiable. Treat them as self-reported:
- "Yesterday must have been a 20-hour workday." Self-reported.
- "I slept for like five hours." Self-reported.
- "We have a ton of users, we have people that are actually engaging with us." Self-reported, with no number given and no analytics shown on screen.
- "Look at this hair. This is indicative of a really difficult work week." Self-reported, and the frames show nothing that corroborates a work week.
The argument underneath those claims is more durable than the claims themselves. The creator says: "Normally, when they develop apps and softwares, it's a team of people. There's a reason for that, right? As a vibe coder, you are now bearing the full weight of your business, of your software." That holds. A solo operator absorbs every role a team would split.
The video's posted stats at capture: 2,740 views, 125 likes, 37 comments.
Where "vibe coding" came from, and when each dictionary picked it up
Andrej Karpathy coined the term in a post on X on February 2, 2025. The original wording:
"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."
The same post goes on to describe the practice in detail: "I 'Accept All' always, I don't read the diffs anymore," and "The code grows beyond my usual comprehension, I'd have to really read through it for a while." Karpathy's own scoping line matters here, because the video omits it: "It's not too bad for throwaway weekend projects."
Two dictionary milestones, with distinct dates that get conflated in secondhand coverage:
- Collins English Dictionary named "vibe coding" its Word of the Year on November 6, 2025, defining it as "the use of artificial intelligence prompted by natural language to write computer code."
- Merriam-Webster first flagged the term as "slang & trending" in March 2025, then added it as a full dictionary entry on September 15, 2026, in a batch of roughly 1,400 new words.
The Merriam-Webster entry is 2026, not 2025. The Word of the Year designation is Collins, not Merriam-Webster.
The post-launch load is older than the tools
The video's implicit framing is that vibe coding uniquely produces this exhaustion. That framing does not survive contact with the record. Google's Site Reliability Engineering book, chapter 5 ("Eliminating Toil," written by Vivek Rau), was published by O'Reilly years before any AI coding assistant shipped, and it exists because running software in production already consumed teams. Rau defines toil as work "tied to running a production service that tends to be manual, repetitive, automatable, tactical, devoid of enduring value, and that scales linearly as a service grows." Google's stated policy is a cap: "keeping operational work (i.e., toil) below 50% of each SRE's time."
That cap is the point. A company with dedicated reliability engineers found it necessary to formally limit how much of a specialist's week gets eaten by operations, because without a limit it consumes everything. A solo founder has no such cap and no one to hand the overflow to. Security review, on-call, support tickets, incident response, billing failures, and data handling all land on one person.
The creator's fatigue is real and the causal story needs correcting. Writing the code faster did not create the operating burden. It removed the gate that used to keep people with no operations plan from reaching a live user base in the first place.
What the evidence says about AI-assisted code and review burden
Three published measurements are worth naming, because each one describes work that arrives after the generation step:
Veracode, 2025 GenAI Code Security Report (published July 30, 2025). Over 100 large language models were run against a set of coding tasks across Java, Python, C#, and JavaScript. 45% of generated code samples introduced an OWASP Top 10 vulnerability. Java failed at 72%, C# at 45%, JavaScript at 43%, Python at 38%. The report's sharpest finding is that security outcomes stayed flat regardless of model size or training sophistication, while functional correctness improved.
METR randomized controlled trial (published July 10, 2025). 16 experienced open-source developers worked 246 real issues on repositories averaging over 22,000 stars and a million lines of code, with each issue randomly assigned to AI-allowed or AI-disallowed. Measured result: developers took 19% longer with AI tools available. They had forecast a 24% speedup beforehand and still reported a 20% speedup afterward. METR explicitly caveats that this is a snapshot of early-2025 tooling with a small sample and does not generalize to all developers.
GitClear, 2025 research on AI assistant code quality. 211 million changed lines of code from January 2020 through December 2024. Copy-pasted lines rose from 8.3% to 12.3% of changed lines. Lines associated with refactoring fell from 25% in 2021 to under 10% in 2024. For the first time in the dataset, copy-paste exceeded moved code.
None of these studies measures founder burnout. They measure the thing that generates it: output that is faster to produce and no cheaper to review, secure, and maintain.
