The Vibe Coding Journey begins in the LLM (Large Language Model) Experience
When AI is asked about the number of Rs in “cranberry”, “elderberry” or “barberry”, different answers might be given each time, ranging from two to four, and only sometimes three. This inconsistency isn’t a glitch, but rather a fundamental aspect of the transformer architecture that underlies language models.
Unlike humans, these models don’t count letters; instead, they rely on probability distributions to sample tokens. So, an inquiry about the number of Rs in a word prompts the system to predict the typical token sequence that follows that question pattern, without examining the word letter by letter.
The predictions are sometimes correct and sometimes not, as there’s no symbolic reasoning or mental model of the word as a sequence of characters at play. The transformer architecture compresses words into embeddings that capture semantic relationships and contextual token patterns, but these embeddings don’t preserve structural information about spelling.
As a result, the model can’t “see” individual letters the way humans do when spelling out a word like C-R-A-N-B-E-R-R-Y. To the model, two Rs are simply a token, not two distinct letters. This limitation has significant implications, as humans can count letters with perfect accuracy every time, simply by holding the word in mind as a sequence and counting systematically. In contrast, large language models (LLMs) require careful orchestration, step-by-step prompting, and external tools, as well as human guidance for every symbolic operation. While it’s possible to work around these limitations using techniques like chain of thought prompting and retrieval systems, these workarounds are essentially patches that cover up fundamental gaps in the models’ capabilities.
The economic promise of general intelligence breaks down when every simple task demands complex scaffolding. The reality is that LLMs excel at pattern matching, but they need constant supervision to perform basic symbolic operations accurately. Moreover, these models tend to prioritize likelihood over truth and are inherently brittle.
Like simple spell checker and calculator technology before them, the user MUST have an instinctive ‘feeling’ for the suggested data, otherwise ‘invested’ may be accepted, when the word needed is ‘invented’ or 2047 instead of 20,047.
But this broader grasp of rudimentary language or mathematics being ‘right’ is something we have mastered leaving primary school.
Coding is something completely different. In 2025, most people finishing secondary school may have done some very basic programming, perhaps in Python.
Serious programming for most people is a professional pursuit they follow afterwards.
Vibe coding is an AI-assisted software development method where users describe their desired software in natural language, and the AI generates the corresponding code. While it makes coding more accessible, it also presents several significant challenges.
They begin by creating prompts in an approach that’s similar to using LLM. In most cases, the inherent skills they have to proof read for a spell check result, or do a quick ‘rough’ mental arithmetic to check a calculator result, is replaced by overview and interrogation skills beyond most vibe-coders.
Software design is a less exact science, and because something looks great, and seems to work, doesn’t mean it’s secure, free of malicious things, efficient, or even right all the time.
Add that many vibe-coders have no clue how to discover and fix subtle or nuanced problems.
Vibe coding places the challenges AI models have in ‘LLM’, ‘on steroids’.
Repeat problems experienced with vibe-coding.
Security. The most concerning risk with AI-coded software is security vulnerabilities due to the code’s learning from public repositories, including insecure patterns. AI models suggest code with known vulnerabilities, such as malicious code, SQL injection, and insecure file handling, as they are an average of all developers’ shared work, including their security failings.
Maintenance and Scaling. AI-generated code from vibe coding can be hard to maintain or scale. It often passes initial tests but is brittle and poorly organized, with inconsistent structure, minimal comments, and ad-hoc logic. This lack of documentation and organization makes the software difficult to understand or extend. The codebase quickly accrues technical debt due to inefficient or overly complex solutions. AI-introduced inconsistencies in naming, coding styles, or logic flows make the codebase harder to navigate. Scaling applications created with vibe coding tools is challenging, as adding new features or handling more users may require a costly and time-consuming rewrite due to the underlying system design.
A Triumph of Style over Substance. AI may over-engineer simple features, introducing unnecessary complexity, convoluted patterns, or extra components. This results in a more complicated app that’s harder to understand or slower to load. Flashy design doesn’t replace well-thought-out user experience design, and relying on vibe coding can prioritize appearance over substance, requiring human judgment to ensure the design serves the product’s goals without unnecessary complexity.
Inadequate and insufficient training content. Vibe coding platforms are optimized for common use cases and standard tech stacks. Their training is often pre 2021 and they struggle with much that’s exotic. This includes much in Web 3.
Tools like Lovable and Bolt have limited integrations and building blocks. If a feature is outside their environment, custom code is required. Integration with less-common frameworks may be limited or impossible.
We’ve seen many people make interesting posts with them on platforms like LinkedIn, but other than an interesting looking support image or video for the post, there’s no evidence any of these things actually work.
There’s probably some good stuff out there, but we’re not yet finding anything we can endorse.
The Verdict on Vibe Coding.
At the moment, Vibe Coding has limited application due to:
- Limited and aged training library
- Insufficient security scrutiny by content selection algorithms
- Insufficient intuitiveness for efficiency and eloquence of code assembly and expression
- Overblown expectations of being an easy tool for everyone
What it can be used for:
- Static and Simple constructs that lack evolving back-ends, interoperability, bridges, integrations, handling personal or valuable data, such as: – Simple Websites, Html Email footer design, Code-rich Social Media visual data, virtual presentations, some testing, some educational.
- ‘Heavy Lifting’ by experienced coders who are authoring AI components in sandboxes and integrating them with manual steps in a project they manage end-to-end.
What to be careful of:
Just like from 2021 we had the bitcoin grafters and 2022 began the Web 3 grafters, we now have these software engineer/product design grafters who have appeared out of thin air claiming they can code anything. Experienced technicians will have a portfolio with credible referees. Validate that they have a substantial body of work pre-2023 (pre vibe-coding), especially if the arrangement will be remote.
It will get better, but ‘we’re still early’
Credit : Veronica Bridgewater, 9ja Cosmos Ambassador focusing on LinkedIn presence – Cocktail of Social Media extracts.
Glideapps Article.
9ja Cosmos is here…
Detoxants – FiendYard Fruit & Veg
Opaque Emotion Pathways Tokenized Artworks
Preview our Sino Amazon/Sinosignia releases (Ente)
