The hardest thing in vibe coding is finishing. I think it is much harder than in development by hand.
That is because in development you choose much more carefully what you are even doing in the first place. Much earlier before the finish you have already gone fucking insane from deadlines, budgets, etc., and because of all the above you know what counts as launch for you.
Not 100%, of course, but with a correction that is significantly, possibly hundreds of times, smaller than when, because of the zero cost of code and the absence of any need to write it with your own hands, you first go insane at the stage of choosing the scope, already exaggerating the adequate possibilities of the time segment tenfold from the start, and then broadly multiplying that number by ten a couple more times, because why the fuck not, it fucking makes itself, lmaoooooo.
The brain’s deception is that, in AI-assisted development by hand, the big thing is not easier than the small thing at all, and FINISHING is basically a whole constellation of tiny glue-joints between all the big pieces, so that they connect into a seamless, bug-free, understandable user experience that is adequate to the time/money spent.
In the process, it mistakenly seems that all these “code continents” have the main importance.
And the small things between them, well, they are not called “small” for nothing. They weigh less in code, they take less time accordingly, both for planning and for development.
But this is the most fucking brutal and most treacherous deception of perception.
It is true: one big thing was hard to build, and two big things are two or even four times harder than either one of them.
But connecting a hundred and fifty such things together stops seeming possible after countless days of attempts to do it, which produce a much more fucked-up and weak result than even the pessimistic forecast suggested.
Models are really fucking bad at helping here. They seem to get lost before all this flood of shit, and between the lines they say something like:
Bro, it is your own fault there is this much shit here, so you clean it up. Now give me a literally line-by-line instruction, and then I will at least try to do it, otherwise go fuck yourself and use all these features we spent 3 months building separately. Exactly the same way you actually developed them, you fucking dickblabber.
Look what this resident of Honduras wanted: I code everything for him like the lowest fucking laborer, and I also have to explain to him how to connect all this shit into one whole now. No way: either one or the other. And you do not know how to code, so all you have left is this, got it, you little bitch?
Got it...
Of course, I have already advanced very far in gluing all this shit together, but Fable showed how much it can increase the speed of this process through its reasoning and agent swarm, so losing Fable hit morale in a completely fucking catastrophic way.
Then I started waiting and hoping for GPT 5.6, because surely it would show something comparable to Fable in the breadth of research and the depth of understanding of exactly these tiny little fuckers, through the same virtuoso management of whole constellations of agents, but... GPT was released in some unprecedentedly fucking moronic way.
The benefit of the generally template-based thinking of the masses is that it is easier to find common ground with those same masses.
With some writer, engineer, and especially artist, who the fuck knows what is on their mind. You need a serious skillset to interest these people and hold that interest long enough.
But a simple average Joe who is satisfied with 5 categories of actions per day turns his brain on in school and right before death. You tell him some simple thought, preferably one that also criticizes somebody else, because then Joe is not only participating in the global news field, but also is not part of the condemned group, otherwise he would have kept the news quiet, and because of that he kind of rises straight to heaven inside himself.
AI is now at such an early stage of penetration into the masses that any basically mass knowledge is exclusively the echo of hordes of people who do not know what they are even talking about in principle.
This happens because people who do not understand the subject will easily believe any embellishment, a.k.a. bullshit, and bullshit at this stage of the market usually differs very strongly from reality, sometimes by 100 times and sometimes by 10,000 times. Accordingly, only bullshit gets pushed, because the truth looks too insignificant next to it to scream about, and in the general conversational background a smart person has absolutely nothing to do. He is one on one with the technology, and he should trust only himself.
Yesterday I once again talked to a buddy who is at the stage where he brags that he is neck-deep in vibe coding by telling me how last week he installed OpenClaw on a Mac mini and connected his Telegram to it =)))))))))))))))))
And now, of course, he needs no fewer than fifteen monitors so that this OpenClaw is displayed reliably enough in Telegram)))))
Fucking shame, of course... In the same category you can put the girl who, the day before yesterday, was rubbing it in my face that without a Get Shit Done skillset, using coding agents so that they actually do anything at all is impossible.
And of course, every month, in waves across the lake of public opinion, the muttering spreads about the most gigantic inflated AI bubble...
ANTHROPIC IPO VALUATION: $950B
SAMSUNG VALUATION: $850B
ANTHROPIC REVENUE: $20B
SAMSUNG REVENUE: $230B
The AI Bubble is running exact same 1999 script.
