What is Vibe-Coding (and does it work?)

Vibe-Coding is the use of AI agents and NLPs to help one generate code for their project, with the majority of the code being written by the AI. Though this can be seen as “cheating”, many programmers and tech hobbyists are starting to vibe-code their software because of how easy it is. Alas, this doesn’t prevent bugs, which still have to be found and troubleshooted, albeit easier with these AI agents.

That’s my take, at least. Here’s Google’s definition of “vibe-coding”:

Wed >> Nov 26

Google just released their Antigravity AI agent last week, and I decided to try it out on my computer for training a YOLO model. Released in November 18, 2025, Antigravity is built to assist programmers and engineers in programming their projects within their own Antigravity-designed Integrated Development Environment (IDE), which is a kind of text editor that’s designed for programming, game development, data science, or anything of the like.

A warning to teen data scientists/programmers: you will need access to a 18+ Google Account in order to use Antigravity (I used my parents, lmao).

Fri >> November 28

Today, I compared Antigravity’s performance versus that of Claude Code’s (another AI coding agent, made by Anthropic). I used them both to prepare data to train a YOLO model that I was going to use on some 3D medical files.

In order to use both effectively, I used some of the same prompts across both AI coding agents, and analyzed how their edits and responses differed. This is a method called “ensembling”, and is also used in AI models where the same input is run through different models that have the same goal, and examining the combined result. In my case, however, I didn’t combine both results of Antigravity and Claude Code’s edits. I just chose the better one, and so far with my experience, they both perform pretty similarly!

One main difference, however, is that Antigravity provides a holistic “implementation plan” before executing any commands or editing anything after every prompt, to make sure that’s what I want Antigravity to do. In Claude Code, however, it only asks for verification of individual steps of edits and executions along the way instead of one big plan that covers the requested tasks.

Sat >> November 29

I took a step back today and delved into some of the theory behind these AI models, specially for images and 3D files. One of the many common denominators I found throughout recent AI models for images and 3D files is that they use a (C, D, H, W) notation to refer to the “shape” of a given file that’s being processed.

Here’s what each of the dimensions (the letters) mean:

C: Channels, which are like the different “perspectives” of the data. A good example of channels is the different RGB channels (the channels would be red, green, and blue) which add up the the full data.

D: Depth, which is the same as the “depth” of a cube or any kind of prism in geometry. I like to refer to it as the “third dimension” of a shape, but that’s up for you to decide which dimension to be as first, second or third.

H: Height, which is the same as the “height” of a shape. In this case it would be the height of the image or a “2d slice” taken from a 3D shape.

W: Width, which is the same as the “width/length” of a shape. In this case it would be the width or length of the image or a “2d slice” taken from a 3D shape.

Sometimes, if the model or algorithm is using video data, there would be a “T” dimension thrown in there as well, which represents the “time frame”, like as in the different frames you would see in a video. The T would measure which frame each one would be (from 1 to 1000000 or something depending on the video quality).