On the burnout statistics I did not use
I looked for survey data on solo founder and indie-hacker burnout with a named source, sample size, and date. I did not find one that meets that bar. The figures circulating online range from 28% to 54% burnout among solopreneurs depending on which marketing blog you land on, with no published methodology and no consistent definition of burnout. I am leaving them out rather than lending them a citation they have not earned. The closest adjacent work is Michael Freeman's 2015 "Are Entrepreneurs Touched with Fire?" manuscript on entrepreneur mental health, and I could not retrieve the primary document to verify its sample figures, so I am not quoting numbers from it either.
A transcription artifact in the tail
The automated transcript ends with the phrase "Peace be upon you" repeated twelve times. That is a Whisper hallucination loop, not content. Two checks confirm it. The audio is 82 seconds long, and the creator is still mid-sentence in frame 41, sampled at roughly the 80-second mark. Twelve repetitions of a four-word phrase cannot fit in the remaining audio. The actual sign-off is the brief "Okay, peace" that precedes the loop. Anyone reading machine transcripts of short-form video should expect this failure mode on quiet outros, where the model has little signal and fills the gap by repeating its last confident output.
What a solo builder can do about it
The useful version of the creator's warning is narrower than "do not start vibe coding." It is closer to this: decide before you launch which operational costs you are willing to carry, and refuse the ones you are not.
- Treat the Veracode number as a review budget, not a verdict. If roughly 45% of generated samples carry an OWASP Top 10 issue, then reading diffs on anything touching auth, user data, or payments is the highest-return hour in the week.
- Pick a toil ceiling and defend it, the way the SRE book does. A number you decided in advance is easier to hold than a boundary you negotiate at 2am.
- Push the work you cannot staff onto services that already solved it: managed auth, a payment processor, a hosted database with backups, an error tracker with alerting. Every one of those is an on-call rotation you do not have to run.
- Read Karpathy's original scoping. He described vibe coding as appropriate for throwaway weekend projects. A shipped product with a real user base is a different commitment, and it was a different commitment before February 2025.
Key Takeaways
- The creator's "20-hour workday," "slept for like five hours," and "a ton of users" are self-reported and unverifiable. No user count, dashboard, or analytics screen appears in any of the 41 sampled frames.
- Andrej Karpathy coined "vibe coding" in a post on X on February 2, 2025, and scoped it in that same post to "throwaway weekend projects."
- Collins English Dictionary named it Word of the Year on November 6, 2025. Merriam-Webster flagged it as "slang & trending" in March 2025 and added a full entry on September 15, 2026. The Word of the Year designation belongs to Collins.
- Correction to the video's framing: the operating burden after launch is not specific to vibe coding. Google's SRE book caps operational toil at 50% of an engineer's time, a policy written years before AI coding assistants existed.
- Veracode found 45% of AI-generated code samples introduced an OWASP Top 10 vulnerability across more than 100 models, with no improvement as models got larger.
- METR's randomized trial of 16 developers on 246 issues measured a 19% slowdown with AI tools, against a self-reported 20% speedup. METR labels this a snapshot of early-2025 tooling.
- GitClear's analysis of 211 million changed lines found copy-pasted code rising from 8.3% to 12.3% and refactoring falling from 25% to under 10% between 2021 and 2024.
- I found no survey on solo-builder burnout with a published sample and methodology, so no burnout percentage appears in this write-up.
- The transcript's twelve repetitions of "Peace be upon you" are a Whisper hallucination on the outro. The real sign-off is "Okay, peace."
Resources
- Vibe coding (Wikipedia) Archives the full text of Karpathy's February 2, 2025 post and tracks the term's dictionary history.
- Collins' Word of the Year 2025 Collins Dictionary's own November 6, 2025 announcement, with its definition of vibe coding.
- Merriam-Webster's 2026 new word list Confirms the September 15, 2026 addition of vibe coding alongside AGI and other AI terms.
- Veracode 2025 GenAI Code Security Report 100+ models tested; 45% of samples introduced an OWASP Top 10 vulnerability.
- METR: Measuring the impact of early-2025 AI on experienced developer productivity Randomized trial, 16 developers, 246 issues, 19% measured slowdown.
- GitClear: AI assistant code quality research 211 million changed lines, 2020 through 2024, tracking duplication and refactoring trends.
- Google SRE Book, Chapter 5: Eliminating Toil Defines toil and sets the 50% ceiling on operational work per engineer.
Published October 1, 2026. Writeup generated from a favorited TikTok.