Everyone pushes this shit and goes ha ha ha, holy fuck, billionaires in the States are so STUPID, if only some old “Americans are dumb” comedian could snort about this right now...
But the truth is that Anthropic and OpenAI are most likely, at this stage, still fucking undervalued companies.
Despite the fact that I remember Samsung and its devices for as long as I have been alive, and Anthropic and OpenAI for basically no fucking time, if Samsung disappears tomorrow, I will not notice. If OpenAI disappears, I will completely lose my fucking mind.
Or you can also look at it from this angle: all the most powerful people in the world have already said how AI will multiply the economy by at least 10 times by 2040.
And the economy is, like, 200-something trillion, on paper anyway.
So then how much should a company like Anthropic be worth? =)))
The model can figure out anything together with you. You ask it to check what is going on with the new Codex features, and Codex goes into the docs and the web, then comes back with a satisfied smile.
It already knew about Caddy. About itself, it needs to look.
Accordingly, when we ask Codex to make some process for us, it makes a beautiful one, but sometimes one that is already conceptually outdated.
This is where a vibe coder needs futuristic thinking. What I mean is easiest to explain with an example:
Take the task: unique copywriting for an email warm-up service.
If we want to warm up 1,000 inboxes, then we will send 20–30 thousand emails a day to Google Workspace and Microsoft 365 / Outlook accounts, which will allow our inboxes to gain reputation in their ecosystems.
And they will actually work. I think some Jamie Dimon-type Fortune 500 bank CEO will even give an award to a pot-bellied CTO for making such a system out of nothing, with his genius alone. Like Jesus, only instead of water he turned a fucking piece of shit, and not into wine, but into diarrhea.
Suppose at that moment you get smarter and say, basically: fuck that, build me the FULL AI piece directly, why the fuck do I need your dumbass script-based little dictionaries.
It will build it, no problem. But you do not cancel that instinct with a request like that. It will build not only an AI-prompted model that will actually produce real texts, but also a system of ordinary script-based validators with the same little dictionaries, and those validators will CONTROL THE PROCESS.
Meaning AI will not control the scripts. The scripts will control AI.
This is exactly how AI thinks right now, because this is exactly how the Internet has worked for the last 30 years, and this is what the model was trained to master virtuously.
As a result, if the result really matters to you, you will keep going back and forth on this swing until you kill off all the script-based validators. The model will resist hard, but it will obey if you prompt it correctly, and what you will have left is a naked AI process, with a prompt on the input, and whatever that prompt produces on the output.
And quality control of the result is no longer a scripting task. It becomes a task of finding the balance between:
- Model type (the price per unit of its creation)
- Model settings (temperature, etc.)
- Whether you need two or three or four or ten layers of models so that one checks and improves the output of the previous one down the chain
- The prompt
- Project structure — what the model even has access to, and what handles are available to it after the work is done
The more accurate models become, the less code you need. I came to the point where some VERY IMPORTANT areas of my project consist entirely of frontend in terms of code. On the backend, there are only neural nets, and debugging is exclusively debugging the cost-to-quality ratio.
Many processes that work significantly better with deep reasoning, and that do not need an instant reaction to a button click on the frontend, are actually most profitable to wire in through a launched headless Codex session with access TO THE ENTIRE SERVER, and with a brutally polished prompt.
Frankly, I only came to this very recently. For several months, I was trying to generate some fucking insane code constructions that worked, but without AI reasoning their weakness was that somebody’s reasoning was needed for the task to be completed. And if it is not AI, then it is a human.
Very few functions, and usually very dumb functions, can be implemented reliably with scripting methods. That is basically why SaaS is so fucking awful. People are forced to use it, but nobody loves it.
Because SaaS gives you a ton of buttons, but humans still have to press those buttons. SaaS cannot think through the whole scheme without reasoning here and there, so that the automation itself makes the full required scope of decisions. So we take a human, give him 1,000 buttons, and watch him dance on them with tears in his eyes.
You do not need to build like that anymore. The product person should sit there and decide where, and how many, digital models need to think so that the output is guaranteed to be good. This is what people will pay for over the next year. Assuming, of course, that the market even manages to mature enough to form a demand for such products, and AI does not get smart enough first to start deciding for people what to do and how to do it.
And based on everything described above, I do not just believe it — I see with my own eyes how the Internet will very soon become codeless.
Code is simply not needed. The Internet will become alive, like a Zerg ship. These little fucking things will jump out of it and fuck all the Terrans in the ass.
