The Future of Vibe Coding, as I See It
language
eng
date
May 6, 2026
slug
future-of-vibe-coding
author
status
Public
tags
AI
Vibe Coding
LLM
AI Agent
summary
Vibe coding makes it seem as if anyone can build software now, but beneath that promise lie new constraints: token costs, platform moderation, market concentration, and relentless quality competition. This essay looks at the future of vibe coding from both optimistic and skeptical angles, and argues that what we really need is not a fantasy of easy money but the ability to adapt to the AI era.
type
Post
thumbnail
category
😎 Daily
updatedAt
May 9, 2026 07:05 AM
Because I set AI and software development as my main interests on Threads, a large share of the posts I see now are about vibe coding. The spectrum is wide: provocative claims that vibe coding can lead to effortless monthly income, reviews about which model or tool is better, and even recycled AI news repeated by multiple accounts in a single day. I also use many different AI models and tools in my own workflow, and I actively rely on them when building real products. So the question is worth asking seriously: where is this wave of vibe coding actually headed?
Vibe coding is already evolving into the AI agent era
One thing is unmistakable: model capability and tool usability are improving at extraordinary speed. Anthropic's Claude Code and OpenAI's Codex keep leapfrogging each other, while China is releasing open-weight models that are only a step behind the frontier. Tooling is getting better almost every week. At this point, these products feel less like coding assistants and more like AI agents. Features that once looked like flashy demos—such as browser control—have quickly become standard, and the frontier is now moving beyond the browser toward full-PC interaction, code debugging, and even non-programming tasks. In other words, vibe coding is no longer just about generating code; it is becoming part of a broader AI agent experience.
The boom is real, but so is the overheating
A world where anyone can build the software they want sounds like a blessing. On YouTube and social platforms, people constantly tell us to jump into vibe coding right now, while leaders at NVIDIA, Anthropic, and OpenAI talk about solo founders and human managers who can suddenly wield the productivity of entire teams. But these are also the people who benefit when clicks, courses, and token consumption rise. This mix of genuine innovation and exaggerated promise is not new.
We saw the same pattern during the rise of the internet and smartphones. At first, these technologies looked like universal opportunity machines. Then came bubbles, oversupply, spam, fraud, phishing, and a long adjustment period before the tools became meaningfully integrated into everyday life. Vibe coding may be on a similar path. Right now, it feels less like the final form of the technology and more like the phase where an important tool is being consumed, copied, and wasted at scale.
There are four reasons I cannot view the future of vibe coding with pure optimism.
Token usage does not automatically translate into knowledge growth
When the internet first became mainstream, many people believed that once all information was online, humanity would naturally become wiser. That did not happen. Information volume grew exponentially, but human understanding did not. In practice, knowledge accumulation looked closer to a logarithmic curve. When information becomes too abundant, people do not necessarily become more open-minded; they often become more selective, more closed off, and more vulnerable to bias. Low-quality data and false information scale just as easily as good information. The AI era may follow the same rule: more tokens consumed does not necessarily mean deeper learning or better judgment.
Tokens may not remain cheap air forever
If corporations still exist to maximize shareholder value, the endgame of the AI model race is not hard to imagine. Eventually, only one or two dominant players may remain. If that happens, will they keep offering generous subscription plans with effectively abundant usage? Or will they raise prices, remove flat-rate subscriptions, and move fully toward usage-based pricing in order to recover infrastructure costs faster? If that shift happens, the promise that anyone can become a builder will quickly thin out. In the world of vibe coding, tokens are the new air—and when air becomes expensive, many builders will struggle to keep breathing.
The geopolitical dimension makes this even more serious. If the world ends up relying on just a few countries for token supply, that is not so different from relying on a few countries for food. Diplomatic, political, or military conflicts could easily translate into supply constraints. Just as modern life becomes fragile when electricity or the internet disappears, an AI-dependent world becomes fragile when token access is disrupted.
LLMs do not automatically produce new algorithms or new languages
Many people still describe LLMs as probability machines that reproduce the statistical average. That description is incomplete, but not entirely wrong. Can thousands of lines of instantly generated code, produced with minimal human intervention, reliably lead to genuinely new algorithms, legendary implementations, or fundamentally new patterns? So far, we have not seen many convincing examples. For now, LLMs seem much better at producing ordinary software faster and at larger volume than at generating truly original breakthroughs.
There is still an interesting shift here. Languages like Rust, for example, have benefited from AI assistance because developers no longer need to master every detail of the language before becoming productive with it. That suggests LLMs may first have their biggest impact by lowering the activation energy of existing languages and tools, rather than by inventing entirely new ones.
Truly finished replacements are still rare
If vibe coding alone can already build almost anything, why have we not yet seen a wave of polished 1:1 replacements for major paid applications and services? Even products like MS Office or Adobe software—which do not necessarily require massive cloud-scale infrastructure—still face little direct pressure from vibe-coded challengers. Open-source alternatives have certainly improved, but relatively few have reached the level of durable commercial products. That suggests that, at its current level, LLM-assisted vibe coding is excellent for prototyping and productivity, but still struggles when the goal is to scale, maintain, and finish truly complex products.
So what should we do now? Adapt
This is not an anti-vibe-coding essay. On the contrary, I think this trend is something people in software-related fields must learn by doing. It is also a rare opportunity for non-developers to build their own tools and services. But it is risky to assume that the window will stay open forever, or to approach it only with fantasies of easy money.
What we need now is not a shortcut to wealth, but the ability to adapt by using AI directly and learning how these tools change the way we work. Instead of obsessing over building the next huge business, it may be more valuable to start with something simple—even a personal homepage—and experience this AI wave firsthand. If I had to summarize the future of vibe coding in one line, it would be this: before chasing profit, prepare your adaptability.
Related posts
talkie: The Language Model Frozen in 1930
May 1, 2026
An experiment in conversing with talkie, a language model trained only on texts written before 1930. From lunar travel and Korean independence to predictions about 2026, it is a fascinating record where the biases and insights of a century-old worldview collide.
Labor song recommendation channel - An era where anyone can create, yet some are still chosen.
Feb 27, 2026
With AI-composed concentration music overflowing in YouTube work playlists, a sense of 'familiar yet unappealing' fatigue has emerged. As this happens, channels that create song selections and videos manually become even rarer. Even in an era where anyone can create, there is a clear reason why certain content gets chosen.
Parenting and business, ultimately 'securing time' is essential.
Feb 11, 2026
Reading Adrian Li Mow Ching's Father's Day Q&A on Clicks Blog made me rethink the meaning of 'securing time.' Time is not something that is left over; it must be secured. These days, I have been prioritizing handling over planning and living reactively. Through the perspective that parenting and leadership are similar, I reflect on the attitude of 'bystanding' and intend to start implementing small routines again.